Biomarkers and prediction models for type 2 diabetes and diabetes related outcomes Abbasi, Ali

Transcription

1 University of Groningen Biomarkers and prediction models for type 2 diabetes and diabetes related outcomes Abbasi, Ali IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below. Document Version Publisher's PDF, also known as Version of record Publication date: 2013 Link to publication in University of Groningen/UMCG research database Citation for published version (APA): Abbasi, A. (2013). Biomarkers and prediction models for type 2 diabetes and diabetes related outcomes Groningen: s.n. Copyright Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons). Take-down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. Downloaded from the University of Groningen/UMCG research database (Pure): For technical reasons the number of authors shown on this cover page is limited to 10 maximum. Download date:

2 Biomarkers and Prediction Models for Type 2 Diabetes and Diabetes Related Outcomes Ali Abbasi

3 The research presented in this thesis was performed at the Department of Epidemiology in close collaboration with the Department of Internal Medicince/Nephrology, University of Groningen, University Medical Center Groningen (UMCG); and the Julius Center for Health Sciences and Primary Care, University of Utrecht, UMC Utrecht. This work was supported by the Netherlands Heart Foundation, Dutch Diabetes Research Foundation and Dutch Kidney Foundation. This research was performed within the framework of the Center for Translational Molecular Medicine; project PREDICCt. Publication of this thesis was financially supported by the University of Groningen (RuG), UMCG, Graduate School of Medical Sciences, Thermo Fisher Scientific/ BRAHMS GmbH, the Netherlands Heart Foundation, Dutch Kidney Foundation, GlaxoSmithKline, Novo Nordisk, Roche and Boehringer Ingelheim. Biomarkers and Prediction Models for Type 2 Diabetes and Diabetes Related Outcomes PhD thesis, University of Groningen, UMCG, the Netherlands. Cover design and layout: Ali Abbasi ISBN: (printed version) ISBN: (electronic version) Published by the University of Groningen, Groningen, the Netherlands. Printed by CPI, Wöhrmann Print Service B.V. Zuphen. Copyright 2013 by Ali Abbasi All rights reserved. No part of this publication may be reproduced, stored in a retrieval system of any nature, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without prior written permission of the author.

4 Biomarkers and Prediction Models for Type 2 Diabetes and Diabetes Related Outcomes Proefschrift ter verkrijging van het doctoraat in de Medische Wetenschappen aan de Rijksuniversiteit Groningen op gezag van de Rector Magnificus, dr. E. Sterken, in het openbaar te verdedigen op woensdag 1 mei 2013 om 14:30 uur door Ali Abbasi geboren op 19 november 1978 te Yazd, Iran

5 Promotores: Copromotores: Beoordelingscommissie: Prof.dr. R.P. Stolk Prof.dr. G.J. Navis Dr. S.J.L. Bakker Dr.ir. E. Corpeleijn Dr.ir. J.W.J. Beulens Prof.dr. R.O.B. Gans Prof.dr. K.G.M. Moons Prof.dr. N. Wareham

6 Paranimfen: Azadeh Zaferani Douwe Postmus To beloved Mom, dear Dad and Dearest Shabnam

7 CONTENTS Chapter 1 Introduction 7 Part I. Prediction models for risk of type 2 diabetes Chapter 2 Prediction models for risk of developing type 2 diabetes: systematic literature search and independent external validation study. BMJ 2012;345:e Chapter 3 Chapter 4 External validation of the KORA S4/F4 prediction models for the risk of developing type 2 diabetes in older adults: the PREVEND study. Eur J Epidemiol 2012;27: Liver function tests and risk prediction of incident type 2 diabetes: evaluation in two independent cohorts. PLoS One. 2012; /journal.pone Part II. Risk factors and predictive value of biomarkers for diabetes and other cardiometabolic outcomes Chapter 5 Maternal and paternal transmission of type 2 diabetes: influence of diet, lifestyle and adiposity. J Intern Med 2011;270: Chapter 5a Chapter 6 Chapter 7 Chapter 8 Chapter 9 Commentary: Both multiplicative and additive components may contribute to parental transmission of type 2 diabetes. J Intern Med 2011;270: Parental history of type 2 diabetes and cardiometabolic biomarkers in offspring. Eur J Clin Invest 2012;42: Plasma procalcitonin is associated with obesity, insulin resistance, and the metabolic syndrome. J Clin Endocrinol Metab 2010;95:E Sex differences in the association between plasma copeptin and incident type 2 diabetes: the Prevention of Renal and Vascular Endstage Disease (PREVEND) study. Diabetologia 2012;55: HDL cholesterol, apolipoprotein A-I and HDL particle composition independently predict incident type 2 diabetes mellitus in the general population: the Prevention of Renal and Vascular End-stage Disease (PREVEND) Study.

8 Submitted 173 Chapter 10 Chapter 11 Plasma Procalcitonin and Risk of Type 2 Diabetes in the General Population. Diabetologia 2011;54: Peroxiredoxin 4, a novel circulating biomarker for oxidative stress, and the risk of incident cardiovascular disease and all-cause mortality. J Am Heart Assoc 2012;1:e Chapter 12 Discussion, future perspectives, conclusions 227 Summary 239 Nederlandse Samenvatting 243 Acknowledgements 246

9

10 Chapter 1 Biomarkers for prediction of type 2 diabetes: clinical and methodological views

11 Introduction Type 2 diabetes (T2D) is a multi-factorial disease which can be diagnosed by levels of plasma glucose ( 7 mmol/l) or HbA1c ( 6.5%) in the circulation 1, 2. In fact, glucose and HbA1c are two major biological clues of which variation below the diagnostic threshold is strongly predictive of future risk of T2D, particularly within a time frame of 5 to 10 years before the actual diagnosis of T2D 3. In clinical practice, all measurable biological clues are called biomarkers, irrespective of whether the biological clue is obtained from physical examination of the body (i.e. blood pressure or blood tests) or anywhere else from medical history 4. From a public health perspective, T2D has become an increasing global health burden, with an increase in the number of people with diabetes by 2-3 fold between 1985 and In 2008, age-standardised (to the WHO reference population) prevalence of T2D was 9.8% in men and 9.2% in women worldwide. The majority of people with diabetes worldwide (40%) come from India and China. For china, this estimate was 9.6% in men and 9.4% in women. For The Netherlands, this estimate was 6.1% in men and 4.1% in women, while, for Iran, these estimates were 9.3% and 10.5%, respectively 5. The emerging pandemic is presumably driven by the combined effects of population ageing, rising rates of obesity and unhealthy lifestyle related to prosperity 6. Meanwhile, the prevalence of obesity and T2D is shifting to younger ages, increasing the lifetime burden of T2D and diabetesassociated co-morbidities, including cardiovascular disease (CVD) and renal disease 7. Good news is that T2D and most likely its complications are preventable 8, 9. To reverse future projections on the increasing prevalence of T2D, early identification of individuals at high risk for T2D is essential for early implementation of targeted prevention strategies 10, 11. Of course, such a targeted strategy should be complementary to a general population strategy of reducing well-known risk factors for T2D, such as smoking, sedentary lifestyle and a high-calorie diet 9. However, given limited healthcare resources, risk classification of the population will aid in avoiding implementation of interventions in a very large number of individuals who are at low-risk for T2D 10, 11. What makes prediction different from etiologic research? From an epidemiological perspective, we can investigate risk factors that might be causally related to disease. In this type of research, the strength of the association is usually expressed in a relative sense: a relative risk, or, if time of onset of disease is known, a hazard ratio 12. This type of study, called etiologic research is performed to investigate the risk of disease by levels of risk factors such as biomarkers 13. For example, parental history of T2D is known to be associated with an increased risk of T2D (2.9 fold higher risk of T2D) 14, 15. A relative risk of two or three tells us that a person with parental diabetes has 2-3 fold higher risk of T2D when compared to another person, of similar age, gender and possibly other characteristics, but without 8

12 parental history of diabetes. Those factors that are etiologically linked to the outcome of disease are good candidates to include in prediction models (or risk scores) Clearly, also factors that are a consequence of the disease may be good predictors, since they may reflect the early stages of a pathological process. Importantly, in prediction modelling it does not matter whether markers are causally related to the development of T2D, or a consequence of disease, as long as the marker has the capacity to distinguish between those who will develop the disease or the complication and those who will not 18, 20. The product of prediction research is a prediction tool providing absolute estimates of disease risk in the future 16. To estimate risk of future T2D, the use of a reliable and practical scoring tool/questionnaire is recommended. Of course, a valid tool can also help to inform patients about the future expected course of their illness, and can guide doctors and patients in decisions on further treatment. How to assess performance of prediction models? For early identification of individuals at high risk for T2D, accurate prediction of future T2D is of utmost importance. To date, many prediction models for T2D have been developed in different settings and populations 10, 11, 19, 21. In general, prediction models perform less well in other populations than the population in which it was developed. The poor performance is because of inherent deficiencies in the development of prediction models, like an over-fitted model or lack of important predictor(s) 22. Therefore, external validation of such models in independent populations and datasets is an essential step to broadly evaluate the performance of such models 11. External validation is also relevant because treatment and characteristics of populations may change over time. To assess the performance of a prediction model, epidemiologists and statisticians propose two essential measures; discrimination and calibration 11, Discrimination describes the ability of the model to distinguish those at high risk of developing T2D (or those who will get the disease) from those at low risk (those who will not get the disease) 11, 12. It is evaluated using the area under the curve (AUC) of a receiver operating characteristic curve (ROC curve) (Figure 1). For a binary outcome, the AUC is identical to the C-statistic 13. This ROC-curve is derived from the specificity and sensitivity of the model. For a certain cut-off point (a score), sensitivity and specificity can be determined. Sensitivity is the percentage of individuals who are correctly identified as cases. Specificity is the percentage of individuals who are correctly identified as non-cases. A ROC-curve plots the sensitivity against 1- specificity for a large number of possible cut-off points, and is thereby a general measure for overall predictability 13. Calibration addresses whether the estimated absolute risk is in agreement with the observed risk (e.g. incident T2D) 13. This can be visualized by calibration plots (Figure 2). In a calibration plot, the estimated risk is plotted against the observed incidence of the outcome. Ideally the estimated risk equals the observed incidence throughout the entire risk spectrum and the calibration plot follows the 45 line 11, 13, 25. Calibration is also tested using the Hosmer 9

13 Lemeshow (H-L) goodness-of-fit statistic (or Chi 2 -test). A lack of significant difference is then interpreted no difference between observed and estimated risks, and indicates a good model fit 13. Even with good discrimination, calibration may be insufficient in an external validation, since the absolute risk strongly depends on the incidence of the disease in the development population and on population characteristics like age 11, 26. In order to overcome such differences between a development and validation population, a first step is recalibration of prediction model by adjusting for the difference in incidence of the disease 11, 23. In this step, the intercept (for logistic regression models) or the baseline survival function (for survival regression models) of the original prediction model will be updated or adjusted to the new circumstance 11, 23, 27, 28. Sensitivity Specificity Figure 1. Hypothetical receiver operating characteristic (ROC) curves for prediction of type 2 diabetes. The C-statistics for models were 0.83 and 0.93, indicating good to excellent discrimination. A C-statistic of 0.5 is considered as threshed of random chance, indicating discrimination not better than tossing a coin. Other prognostic measures such as the net reclassification improvement (NRI) or the integrated discrimination improvement (IDI) will be calculated to address the improvement in prediction of disease by adding a (bio)marker to a validated model 12, 13. For the NRI, it is necessary to have a validated prediction model that uses risk 10

14 categories, such as low, intermediate or high risk. The NRI measures to what extent the new model that incorporates new biomarker(s) leads to improvement in classification. Improved classification is determined by upward movement in categories for individuals who will get the disease, and downward movement for individuals who will not get the disease. The IDI is a continuous version of the NRI without a priori defined risk categories 12, 29. Observed risk Under-estimation Over-estimation Ideal calibration Predicted risk Figure 2. Hypothetical calibration plot depicting predicted risk against observed risk of type 2 diabete. The red line (the 45 line) from zero denotes ideal calibration line (slope=1, intercept=0), the solid line denotes smooth calibration curve for a model with under-estimation of the risk and the dotted line for a model with over-estimation of risk. To what extent biomarkers could improve prediction of T2D? Observational epidemiological studies have identified associations of many wellknown or emerging biomarkers with risk of development of T2D and its associated complications 20, In principle, prior studies suggest that the risk of future T2D can differ by levels of a given biomarker. Whether intervention at the level of a given biomarker is useful will depend on whether the biomarker is causal to T2D or not. 11

15 If the biomarker is not causally related, the process of developing T2D may cause an increase or decrease in the levels of biomarker, as one of the consequences of T2D. Etiologic research for causal biomarkers provides better insights into mechanism of T2D and perhaps introduces targets for (pharmacological) intervention and treatment. Moreover, research for prognostic biomarkers might improve prediction of T2D and have implications for early prevention of T2D. From clinical perspectives, the main application of a biomarker lies within risk stratification and guided preventive strategies for the outcome 12, 34, 38. In this phase, candidate biomarkers, either causal or not, are added to the validated prediction models to examine which biomarkers have incremental predictive value on top of the models. For example, in the Framingham Offspring Study, several risk scores have been developed for the prediction of T2D 39. A simple model, incorporating data on age, sex, parental diabetes and body mass index, correctly classified 72% of cases and non-cases. Addition of well-known biomarkers, including fasting glucose, blood pressure, highdensity lipoprotein cholesterol (HDL-C) level and triglyceride level yielded a significant improvement in discrimination to a level of 85% of correct classification. Next, addition of C-reactive protein, insulin sensitivity and resistance indices showed no further improvement. In the Data from an Epidemiological Study on the Insulin Resistance Syndrome (DESIR) cohort 40, the clinical prediction model that included waist circumference, hypertension, family history of diabetes (in women) and smoking (in men) showed C statistics of 71.3% and 82.7% for men and for women, respectively. The addition of known biomarkers such as fasting glucose, triglycerides and gamma-glutamyl transferase considerably increased the discriminative power (up to 85% for men and 91.7% for women). Further addition of genetic risk factors added little to the prediction of diabetes in the DESIR study. This is consistent with another study in which a genotype score provided slightly better prediction for risk of T2D on top of common diabetes risk factors 41. One explanation is that the contribution of some genetic variants on T2D risk are effectuated through the intermediate risk factors 41. So far, the incremental predictive value of most known and novel biomarkers above validated model(s) for T2D is still unclear. Independent studies evaluating utility of the measured markers when incorporated in validated prediction model(s) are needed to answer this question 20, 38, 42. Outline and aims of the thesis This thesis is conducted within the framework of the Center for Translational Molecular Medicine (CTMM), a public-private consortium dedicated to the development of medical technologies that enable the design of new and personalized treatments for the main causes of mortality and diminished quality of life and the rapid translation of these treatments to the patient. The aim of the PREDICCt project, a CTMM project, entitled Biomarkers for the PREdiction and early diagnosis of DIabetes and diabetes-related Cardiovascular Complications, is to identify biomarker for the prediction and early detection of diabetes and its 12

16 complications. The aim of work package 7 of Predicct is the validation of biomarkers in large prospective cohorts. The aim of this thesis, within the work package 7 of the PREDICCt project, is to identify biomarkers with added predictive value for risk of future T2D. The first part is aimed at validating and updating existing prediction models for the risk of T2D in Dutch population-based cohorts. In Chapter 2, we will systematically search the literature to identify existing prediction models for the risk of future T2D and to externally validate the retrieved models in the Dutch contribution of the large European Prospective Investigation Into Cancer and Nutrition cohort study (EPIC-NL). Chapter 3 describes the validation and update of the prediction models (from the KORA S4 F4 study) for T2D in the Prevention of Renal and Vascular Endstage Disease (PREVEND) study. Chapter 4 will focus on incremental predictive value of liver function tests compared with the KORA models in EPIC-NL and PREVEND, separately. The second part of thesis focuses on associations of conventional risk factors with risk of T2D and investigation of novel biomarkers for risk of T2D (etiologic research). Next, we evaluate incremental predictive value of novel biomarkers for risk of T2D and CVD (prediction research). In Chapters 5 and 6, we will investigate the associations of parental history of diabetes specified to maternal and/or paternal transmission with risk of incident T2D and cardio-metabolic biomarkers. We also explore associations of novel risk factors, such as inflammatory biomarkers, lipid markers, CRP and procalcitonin with the determinants of obesity, the metabolic syndrome, insulin resistance, kidney function and risk of incident T2D in Chapters 7, 9 and 10. In Chapter 8, we aim to test whether the association of a novel stress system biomarker, named copeptin, with T2D is independent of other predictors including clinical variables and more established biomarkers like glucose, CRP and urine albumin excretion. Chapter 9 describes the associations of HDL-cholesterol, apolipoproteins and HDL particle composition with risk of incident T2D in the PREVEND study. In Chapter 11, we will study, a novel circulating biomarker with antioxidant properties, to be associated with the common risk factors of CVD and risk of incident CVD and all-cause mortality. We also examine the incremental predictive value of the marker compared with the Framingham risk score in terms of discriminative ability and net reclassification. In general discussion chapter, we provide an overview of concepts around screening and prediction of T2D, validation of prediction models and the incremental value of known and novel biomarkers for mid-term risk of T2D and CVD. Future perspectives focus on direction of observational studies for the estimated risk of T2D and its complications over the lifetime. 13

17 References 1. American Diabetes Association. Standards of medical care in diabetes Diabetes Care 2011;34 Suppl 1:S Stolk RP, Rosmalen JG, Postma DS, et al. Universal risk factors for multifactorial diseases: LifeLines: a three-generation population-based study. Eur J Epidemiol 2008;23(1): Tabak AG, Herder C, Rathmann W, Brunner EJ, Kivimaki M. Prediabetes: a high-risk state for diabetes development. Lancet 2012;379(9833): Gerszten RE, Wang TJ. The search for new cardiovascular biomarkers. Nature 2008;451(7181): Danaei G, Finucane MM, Lu Y, et al. National, regional, and global trends in fasting plasma glucose and diabetes prevalence since 1980: systematic analysis of health examination surveys and epidemiological studies with 370 country-years and 2.7 million participants. Lancet 2011;378(9785): Hu FB. Globalization of diabetes: the role of diet, lifestyle, and genes. Diabetes Care 2011;34: Seshasai SR, Kaptoge S, Thompson A, et al. Diabetes mellitus, fasting glucose, and risk of cause-specific death. N Engl J Med 2011;364(9): Group DPPR, Fowler SE, Hamman RF, et al. 10-year follow-up of diabetes incidence and weight loss in the Diabetes Prevention Program Outcomes Study. Lancet 2009;374(9702): Ezzati M, Riboli E. Can noncommunicable diseases be prevented? Lessons from studies of populations and individuals. Science 2012;337(6101): Noble D, Mathur R, Dent T, Meads C, Greenhalgh T. Risk models and scores for type 2 diabetes: systematic review. BMJ 2011;343:d Abbasi A, Peelen LM, Corpeleijn E, et al. Prediction models for risk of developing type 2 diabetes: systematic literature search and independent external validation study. BMJ 2012;345:e Cook NR. Statistical evaluation of prognostic versus diagnostic models: beyond the ROC curve. Clin Chem 2008;54(1): Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010;21(1): Abbasi A, Corpeleijn E, van der Schouw YT, et al. Maternal and paternal transmission of type 2 diabetes: influence of diet, lifestyle and adiposity. J Intern Med 2011;270(4): Samocha-Bonet D, Campbell LV, Viardot A, et al. A family history of type 2 diabetes increases risk factors associated with overfeeding. Diabetologia 2010;53(8): Moons KG, Altman DG, Vergouwe Y, Royston P. Prognosis and prognostic research: application and impact of prognostic models in clinical practice. BMJ 2009;338:b Moons KG, Royston P, Vergouwe Y, Grobbee DE, Altman DG. Prognosis and prognostic research: what, why, and how? BMJ 2009;338:b Pepe MS, Cai T, Longton G. Combining predictors for classification using the area under the receiver operating characteristic curve. Biometrics 2006;62(1): Collins GS, Mallett S, Omar O, Yu LM. Developing risk prediction models for type 2 diabetes: a systematic review of methodology and reporting. BMC Med 2011;9: Abbasi A, Bakker SJ, Corpeleijn E, et al. Liver function tests and risk prediction of incident type 2 diabetes: evaluation in two independent cohorts. PLoS One 2012;7(12):e Buijsse B, Simmons RK, Griffin SJ, Schulze MB. Risk assessment tools for identifying individuals at risk of developing type 2 diabetes. Epidemiol Rev 2011;33(1): Altman DG, Vergouwe Y, Royston P, Moons KG. Prognosis and prognostic research: validating a prognostic model. BMJ 2009;338:b

18 23. Vergouwe Y, Moons KG, Steyerberg EW. External validity of risk models: Use of benchmark values to disentangle a case-mix effect from incorrect coefficients. Am J Epidemiol 2010;172(8): Collins GS, Moons KG. Comparing risk prediction models. BMJ 2012;344:e Harrell FE, Jr., Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996;15(4): Abbasi A, Corpeleijn E, Peelen LM, et al. External validation of the KORA S4/F4 prediction models for the risk of developing type 2 diabetes in older adults: the PREVEND study. Eur J Epidemiol 2012;27(1): Toll DB, Janssen KJ, Vergouwe Y, Moons KG. Validation, updating and impact of clinical prediction rules: a review. J Clin Epidemiol 2008;61(11): van Houwelingen HC. Validation, calibration, revision and combination of prognostic survival models. Stat Med 2000;19(24): Pencina MJ, D'Agostino RB, Sr., D'Agostino RB, Jr., Vasan RS. Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Stat Med 2008;27(2):157-72; discussion Abbasi A, Corpeleijn E, Meijer E, et al. Sex differences in the association between plasma copeptin and incident type 2 diabetes: the Prevention of Renal and Vascular Endstage Disease (PREVEND) study. Diabetologia 2012;55(7): Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin and risk of type 2 diabetes in the general population. Diabetologia 2011;54(9): Sattar N, Wannamethee SG, Forouhi NG. Novel biochemical risk factors for type 2 diabetes: pathogenic insights or prediction possibilities? Diabetologia 2008;51(6): Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin is associated with obesity, insulin resistance, and the metabolic syndrome. J Clin Endocrinol Metab 2012;95(9):E Abbasi A, Corpeleijn E, Postmus D, et al. Peroxiredoxin 4, A Novel Circulating Biomarker for Oxidative Stress and the Risk of Incident Cardiovascular Disease and All-Cause Mortality. Journal of the American Heart Association 2012;1(5). 35. Abbasi A, Corpeleijn E, van der Schouw YT, et al. Parental history of type 2 diabetes and cardiometabolic biomarkers in offspring. Eur J Clin Invest 2012;42(9): Enhorning S, Wang TJ, Nilsson PM, et al. Plasma copeptin and the risk of diabetes mellitus. Circulation 2010;121(19): Salomaa V, Havulinna A, Saarela O, et al. Thirty-one novel biomarkers as predictors for clinically incident diabetes. PLoS One (4):e Hlatky MA, Greenland P, Arnett DK, et al. Criteria for evaluation of novel markers of cardiovascular risk: a scientific statement from the American Heart Association. Circulation 2009;119(17): Wilson PW, Meigs JB, Sullivan L, Fox CS, Nathan DM, D'Agostino RB, Sr. Prediction of incident diabetes mellitus in middle-aged adults: the Framingham Offspring Study. Arch Intern Med 2007;167(10): Balkau B, Lange C, Fezeu L, et al. Predicting diabetes: clinical, biological, and genetic approaches: data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR). Diabetes Care 2008;31(10): Meigs JB, Shrader P, Sullivan LM, et al. Genotype score in addition to common risk factors for prediction of type 2 diabetes. N Engl J Med 2008;359(21): Moons KG. Criteria for scientific evaluation of novel markers: a perspective. Clin Chem 2010;56(4):

19 16

20 Chapter 2 Prediction models for risk of developing type 2 diabetes: systematic literature search and independent external validation study Ali Abbasi 1,2,3 ; Linda M. Peelen 3 ; Eva Corpeleijn 1 ; Yvonne T. van der Schouw 3 ; Ronald P. Stolk 1 ; Annemieke M.W. Spijkerman 5 ; Daphne L van der A 4 ; Karel G M Moons 3 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 ; Joline W.J. Beulens 3 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands 4 Center for Prevention and Health Services Research, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands 5 Center for Nutrition and Health, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands BMJ 2012; 345: e5900

21 Abstract Background The aim of this study was to identify existing prediction models for the risk of development of type 2 diabetes and to externally validate them in a large independent cohort. Methods We conducted systematic search of English, German, and Dutch literature in PubMed until February 2011 to identify prediction models for diabetes. Performance of the models was assessed in terms of discrimination (C statistic) and calibration (calibration plots and Hosmer-Lemeshow test). The validation study was a prospective cohort study, with a case cohort study in a random subcohort. Models were applied to the Dutch cohort of the European Prospective Investigation into Cancer and Nutrition cohort study (EPIC-NL). Participants were 38,379 people aged with no diabetes at baseline, 2,506 of whom made up the random subcohort. Main outcome measure was incident type 2 diabetes. Results The review identified 16 studies containing 25 prediction models. We considered 12 models as basic because they were based on variables that can be assessed non-invasively and 13 models as extended because they additionally included conventional biomarkers such as glucose concentration. During a median follow-up of 10.2 years there were 924 cases in the full EPIC-NL cohort and 79 in the random subcohort. The C statistic for the basic models ranged from 0.74 (95% confidence interval 0.73 to 0.75) to 0.84 (0.82 to 0.85) for risk at 7.5 years. For prediction models including biomarkers the C statistic ranged from 0.81 (0.80 to 0.83) to 0.93 (0.92 to 0.94). Most prediction models overestimated the observed risk of diabetes, particularly at higher observed risks. After adjustment for differences in baseline incidence of diabetes, calibration improved considerably. Conclusions Most basic prediction models can identify people at high risk of developing diabetes in a time frame of five to 10 years. Models including biomarkers classified cases slightly better than basic ones. Most models overestimated the actual risk of diabetes. Existing prediction models therefore perform well to identify those at high risk, but cannot sufficiently quantify actual risk of future diabetes. 18

22 Introduction Type 2 diabetes is a large burden in healthcare worldwide 1. Studies on lifestyle modifications and drug intervention have convincingly shown that these measures can prevent diabetes 2, 3. Early identification of populations at high risk for diabetes is therefore important for targeted prevention strategies and is necessary to enable proper efforts to be taken for prevention in the large number of individuals at high risk, while avoiding the burden of prevention and treatment for the even larger number of individuals at low risk, both for the individual and for society. The professional practice committee of the American Diabetes Association recommends screening for all overweight or obese adults (body mass index (BMI) 25) of any age who have one or more additional risk factors for diabetes such as family history or hypertension 4. The European evidence based guidelines for the prevention of type 2 diabetes 5 and the International Diabetes Federation 6 recommend the use of a reliable, simple, and practical risk scoring system or questionnaire to identify people at high risk of future diabetes. During the past two decades, many such prediction models have been developed Three recent reviews on this topic described existing prediction models and the predictive value of specific risk factors (such as metabolic syndrome) over a wide range of populations 7-9. Surprisingly, however, the performance of less than a quarter of the prediction models was externally validated Because the performance of a prediction model is generally overestimated in the population in which it was developed, external validation of such models in an independent population, ideally by researchers not involved in the development of the models, is essential to broadly evaluate the performance and thus the potential utility of such models in different populations and settings Consequently, certain prediction models to identify those at high risk of diabetes cannot be recommended when external validity of available models is unknown 12, 16. Moreover, a direct comparison of the performance of the existing models in the same (external) validation cohort is essential to bridge the gap between the development of models and the conduct of studies for clinical utility. The recent systematic reviews highlighted the need for an independent study to identify the existing prediction models and subsequently validate and compare their performance to support the current recommendations 7-9. Few studies have externally validated such models, commonly not more than two or three at once, and almost always in medium sized cohorts 10, 11, 14, 17. We applied a more comprehensive approach as recently suggested 14, 15. Firstly, we carried out a systematic review to identify the most relevant existing models for predicting the future risk of type 2 diabetes. Then we used various analytical measures for validating 18 and comparing their predictive performance in a large independent general population based cohort the Dutch cohort of the European Prospective Investigation into Cancer and Nutrition (EPIC-NL)

23 Methods Systematic literature search We performed a systematic literature search according to the PRISMA guidelines, 20 when applicable. We searched PubMed for all published cohort studies that reported prediction models for the risk of type 2 diabetes until February 2011 using the following search string: (( diabetes OR diabetes mellitus OR type 2 diabetes ) AND ( risk score OR prediction model OR predictive model OR predicting OR prediction rule OR risk assessment OR algorithm )) NOT review [pt] AND English [LA]. We repeated this search for publications in German and Dutch. Finally, we checked systematic reviews and validation studies of prediction models to identify other relevant articles for our validation study. Because we did not perform a formal meta-analysis, the PRISMA items related to protocol and registration and synthesis of results for meta-analyses are not applicable to our study. Studies were included if they met the following criteria: the study presented at least one formal prediction model or an update on a previously developed model; the endpoint was incident type 2 diabetes in a longitudinal design; and the population had to be at least partly white because the EPIC-NL cohort to be used for validation consists predominantly of white adults. We excluded studies using data on individuals with impaired glucose tolerance or impaired fasting glucose. Furthermore, we excluded models that used the two hour oral glucose tolerance test as a predictor variable because this was not available in our validation dataset and there was no reliable proxy variable available that could be taken as a substitute. After review of the retrieved titles, two authors (AA and JWJB) independently reviewed the abstracts to select the relevant papers for full text review and subsequently reviewed and assessed the full papers. Discrepancies between the two reviewers were solved by having a third author (EC) review to reach consensus. For included studies, we made a primary plan to extract necessary data from the original studies to validate the models or contact the authors to obtain this information. Table 1 summarises characteristics of the included studies. The extracted data included the first author s name, year of publication, country, name of study/score, number of cases and population, ascertainment of diabetes, duration of follow-up, statistical model, number of predictors, and reported performance of the model. The retrieved models were divided into models that contained only non-invasive predictors ( basic models ) and models that also included conventional biomarkers, such as glucose, HbA1c lipids, uric acid, or γ-glutamyltransferase ( extended models ). 20

24 Validation cohort The EPIC-NL cohort (n=40,011) includes the Monitoring Project on Risk Factors for Chronic Diseases (MORGEN-EPIC) and Prospect-EPIC cohorts, initiated between 1993 and The Prospect-EPIC cohort comprises 17,357 women aged who participated in a breast cancer screening programme. The MORGEN cohort comprises 22,654 men and women aged who were recruited through random population sampling in three Dutch towns (Amsterdam, Maastricht, and Doetinchem). At baseline, all participants were sent a general questionnaire and a food frequency questionnaire; these were returned when they visited the study centre for a medical examination. Reporting of the study results conforms to STROBE along with references to STROBE 21. We excluded 615 individuals with prevalent type 2 diabetes and 1,017 with missing follow-up or who did not consent to linkage with disease registries. The 38,379 remaining participants were used to validate the basic models in a full cohort design. We applied similar exclusion criteria in a 6.5% baseline random sample (n=2,604) in which measurements of conventional biomarkers were available, 19, leaving 2506 individuals. We used this random sample and all incident cases of type 2 diabetes to validate the extended models in a case cohort design 22. Table S1 in appendix provides baseline characteristics for the entire cohort, the random sample, and the people with incident type 2 diabetes. Assessment of predictor variables Variables in the prediction models included in this study were assessed with a baseline general questionnaire for disease history and lifestyle variables. A validated food frequency questionnaire filled in at baseline was used to assess nutritional variables 23. During the baseline visit, body weight, height, waist, and hip circumference, and blood pressure were measured and blood samples were drawn. Details of these procedures have been described elsewhere 19 and are shown in appendix. Assessment of type 2 diabetes Occurrence of diabetes during follow-up was self reported via two follow-up questionnaires at three to five year intervals in the MORGEN and Prospect cohort. In the Prospect cohort, incident cases of diabetes were also detected as glucosuria via a urinary glucose strip test, which was sent out with the first follow-up questionnaire. Diagnoses of diabetes were also obtained from the Dutch Center for Health Care Information, which holds a standardised computerised register of diagnoses at hospital discharge. Follow-up was complete up to 1 January Potential cases identified by these methods were verified against general practitioner (MORGEN and Prospect) or pharmacist records (Prospect only). Diabetes was defined as present when the diagnosis was confirmed by either of these methods. For 89% (n=1142) of participants with potential diabetes, verification information was available, and 72% 21

25 (n=924) were verified as having type 2 diabetes and were included as cases of type 2 diabetes in this analysis 24. Data analysis To evaluate the predictive performance of the retrieved prediction models, we used the original prediction models (regression coefficients with intercept or baseline hazard) as published. If the paper did not contain sufficient information, we asked the authors to provide us with the original model 25, 26. Particularly, we obtained regression coefficients 26 and the intercept of the model 25 by asking for complementary information. Using these original (regression) model formulas, we calculated the probability of developing type 2 diabetes per model for each individual in our study sample. Two authors (AA and JWJB) first matched the predictors of the original models with the variables available in our data. A direct match was available in our data for most variables. If a direct match was not possible, we replaced the original predictor with a proxy variable to avoid having to drop the model from our validation study. For example, we used non-fasting glucose values because fasting glucose values were not available in our data. Also, nutritional variables were collected with our food frequency questionnaire as continuous variables (g/day) and were re-coded into corresponding categories used in the prediction models by using Dutch portion sizes. Table S2 (in appendix) provides an overview of the variables used in each of the prediction models, and Text S2 (in appendix) gives the exact details on the proxy variables that were used. We assessed performance of the models using measures of discrimination and calibration.[13] Discrimination describes the ability of the model to distinguish those at high risk of developing diabetes from those at low risk. The discrimination was examined by calculating Harrell s C (comparable with the area under the ROC curve), accounting for censored data 27. Calibration indicates the ability of the model to correctly estimate the absolute risks and was examined by calibration plots. In a calibration plot, the predicted risk is plotted against the observed incidence of the outcome. Ideally the predicted risk equals the observed incidence throughout the entire risk spectrum and the calibration plot follows the 45 line. The calibration plot was extended to a validation plot as a summary tool 18, 27. Text S4 (in appendix) gives more details on information provided by this plot. Calibration was also tested with the Hosmer-Lemeshow goodness of fit statistic for time to event data 18, 27. Follow-up of our cohort was almost complete until about eight years: 3% were censored at 5 years, 5% at 7.5 years, and 44.6% at 10 years. To account for censoring when obtaining the observed probabilities for assessing calibration over, say, 10 years of follow-up, we first calculated for each individual the linear predictor and subsequently 10 year predicted outcome probability by the original survival models. This predicted probability was then divided into tenths, and we performed a Kaplan- Meier analysis per tenth, which accounts for the observed censoring. Per tenth, we obtained at the 10 year time point the observed outcome percentages, which in turn were compared with the 10 mean predicted outcome probabilities to obtain the 22

26 calibration plot and measure of goodness of fit. This was done for each model and for the other time points (5 and 7.5 years) 28. Moreover, we reported the calibration slope for the logistic regression models 18 and calculated observed over predicted (expected) outcomes (O/E ratio) with 95% confidence intervals 18, 29. A ratio below 1.0 indicates overestimation of risk, and a ratio more than 1.0 indicates underestimation of risk. Differences in the incidence of diabetes in our cohort and in the development populations led to significant deviation between observed risk in our cohort and predicted risk estimated by the prediction model. To reduce this source of miscalibration, we recalibrated each prediction model by adjusting the intercept (for logistic regression models) or the baseline survival function (for survival regression models) 28, 30, 31. The original models were developed for different time periods of risk prediction (different prediction horizons ) for instance, some models estimate 5 year risk and others 10 year risk. We therefore assessed the performance of each model for prediction of risk at 5, 7.5, and 10 years to account for the different time periods. For example, for 5 year risk, we considered individuals as incident cases if they had developed diabetes within the first five years of follow-up. Participants who developed diabetes after more than five years of follow-up were included in five year prediction as non-cases. A similar approach was followed for 7.5 and 10 year predictions. In addition, we performed a sensitivity analysis using the prediction horizon for which each model was developed in case this differed from 5, 7.5, or 10 years. For the basic prediction models, which included only data from non-invasive clinical variables, we quantified their performance in the full dataset. The extended models were validated in the case cohort data. To account for this design, we applied an extrapolation approach that extends the case cohort data to the size of the full cohort 22. This is achieved by extrapolating the non-cases of the random sample (that is, the total random sample of 2,506 individuals minus 79 cases) to the number of non-cases in the full cohort (that is, the total sample of 38,379 individuals minus 924 cases). To do so, we substituted the non-cases in the full cohort (n=37,455) with a random multiplication of non-cases of the random sample (n=2,427). On average, we multiplied the non-cases in the random sample by 15.4 (that is, 37,455 divided by 2,427). Next, we merged the extrapolated data from non-cases to those from all the cases (total non-cases of 37,455 individuals plus 924 cases), recreating the size and composition of the full cohort. In sensitivity analyses, we estimated the performance of the basic models in the re-sampled data from the case cohort and compared these results with those obtained from the full cohort. This allowed us to confidently use the extrapolation approach for the extended models in the case cohort design. For most predictors data from less than 1% of the values were missing, although missing values occurred in 5% for family history of diabetes, about 15% for physical activity, and 20.5% for non-fasting glucose concentrations. Because an analysis of only the completely observed participants could lead to biased results 32-34, we imputed these missing values using single imputation and predictive mean 23

27 matching. As the percentage of missing values for the non-fasting glucose concentration was high, we repeated our analyses using only data from the MORGEN cohort, in which less than 10% of values for non-fasting glucose concentration were missing, as a sensitivity analysis. Table S3 (in appendix) shows the number of missing values for all variables incorporated in the original model. We carried out a third sensitivity analysis to account for the use of non-fasting glucose values, as we had to approximate the fasting glucose values included in the models by the non-fasting glucose values in our data. In this analysis, we excluded individuals with a non-fasting glucose of 11.1 mmol/l (n=130), as this cut point is considered as a high blood glucose concentration at which diabetes is suspected especially if it is accompanied by the classic symptoms of hyperglycaemia 4. In another sensitivity analysis, we excluded 19,295 individuals (including 537 incident cases of diabetes) with fasting period of under two hours. In a fifth sensitivity analysis we excluded 255 individuals for whom we had no verification information of diabetes status. All statistical analyses were conducted with SPSS version 18 (SPSS, Chicago, IL) and R version (Vienna, Austria) for Windows ( Results Systematic literature search We scanned 7756 titles and selected 134 abstracts for review. Figure 1 depicts the flow of the study selection process. We selected 46 articles for full text review and added six that were identified from other sources such as recent systematic reviews 7-9. After full review of these 52 articles, we excluded 36 as they did not meet all inclusion criteria (appendix 3). The main reasons for exclusion were no prediction of the future risk of diabetes (n=15); validation study (n=10); no formal prediction models provided (n=6); and incomparable derivation populations (n=2) or unavailable data of predictors (n=3). Of three studies that used data from two hour oral glucose tolerance tests, we excluded two because they were cross sectional and one because it did not provide any prediction model. Table 1 summarises the characteristics of the 16 studies included in this validation study 25, 26, Eleven studies described 34 basic models based on data that can be assessed non-invasively, including demographics, family history of diabetes, measures of obesity, diet, and lifestyle factors, blood pressure, and use of antihypertensive drugs. Of these 34 basic models, 12 models were presented as the final model 8. 24

28 Based on search term (appendix 1) for the risk prediction of type 2 diabetes, 7,756 titles were identified in PubMed. 7,622 articles were excluded in title review (not relevant to the study objectives) 134 abstracts were selected for review. 88 articles were excluded in abstract review (not performed for the risk prediction of diabetes). 6 articles were identified by other sources (reference lists, reviews experts view). 46 abstracts were selected for full article review. 52 articles were fully reviewed. 36 studies were excluded: No prediction of future risk of developing diabetes (n=13) Validation study (n=10) No formal prediction models (n=6) Others (e.g., unavailable data on predictors, incomparable derivation population) (n=7) 16 studies were evaluated for predictive performance. Figure 1: Overview of systematic literature search of studies that derived prediction models for risk of type 2 diabetes Nine studies described 42 extended models including data on one to three conventional biomarkers such as glucose, HbA1c, lipids, uric acid, or γ- glutamyltransferase. Of these 42 extended models, 13 models were presented as the final model. The C statistics in the development datasets ranged from 0.71 for the Atherosclerosis Risk in Communities (ARIC) model to 0.86 for the FINDRISC full model. Only half of the studies reported measures of calibration, and almost all showed good calibration in the development datasets. Table S2 (in appendix) shows the variables that are part of the prediction models. Validation of prediction models Table S1 (in appendix) summarises the baseline characteristics of participants in the EPIC-NL study (for the full cohort, random sample, and incident cases of type 2 diabetes). During a median follow-up of 10.2 years (over 387,000 person years), we observed 924 incident cases (rate of 2.2 per 1000 person years). The observed 5, 7.5, and 10 year risks of incident diabetes were 1.3%, 1.8%, and 2.3%, respectively. 25

29 Table 1 General characteristics of models to predict risk of incident type 2 diabetes included in study Study and prediction Cases/total Ascertainment of incident Statistical risk model sample size diabetes* model Alssema, 2011, Netherlands 25 DETECT-2 844/ Self report, 2 h plasma glucose Wannamethee, 2011, UK 46 BRHS: Simple clinical Fasting bio Non-fasting bio Rathmann, 2010, Germany 35 KORA: Basic Clinical 297/6927 Self report, review of patients notes 91/873 Self report, fasting glucose, non-fasting glucose Chen, 2010, Australia 36 AUSDRISK 362/ 6060 Drug use, fasting glucose, 2 h plasma glucose Rosella, 2010, Canada 37 DPoRT 1410/ Physician diagnosed in survey data Joseph, 2010, Norway 47 Tromsø 522/ Self report, HbA1c, medical record, fasting glucose Kahn, 2009, US 26 ARIC: Basic Enhanced Hippisley-Cox, 2009, UK 38 QDScore / / 9587 Self report, fasting glucose, non-fasting glucose, hospital records, questionnaire General practice computer records Prediction horizon (years) No of predictors Discrimination C statistic (95% CI) Calibration P value 5 Logistic (0.746 to 0.783) Logistic (0.740 to 0.791) (0.793 to 0.840) (0.785 to 0.833) 7.5 Logistic (0.713 to 0.812) (0.801 to 0.887) Logistic (0.76 to 0.81) Weibull 7 M: 0.77 (0.76 to 0.79); F: 0.78 (0.76 to 0.79) < Cox 10 M: 0.87; F: 0.88 NR 15 Weibull (0.69 to 0.73) 0.79 (0.77 to 0.81) 10 Cox 9 M: (0.831 to 0.836); F: (0.850 to 0.856) NR Almost perfect calibration 26

30 Balkau, 2008, France 39 DESIR: Clinical Clinical+bio Wilson, 2007, US 42 Framingham: Model 1 Model 2 Model 3 Continuous 203/ 3814 Drug use, fasting glucose 9 Logistic /3140 Drug use, fasting glucose 7 Logistic Simmons, 2007, UK 41 EPIC-Norfolk 209/ Hospital and general practice resisters, drug use, HbA1c >7% Schulze, 2007, Germany 40 GDRS (EPIC-Potsdam) Lindstrom, 2003, Finland 43 FINDRISC: Concise Full 849/ Self report verified by diagnosing physician 182/4435 Fasting glucose, non-fasting glucose Stern, 2002, US 44 San Antonio, clinical 275/3004 Drug use, fasting glucose, 2 h plasma glucose 8 8 M: 0.733; F: M: 0.850; F: reported M: 0.7; F 0.6 M: 0.8; F: Logistic (0.730 to 0.790) NR 5 Cox to 0.84 NR 10 Logistic Logistic (0.818 to 0.867) >0.20 Von Eckardstein, 2000, Germany 48 PROCAM 200/3737 Self report, fasting glucose 6.3 Logistic (0.780 to 0.806) NR Stern, 1993, US 45 San Antonioreduced model 79/1453 Drug use, fasting glucose, 2 h plasma glucose 8 Logistic 5 NR Excellent calibration reported NR NR 27

31 NR=not reported. DETECT-2=Detection Strategies for Type 2 Diabetes and Impaired Glucose Tolerance, KORA=Cooperative Health Research in Region of Augsburg; BRHS=British Regional Heart Study; AUDSRISK=Australian Type 2 Diabetes Risk Assessment Tool; DPoRT=Diabetes Population Risk Tool; ARIC=Atherosclerosis Risk in Communities; QDScore=diabetes risk algorithm; DESIR=Data from Epidemiological Study on Insulin Resistance Syndrome; EPIC=European Prospective Investigation Into Cancer and Nutrition; GDRS=German Diabetes Risk Score; FINDRISC=Finnish diabetes risk score; PROCAM=Prospective Cardiovascular Münster Study. *Criteria for plasma glucose concentrations were as fasting glucose (7.0 mmol/l (126 mg/dl) and non-fasting or two hour glucose 11.1 mmol/l (200 mg/dl). Hosmer-Lemeshow χ 2. 28

32 Tables 2 and 3 show the performance of the basic models and the extended models, respectively. The basic models performed well in terms of discrimination, with C statistics ranging from 0.74 (95% confidence interval 0.73 to 0.75) to 0.84 (0.82 to 0.85) for the prediction of risk of diabetes at 7.5 years. Similar but slightly higher C statistics were found for the 5 year risk prediction and slightly lower for the 10 year risk prediction of incident diabetes. For the extended models, the discrimination was higher, with C statistics ranging from 0.81 (0.80 to 0.83) to 0.93 (0.92 to 0.94) for the risk at 7.5 years. Similar, but again slightly higher, C statistics were found for the 5 year risk prediction and slightly lower for the 10 year risk prediction of incident diabetes. Both basic and extended models showed a poor calibration based on the Hosmer-Lemeshow test (P<0.001). Except for the EPIC-Norfolk and PROCAM models, all models overestimated the predicted against the observed 7.5 year risk of diabetes by 38.9% to more than 100%. Similarly, all observed to expected ratios were different from 1.0 (tables 2 and 3). The EPIC-Norfolk model underestimated the 7.5 year risk of incident diabetes by 73.9%. Figure S1 (in appendix) shows the calibration plots for the original models. After adjustment for differences in the incidence of diabetes between our cohort and the development populations, all prediction models showed better calibration (figure 2, panels A and B). For some of the models (such as the ARIC basic model) the calibration plot stayed close to the ideal line throughout the risk spectrum, whereas others showed severe overestimation, especially at higher predicted risks (such as Framingham continuous, DESIR, and BRHS models). Compared with the original models, the models adjusted for differences in the incidence of diabetes between the development and validation cohort performed better, with lower Hosmer-Lemeshow statistics, but deviation of calibration from ideal was still significant for all models, except for the KORA basic model (Hosmer-Lemeshow test P=0.17). For the KORA basic model, AUSDRISK, and EPIC-Norfolk model, calibration slopes were close to 1.0, but those were smaller or larger than 1.0 for other logistic regression models (tables 2 and 3). To further investigate the different effect size for each predictor, we compared hazard ratios for predictors between the validation cohort and one development cohort 49 as an example. We used data from the EPIC-Potsdam study 40 because the model was developed in the German cohort of EPIC using Cox proportional-hazards regression. Table S4 (in appendix) presents the hazard ratios of the diabetes predictors incorporated in this risk score compared with those obtained in our validation cohort. The hazard ratios for age, intake of red meat, physical activity, and current heavy smoking differed significantly (P<0.05) between both cohorts. Sensitivity analyses Tables 4 and 5 show the results of sensitivity analyses. Our results using the extrapolation approach for the case cohort design were similar when we looked at C statistics and Hosmer-Lemeshow statistics of 13 basic models obtained from the 29

33 Table 2 Discrimination and calibration of 12 basic models for prediction of risk of incident type 2 diabetes in validation cohort* C statistic (95% CI) Hosmer-Lemeshow χ 2 * Risk prediction model DETECT-2, (0.82 to 0.85) KORA, 2010, basic 0.83 (0.82 to 0.85) BRHS, 2011, 0.80 (0.78 to simple clinical 0.82) AUSDRISK, (0.83 to 0.86) DPoRT, (0.73 to 0.76) ARIC, 2009, basic 0.83 (0.82 to 0.85) QDScore, (0.75 to 0.79) DESIR, 2008, clinical 0.82 (0.80 to 0.84) EPIC-Norfolk, (0.80 to 0.84) EPIC-Potsdam, 0.84 (0.82 to 2007, GDRS 0.85) FINDRISC, 2003: 0.83 (0.82 to Concise 0.85) Full 0.83 (0.81 to At 5 years At 7.5 years At 10 years 0.85) 0.82 (0.81 to 0.84) 0.83 (0.81 to 0.84) 0.79 (0.78 to 0.81) 0.84 (0.82 to 0.85) 0.74 (0.73 to 0.75) 0.83 (0.81 to 0.84) 0.76 (0.74 to 0.78) 0.81 (0.80 to 0.83) 0.81 (0.80 to 0.82) 0.84 (0.82 to 0.85) 0.82 (0.80 to 0.83) 0.82 (0.80 to 0.83) 0.82 (0.81 to 0.83) 0.82 (0.81 to 0.83) 0.79 (0.78 to 0.80) 0.83 (0.82 to 0.84) 0.74 (0.73 to 0.75) 0.82 (0.81 to 0.84) 0.74 (0.72 to 0.76) 0.81 (0.79 to 0.82) 0.81 (0.79 to 0.82) 0.83 (0.82 to 0.84) 0.81 (0.80 to 0.82) 0.81 (0.80 to 0.82) At 5 years At 7.5 years At 10 years O/E ratio at 7.5 years (95% CI) (0.281 to 0.327) (0.468 to 0.545) (0.685 to 0.779) (0.290 to 0.337) (0.056 to 0.065) (0.127 to 0.147) Calibration slope Recalibrated Hosmer-Lemeshow χ 2 at 7.5 years (P value) (<0.001) (0.17) (<0.001) (<0.001) (<0.001) (<0.001) (<0.001) (0.365 to 0.375) (0.232 to (<0.001) 0.270) (3.450 to 4.016) (0.746 to 0.868) (0.746 to 0.868) (0.661 to 0.769) (<0.001) 67.8 (<0.001) (<0.001) O/E=observed to expected. DETECT-2=Detection Strategies for Type 2 Diabetes and Impaired Glucose Tolerance; KORA=Cooperative Health Research in Region of Augsburg; BRHS=British Regional Heart Study; AUDSRISK=Australian Type 2 Diabetes Risk Assessment Tool; DPoRT=Diabetes Population Risk Tool; ARIC=Atherosclerosis Risk in Communities; QDScore=diabetes risk algorithm; DESIR=Data from Epidemiological Study on Insulin Resistance Syndrome; EPIC=European Prospective Investigation Into Cancer and Nutrition; GDRS=German Diabetes Risk Score, FINDRISC Finnish diabetes risk score. *All P< After recalibration. 30

34 Table 3 Discrimination and calibration of 13 extended models for prediction of risk of incident type 2 diabetes in validation cohort* Risk prediction model Hosmer-Lemeshow χ C statistic (95% CI) 2 BRHS, 2011: Fasting bio Non-fasting bio At 5 years At 7.5 years At 10 years At 5 years At 7.5 years At 10 years 0.87 (0.85 to 0.88) 0.82 (0.81 to 0.84) KORA, 2010, clinical 0.94 (0.93 to 0.95) Tromsø, (0.81 to 0.84) ARIC, 2009, enhanced 0.90 (0.88 to 0.91) DESIR, 2008, clinical+bio Framingham, 2007: Model 1 Model 2 Model 3 Continuous 0.89 (0.87 to 0.90) 0.82 (0.81 to 0.84) 0.82 (0.80 to 0.84) 0.83 (0.81 to 0.84) 0.89 (0.88 to 0.90) San Antonio, (0.91 to 0.93) PROCAM, (0.83 to 0.86) San Antonio, (0.89 to 0.92) 0.86 (0.84 to 0.87) 0.81 (0.80 to 0.83) 0.93 (0.92 to 0.94) 0.81 (0.80 to 0.83) 0.89 (0.87 to 0.90) 0.88 (0.87 to 0.89) 0.82 (0.80 to 0.83) 0.81 (0.80 to 0.83) 0.82 (0.81 to 0.83) 0.88 (0.87 to 0.89) 0.91 (0.90 to 0.92) 0.84 (0.83 to 0.85) 0.89 (0.88 to 0.90) 0.86 (0.84 to 0.87) 0.81 (0.80 to 0.82) 0.92 (0.91 to 0.93) 0.81 (0.80 to 0.82) 0.88 (0.87 to 0.89) 0.88 (0.87 to 0.89) 0.81 (0.80 to 0.83) 0.81 (0.80 to 0.83) 0.82 (0.81 to 0.83) 0.88 (0.86 to 0.89) 0.90 (0.89 to 0.91) 0.83 (0.82 to 0.84) 0.88 (0.87 to 0.90) O/E at 7.5 years (95% CI) (0.794 to 0.924) (0.334 to 0.389) Calibration slope Recalibrated Hosmer- Lemeshow χ 2 at 7.5 years (0.507 to 0.590) (0.055 to 0.064) (0.159 to 0.185) (0.153 to 0.178) (0.451 to 0.525) (0.441 to 0.513) (0.391 to 0.456) (0.079 to 0.092) (0.104 to 0.122) ( to ) (0.174 to 0.202) O/E=observed to expected; BRHS=British Regional Heart Study; KORA=Cooperative Health Research in Region of Augsburg; ARIC=Atherosclerosis Risk in Communities; DESIR=Data from Epidemiological Study on Insulin Resistance Syndrome; PROCAM=Prospective Cardiovascular Münster Study. *Extrapolation approach applied for case cohort design. All P< After recalibration. 31

35 Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax DETECT-2 Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q KORA (Basic) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q BRHS, simple clinical Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q AUSDRISK Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Recalibration: Panel A 32

36 DPoRT Predicted probability Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax ARIC (Basic) Ideal Nonparametric Predicted Risk (7.5yr) Observed probability Observed Risk (7.5yr) QDScore Predicted probability Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q DESIR (Clinical) Ideal Nonparametric Predicted Risk (7.5yr) Observed probability Observed Risk (7.5yr)

37 Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax EPIC-Norfolk Ideal Nonparametric Predicted Risk (7.5yr) GDRS Predicted probability Observed Risk (7.5yr) Observed probability Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q FINDRISC (Full) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr)

38 Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax KORA (Clinical) Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q ARIC (Enhanced) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q BRHS, fasting bio (2011) Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q BRHS, non-fasting bio (2011) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Recalibration: Panel B 35

39 TROMO Predicted probability Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q DESIR (Clinical+bio) Ideal Nonparametric Predicted Risk (7.5yr) Observed probability Observed Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q Framingham (Continuous ) Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax San Antonio (2002) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr)

40 Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax PROCAM Observed Risk (7.5yr) Ideal Nonparametric Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax San Antonio (1993) Observed Risk (7.5yr) Ideal Nonparametric Predicted Risk (7.5yr) Predicted Risk (7.5yr) Figure 2. Calibration plots for the 7.5-year risk of diabetes depicting the predicted risk against observed risk of developing type 2 diabetes in the validation data set. Panel A (the recalibrated basic models), Panel B (the recalibrated extended models). In those plots, depicting the ideal and nonparametric terms, the dashed line (the 45 line) from zero denotes ideal calibration line (slope=1, intercept=0) and the dotted lines (or the solid lines in other plots) denotes smooth calibration curve for each models. 37

41 Table 4 Performance of 12 basic models for prediction of risk of incident type 2 diabetes in sensitivity analyses Risk prediction C statistic (95% CI) at 7.5 years model Case cohort design in extrapolated dataset* Random glucose <11.1 mmol/l Verified diabetes status C statistic (95% CI) for original prediction horizon DETECT-2, (0.81 to 0.84) 0.82 (0.81 to 0.84) 0.83 (0.81 to 0.84) KORA, 2010, basic 0.82 (0.81 to 0.84) 0.82 (0.81 to 0.84) 0.83 (0.81 to 0.84) BRHS, 2011, 0.79 (0.77 to 0.80) 0.79 (0.77 to 0.80) 0.79 (0.78 to 0.81) 0.79 (0.78 to 0.81) simple clinical AUSDRISK, (0.82 to 0.85) 0.83 (0.81 to 0.84) 0.83 (0.82 to 0.85) DPoRT, (0.72 to 0.75) 0.74 (0.72 to 0.75) 0.74 (0.72 to 0.76) 0.74 (0.73 to 0.75) ARIC, 2009, basic 0.82 (0.81 to 0.84) 0.83 (0.81 to 0.84) 0.83 (0.82 to 0.84) 0.82 (0.81 to 0.84) QDScore, (0.74 to 0.78) 0.76 (0.74 to 0.79) 0.76 (0.74 to 0.78) DESIR, 2008, 0.80 (0.79 to 0.82) 0.81 (0.79 to 0.82) 0.81 (0.80 to 0.83) 0.81 (0.79 to 0.82) clinical EPIC-Norfolk, 0.81 (0.79 to 0.82) 0.81 (0.79 to 0.82) 0.81 (0.80 to 0.83) 2007 EPIC-Potsdam, 0.83 (0.82 to 0.85) 0.83 (0.82 to 0.85) 0.84 (0.82 to 0.85) 2007, GDRS FINDRISC, 2003: Concise Full 0.81 (0.80 to 0.83) 0.81 (0.79 to 0.82) 0.81 (0.80 to 0.83) 0.81 (0.80 to 0.83) 0.82 (0.80 to 0.83) 0.82 (0.80 to 0.83) DETECT-2=Detection Strategies for Type 2 Diabetes and Impaired Glucose Tolerance, KORA=Cooperative Health Research in Region of Augsburg; BRHS=British Regional Heart Study; AUDSRISK=Australian Type 2 Diabetes Risk Assessment Tool; DPoRT=Diabetes Population Risk Tool; ARIC=Atherosclerosis Risk in Communities; QDScore=diabetes risk algorithm; DESIR=Data from Epidemiological Study on Insulin Resistance Syndrome; EPIC=European Prospective Investigation Into Cancer and Nutrition; GDRS=German Diabetes Risk Score; FINDRISC=Finnish diabetes risk score. *All Hosmer-Lemeshow χ 2 P< At 7 years. At 9 years. At 15 years. 38

42 Table 5 Performance of 13 extended models for prediction of risk of incident type 2 diabetes in sensitivity analyses* Risk prediction model C statistic (95% CI) at 7.5 years BRHS, 2011: Fasting bio Non-fasting bio Random glucose <11.1 mmol/l 0.86 (0.84 to 0.88) 0.81 (0.80 to 0.83) MORGEN dataset Fasting for >2 hours 0.89 (0.87 to 0.91) 0.84 (0.82 to 0.86) 0.86 (0.84 to 0.88) 0.80 (0.79 to 0.82) C statistic (95% CI) for original prediction horizon 0.86 (0.84 to 0.87) 0.81 (0.80 to 0.83) KORA, 2010, clinical 0.93 (0.92 to 0.94) 0.92 (0.90 to 0.93) 0.93 (0.92 to 0.94) Tromsø, (0.80 to 0.83) 0.84 (0.83 to 0.86) 0.81 (0.79 to 0.83) 0.81 (0.80 to 0.82) ARIC, 2009 enhanced 0.88 (0.87 to 0.89) 0.91 (0.89 to 0.92) 0.90 (0.88 to 0.91) 0.88 (0.87 to 0.89) DESIR, 2008, 0.87 (0.86 to 0.88) 0.89 (0.87 to 0.91) 0.90 (0.88 to 0.91) 0.88 (0.87 to 0.89) clinical+bio Framingham, 2007: Model 1 Model 2 Model 3 Continuous 0.82 (0.80 to 0.83) 0.82 (0.80 to 0.83) 0.82 (0.81 to 0.84) 0.87 (0.86 to 0.88) 0.79 (0.77 to 0.81) 0.79 (0.76 to 0.81) 0.80 (0.77 to 0.82) 0.89 (0.87 to 0.91) 0.82 (0.80 to 0.84) 0.82 (0.80 to 0.84) 0.82 (0.80 to 0.84) 0.90 (0.88 to 0.91) 0.81 (0.80 to 0.83) 0.81 (0.80 to 0.83) 0.82 (0.81 to 0.83) 0.88 (0.87 to 0.89) San Antonio, (0.89 to 0.91) 0.91 (0.87 to 0.92) 0.92 (0.91 to 0.93) PROCAM, (0.82 to 0.85) 0.87 (0.85 to 0.88) 0.83 (0.81 to 0.85) 0.84 (0.83 to 0.85)** San Antonio, (0.87 to 0.90) 0.89 (0.86 to 0.91) 0.91 (0.90 to 0.92) 0.89 (0.87 to 0.90) BRHS=British Regional Heart Study; KORA=Cooperative Health Research in Region of Augsburg; ARIC=Atherosclerosis Risk in Communities; DESIR=Data from Epidemiological Study on Insulin Resistance Syndrome; PROCAM=Prospective Cardiovascular Münster Study. *Extrapolation approach applied for case cohort design. At 7 years. At 10.8 years. At 15 years. At 9 years. **At 6.3 years. At 8 years. 39

43 extrapolation approach compared with those from the full cohort design (for example, C statistics ranging from 0.74 (95% confidence interval 0.72 to 0.76) to 0.84 (0.82 to 0.86), and Hosmer-Lemeshow test P<0.001). Additionally, our results using data only from the MORGEN cohort with less than 10% missing values for nonfasting glucose were comparable with our results using both cohorts; C statistics ranged from 0.79 (0.76 to 0.81) to 0.92 (0.90 to 0.93) for 13 extended models. Exclusion of individuals with a non-fasting glucose of 11.1 mmol/l did not influence the results, both for the basic (C statistics ranged from 0.74 (0.72 to 0.75) to 0.83 (0.81 to 0.84)) and the extended models (C statistics ranged from 0.81 (0.80 to 0.83) to 0.93 (0.92 to 0.94)). Moreover, when we excluded the individuals with less than two hours fasting or those without verified diabetes status, the C statistics were similar to those of the full cohort analysis. Finally, use of the prediction horizon for which the original models were developed hardly affected the results. Discussion An evaluation of the performance of 25 prediction models for type 2 diabetes in an independent Dutch cohort with over 10 years of follow-up showed that basic models perform similarly well in identifying individuals at high and low risk of developing diabetes. The performance was slightly better for extended models that included conventional biomarkers. With regard to the actual values of the predicted risks, all but two models overestimated the risk of developing diabetes, which improved slightly, but not sufficiently, after correction of the models for differences in incidence of diabetes between development and validation populations. Strengths and limitations of the study All models were identified through a systematic literature search, and we included most existing prediction models in the validation study. Other strengths included the study s large sample size, prospective design, verification of incident diabetes, and extensive information on individuals characteristics. Nevertheless, some limitations of our study need to be mentioned. Nearly all participants in the EPIC-NL cohort are white adults, and further studies are warranted to validate the models in other populations. In addition, the participation rate was about 40% 19, 50. We previously showed that such a low response rate might affect prevalence estimates of baseline characteristics of participants but does not cause bias in the examined associations 50. We therefore consider that our cohort is appropriate for the purpose of our study. Although our data had certain limitations regarding availability of the variables, we made an effort to assign all variables and applied definitions as closely as possible. To handle missing variables, we performed single imputation and repeated the analysis in one of the two cohorts with lower missing values for glucose concentration, which gave similar results. It is therefore unlikely that these limitations influenced our results to a large extent. Next, we used data for non-fasting glucose concentration. We cannot rule out that this affected our results because glucose is an important predictor 40

44 of diabetes. We therefore performed sensitivity analyses in which we excluded individuals with a non-fasting glucose of 11.1 mmol/l 4 and those who fasted for less than two hours, which again yielded similar results. This is in line with previous studies showing that using non-fasting lipid concentrations does not influence prediction of, for example, cardiovascular events 51, 52. Because we used data only from verified potential cases we could have missed false negative cases in the remainder of the cohort as type 2 diabetes can remain undiagnosed for several months to years. False negatives can lead to an underestimation of the C statistic as the linear predictor resulting from the predictor variables will be high, whereas their event status is that of a non-case. Given the large size of our cohort in combination with the low incidence of diabetes we do not expect this to largely change our findings. Similarly, false negative cases lead to underestimation of the observed risk in our cohort and this influences calibration. We adjusted for this effect, however, by correcting the intercept of the models to the incidence observed in our cohort. In addition, as the incidence is expected to be low 53, potential false negative cases cannot account for the large overestimations of risk in the models observed in our study. Moreover, certain development cohorts used similar methods for verification of diabetes. External validation of prediction models The retrieved prediction models differed considerably in terms of type and number of predictors, age ranges, type of model, duration of follow-up, and outcome measure. Three recent systematic reviews presented overviews of studies that developed these models or validated some selected models 7-9. These reviews, however, also indicated that most of these models were never validated in an external population. Our study has now evaluated performance of most developed prediction models for future diabetes in an external population and shows that most basic models perform well to identify those at high risk of diabetes and that extended models perform slightly better. Generally, the performance of a prediction model decreases when it is applied in a validation dataset. Despite this, our study showed that most of the basic models identified those at high absolute risk well, with C statistic over This discrimination further improved for the extended models with C statistic of about Surprisingly, the C statistics in our validation study were, in some cases, even higher than in their original development populations. This might be explained by differences in heterogeneity between the populations 30 : larger heterogeneity between individuals in a validation study can in some situations lead to a higher C statistic than in the development study. For example, variables like age, sex, and BMI might have larger heterogeneity in our study compared with the older population of the KORA study 35. Although it would be of interest to explore whether performance of diabetes risk scores differs by age or sex, larger studies are warranted for these subgroup analyses. Another aspect that could influence model performance is the type of regression analysis used to derive the prediction model 7. Most studies used logistic regression rather than survival models 7, 8 and therefore do not account for 41

45 censoring 54. Similar to the results of the Framingham Offspring Study 42, however, our results showed that the survival models do not necessarily perform better than the logistic ones. Quantification of actual risk of future diabetes All except two prediction models overestimated the absolute risk of diabetes in our validation dataset, which can partly be explained by the difference in incidence of diabetes between development and validation populations. To account for this, we adjusted the models for difference in incidence, resulting in much better calibration. Significant deviations between the predicted and observed risks, however, remained for most models. There are various other explanations for the deviation in predicted versus observed risks. Firstly, in large cohorts the Hosmer-Lemeshow test is sensitive to small differences between the predicted and observed risks, so calibration can be indicated as significantly deviant by statistical tests even when the calibration plots indicate good calibration based on visual inspection and for practical purposes 55. So, in large cohorts significant deviations on the Hosmer-Lemeshow test should be interpreted cautiously. Secondly, mis calibration can be caused by differences in how certain predictors, the outcome variable, or baseline characteristics of the study populations are measured, which can lead to different predictive effects 13, 30. For example, if the two hour oral glucose tolerance test is used to determine the presence of diabetes in a population, the incidence is likely to be higher, and among the cases there will be patients with a less severe form of the disease and different values for the potential predictors. This is also illustrated by comparing the effect sizes of the predictors of the German Diabetes Risk Score in our cohort, which showed significant differences for important predictors like age. It is important to note, however, that most prediction models showed overprediction, particularly at higher absolute risk. Some models might not have been well calibrated in the original populations 7, 9. Furthermore, the overestimation of risk at the higher end could be caused by overestimation of certain predictors in development populations with high risk individuals. Although it is important to accurately estimate the risk for people at high risk, it might not directly influence the effects of screening and public health strategies: interventions are often initiated beyond a certain threshold of absolute risk and overprediction beyond this threshold might therefore not necessarily lead to different treatment decisions. Certain models in our study, however, also overestimated the absolute risk in the lower ranges around 10%. Although decision thresholds for type 2 diabetes have not been determined, this prediction could be in the range of a threshold for a clinical decision. To use such models in clinical practice, calibration needs to be further improved. Prior external validation of existing prediction models Although the importance of external validation of prediction models is now widely acknowledged, only a quarter of existing prediction models have been externally validated, mostly in studies including only a single model and not reporting any 42

46 measures of calibration 7, 9. To date, four studies have been published that performed a comparative external validation of several different models 10, 11, 17, 56. Two of these studies validated models for presence of diabetes rather than future risk of diabetes 11, 56. One prospective validation of three extended models 42, 44, 57 has been performed and showed C statistics ranging from 0.78 to 0.84 with underestimation or overestimation of the risk 10. Another prospective validation study showed C statistics ranging from 0.74 to 0.90, without reporting calibration and performing adjustments 17. These results are in line with the discrimination observed in our study. Altogether, the results from the previous reviews and our study suggest that most of the basic models performed similarly in terms of discrimination, whereas the Diabetes Population Risk Tool (DPoRT) showed slightly lower discrimination. The latter model was primarily developed to predict risk of diabetes at a population level, which could explain its slightly worse performance when it is applied on an individual level 37. Implications for use of prediction models in practice Results from our study show that prediction models perform well to identify those at high risk of future diabetes, being a first prerequisite for use of such models in practice as currently recommended 5, 6. As expected 18, 30 and observed in our results, however, the model should possibly be adapted to the local setting and purpose of the model and at least corrected for the incidence of diabetes of the population in which it is to be applied. The main relevance of prediction models is to correctly identify individuals at high risk, while avoiding the burden of treatment for individuals at low risk. This requires adequate discriminative power in the general population, as well as in populations characterised by a somewhat higher risk, such as those with excess weight. In public health practice, one would perhaps prefer to use a model including only a limited number of predictors based on non-invasive tests with the highest performance, which would favour use of a basic model. Noble et al 18 [8] suggested seven models as most promising for use in clinical or public health practice, of which three were extended models (ARIC enhanced, Framingham, and San Antonio) 42, 44, 57 and four were basic models (AUSDRISK, QDScore, FINDRISC, and Cambridge Risk Score) 36, 38, 43, 58, 59. According to the current validation, it seems that this judgment is likely to be correct in statistical terms. The basic DESIR model that we additionally evaluated consisted of four predictors 39, while most models such as QDScore and AUSDRISK consist of seven to 10 predictors. Interestingly, the models including only four to six predictors35, 39, 43 performed similarly to the more extensive models 36, 38. We found that discrimination of two other basic models KORA basic 35 and DESIR clinical equation 39 which were not included in the list of Noble and colleagues 8, approximated performance of the models incorporating more predictors. Moreover, the KORA basic model performed sufficiently to quantify absolute risk after recalibration. This suggests that a basic model like the KORA, which uses a limited set of non-invasive predictors, already provides good discrimination and good calibration and could therefore be useful in 43

47 practice after appropriate adaptation of the model to the setting. The extended models including biomarkers could then perhaps be used only for those at high risk based on a basic prediction model. Finally, a model developed in one setting (such as public health data) or in a particular country does not necessarily need to be useful in another setting (such as secondary care) or country. As a next step, the utility of such models needs to be further investigated in clinical and public health practice. In conclusion, most of the basic prediction models including data on noninvasive variables performed well to identify those at high risk of developing type 2 diabetes in an independent population. The discriminative performance was slightly better for the extended models with additional data on conventional biomarkers. Most models, however, overestimated the actual risk of diabetes. Whether this influences treatment decisions needs to be further investigated. Hence, existing prediction models, even with only limited information, are valid tools to identify those at high risk but do not perform well enough to quantify the actual risk of future diabetes. 44

48 Acknowledgments We thank Statistics Netherlands and the PHARMO Institute for follow-up data on cancer, cardiovascular disease and vital status. This study was funded by the Netherlands Heart Foundation, the Dutch Diabetes Research Foundation and the Dutch Kidney Foundation, the Center for Translational Molecular Medicine (project PREDICCt, grant 01C ), Europe against Cancer Programme of the European Commission (SANCO), the Dutch Ministry of Health, the Dutch Cancer Society, the Netherlands Organization for Health Research and Development (ZonMW), and World Cancer Research Fund (WCRF), and the Netherlands Organization for Scientific Research project ( and ). None of the study sponsors had a role in the study design, data collection, analysis and interpretation, report writing, or the decision to submit the report for publication. 45

49 References 1. Wild S, Roglic G, Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the year 2000 and projections for Diabetes Care 2004;27(5): Knowler WC, Barrett-Connor E, Fowler SE, et al. Reduction in the incidence of type 2 diabetes with lifestyle intervention or metformin. N Engl J Med 2002;346(6): Diabetes Prevention Program Research Group, Fowler SE, Hamman RF, et al. 10-year follow- up of diabetes incidence and weight loss in the Diabetes Prevention Program Outcomes Study. Lancet 2009;374(9702): American Diabetes Association. Standards of medical care in diabetes Diabetes Care 2011;34 Suppl 1:S Paulweber B, Valensi P, Lindstrom J, et al. A European evidence-based guideline for the prevention of type 2 diabetes. Horm Metab Res 2010;42 Suppl 1:S Alberti KG, Zimmet P, Shaw J. International Diabetes Federation: a consensus on Type 2 diabetes prevention. Diabet Med 2007;24(5): Collins GS, Mallett S, Omar O, Yu LM. Developing risk prediction models for type 2 diabetes: a systematic review of methodology and reporting. BMC Med 2011;9: Noble D, Mathur R, Dent T, Meads C, Greenhalgh T. Risk models and scores for type 2 diabetes: systematic review. BMJ 2011;343:d Buijsse B, Simmons RK, Griffin SJ, Schulze MB. Risk assessment tools for identifying individuals at risk of developing type 2 diabetes. Epidemiol Rev 2011;33(1): Mann DM, Bertoni AG, Shimbo D, et al. Comparative validity of 3 diabetes mellitus risk prediction scoring models in a multiethnic US cohort: the Multi-Ethnic Study of Atherosclerosis. Am J Epidemiol 2010;171(9): Lin JW, Chang YC, Li HY, et al. Cross-sectional validation of diabetes risk scores for predicting diabetes, metabolic syndrome, and chronic kidney disease in Taiwanese. Diabetes Care 2009;32(12): Moons KG, Royston P, Vergouwe Y, Grobbee DE, Altman DG. Prognosis and prognostic research: what, why, and how? BMJ 2009;338:b Altman DG, Vergouwe Y, Royston P, Moons KG. Prognosis and prognostic research: validating a prognostic model. BMJ 2009;338:b Collins GS, Moons KG. Comparing risk prediction models. BMJ 2012;344:e Siontis GC, Tzoulaki I, Siontis KC, Ioannidis JP. Comparisons of established risk prediction models for cardiovascular disease: systematic review. BMJ 2012;344:e Reilly BM, Evans AT. Translating clinical research into clinical practice: impact of using prediction rules to make decisions. Ann Intern Med 2006;144(3): Schmid R, Vollenweider P, Bastardot F, Waeber G, Marques-Vidal P. Validation of 7 type 2 diabetes mellitus risk scores in a population-based cohort: CoLaus study. Arch Intern Med 2012;172(2): Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010;21(1): Beulens JW, Monninkhof EM, Verschuren WM, et al. Cohort profile: the EPIC-NL study. Int J Epidemiol 2010;39(5): Liberati A, Altman DG, Tetzlaff J, et al. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration. PLoS Med 2009;6(7):e von Elm E, Altman DG, Egger M, Pocock SJ, Gotzsche PC, Vandenbroucke JP. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. BMJ 2007;335(7624): Volovics A vdbp. Methods for the Analyses of Case-Cohort Studies. Biom J 1997;39:

50 23. Ocke MC, Bueno-de-Mesquita HB, Goddijn HE, et al. The Dutch EPIC food frequency questionnaire. I. Description of the questionnaire, and relative validity and reproducibility for food groups. Int J Epidemiol 1997;26 Suppl 1:S Sluijs I, van der AD, Beulens JW, et al. Ascertainment and verification of diabetes in the EPIC- NL study. Neth J Med 2010;68(1): Alssema M, Vistisen D, Heymans MW, et al. The Evaluation of Screening and Early Detection Strategies for Type 2 Diabetes and Impaired Glucose Tolerance (DETECT-2) update of the Finnish diabetes risk score for prediction of incident type 2 diabetes. Diabetologia 2011;54(5): Kahn HS, Cheng YJ, Thompson TJ, Imperatore G, Gregg EW. Two risk-scoring systems for predicting incident diabetes mellitus in U.S. adults age 45 to 64 years. Ann Intern Med 2009;150(11): Harrell FE, Jr., Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996;15(4): van Houwelingen HC. Validation, calibration, revision and combination of prognostic survival models. Stat Med 2000;19(24): Liddell FD. Simple exact analysis of the standardised mortality ratio. J Epidemiol Community Health 1984;38(1): Vergouwe Y, Moons KG, Steyerberg EW. External validity of risk models: Use of benchmark values to disentangle a case-mix effect from incorrect coefficients. Am J Epidemiol 2010;172(8): Janssen KJ, Moons KG, Kalkman CJ, Grobbee DE, Vergouwe Y. Updating methods improved the performance of a clinical prediction model in new patients. J Clin Epidemiol 2008;61(1): Sterne JA, White IR, Carlin JB, et al. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. BMJ 2009;338:b Donders AR, van der Heijden GJ, Stijnen T, Moons KG. Review: a gentle introduction to imputation of missing values. J Clin Epidemiol 2006;59(10): Marshall A, Altman DG, Royston P, Holder RL. Comparison of techniques for handling missing covariate data within prognostic modelling studies: a simulation study. BMC Med Res Methodol 2010;10: Rathmann W, Kowall B, Heier M, et al. Prediction models for incident type 2 diabetes mellitusin the older population: KORA S4/F4 cohort study. Diabet Med 2010;27(10): Chen L, Magliano DJ, Balkau B, et al. AUSDRISK: an Australian Type 2 Diabetes Risk Assessment Tool based on demographic, lifestyle and simple anthropometric measures. Med J Aust 2010;192(4): Rosella LC, Manuel DG, Burchill C, Stukel TA. A population-based risk algorithm for the development of diabetes: development and validation of the Diabetes Population Risk Tool (DPoRT). J Epidemiol Community Health 2011;65(7): Hippisley-Cox J, Coupland C, Robson J, Sheikh A, Brindle P. Predicting risk of type 2 diabetes in England and Wales: prospective derivation and validation of QDScore. BMJ 2009;338:b Balkau B, Lange C, Fezeu L, et al. Predicting diabetes: clinical, biological, and genetic approaches: data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR). Diabetes Care 2008;31(10): Schulze MB, Hoffmann K, Boeing H, et al. An accurate risk score based on anthropometric, dietary, and lifestyle factors to predict the development of type 2 diabetes. Diabetes Care 2007;30(3): Simmons RK, Harding AH, Wareham NJ, Griffin SJ. Do simple questions about diet and physical activity help to identify those at risk of Type 2 diabetes? Diabet Med 2007;24(8): Wilson PW, Meigs JB, Sullivan L, Fox CS, Nathan DM, D'Agostino RB, Sr. Prediction of incident diabetes mellitus in middle-aged adults: the Framingham Offspring Study. Arch Intern Med 2007;167(10): Lindstrom J, Tuomilehto J. The diabetes risk score: a practical tool to predict type 2 diabetes risk. Diabetes Care 2003;26(3):

51 44. Stern MP, Williams K, Haffner SM. Identification of persons at high risk for type 2 diabetes mellitus: do we need the oral glucose tolerance test? Ann Intern Med 2002;136(8): Stern MP, Morales PA, Valdez RA, et al. Predicting diabetes. Moving beyond impaired glucose tolerance. Diabetes 1993;42(5): Wannamethee SG, Papacosta O, Whincup PH, et al. The potential for a two-stage diabetes risk algorithm combining non-laboratory-based scores with subsequent routine non-fasting blood tests: results from prospective studies in older men and women. Diabet Med 2011;28(1): Joseph J, Svartberg J, Njolstad I, Schirmer H. Incidence of and risk factors for type-2 diabetes in a general population: the Tromso Study. Scand J Public Health 2010;38(7): von Eckardstein A, Schulte H, Assmann G. Risk for diabetes mellitus in middle-aged Caucasian male participants of the PROCAM study: implications for the definition of impaired fasting glucose by the American Diabetes Association. Prospective Cardiovascular Munster. J Clin Endocrinol Metab 2000;85(9): D'Agostino RB, Sr., Grundy S, Sullivan LM, Wilson P. Validation of the Framingham coronary heart disease prediction scores: results of a multiple ethnic groups investigation. JAMA 2001;286(2): Van Loon AJ, Tijhuis M, Picavet HS, Surtees PG, Ormel J. Survey non-response in the Netherlands: effects on prevalence estimates and associations. Ann Epidemiol 2003;13(2): Herder C, Baumert J, Zierer A, et al. Immunological and cardiometabolic risk factors in the prediction of type 2 diabetes and coronary events: MONICA/KORA Augsburg case-cohort study. PLoS One 2011;6(6):e van Dieren S, Nothlings U, van der Schouw YT, et al. Non-fasting lipids and risk of cardiovascular disease in patients with diabetes mellitus. Diabetologia 2011;54(1): Langenberg C, Sharp S, Forouhi NG, et al. Design and cohort description of the InterAct Project: an examination of the interaction of genetic and lifestyle factors on the incidence of type 2 diabetes in the EPIC Study. Diabetologia 2011;54(9): Harrell FE, Jr., Lee KL, Califf RM, Pryor DB, Rosati RA. Regression modelling strategies for improved prognostic prediction. Stat Med 1984;3(2): McGeechan K, Macaskill P, Irwig L, Liew G, Wong TY. Assessing new biomarkers and predictive models for use in clinical practice: a clinician's guide. Arch Intern Med 2008;168(21): Rathmann W, Martin S, Haastert B, et al. Performance of screening questionnaires and risk scores for undiagnosed diabetes: the KORA Survey Arch Intern Med 2005;165(4): Schmidt MI, Duncan BB, Bang H, et al. Identifying individuals at high risk for diabetes: The Atherosclerosis Risk in Communities study. Diabetes Care 2005;28(8): Griffin SJ, Little PS, Hales CN, Kinmonth AL, Wareham NJ. Diabetes risk score: towards earlier detection of type 2 diabetes in general practice. Diabetes Metab Res Rev 2000;16(3): Rahman M, Simmons RK, Harding AH, Wareham NJ, Griffin SJ. A simple risk score identifies individuals at high risk of developing Type 2 diabetes: a prospective cohort study. Fam Pract 2008;25(3):

52 Appendix Text S1. Brief description of the measurements The general questionnaire contained questions on demographic characteristics and risk factors for the presence of chronic diseases. Body weight, height and waist and hip circumference were measured according to standard procedures. Smoking status was categorized into current, past and never smoker. Blood pressure was measured twice on the left arm. The mean of the two blood pressure measurements was used in the analysis 1. In the Prospect study, systolic and diastolic blood pressures were measured with the participants in the supine position using a Boso Oscillomat (Bosch & Sohn, Jungingen, Germany), whereas a random-zero sphygmomanometer (Hawksley & Sons, Lancing, UK) with the participant in the sitting position was used in the MORGEN cohort. The comparability of these different measurement procedures has been described in more detail previously 2. The assessment of the Prospect cohort slightly overestimated blood pressure compared with the MORGEN cohort. Hypertension was defined based on self-report of diagnosis by a physician, measured hypertension ( 140 mmhg systolic blood pressure or 90 mmhg diastolic blood pressure) or the use of blood pressure-lowering medication. In both cohorts, daily food intake was determined using the same validated FFQ 3, 4, which contains questions on the usual frequency of consumption of 79 main food groups during the year preceding enrolment. Overall, the questionnaire enables estimation of the average daily consumption of 178 foods. Non-fasting blood samples were collected at baseline from all participants. Blood samples were fractionated into aliquots and stored at -196 C, for future use. HbA1c was measured in erythrocytes using an immunoturbidimetric latex test. Biomarkers were assessed in EDTA (in MORGEN cohort) or citrate (in Prospect cohort) plasma. We compared EDTA and citrate measurements and validated these against serum in a sample of 50 participants, observing good correlation in this validation 1. Gamma-glutamyltransferase (GGT), total cholesterol, triglycerides, glucose and uric acid were measured using enzymatic methods, whereas hscrp was measured with a turbidimetric method. LDL-cholesterol was measured using a homogeneous assay with enzymatic endpoint. These assays were all performed on an autoanalyser (LX20, Beckman Coulter, Mijdrecht, the Netherlands). Technicians were blinded to the participants characteristics. Table S1. Baseline participants characteristics in the EPIC-NL study* Variables Full cohort Random sample Cases with incident type 2 diabetes No. of individuals 38,379 2, Age yr 49.1 (11.9) 49.2 (11.9) 56.6 (7.3) Female gender no. (%) (74.3) 1872 (74.7) 722 (78.1) Parental history of diabetes no. (%) 7379 (19.2) 498 (19.9) 377 (40.8) History of cardiovascular disease no. (%) 1075 (2.8) 60 (2.4) 69 (7.5) 49

53 Hypertension no. (%) (36.8) 953 (38.0) 626 (67.7) Antihypertensive medication no. (%) 3736 (9.7) 259 (10.3) 275 (29.8) Current smoker no. (%) Exsmoker no. (%) Heavy smoker no. (%) (30.6) (31.3) 4070 (10.6) 789 (31.5) 769 (30.7) 250 (10.0) 322 (34.8) 241 (26.1) 100 (10.8) Alcohol use g/d 11.1 (15.5) 11.3 (15.9) 7.9 (13.2) Non- drinker no. (%) (27.7) 697 (27.8) 422 (45.7) Weight kg 72.7 (12.9) 72.8 (13.0) 82.3 (14.5) Height cm (8.9) (8.6) (8.3) Body-mass index 25.6 (4.0) 25.7 (4.0) 29.9 (4.7) Waist cimrcumfernce cm 85.1 (11.4) 85.3 (11.6) 97.0 (11.6) Systolic blood pressure mm Hg (18.8) (18.6) (21.6) Pulse rate beats/min 73.4 (10.8) 73.1 (10.7) 75.4 (11.8) Physical activity h/wk 18.2 (15.1) 18.0 (15.0) 15.4 (13.2) Coffee g/d (308.5) (304.1) (317.6) Red meat g/d 89.2 (51.3) 89.7 (51.5) 95.0 (51.2) Whole bread g/d 64.3 (65.7) 62.9 (65.1) 57.3 (64.1) Eat fruit g/d (167.9) (173.0) (166.5) Eat vegetable g/d (54.7) (55.1) (53.5) Biomarkers Glucose mmol/liter 4.9 (1.2) 4.89 (1.17) 6.75 (2.48) HbA1c % (0.58) 6.5 (1.4) HDL-cholesterol mmol/liter (0.35) 1.05 (0.28) Triglyceride mmol/liter (1.01) 2.27 (1.34) Uric acid µmol/liter (68.5) (71.0) GGT U/liter (20.0) 36.6 (28.9) * Data were shown as mean (SD) for continuous variables, and numbers (percentage) for categorical variables. EPIC-NL denotes Dutch contribution of the European Prospective Investigation Into Cancer and Nutrition, HbA1c glycated hemoglobin, HDL high-density lipoprotein, GGT γ- glutamyltranspeptidase. Hypertension was ascertained on the basis of self-reported diagnosis by a physician, antihypertensive medication use, systolic blood pressure 140 mm Hg, or diastolic blood pressure 90 diastolic blood pressure, or a combination of these. Body mass index is the weight in kilograms divided by the square of the height in meters. To convert values for glucose to milligrams per deciliter, divide by To convert values for HDL cholesterol to milligrams per deciliter, divide by To convert values for triglycerides to milligrams per deciliter, divide by To convert values for uric acid to milligrams per deciliter, divide by

54 Table S2. Characteristics of the original prediction models* DETECT BRHS KORA AUSDRISK (2011) (2011) (2010) (2010) DPoRT (2010) Tromsø (2010) ARIC (2009) QDScore (2009) DESIR (2008) EPIC- Norfolk (2007) Predictors C C B C B C C B C B C C B C Age Sex Immigrant status Ethnicity Education Townsend score Waist circumference, Weight Height BMI Obesity Smoking Family history of diabetes Gestational diabetes History of high blood glucose Blood pressure Pulse pressure 51

55 Hypertension Antihypertensive medication Cardiovascular disease Corticosteroids use Resting pulse Alcohol use Coffee Red meat Whole-grain bread Use of vegetables, fruits, or berries Physical activity Glucose level HbA1c Triglyceride level Total Cholesterol level HDL cholesterol level Uric acid level GGT level Risk score n.r. n.r. n.r. n.r. n.r. n.r. Prospective Validation n.r. n.r. n.r. n.r. n.r. n.r. n.r. n.r. n.r. n.r. n.r. n.r. n.r. 52

56 Table S2. Continued Framingham (2007) GDRS (2007) FINDRISC (2003) San Antonio (2002) PROCAM (2000) San Antonio (1993) Predictors B C C B B B Age Sex Immigrant status Ethnicity Education Townsend score Waist circumference, Weight Height BMI Obesity Smoking Family history of diabetes Gestational diabetes History of high blood glucose Blood pressure Pulse pressure 53

57 Hypertension Antihypertensive medication Cardiovascular disease Corticosteroids use Resting pulse Alcohol use Coffee Red meat Whole-grain bread Use of vegetables, fruits, or berries Physical activity Glucose HbA1c Triglyceride HDL cholesterol level Uric acid GGT Risk score n.r. n.r. n.r. Prospective Validation n.r. n.r. * DETECT-2 denotes Detection Strategies for Type 2 Diabetes and Impaired Glucose Tolerance, KORA Cooperative Health Research in the Region of Augsburg, BRHS, British Regional Heart Study, AUDSRISK Australian Type 2 Diabetes Risk Assessment Tool, DPoRT Diabetes Population Risk Tool, ARIC Atherosclerosis Risk in Communities, QDScore diabetes risk algorithm, DESIR Data from the Epidemiological Study on the Insulin Resistance Syndrome, EPIC European Prospective Investigation Into Cancer and Nutrition, GDRS German Diabetes Risk Score, FINDRISC Finnish diabetes risk score, PROCAM, Prospective Cardiovascular Münster Study, C clinical, B biomarker-added, BMI body mass index, HbA1c glycated hemoglobin, HDL high-density lipoprotein, GGT γ - glutamyltranspeptidase and n.r. not reported. Sex specific prediction models have been developed. 54

58 Table S3. Missing data pattern in the EPIC-NL study Missing values Percent Age 0 0 Sex 0 0 Incident type 2 diabetes 0 0 Ethnicity 0 0 Educational level Occupational status Waist circumference Hip circumference Weight Height Body Mass Index Obesity Currently smoke Pack years smoking Family history of diabetes Gestational diabetes 0 0 Systolic blood pressure Diastolic blood pressure History of hypertension Antihypertensive medication 0 0 History of cardiovascular disease Resting pulse Alcohol use Coffee use Intake of red meat Whole bread Intake of vegetables Intake of fruit Physical activity variables Walking in summer Walking in winter Cycling in summer Cycling in winter Physical exercise in summer Physical exercise in winter Gardening in summer Gardening in winter Glucose HbA1c Triglycerides Total cholesterol HDL-cholesterol Uric acid GGT EPIC-NL denotes European Prospective Investigation Into Cancer and Nutrition, Netherlands, HbA1c glycated hemoglobin, HDL highdensity lipoprotein, GGT γ glutamyltranspeptidase. 55

59 Text S2. Use of proxy variables in applying the prediction models The following variables were available in our data similar to the original prediction models: age, gender, parental history of diabetes, history of cardiovascular disease, hypertension, antihypertensive medication, smoking status, alcohol use, weight, height, body mass index, waist circumference, systolic blood pressure, pulse rate, physical activity, glucose, HbA1c, triglyceride, uric acid, and γ glutamyltranspeptidase. For some original variables we replaced the original predictor by an approximation. Fasting glucose was not available in our data set and we therefore used non-fasting glucose levels. The Australian Type 2 Diabetes Risk Assessment Tool (AUSDRISK) and Finish diabetes risk score (FINDRISC) included history of high blood glucose as a predictor, which is unavailable for our cohort. Because the self-reported prevalent cases of diabetes had already been excluded in our study, we set this variable to zero, assuming we already excluded those with a history of high blood glucose. For the Townsend deprivation score that was included in the QDScore, we used a Z-score of a combination of education levels and occupation status as a proxy of social economic status. In the QDScore, systemic corticosteroid use at baseline was used as a predictor. Because this information was only available for the Prospect cohort with only a very small number of individuals (<1%) using corticosteroids, we set this value to zero for the entire cohort. For the nutritional variables, we used data from the FFQ, which provided daily intake of coffee, red meat, fruits, vegetables and whole bread as continuous variables in g/day. Based on average Dutch portion sizes, corresponding categorical variables were created for these foods. Text S3. Brief protocol for the systematic review Inclusion criteria: 1) Predicting type 2 diabetes in the general population 2) Presenting a specific prediction rule/model with sufficient information (beta coefficients of the model or otherwise a scoring system/graph/score card/nomogram) for all variables to calculate the prediction rule in a different population 3) A sample included at least partly Caucasians and a beta for ethnicity in the prediction model 4) Prospective study 5) 'Real' prediction study presenting at least a measure of discrimination of the prediction rule or measure of calibration (Hosmer-Lemeshow statistic) or other measures of performance (e.g., Brier score). 6) Data are available in the EPIC-NL study to validate the rule Exclusion criteria: 1) 2 hour- GTT as one of the major predictive variables 56

60 2) Validation of an existing risk score 3) Others like not finding a proxy variable in the EPIC-NL study unavailable original risk function and incomparable derivation populations 4) Prior review articles 5) References from the prior papers 6) Experts' view (such as evaluation of FINDRISC in the PREVEND, NTVG) 7) Prior validation studies Text S4. Brief description for information provided by validation plot Validation plot provides additional information to examine distance between predicted probability and observed outcome This plot gives a summary of information including: 1) Dxy, Somers D rank correlation index which is calculated as 2(C statistic -0.5); Dxy=0 infers a random prediction, while Dxy=1 infers a perfect discrimination. 2) C statistic; a rank order statistic to indicate discrimination performance of model; The C statistic is identical to the area under the ROC curve for a binary outcome; C statistic=1 infers a perfect discrimination. 3) Nagelkerke R 2 index, the squared multiple correlation coefficient as a measure of explained variation; R 2 =1 infers a perfect prediction. 4) D, discrimination index which is calculated as model χ 2-1 divided by number of individuals, on the log-likelihood scale. 5) U, the Unreliability index; U statistic compares the original intercept and the slope of linear predictor to the perfect calibration line, i.e, slope=1 and intercept=0 6) Q, the overall quality index; this is calculated as difference between discrimination D index and U index. 7) Brier score denotes an average squared difference in predicted probability and observed outcome; varies from 0 for a perfect prediction model to 0.25 noninoformative model with 50% incident outcome. 8) Intercept and Slope, the intercept and the slope of linear predictor (lp) as measure of overall calibration; αcalibration + βcalibration lp. 9) Emax, a maximum absolute difference in predicted and calibrated probabilities; Emax value is close to zero in a well-calibrated model. 57

61 Table S4. Comparison of the effect size for each predictor in the EPIC-Potsdam and the EPIC-NL populations No. of incident diabetes/participants EPIC-NL 499/38,379 EPIC-Potsdam 849/25,167 Predictors Hazard ratio (95% CI) * Waist cimrcumfernce (cm) ( ) ( ) Height (cm) ( ) ( ) Age (years) ( ) ( ) Hypertension (self-report) ( ) ( ) Intake of red meat (each 150 g/day) ( ) ( ) Intake of whole-grain bread (each 150 g/day) ( ) ( ) Consumption of coffee (each 150 g/day) ( ) ( ) Moderate alcohol consumption (10-40 g/day) ( ) ( ) Sport, biking or gardening (h/week) ( ) ( ) Former smoker ( ) ( ) Current heavy smoker ( ) ( ) EPIC-NL denotes Dutch contribution of the European Prospective Investigation Into Cancer and Nutrition (EPIC). * Accordingly, the hazard ratio of each predictor was calculated over a follow-up period of 5 years in the EPIC-NL study. The difference between the HRs was examined by the z statistic tests, where z = (β EPIC-Potsdam - β EPIC-NL)/ (SE 2 EPIC-Potsdam + SE 2 EPIC-NL ). HR was significantly different from that of the EPIC-NL study (P<0.05). Reference S2: Excluded studies in the full review of articles 1. Abdul-Ghani MA, Abdul-Ghani T, Ali N, Defronzo RA. One-hour plasma glucose concentration and the metabolic syndrome identify subjects at high risk for future type 2 diabetes. Diabetes Care 2008;31: Abdul-Ghani MA, Williams K, DeFronzo RA, Stern M. What is the best predictor of future type 2 diabetes? Diabetes Care 2007;30: Baan CA, Ruige JB, Stolk RP, Witteman JC, Dekker JM, et al.performance of a predictive model to identify undiagnosed diabetes in a health care setting. Diabetes Care 1999;22: Bang H, Edwards AM, Bomback AS, Ballantyne CM, Brillon D, et al. Development and validation of a patient self-assessment score for diabetes risk. Ann Intern Med 2009;151: Cabrera de León A, Coello SD, Rodríguez Pérez Mdel C, Medina MB, Almeida González D, et al. A simple clinical score for type 2 diabetes mellitus screening in the Canary Islands. Diabetes Res Clin Pract 2008;80: Cameron AJ, Magliano DJ, Zimmet PZ, Welborn TA, Colagiuri S, et al. The metabolic syndrome as a tool for predicting future diabetes: the AusDiab study. J Intern Med 2008;64: Dankner R, Abdul-Ghani MA, Gerber Y, Chetrit A, Wainstein J, et al. Predicting the 20-year diabetes incidence rate. Diabetes Metab Res Rev 2007; 23: Ferrannini E, Massari M, Nannipieri M, Natali A, Ridaura RL, et al. Plasma glucose levels as predictors of diabetes: the Mexico City diabetes study. Diabetologia 2009;52: Franciosi M, De Berardis G, Rossi MC, Sacco M, Belfiglio M, et al. Use of the diabetes risk score for opportunistic screening of undiagnosed diabetes and impaired glucose tolerance: the IGLOO (Impaired Glucose Tolerance and Long-Term Outcomes Observational) study. Diabetes Care 2005;28: Gao WG, Qiao Q, Pitkäniemi J, Wild S, Magliano D, et al. Risk prediction models for the development of diabetes in Mauritian Indians. Diabet Med 2009;26:

62 19. Glümer C, Carstensen B, Sandbaek A, Lauritzen T, Jørgensen T, et al; inter99 study. A Danish diabetes risk score for targeted screening: the Inter99 study. Diabetes Care 2004;27: Hanley AJ, Karter AJ, Williams K, Festa A, D'Agostino RB Jr, et al. Prediction of type 2 diabetes mellitus with alternative definitions of the metabolic syndrome: the Insulin Resistance Atherosclerosis Study. Circulation 2005;112: Heikes KE, Eddy DM, Arondekar B, Schlessinger L. Diabetes Risk Calculator: a simple tool for detecting undiagnosed diabetes and pre-diabetes. Diabetes Care 2008;31: Herman WH, Smith PJ, Thompson TJ, Engelgau MM, Aubert RE. A new and simple questionnaire to identify people at increased risk for undiagnosed diabetes. Diabetes Care 1995;8: Ruige JB, de Neeling JN, Kostense PJ, Bouter LM, Heine RJ. Performance of an NIDDM screening questionnaire based on symptoms and risk factors. Diabetes Care 1997;20: Kanaya AM, Wassel Fyr CL, de Rekeneire N, Shorr RI, Schwartz AV, et al. Predicting the development of diabetes in older adults: the derivation and validation of a prediction rule. Diabetes Care 2005;28: Kolberg JA, Jørgensen T, Gerwien RW, Hamren S, McKenna MP, et al. Development of a type 2 diabetes risk model from a panel of serum biomarkers from the Inter99 cohort. Diabetes Care 2009;32: Li J, Bornstein SR, Landgraf R, Schwarz PE. Validation of a simple clinical diabetes prediction model in a middle-aged, white, German population. Arch Intern Med 2007;167: Manley SE, Sikaris KA, Lu ZX, Nightingale PG, Stratton IM, et al. Validation of an algorithm combining haemoglobin A(1c) and fasting plasma glucose for diagnosis of diabetes mellitus in UK and Australian populations. Diabet Med 2009;26: Meisinger C, Thorand B, Schneider A, Stieber J, Döring A, et al. Sex differences in risk factors for incident type 2 diabetes mellitus: the MONICA Augsburg cohort study. Arch Intern Med 2002;162: Nichols GA, Brown JB.Validating the Framingham Offspring Study equations for predicting incident diabetes mellitus. Am J Manag Care 2008; 14: Park PJ, Griffin SJ, Sargeant L, Wareham NJ. The performance of a risk score in predicting undiagnosed hyperglycemia. Diabetes Care 2002; 25: Pearson TL, Pronk NP, Tan AW, Halstenson C. Identifying individuals at risk for the development of type 2 diabetes mellitus. Am J Manag Care 2003;9: Pires de Sousa AG, Pereira AC, Marquezine GF, Marques do Nascimento-Neto R, Freitas SN, et al. Derivation and external validation of a simple prediction model for the diagnosis of type 2 diabetes mellitus in the Brazilian urban population. Eur J Epidemiol 2009;24: Rahman M, Simmons RK, Harding AH, Wareham NJ, Griffin SJ. A simple risk score identifies individuals at high risk of developing Type 2 diabetes: a prospective cohort study. Fam Pract 2008;25: Rathmann W, Martin S, Haastert B, Icks A, Holle R, et al; KORA Study Group. Performance of screening questionnaires and risk scores for undiagnosed diabetes: the KORA Survey Arch Intern Med 2005;28;165: Schulze MB, Weikert C, Pischon T, Bergmann MM, Al-Hasani H, et al. Use of multiple metabolic and genetic markers to improve the prediction of type 2 diabetes: the EPIC-Potsdam Study. Diabetes Care 2009;32: Spijkerman AM, Yuyun MF, Griffin SJ, Dekker JM, Nijpels G, et al. The performance of a risk score as a screening test for undiagnosed hyperglycemia in ethnic minority groups: data from the 1999 health survey for England. Diabetes Care 2004;27: Stern M, Williams K, Eddy D, Kahn R. Validation of prediction of diabetes by the Archimedes model and comparison with other predicting models. Diabetes Care 2008;31: Tuomilehto J, Lindström J, Hellmich M, Lehmacher W, Westermeier T, et al. Development and validation of a risk-score model for subjects with impaired glucose tolerance for the assessment of the risk of type 2 diabetes mellitus-the STOP-NIDDM risk-score. Diabetes Res Clin Pract 2010;87:

63 39. Thomas C, Hyppönen E, Power C. Type 2 diabetes mellitus in midlife estimated from the Cambridge Risk Score and body mass index. Arch Intern Med 2006;166: Urdea M, Kolberg J, Wilber J, Gerwien R, Moler E, et al. Validation of a multimarker model for assessing risk of type 2 diabetes from a five-year prospective study of 6784 Danish people (Inter99). J Diabetes Sci Technol 2009; 3: Wannamethee SG, Papacosta O, Whincup PH, Carson C, Thomas MC, et al. Assessing prediction of diabetes in older adults using different adiposity measures: a 7 year prospective study in 6,923 older men and women. Diabetologia 2010;53: Talmud PJ, Hingorani AD, Cooper JA, Marmot MG, Brunner EJ, et al. Utility of genetic and non-genetic risk factors in prediction of type 2 diabetes: Whitehall II prospective cohort study. BMJ 2010;340:b Witte DR, Shipley MJ, Marmot MG, Brunner EJ. (2010) Performance of existing risk scores in screening for undiagnosed diabetes: an external validation study. Diabet Med 2010;27: Griffin SJ, Little PS, Hales CN, Kinmonth AL, Wareham NJ. Diabetes risk score: towards earlier detection of type 2 diabetes in general practice. Diabetes Metab Res Rev 2000;16:

64 Panel A Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax KORA (Basic) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q DETECT-2 Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q BRHS, simple clinical Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax AUSDRISK Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr)

65 DPoRT Predicted Risk (7.5yr) Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax ARIC (Basic) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) QDScore Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q DESIR (Clinical) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr)

66 Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax EPIC-Norfolk Ideal Nonparametric Predicted Risk (7.5yr) GDRS Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q FINDRISC (Full) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr)

67 Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax KORA (Clinical) Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q Ideal Nonparametric Predicted Risk (7.5yr) Panel B Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax ARIC (Enhanced) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) BRHS, fasting bio (2011) Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax BRHS, non-fasting bio (2011) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr)

68 TROMO Predicted Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q DESIR (Clinical+bio) Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax Framingham (Continuous) Ideal Nonparametric Predicted Risk (7.5yr) Dxy C (ROC) R2 D U Q Brier Intercept Slope Emax Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) Observed Risk (7.5yr) 65

69 Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q PROCAM Ideal Nonparametric Observed Risk (7.5yr) Dxy C (ROC) R2 Brier Intercept Slope Emax D U Q San Antonio (1993) Predicted Risk (7.5yr) Ideal Nonparametric Predicted Risk (7.5yr) Figure S1. Calibration plots for the 7.5-year risk of diabetes depicting the predicted risk against observed risk of developing type 2 diabetes in the validation data set. Panel A (the basic models), Panel B (the extended models). In those plots, depicting the ideal and non-parametric terms, the dashed line (the 45 line) from zero denotes ideal calibration line (slope=1, intercept=0) and the dotted lines (or the solid lines in other plots) denotes smooth calibration curve for each models. Observed Risk (7.5yr) 66

70 References 1. Beulens JW, Monninkhof EM, Verschuren WM, et al. Cohort profile: the EPIC-NL study. Int J Epidemiol 2010;39(5): Schulze MB, Hoffmann K, Boeing H, et al. An accurate risk score based on anthropometric, dietary, and lifestyle factors to predict the development of type 2 diabetes. Diabetes Care 2007;30(3): Ocke MC, Bueno-de-Mesquita HB, Goddijn HE, et al. The Dutch EPIC food frequency questionnaire. I. Description of the questionnaire, and relative validity and reproducibility for food groups. Int J Epidemiol 1997;26 Suppl 1:S Ocke MC, Bueno-de-Mesquita HB, Pols MA, Smit HA, van Staveren WA, Kromhout D. The Dutch EPIC food frequency questionnaire. II. Relative validity and reproducibility for nutrients. Int J Epidemiol 1997;26 Suppl 1:S Tjur T. Coefficients of Determination in Logistic Regression Models A New Proposal: The Coefficient of Discrimination. Am Stat 2009;63(4): Rufibach K. Use of Brier score to assess binary predictions. J Clin Epidemiol 2010;63(8):938-9; author reply Spiegelhalter DJ. Probabilistic prediction in patient management and clinical trials. Stat Med 1986;5(5): Miller ME, Hui SL, Tierney WM. Validation techniques for logistic regression models. Stat Med 1991;10(8): Harrell FE, Jr., Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 1996;15(4): Stallard N. Simple tests for the external validation of mortality prediction scores. Stat Med 2009;28(3): Steyerberg EW, Harrell FE, Jr., Borsboom GJ, Eijkemans MJ, Vergouwe Y, Habbema JD. Internal validation of predictive models: efficiency of some procedures for logistic regression analysis. J Clin Epidemiol 2001;54(8): Harrell FE, Lee KL. A comparison of the discrimination of discriminant analysis and logistic regression under multivariate normality. In: Biostatistics: Statistics in Biomedical, Public Health, and Environmental Sciences. The Bernard G. Greenberg Volume, ed. New York, US: NorthHolland, 1985:

71 68

72 Chapter 3 External validation of the KORA S4/F4 prediction models for the risk of developing type 2 diabetes in older adults: the PREVEND study Ali Abbasi 1,2,3 ; Eva Corpeleijn 1 ; Linda M. Peelen 3 ; Ron T. Gansevoort 2 ; Paul E. de Jong 2 ; Rijk O.B. Gans 2 ; Wolfgang Rathmann 4 ; Bernd Kowall 4 ; Christine Meisinger 5 ; Hans L. Hillege 1 ; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Joline W.J. Beulens 3 ; Stephan J.L. Bakker 2 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands 4 Institute of Biometrics and Epidemiology, German Diabetes Center, Leibniz Center for Diabetes Research at Heinrich Heine University, Düsseldorf, Germany 5 Helmholtz Zentrum München, German Research Center of Environmental Health, Institute of Epidemiology II, Neuherberg, Germany Eur J Epidemiol 2012;27:47-52

73 Abstract Background Recently, prediction models for type 2 diabetes mellitus (T2DM) in older adults (aged 55 year) were developed in the KORA S4/F4 study, Augsburg, Germany. We aimed to externally validate the KORA models in a Dutch population. Methods We used data on both older adults (n = 2,050; aged 55 year) and total nondiabetic population (n = 6,317; aged year) for this validation. We assessed performance of base model (model 1: age, sex, BMI, smoking, parental diabetes and hypertension) and two clinical models: model 1 plus fasting glucose (model 2); and model 2 plus uric acid (model 3). For 7-year risk of T2DM, we calculated C-statistic, Hosmer Lemeshow χ 2 -statistic, and integrated discrimination improvement (IDI) as measures of discrimination, calibration and reclassification, respectively. Results After a median follow-up of 7.7 years, 199 (9.7%) and 374 (5.9%) incident cases of T2DM were ascertained in the older and total population, respectively. In the older adults, C-statistic was 0.66 for model 1. This was improved for model 2 and model 3 (C-statistic = 0.81) with significant IDI. In the total population, these respective C-statistics were 0.77, 0.85 and All models showed poor calibration (P< 0.001). After adjustment for the intercept and slope of each model, we observed good calibration for most models in both older and total populations. Conclusions We validated the KORA clinical models for prediction of T2DM in an older Dutch population, with discrimination similar to the development cohort. However, the models need to be corrected for intercept and slope to acquire good calibration for application in a different setting. 70

74 Introduction Type 2 diabetes is one of the major concerns in public health, becoming more prevalent worldwide in parallel with increasing rate of obesity and ageing 1. There is evidence suggesting that diabetes can be prevented by diet and life-style modifications 2. For this, individuals at risk of developing diabetes need to be accurately identified 3,4. In the rapidly growing group of older subjects, prediction and primary prevention of chronic complex diseases such as diabetes are important to aid in healthy aging 5. Risk for development of diabetes is appreciably higher in older subjects than in younger subjects. Several pre- diction models, including the Finnish (FINDRISC), the Atherosclerosis Risk in Communities (ARIC), the Framingham, Data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR) and Cambridge diabetes risk scores, have been developed in middle-aged populations and validated in other populations 4,6-11. As there are indications of reverse epidemiology in older populations 12,13, it is questionable whether risk scores developed in middle-aged populations can be extrapolated to older subjects. Moreover, it has been shown that predictive value of common risk factors declines with ageing 14. Recently, the Cooperative Health Research in the Region of Augsburg (KORA) S4/F4 models have been developed to specifically predict the risk of type 2 diabetes in older Subjects 15. Because some risk scores that showed less performance when external validation was attempted, it is important that risk scores are validated in an independent population before they are brought into clinical practice 15,16. Therefore, we aimed to investigate the performance of the KORA models to predict incident type 2 diabetes in a large sample of non-diabetic Dutch adults, in particular older adults. We assessed performance of the model in terms of discrimination, calibration, recalibration and reclassification. Methods Population and design of the derivation study The KORA S4/F4 study is a community-based cohort of 2,656 individuals (aged years) living in the area of Augsburg, Germany in Details of the study design, recruitment, and procedures have been published elsewhere 17. Among 887 individuals who participated for a median follow-up of 7-years, 91 (10.5%) incident cases of type 2 diabetes were observed in the KORA cohort 15,17. The KORA data set was used to compare baseline characteristics with those of the validation cohort. Population and design of the validation study We used data from the Prevention of Renal and Vascular Endstage Disease (PREVEND) study. The PREVEND study is a community-based prospective cohort of 8,592 inhabitants (aged years) of the city of Groningen, The Netherlands 71

75 who were screened for baseline measurements between 1997 and Details of the study design, recruitment, and measurements have been published elsewhere 18. From the baseline cohort, we excluded 295 participants who had diabetes and 1,980 with missing data on clinical characteristics or data on follow- up, leaving 6,317 nondiabetic total population and a sample of 2,050 older adults (aged 55 years old) for this prospective validation analysis. The latter was used for primary validation while the former was used for secondary validation in a population with a much larger age range. All participants gave written informed consent prior to study inclusion. The PREVEND cohort complied with the Declaration of Helsinki and was approved by the medical ethics committee in The Netherlands. Outcome, predictors and measurements The main outcome was incidence of type 2 diabetes which was classified if one or more of the following criteria were met: fasting plasma glucose 7.0 mmol/l (126 mg/dl); non-fasting sample plasma glucose 11.1 mmol/l (200 mg/dl); self-report of a physician diagnosis of type 2 diabetes; pharmacy-registered use of glucoselowering agents 19. To estimate the predicted 7-year risk for type 2 diabetes in our cohort, we calculated the linear predictors of the KORA prediction models 15. The base model (model 1) included data on age, sex, parental diabetes, body mass index (BMI), smoking status and hypertension. The clinical KORA models included additional data on fasting glucose, serum uric acid and HbA1c 15. As data on HbA1c was unavailable, we validated a reported clinical model with data on fasting glucose (model 2) 15. Moreover, the authors were asked to provide a clinical model with data on fasting glucose and uric acid (model 3), presented in Table S1. Data analysis To externally validate these models, we assessed the discrimination and calibration performances in our cohort 20. The discrimination performance denotes to what extent the model distinguishes between individuals with and without the outcome. Discrimination was expressed as the C-statistic with 95% confidence interval, where a value of 1 implies a perfect discrimination and a value of 0.5 implies performance no better than chance. We compared the C-statistics of the clinical models to that of the base model as reference. The calibration compares predicted risks with observed risks. We applied the Hosmer Lemeshow χ2 test to evaluate the calibration performance. A lower χ2 value with a non-significant P value represents good calibration. Also, calibration was visually checked by comparing the predicted probabilities versus observed incident cases of diabetes in each decile of predictions 20. To recalibrate the prediction models, we used the original KORA models and applied logistic regression to derive the intercept and the calibration slope of each model in the total and older populations, and separately for women and men (Supplementary method part 1) 21. We used these intercepts and slopes by fitting a model with the original linear predictor as the only covariate in the PREVEND data set. We 72

76 multiplied each linear predictor by the calibration slope and added the calibration intercept to each original model 21. Thereafter, we added data on waist circumference, a non-invasive risk factor for diabetes 11, to the re-calibrated models and assessed if this could improve predictive ability. We examined improvement of diabetes prediction in terms of discrimination, calibration and integrated discrimination improvement (IDI), a measure of reclassification (Supplementary method part 2) 20. The analyses were performed separately in the older adults and in total PREVEND population. All the statistical analyses were carried out using Statistical Package for Social Sciences version 18 (SPSS Inc, Chicago, Illinois, USA), Stata software version 10.0 (Stata-Corp LP, College Station, TX, USA) and R for Windows ( Results In the older adults, we observed 199 (9.7%) incident cases of type 2 diabetes during follow-up for a median of 7.7 years. In the total population, we observed 374 (5.9%) cases during this follow-up. Baseline participants characteristics of the KORA and PREVEND cohorts (aged 55 years) are shown in Table S2. Participants of PREVEND were more likely to be male, older, more likely to be smoker and to have hypertension, but had lower BMI, lower parental history of diabetes and had lower fasting glucose and serum uric acid than participants of KORA. Table 1 depicts the performance of the KORA models in terms of discrimination and calibration. In the older adults, a relatively low discriminative ability was observed for the base model (C-statistic=0.66), being lower than the original C-statistic of This was significantly improved for both models 2 and 3 (C-statistic=0.81), being comparable with the original C-statistic of The discriminative ability was not significantly different between these clinical models (P=0.78). The base and both clinical models did not show good calibration (P<0.001 for 7-year risk). When we tested the performance of each model in the total population, we observed a better discriminative ability for the base (C-statistic of 0.77; P<0.001) and both clinical models (both C-statistics=0.85; P<0.001). Similarly, the base and both clinical models did not show good calibration (P<0.001 for 7-year risk) in the total population. After adjustment for the calibration intercept and the calibration slope of each model, good calibration was observed for the clinical models (P>0.05 for 7-year risk, Table 1), but not for the base model. Figure 1 (A, B) depicts the agreement between the predicted 7-year risk and observed risk of type 2 diabetes in each decile of predictions before and after recalibration. Of note, the predictive probability of model 3 was deviated from the ideal line in the older adults above 15% risk (Figure 1 A); indeed, the 7-year risk was underestimated for this risk category. The IDIs were significant when we compared the prediction performance of the clinical models to that of the base model (P<0.001). In a subsequent analysis, we stratified total population by gender. We observed that all KORA models showed better discrimination performance in women than in men (Table 1). Addition of waist circumference improved predictive ability of the base model in the total population 73

77 Table 1. Predictive performance of the KORA models in the PREVEND study a Prediction Model Cases/Total No. of predictors C-statistic (95%CI) IDI (p value) (Not recalibrated model) b χ2 statistic (p value) (Not recalibrated model) Age 55: 199/2050 Base Model ( ) Ref (<0.001) Model ( ) (<0.001) (<0.001) Model ( ) (<0.001) (<0.001) Whole population: 374/6317 Base Model Model 2 Model ( ) 0.85 ( ) 0.85 ( ) Ref (<0.001) (<0.001) (<0.001) (<0.001) (<0.001) Women: 155/3282 Base Model ( ) Ref (<0.001) Model ( ) (<0.001) (<0.001) Model ( ) (<0.001) (0.01) Calibration intercept Calibration slope χ2 statistic (p value) (Recalibrated model) (0.02) (0.35) 3.03 (0.96) (0.004) 4.58 (0.87) (0.21) (<0.001) 5.35 (<0.80) 8.46 (0.49) IDI (p value) (Recalibrated model) b Ref (<0.001) (<0.001) Ref (<0.001) (<0.001) Ref (<0.001) (<0.001) 74

78 Men: 219/3035 Base Model ( ) Ref (<0.001) (0.02) Ref. Model ( ) (<0.001) (<0.001) (0.79) (<0.001) Model ( ) (<0.001) (<0.001) (0.74) (<0.001 a Base model (model 1) included data on age, sex, parental diabetes, BMI, smoking status and hypertension. Model 2 additionally included fasting glucose, and model 3 additionally included fasting glucose and uric acid. A higher discrimination C-statistic, non-significant p value of the calibration χ 2 statistic and significant p value of IDI represent better performance for each model. b IDI denotes integrated discrimination improvement, and is calculated as difference of the mean predicted risk between two models for those who developed outcome and those who did not develop outcome. The base model was considered the reference model. 75

79 (C-statistic=0.79; P<0.001), but not in the older population separately (C-statistic= 0.67; P=0.30). Addition of waist circumference did not improve predictive ability of model 3, neither in the total population (C-statistic=0.85; P=0.11) nor in the older population separately (C-statistic=0.81; P=0.41). Observed frequency A Ideal Nonparametric Ideal Not-recalibrated Recalibrated Observed frequency B Ideal Nonparametric Ideal Not-recalibrated Recalibrated Predicted probability Predicted probability Figure 1. Comparison between predicted risk versus observed diabetes frequency in the PREVEND cohort according to the KORA model 3, Model 3, included data on the base KORA model plus glucose and uric acid for the risk prediction of diabetes in the PREVEND cohort. A depicts calibration plots in the older adults, aged C55 years (n = 2,050). B depicts calibration plots in the total population (n = 6,317). The dashed line represents an ideal calibration (with intercept 0 and slope 1); the dotted line is for the not-recalibrated model and the solid line is after adjustment for the intercept and slope Discussion In this external validation study, we prospectively assessed performance of the KORA models to predict the risk of developing type 2 diabetes in an independent Dutch population. We found that the prediction models with clinical data on glucose with or without uric acid performed well and this was much better than the base model in terms of discrimination and reclassification in the older adults. Moreover, we observed a similar pattern but with higher discriminative abilities in the total population. All models showed poor calibration performance. The calibration was good after adjustment for the intercept and the slope of clinical models, but not for the base model. To our best of knowledge, there are few studies that derive and validate prediction models for of the risk of developing type 2 diabetes in the older adults. The main strengths of our study were including a large population- based cohort, available data on blood sampling in each screening visit and pharmacy registry, and applying the latest standards of prediction research. Some limitations should be addressed. First, we excluded the individuals with missing data at baseline or during 76

80 follow-up. However, the baseline characteristics of excluded participants were similar to those who were included in our analysis. Therefore, it is less likely that this might have led to selection bias. Moreover, both derivation and validation data sets were gathered among Whites and our findings need to be further evaluated in other populations. Our findings were consistent with previous validation studies in which performance of other prediction models for type 2 diabetes were tested in independent populations 22, 23. Of note, predictive performance of models is often decreased in the validation sample. Several differences between derivation and validation samples might explain this change of performance. These include differences in healthcare systems, methods of measurement, and patients characteristics 22. Our protocol to measure the predictors and incident cases of type 2 diabetes was very comparable with the KORA study during similar follow-up time and period. Regardless possible differences in healthcare systems, characteristics of the participants of PREVEND were remarkably different from the participants of KORA. When we calculated the C-statistic as a discrimination measure, we observed good ability (C-statistic 0.81) of clinical KORA models to distinguish between incident cases of type 2 diabetes and those who remained free of diabetes in both older and total populations. This was comparable for both clinical models and much higher than the base model. This cannot be explained by differences in the incidence of type 2 diabetes, as the C-statistic is hardly affected by different incidences of the outcome. A better discriminative ability of KORA models in the total population than in the older adults might be explained by a difference in case mix between KORA and PREVEND cohorts, less heterogeneity among the older adults of PREVEND, effect of predictors and difference in regression coefficients of KORA models 23. For example, variables like age, BMI, or hypertension discriminate better between cases and non-cases in the total population 14 because low BMI or hypertension is more frequent in the younger adults who develop diabetes less often. Both base and clinical KORA models showed poor calibration both in the older and total populations. In other words, the mean predicted risk by the KORA models were significantly different from the observed risk of type 2 diabetes in the PREVEND cohort. One explanation for this is different incidence of type 2 diabetes between KORA and PREVEND cohorts. To further assess the calibration performance of each models, we used logistic recalibration of the original prediction models 21. After this adjustment, both KORA clinical models showed good calibration. In conclusion, addition of fasting glucose improved performance of KORA base model in both older and total populations in terms of discrimination and reclassification. Further addition of uric acid did not matter much. Both base and clinical models showed poor calibration. After correction for the intercept and the slope, most KORA models showed good calibration, indicating that there is often a need to adapt prediction models before application in a different setting. 77

81 Acknowledgments This work was supported by the Netherlands Heart Foundation, Dutch Diabetes Research Foundation and Dutch Kidney Foundation. This research was performed within the frame- work of CTMM, the Center for Translational Molecular Medicine ( project PREDICCt (grant 01C ). We thank Prof. Dr. L.T.W. de Jong-van den Berg and Dr. S.T. Visser from the Department of Social Pharmacy, Pharmacoepidemiology and Pharmacotherapy, Groningen University Institute for Drug Exploration, University Medical Center Groningen and University of Groningen, The Netherlands for providing the data on use of glucose-lowering agents according to central pharmacy registration. The Diabetes Cohort Study was funded by a German Research Foundation project grant to W. Rathmann (DFG;RA 459/2 1). The KORA research platform and the KORA Augsburg studies are financed by the Helmholtz Zentrum München, German Research Center for Environmental Health, which is funded by the German Federal Ministry of Education, Science, Research and Technology and by the State of Bavaria. We thank the field staff in Augsburg who were involved in the conduct of the studies. The German Diabetes Center is funded by the German Federal Ministry of Health and the Ministry of Innovation, Science, Research and Technology of the State of North Rhine Westphalia. None of the study sponsors had a role in study design; in data collection, analysis, or interpretation; in writing the report; or in the decision to submit for publication. No duality of interest relevant to this manuscript is declared. 78

82 References 1. Wild S, Roglic G, Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the year 2000 and projections for Diabetes Care 2004;27: Diabetes Prevention Program Research Group. 10-year follow-up of diabetes incidence and weight loss in the diabetes prevention program outcomes study. Lancet 2009;374: Schmidt MI, Duncan BB, Bang H, et al. Identifying individuals at high risk for diabetes: the atherosclerosis risk in communities study. Diabetes Care 2005;28: Buijsse B, Simmons RK, Griffin SJ, Schulze MB. Risk assessment tools for identifying individuals at risk of developing type 2 diabetes. Epidemiol Rev 2011;33: Stolk RP, Hutter I, Wittek RP. Population ageing research: a family of disciplines. Eur J Epidemiol. 2009;24: Lindströ m J, Tuomilehto J. The diabetes risk score: a practical tool to predict type 2 diabetes risk. Diabetes Care 2003;26: Kahn HS, Cheng YJ, Thompson TJ, Imperatore G, Gregg EW. Two risk-scoring systems for predicting incident diabetes melli- tus in U.S. adults age years. Ann Intern Med 2009;150: Wilson PW, Meigs JB, Sullivan L, Fox CS, Nathan DM, D Agostino RB Sr. Prediction of incident diabetes mellitus in middle- aged adults: the framingham offspring study. Arch Intern Med 2007;167: Balkau B, Lange C, Fezeu L, et al. Predicting diabetes: clinical, biological, and genetic approaches: data from the epidemiological study on the insulin resistance syndrome (DESIR). Diabetes Care 2008;31: Griffin SJ, Little PS, Hales CN, Kinmonth AL, Wareham NJ. Diabetes risk score: towards earlier detection of type 2 diabetes in general practice. Diabetes Metab Res Rev 2000;16: Noble D, Mathur R, Dent T, Meads C, Greenhalgh T. Risk models and scores for type 2 diabetes: systematic review. BMJ 2011;343:d Vischer UM, Safar ME, Safar H, et al. Cardiometabolic determinants of mortality in a geriatric population: is there a reverse metabolic syndrome? Diabetes Metab 2009;35: Weiss A, Boaz M, Beloosesky Y, Kornowski R, Grossman E. Body mass index and risk of allcause and cardiovascular mor- tality in hospitalized elderly patients with diabetes mellitus. Diabet Med 2009;26: de Ruijter W, Westendorp RG, Assendelft WJ, et al. Use of framingham risk score and new biomarkers to predict cardiovascular mortality in older people: population based observational cohort study. BMJ 2010;338:a Rathmann W, Kowall B, Heier M, et al. Prediction models for incident type 2 diabetes mellitus in the older population: KORA S4/F4 cohort study. Diabet Med 2010;27: Rathmann W, Martin S, Haastert B, et al. KORA study group. Performance of screening questionnaires and risk scores for undiagnosed diabetes: the KORA survey 2000 Arch Intern Med. 2005;165: Rathmann W, Haastert B, Icks A, et al. High prevalence of undiagnosed diabetes mellitus in Southern Germany: target populations for efficient screening. The KORA survey Diabetologia 2003;46: Lambers Heerspink HJ, Brantsma AH, de Zeeuw D, et al. Albuminuria assessed from first-morning-void urine samples versus 24-h urine collections as a predictor of cardiovascular morbidity and mortality. Am J Epidemiol 2008;168: Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin and risk of type 2 diabetes in the general population. Diabetologia 2011;54: Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010;21:

83 21. Janssen KJ, Moons KG, Kalkman CJ, Grobbee DE, Vergouwe Y. Updating methods improved the performance of a clinical pre- diction model in new patients. J Clin Epidemiol 2008;61: Altman DG, Vergouwe Y, Royston P, Moons KG. Prognosis and prognostic research: validating a prognostic model. BMJ 2009;338:b Vergouwe Y, Moons KG, Steyerberg EW. External validity of risk models: use of benchmark values to disentangle a case-mix effect from incorrect coefficients. Am J Epidemiol 2010;172:

84 Appendix Table S1. Characteristics of the KORA models (regression coefficients of multivariable logistic models) Base model (model 1) Clinical model (model 2) Clinical model (model 3) Intercept Sex (male) Age (years) BMI (kg/m 2 ) Parental diabetes (yes) Former smoker (yes) Current smoker (yes) Hypertension (yes) a Fasting glucose (mg dl) Serum uric acid (mg dl) a Hypertension was defined as Systolic diastolic blood pressure mmhg or use of medication for hypertension. TableS2. Baseline characteristics of individuals in KORA and PREVEND cohorts (aged 55 yr) Characteristics KORA (n = 873) PREVEND (n=2050) p value a Sex (males) (%) Age (years) (5.4) 64.2 (5.6) <0.001 BMI (kg/m 2 ) 28.1 (4.0) 27.2 (3.8) <0.001 Parental diabetes (%) <0.001 Smoking status Never smoker (%) Former smoker (%) <0.001 Current smoker (%) Hypertension (%) Fasting glucose (mmol/l) 5.48 (0.52) 4.97 (0.63) <0.001 Serum uric acid (µmol/l) (81.5) (80.7) 0.03 Data are percent for categorical variables and mean±sd for continuous variables. a Comparison between KORA and PREVEND cohorts in each groups by incident diabetes: χ2 tests for categorical variables; t tests for continuous variables. Text S1. The formula of the KORA base model is: Risk of T2D log = KORA linear predictor = 1 Risk of T2D Sex(male) Age(y) BMI(kg/m ) Parental diabetes Former smoker Current smoker Hypertension 81

85 To do logistic recalibration, logistic regression models were fitted with the KORA linear predictors as the only covariate in the PREVEND data set: Risk of T2D log = α calibratio n+ βcalibration KORA linear predictor 1 Risk of T2D Text S2. Where Risk model 2cases and Risk model 1cases is the mean predicted risk of model 2 and model 1, respectively, for those who developed diabetes, and the rest of values denote the corresponding risks for those who did not develop diabetes, the integrated discrimination improvement (IDI) is calculated as follow: IDI= (Risk model 2 cases - Risk model1 cases ) - (Risk model 2 non-casenon-cases - Risk model1 ) 82

86 Chapter 4 Liver function tests and risk prediction of incident type 2 diabetes: evaluation in two independent cohorts Ali Abbasi 1,2,3 ; Stephan J.L. Bakker 2 ; Eva Corpeleijn 1 ; Daphne L. van der A 4 ; Ron T. Gansevoort 2 ; Rijk O.B. Gans 2 ; Linda M. Peelen 3 ; Yvonne T. van der Schouw; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Annemieke M.W. Spijkerman 5 ; Joline W.J. Beulens 3 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands 4 Center for Nutrition and Health, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands 5 Center for Prevention and Health Services Research, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands PLoS One. 2012;7(12):e51496

87 Abstract Background Liver function tests might predict the risk of type 2 diabetes. An independent study evaluating utility of these markers compared with an existing prediction model is yet lacking. Methods We performed a case-cohort study, including random subcohort (6.5%) from 38,379 participants with 924 incident diabetes cases (the Dutch contribution to the European Prospective Investigation Into Cancer and Nutrition, EPIC-NL, the Netherlands), and another population-based cohort study including 7,952 participants with 503 incident cases (the Prevention of Renal and Vascular End-stage Disease, PREVEND, Groningen, the Netherlands). We examined predictive value of combination of the liver function tests (gamma-glutamyltransferase, alanine aminotransferase, aspartate aminotransferase and albumin) above validated models for 7.5-year risk of diabetes (the Cooperative Health Research in the Region of Augsburg, the KORA study). Basic model includes age, sex, BMI, smoking, hypertension and parental diabetes. Clinical models additionally include glucose and uric acid (model1) and HbA1c (model2). Results In both studies, addition of Liver function tests to the basic model improved the prediction (C-statistic by~0.020; NRI by~9.0%;p<0.001). In the EPIC-NL casecohort study, addition to clinical model1 resulted in statistically significant improvement in the overall population (C-statistic=+0.009; P<0.001; NRI=8.8%; P<0.001), while addition to clinical model 2 yielded marginal improvement limited to men (C-statistic=+0.007; P=0.06; NRI=3.3%; P=0.04). In the PREVEND cohort study, addition to clinical model 1 resulted in significant improvement in the overall population (C-statistic change=0.008; P=0.003; NRI=3.6%; P=0.03), with largest improvement in men (C-statistic change=0.013; P=0.01; NRI=5.4%; P=0.04). In PREVEND, improvement compared to clinical model 2 could not be tested because of lack of HbA1c data. Conclusions Liver function tests modestly improve prediction for medium-term risk of incident diabetes above basic and extended clinical prediction models, only if no HbA1c is incorporated. If data on HbA1c are available, Liver function tests have little incremental predictive value, although a small benefit may be present in men. 84

88 Introduction Change in liver function tests is considered as surrogate marker of liver injury and nonalcholic fatty liver disease (NAFLD) 1. Previous studies have demonstrated that circulating concentration of liver function tests like gamma-glutamyltransferase (GGT), alanine aminotransferase (ALT) and aspartate aminotransferase (AST) are increased in individuals with insulin resistance and the metabolic syndrome 1-3. In addition, these components of liver function tests have been shown to be positively associated with the risk of future type 2 diabetes 1, 4. A recent meta-analysis on this topic showed that both elevated ALT and GGT were associated with increased risk of diabetes, while GGT might be a stronger risk factor than ALT 4. However, there is only a limited number of studies that examine the predictive value of liver function tests for the risk of future diabetes in terms of essential measures of prediction, such as the C-statistic to assess discrimination between people who develop diabetes and those who don t 5-8. These studies mainly developed 6, 7 or updated 5 clinical prediction models by incorporating one or two components of liver function tests in each models. It is important to note that the predictive value of liver function tests was examined in combination with other (bio)markers and in the same data set that was used to develop the original models 5-7. Of these, 2 studies showed improvement in prediction when GGT plus glycaemia indices were added to a basic model consisting only of data that can be derived without need for taking blood samples 6, 7. In another study, a combination of GGT, ALT, triglycerides and HDL cholesterol improved discrimination above a diabetes risk score including HbA1c and glucose 5. So, whether liver function tests have incremental predictive value above validated model(s) is still unclear. An independent study evaluating utility of these markers of liver function when incorporated in an existing prediction model is needed to answer this question 9, 10. Recently, we validated and updated German prediction models from the Cooperative Health Research in the Region of Augsburg (KORA) study in a Dutch general population cohort 11, 12. In the current study, we addressed the incremental predictive value of liver function tests for the risk of future type 2 diabetes when compared with the KORA models 12. To do so, we analysed data from two independent cohorts separately. In each cohort, we performed analyses in the total population and sex-stratified subgroups to account for potential sex differences in the prediction performance of each model 6, 11, 13. Methods Study design and populations We used data from two cohorts of general population in the Netherlands: 1) the Dutch contribution to the European Prospective Investigation Into Cancer and Nutrition (EPIC-NL) study; and 2) the Prevention of Renal and Vascular End-stage 85

89 Disease (PREVEND) study. Details of each study design and recruitment of participants have been published previously 14, 15. In brief, the EPIC-NL cohort (n = 40,011) includes the Monitoring Project on Risk Factors for Chronic Diseases (MORGEN) and Prospect cohorts, initiated between 1993 and Prospect is a prospective cohort study of 17,357 women aged years who participated in a breast cancer screening programme. The MORGEN cohort consists of 22,654 men and women aged years who were recruited through random population sampling in three Dutch towns (Amsterdam, Maastricht and Doetinchem). A new random sample of about 5,000 participants was examined each year. We excluded 615 individuals with prevalent type 2 diabetes and 1,017 with missing follow-up or who did not consent to linkage with disease registries, leaving 38,379 individuals in the full cohort. In a 6.5% baseline random sample (n=2,604) with biochemical measurements 14, similar exclusion criteria were applied. After exclusions, 2,506 individuals (including 79 incident diabetes cases) from the random sample and 924 incident diabetes cases in the full cohort remained for the case-cohort study 16. We used this case-cohort sample for all analyses. In brief, the baseline PREVEND cohort (n=8,592) was recruited from inhabitants (aged years) of the city of Groningen, the Netherlands. Baseline measurements were performed between 1997 and The PREVEND cohort included a total of 6,000 individuals with a morning urinary albumin concentration of 10 mg/l or greater and a random control sample of individuals with a urinary albumin concentration of less than 10 mg/l (n=2,592). Overall, we excluded 336 individuals with prevalent type 2 diabetes and 277 with missing data on follow-up, leaving 7,979 individuals for the full cohort study. We used this full cohort sample for all analyses. Ethics Statement All participants gave written informed consent prior to study inclusion. All cohort studies complied with the Declaration of Helsinki and were approved by local medical ethics committees. Measurements of biomarkers In the EPIC-NL study, the general questionnaire contained questions on demographic characteristics and risk factors for the presence of chronic diseases. Body weight, height and waist and hip circumference were measured according to standard procedures. Hypertension was defined based on self-report of diagnosis by a physician, measured hypertension ( 140 mmhg systolic blood pressure or 90 mmhg diastolic blood pressure) or the use of blood pressure-lowering medication. Non-fasting blood samples were collected at baseline from all participants. HbA1c was measured in erythrocytes using an immunoturbidimetric latex test. Glucose and uric acid were measured using enzymatic methods. AST, ALT and GGT were measured using enzymatic methods and albumin by a colorimetric method

90 In the PREVEND study, the participants underwent two outpatient visits to assess baseline data on demographics, anthropometric measurements, cardiovascular risk factors, health behaviours, and medical family history and to collect two 24-hour urine samples on 2 consecutive days. Blood pressure values are given as the mean of the last two recordings of both visits as this provides the values after stabilization of blood pressure. Plasma glucose was measured by dry chemistry (Eastman Kodak, Rochester, New York). All liver function tests were measured by a standardized enzymatic method (Modular P; Roche Diagnostics, Indianapolis, IN). Definition of main outcome In the EPIC-NL study, potential incident type 2 diabetes was self-reported via two follow-up questionnaires at 3- to 5-year intervals in the MORGEN and Prospect cohort. In the Prospect cohort, a urinary glucose strip test was sent along with the first follow-up questionnaire as a screening method. Diagnoses of type 2 diabetes were also obtained from the Dutch Center for Health Care Information, which holds a standardized computerized register of hospital discharge diagnoses. Follow-up was complete until January 1, Potential cases identified by these methods were verified against general practitioner (MORGEN and Prospect) or pharmacist records (Prospect only). We defined type 2 diabetes as being present when the diagnosis was confirmed by either of these methods. For 89% of participants with potential diabetes, verification information was available, and 72% were verified as having type 2 diabetes and were included as cases of type 2 diabetes in this analysis 17. The rest of individuals were considered as non-cases. In the PREVEND study, incident cases of diabetes were ascertained as described previously 18. In brief, incident diabetes was considered present if one or more of the following criteria were met: 1) a fasting plasma glucose of 7.0mmol/l (126mg/dl) or random sample plasma glucose 11.1 mmol/l (200mg/dl); 2) selfreported physician s diagnosis; 3) use of glucose-lowering agents according to a central pharmacy registration. Statistical analysis First, we examined the association between the components of liver function tests (including GGT, ALT, AST and albumin) and the risk of future diabetes. For liver function tests, we used logarithm transformation with base 2 (log2) to allow for interpretation of results per increase of 100% of values of each component. We used Cox proportional-hazards regression in the EPIC-NL study which was adapted for case-cohort analysis. We used logistic regression in the PREVEND full cohort study, because the events have been detected at regular screening visits or shortly thereafter. Thus, estimated survival and hazards can not be accurately calculated by this type of follow-up. In step 1, we calculated age and sex-adjusted hazard ratios (HRs) and odd ratios with 95% CIs for the risk of diabetes by doubling of concentrations of each liver function tests (per log2 unit increase). In step 2, we adjusted for age, sex, parental 87

91 diabetes, body mass index (BMI), smoking status, hypertension, glucose and uric acid. In step 3, we further adjusted for HbA1c. This could only be done in the EPIC- NL case-cohort study, because data on HbA1c were not available in the PREVEND study. To account for the case-cohort design in the EPIC-NL study, we applied an extrapolation approach which extends the case-cohort data to the size of the entire cohort 19. This is achieved by extrapolating the non-cases of the random sample (i.e., total random sample of 2,506 individuals minus 79 cases) to the number of non-cases in the full cohort (i.e., total sample of 38,379 individuals minus 924 cases). To do so, we substituted the non-cases of the full cohort (n = 37,455) by a random multiplication of non-cases of the random sample (n = 2,427). We have previously described and validated this approach 20. In the second part of this study, we computed the probability of getting diabetes using the KORA basic model which was previously validated and updated in the PREVEND cohort 11. As previously described 11, we recalibrated the original KORA model by means of logistic regression to derive the intercept and the calibration slope in the PREVEND cohort study. We also adjusted for the difference in incident diabetes between the KORA and the EPIC-NL cohorts by fitting the original KORA model in the EPIC-NL case-cohort study 21. Figure S1 (a, b) depicts the agreement between the predicted 7.5-year risk and observed risk of type 2 diabetes after recalibration in each cohort. The basic model included data on age, sex, parental diabetes, body mass index (BMI), smoking status and hypertension 12. Clinical model 1 included additional data on glucose plus serum uric acid; and in clinical model 2 we further added HbA1c. As the original KORA models have been developed for a time period of risk prediction of 7.5 years, we examined the incremental predictive value of liver function tests also for the 7.5-year risk of developing type 2 diabetes. Therefore, participants who developed diabetes after more than 7.5 years of followup were included in 7.5-year prediction as non-cases. We examined added predictive value of 1) each component alone, 2) combination of GGT+ALT and 3) a panel of GGT, ALT, AST and albumin. We assessed improvement of type 2 diabetes prediction in terms of discrimination by calculating the C-statistic with 95%CI, and reclassification by calculating integrated discrimination improvement (IDI) and net reclassification improvement (NRI) 22, 23. To calculate the NRI, cut-off values for risk categories have to be defined. In previous studies, a number of risk categories for the 10-year risk of cardiovascular disease 23, 24 or type 2 diabetes 25, 26 have been reported. In the present study, we slightly modified these cut-off values according to the shorter time period (and hence the lower average observed risk) 25, 26, using cut-off values of <4% for low-risk, 4%-8% for intermediate-risk and 8% for high-risk. In the EPIC-NL case-cohort study, for most predictors <1% of data were missing; however, missing values occurred in 5% for parental history of diabetes, and 20.5% for glucose levels. Because an analysis of only the completely observed data may often lead to biased results, we imputed these missing values using single imputation and predictive mean matching 27. As the percentage of missing values for 88

92 the non-fasting glucose concentration was relatively high, we repeated our analyses using only data from the MORGEN cohort, in which less than 10% of values for nonfasting glucose concentration were missing, as a sensitivity analysis. In the PREVEND cohort study, for most variables, <1% were missing, whereas this was up to 7.5% for self-reported variables. To account for missing values, we used a similar approach to that of the EPIC-NL study. Table S1 in supporting information shows the number of missing values for all variables incorporated in each model. We also used a weighted method to compensate for baseline enrichment of the PREVEND participants with high urinary albumin concentration (>10 mg/l). All the statistical analyses were carried out using IBM SPSS 19.0 and R version (Vienna, Austria) for Windows ( Results Baseline clinical characteristics We summarize baseline characteristics of the participants of each study in Table 1. Participants of the EPIC-NL study were more likely to be women, and to have hypertension and parental history of diabetes, whereas participants of the PREVEND study were more likely to be smoker and had slightly higher uric acid and albumin on average. In the EPIC-NL cohort study, we ascertained and validated 924 (2.4%) incident cases of type diabetes during a median follow-up of 10.2 years (over 387,000 person-years). In the PREVEND cohort study, we ascertained 503 (6.3%) incident cases during a median follow-up of 7.7 years (over 60,186 person-years). Liver function tests and type 2 diabetes Table 2 depicts the associations between components of liver function tests and the risk of diabetes, calculated per 100% increase of marker concentrations in total populations and in sex-stratified subgroups. In the EPIC-NL case-cohort study, the multivariable-adjusted HRs (95%CI) for the risk of diabetes were 1.49 ( ), 1.22 ( ), 0.97 ( ) and 0.34 ( ) per doubling concentrations of GGT, ALT, AST and albumin, respectively. In the PREVEND cohort study, the multivariable-adjusted ORs (95%CI) for the risk of diabetes were 1.22 ( ), 1.29 ( ), 1.16 ( ) and 0.31 ( ) per 100% increase of concentrations of GGT, ALT, AST and albumin, respectively. The associations between liver function tests and the risk of diabetes did not significantly differ by sex in both cohorts (P>0.1 for interaction). In the EPIC-NL case-cohort study, stratified analysis by sex showed that the direction of the association between albumin and diabetes risk was changed in men after adjustment for age, BMI with family history of diabetes (also for the KORA basic model plus glucose) (data not shown). Predictive value of liver function tests In the EPIC-NL case-cohort study, the basic model showed a C-statistic of ( ) for the 7.5-year risk of diabetes (Table 2). Addition of liver function tests 89

93 Table 1. Baseline participants characteristics in each study * EPIC-NL study PREVEND study Variables Full cohort Random sample Cases with incident Full cohort Cases with incident type 2 diabetes type 2 diabetes No. of individuals Age yr 49.1 (11.9) 49.2 (11.9) 56.6 (7.3) 48.9 (12.5) 56.5 (10.7) Female gender no. (%) (74.3) 1872 (74.7) 722 (78.1) 4065 (50.9) 210 (41.7) Parental history of diabetes no. (%) 7379 (19.2) 498 (19.9) 377 (40.8) 1252 (15.7) 143 (28.4) Hypertension no. (%) (36.8) 953 (38.0) 626 (67.7) 2248 (28.2) 275 (54.7) Antihypertensive medication no. (%) 3736 (9.7) 259 (10.3) 275 (29.8) 1036 (13.0) 133 (26.4) Current smoker no. (%) (30.6) 789 (31.5) 322 (34.8) 2742 (34.4) 174 (34.6) Exsmoker no. (%) (31.3) 769 (30.7) 241 (26.1) 2896 (36.3) 205 (40.8) Body-mass index 25.6 (4.0) 25.7 (4.0) 29.9 (4.7) 26.0 (4.2) 29.5 (4.7) Waist cimrcumfernce cm 85.1 (11.4) 85.3 (11.6) 97.0 (11.6) 88.1 (12.9) 99.2 (12.2) Systolic blood pressure mm Hg (18.8) (18.6) (21.6) (19.3) (20.4) Biomarkers Glucose mmol/liter 4.9 (1.2) 4.89 (1.17) 6.75 (2.48) 4.7 (0.6) 5.6 (0.7) HbA1c % (0.58) 6.5 (1.4) - - Uric acid µmol/l (68.48) (71.04) (80.16) (81.94) GGT U/liter (20.0) 36.6 (28.9) 32.6 (45.6) 53.9 (126.7) ALT U/liter (11.9) 20.3 (11.8) 23.9 (20.7) 31.3 (40.6) AST U/liter (9.1) 22.7 (9.0) 25.7 (10.4) 28.7 (19.4) Albumin g/l (4.9) 37.1 (4.9) 45.8 (2.7) 45.4 (3.0) * Data were shown as mean (SD) for continuous variables, and numbers (percentage) for categorical variables. EPIC-NL denotes Dutch contribution of the European Prospective Investigation Into Cancer and Nutrition, PREVEND denotes Prevention of Renal and Vascular End-stage Disease, HbA1c glycated hemoglobin, GGT gamma-glutamyltransferase, ALT alanine aminotransferase and AST aspartate aminotransferase. Hypertension was ascertained on the basis of self-reported diagnosis by a physician, antihypertensive medication use, systolic blood pressure 140 mm Hg, or diastolic blood pressure 90 diastolic blood pressure, or a combination of these. Body mass index is the weight in kilograms divided by the square of the height in meters. To convert values for glucose to milligrams per deciliter, divide by To convert values for uric acid to milligrams per deciliter, divide by

94 improved the C-statistic of the basic model (C-statistic change= 0.024; P<0.001) and led to an IDI of (P<0.001) and NRI of 9.5% (P<0.001). After addition of each component of liver function tests alone to the basic models, the C-statistic changes were (P<0.001), (P<0.001), (P=0.15) and (P=0.13) for GGT, ALT, AST and albumin, respectively. Addition of liver function tests also improved prediction for clinical model 1 (C-statistic change=0.009; P<0.001; NRI=8.8%; P<0.001). Although addition of liver function tests did not improve prediction for clinical model 2 in the total population (C-statistic change=0.002; P=0.61; NRI=1.2%; P=0.3), a slight improvement, although not statistically significant in terms of discrimination, was observed when men were considered separately (C-statistic change=0.007; P=0.06; NRI=3.3%; P=0.04). In women, addition of liver function tests improved prediction for clinical model 1, but did not improve for clinical model 2 (Table 3). In the PREVEND cohort study, the basic model showed a C-statistic of ( ). Addition of liver function tests improved the C-statistic of the basic model (C-statistic change= 0.019; P<0.001) and led to an IDI of 0.01 (P<0.001) and NRI of 8.7% (P<0.001). After addition of each component of liver function tests alone to the basic models, the C-statistic changes were (P<0.001), (P=0.002), (P=0.29) and (P=0.98) for GGT, ALT, AST and albumin, respectively. Addition of liver function tests improved prediction for clinical model 1 in the total population (change of C-statistic=0.008; P=0.003; NRI=3.6%; P=0.03), with the largest change in men (C-statistic change=0.013; P=0.01; NRI=5.4%; P=0.04) (Table 1). In both cohorts, predictive power slightly increased when we added more liver function tests to the KORA model. For example, in the EPIC-NL study, NRI increased from 6% to 9.5% when we added the panel of all four available liver function tests to the KORA model rather than only the combination of GGT+ALT (Table S2). In both cohorts, the basic and clinical models provided slightly better discrimination in women than in men. For example, in the EPIC-NL study, the C- statistic of the basic model was ( ) in women while this was ( ) in men. For clinical model 2, the C-statistic was ( ) in women while this was ( ) in men. In the PREVEND study, the C- statistic of the basic model was ( ) in women while this was ( ) in men. For clinical model 1, the C-statistic was ( ) in women while this was ( ) in men (Table 3). In a sensitivity analysis, our results using data only from the MORGEN cohort with less than 10% missing values for non-fasting glucose were comparable with our results using both cohorts of the EPIC-NL study. Addition of liver function tests improved the C-statistic of the basic KORA model (C-statistic change= 0.020; P<0.001) and led to an IDI of (P<0.001) and NRI of 9.3% (P<0.001). Addition of liver function tests did not improve prediction for clinical model 2 (C-statistic change=0.004, P=0.10; IDI=0.003, P=0.11; NRI=2.2%, P=0.20). 91

95 Table 2. Associations of liver function tests with the risk of type 2 diabetes EPIC-NL case-chort study HR (95%) per log2 unit increase PREVEND cohort study OR (95% CI) per log2 unit increase Liver markers Step 1 Step 2 Step 3 Step 1 Step 2 Total GGT U/liter 2.08 ( ) 1.49 ( ) 1.45 ( ) 1.76 (1.60,1.93) 1.22 ( ) ALT U/liter 2.05 ( ) 1.22 ( ) 1.05 ( ) 1.88 (1.66, 2.14) 1.29 ( ) AST U/liter 1.81 ( ) 0.97 ( ) 1.04 ( ) 1.63 (1.31, 2.02) 1.16 ( ) Albumin g/l 0.56 ( ) 0.34 ( ) 0.42 ( ) 0.50 (0.17, 1.49) 0.31 ( ) Women GGT U/liter 2.03 ( ) 1.45 ( ) 1.40 ( ) 1.86 ( ) 1.22 ( ) ALT U/liter 1.85 ( ) 1.07 ( ) 0.92 ( ) 1.95 ( ) 1.29 ( ) AST U/liter 1.71 ( ) 0.80 ( ) 0.85 ( ) 1.44 ( ) 1.20 ( ) Albumin g/l 0.49 ( ) 0.23 ( ) 0.53 ( ) 0.45 ( ) 0.36 ( ) Men GGT U/liter 1.98 ( ) 1.43 ( ) 1.33 ( ) 1.67 ( ) 1.22 ( ) ALT U/liter 2.62 ( ) 1.87 ( ) 1.64 ( ) 1.80 ( ) 1.29 ( ) AST U/liter 2.08 ( ) 1.66 ( ) 1.73 ( ) 1.62 ( ) 1.15 ( ) Albumin g/l 0.44 ( ) 4.73 ( ) 4.23 ( ) 0.38 ( ) 0.22 ( ) In step, we adjusted for age and sex (in total populations); step 2, further adjusted for BMI (kg m2), ex-smoker (yes = 1, no = 0), current smoking (yes = 1, no = 0), parental diabetes (yes = 1, no = 0), hypertension (yes = 1, no = 0), glucose (mmol/l) and uric acid (µmol/l); and model 3 further adjusted for HbA1c (only in the EPIC-NL case-cohort study). EPIC-NL denotes European Prospective Investigation Into Cancer, PREVEND Prevention of Renal and Vascular End-stage Disease, CI confidence interval. 92

96 Table 3. Incremental predictive value of liver function tests for the risk of type 2 diabetes EPIC-NL case-cohort study PREVEND cohort study Prediction Models C value (95% CI) Total sample KORABasic ( ) KORABasic + LFTs ( ) KORABasic + Glucose+ UAS ) KORABasic + Glucose+ UAS + LFTs ( ) KORABasic + Glucose+ UAS +HbA1c ( ) KORABasic + Glucose+ UAS +HbA1c+ LFTs ( ) Women KORABasic ( ) KORABasic + LFTs ( ) KORABasic + Glucose+ UAS ( ) KORABasic + Glucose+ UAS + LFTs ( ) KORABasic + Glucose+ UAS +HbA1c ( ) KORABasic + Glucose+ UAS +HbA1c+ LFTs ( ) Men KORABasic ( ) P value IDI (P value) NRI (%) C value (P value) (95% CI) Ref. Ref. Ref ( ) < < , < ( ) Ref. Ref. Ref ( ) < , < , < ( ) P value IDI, P value NRI (%), P value Ref. Ref. Ref. < , < , <0.001 Ref. Ref. Ref , <0.001 Ref. Ref. Ref , , 0.3 Ref. Ref. Ref ( ) < , < , < ( ) Ref. Ref. Ref ( ) < , < , < ( ) 3.6, Ref. Ref. Ref , < , 0.03 Ref. Ref. Ref , Ref. Ref. Ref , , 0.40 Ref. Ref. Ref ( ) 2.2, Ref. Ref. Ref. 93

97 KORABasic + LFTs ( ) KORABasic + Glucose+ UAS ( ) KORABasic + Glucose+ UAS + LFTs ( ) KORABasic + Glucose+ UAS +HbA1c ( ) KORABasic + Glucose+ UAS +HbA1c+ LFTs ( ) < , < , < ( ) Ref. Ref. Ref ( ) , < , ( ) < , < , Ref. Ref. Ref , <0.001 Ref. Ref. Ref , , , KORABasic model included data on age, BMI (kg m2), ex-smoker (yes = 1, no = 0), current smoking (yes = 1, no = 0), parental diabetes (yes = 1, no = 0), hypertension (yes = 1, no = 0). EPIC-NL denotes European Prospective Investigation Into Cancer, PREVEND Prevention of Renal and Vascular End-stage Disease, CI confidence interval, IDI integrated discrimination improvement, NRI, net reclassification improvement, LFTs, liver function tests (including aspartate aminotransferase, alanine aminotransferase, γ-glutamyl transpeptidase and albumin), KORA Cooperative Health Research in the Region of Augsburg, HbA1c glycated haemoglobin, UAS serum uric acid. 94

98 Discussion In this prospective analysis, we examined whether addition of liver function tests could be useful to improve prediction of developing type 2 diabetes above the basic and clinical models in two independent large population-based cohorts. We observed that addition of liver function tests improved prediction modestly only for a basic model without biomarkers in terms of discrimination and reclassification in each cohort. Furthermore, addition of liver function tests led to small but statistically significant improvements in prediction based on a clinical model incorporating glucose and serum uric acid, but not if the clinical model also includes HbA1c. However, there was a slightly better improvement in prediction for men. Several studies have been performed to investigate the associations of liver function tests with type 2 diabetes and its related outcomes 1, 4. However, just a limited number of studies aimed to examine the incremental predictive value of these markers over available prediction models. A analysis of the EPIC-Potsdam cohort showed that a combination of triglycerides, HDL- cholesterol, GGT and ALT further improved prediction based on the German diabetes risk score incorporating glucose and HbA1c in the same population in which they had previously developed that risk score 5. In the DESIR study, a model with GGT and glucose showed an improved prediction in men compared with a basic model incorporating data on smoking, waist circumference and hypertension 6. Recently, the British Heart Study showed that a clinical model incorporating GGT plus HbA1c improved prediction compared with a basic simple model. However, addition of GGT itself had little improvement above a clinical model incorporating glucose, HDL cholesterol and triglyceride 7. Of note, it is particularly important for the value of biomarkers to be examined in an independent setting, because the improvement in measures of prediction can be overestimated if the same population is used for development and evaluation of the incremental value of new biomarkers 10, 28, 29. In our study, we scientifically evaluated incremental predictive value of liver function tests in two independent Dutch populations because we intended to validate our findings in another setting as well. In this way, we took advantage of using a different case mix and slightly different measurement of diabetes between two cohorts 28. Furthermore, we also did this analysis for women and men separately to take into account potential sex differences in the risk prediction of diabetes 6, 30. For example, we and others have shown that prediction models might have a slightly better performance to identify women at high risk 6, 11, 13. We observed no differences in the incremental predictive value of liver function tests above the basic and clinical model incorporating glucose plus uric acid between women and men. However, there was a statistically significant improvement in prediction only in men when we added liver function tests to a clinical model incorporating glucose plus uric acid plus HbA1c. At population level, it is true that a prediction model, like the KORA basic model, incorporating 6 predictors, performs well to identify the individuals at high risk of future diabetes for 7.5 years. In our study, addition of liver function tests did 95

99 hardly result in any improvement of prediction once additional data on glycaemia indices were included. The reason why improvements in predictions were limited in the latter clinical models is that the glycaemia indices are integral parts of the clinical outcome of interest, i.e., diabetes. Diabetes itself is defined by certain cut-offs for glucose and/or HbA1c 31. Like previous studies 1, 4, we demonstrated significant associations of some components of liver function tests with the risk of type 2 diabetes. The associations were independent of common risk factors but addition of liver function tests only minimally to modestly improve the risk prediction of disease. In other words, the absolute difference of certain (bio)markers between individuals who develop and those who remain free of diabetes at a population level is not likely to resolve whether a (bio)marker can be useful for prediction 29. In fact, on an individual level, the range of marker levels between cases and non-cases overlap, limiting its incremental predictive value 29, 32. In contrast, although a certain (bio)marker does not show statistical significance in an etiologic relation, it might still have incremental predictive value in combination with other predictors. So from that point of view it is reasonable to examine all four components of liver function tests in each model. All the basic and clinical models showed slightly better discrimination performance in the EPIC-NL case-cohort study than in the PREVEND cohort study overall and particularly in men. This difference might be explained by differences in heterogeneity between these two populations 33. Larger heterogeneity between individuals make it easier to differentiate between those at high and low risk and may thus lead to higher C-statistics. For example, variables like age and sex may have larger heterogeneity in the EPIC-NL cohort when compared with the PREVEND cohort. Another explanation for this is that we ascertained incident cases differently in each cohort. Therefore, we adjusted the KORA basic model for this difference in incidence of diabetes between development population and both of our populations. Although C-statistics are insensitive to error in average outcome, different ascertainment of outcome might have affected discrimination performance of models 22. However, the incremental predictive value of liver function tests was comparable above each model for both cohorts It is worthy to mention that our findings are in line with prior evidence on this topic showing minimal to modest prediction improvement for risk of future diabetes 5, 7. As a general limitation, we should mention that the reclassification improvement is strongly determined by the cut-off values for the risk categories. As we have previously explained 34, in the diabetes prediction the clinically-relevant cut-off values are not clearly stated yet. In fact, it is hard to judge the clinical utility of liver function test at this time. Diagnosis of diabetes is always challenging in observational studies because indivduals with type 2 diabetes may remain undiagnosed for several months to years. Since we used data of self-reports, some cases of type 2 diabetes may have been undetected. Finally, the PREVEND cohort was enriched with individuals with a higher urinary albumin concentration. Therefore, we performed weighted analysis to be able to generalize 96

100 our findings to the general population. Further studies are warranted to replicate current findings for long-term risk and subsequently evaluate the incremental value of liver function tests 10, 28. We conclude that a combination of liver function tests can modestly improve prediction of medium-term risk of type 2 diabetes above the basic risk model and the clinical model incorporating data on glucose and serum uric acid. If data on HbA1c are available, these markers of liver injury are of little added predictive value. A slightly better improvement in prediction may be present in men. 97

101 Acknowledgments This work was supported by the Netherlands Heart Foundation, the Dutch Diabetes Research Foundation and the Dutch Kidney Foundation. This research was performed within the framework of the Center for Translational Molecular Medicine (CTMM; project PREDICCt (grant 01C ). The EPIC-NL study was funded by Europe against Cancer Programme of the European Commission (SANCO), the Dutch Ministry of Health, the Dutch Cancer Society, the Netherlands Organization for Health Research and Development (ZonMW), and World Cancer Research Fund (WCRF). We thank Statistics Netherlands and the PHARMO Institute for follow-up data on cancer, cardiovascular disease and vital status. None of the study sponsors had a role in the study design, data collection, analysis or interpretation, report writing, or the decision to submit the report for publication. No potential conflicts of interest relevant to this article are reported. 98

102 References 1. Hanley AJ, Williams K, Festa A, et al. Elevations in markers of liver injury and risk of type 2 diabetes: the insulin resistance atherosclerosis study. Diabetes 2004;53(10): Nakanishi N, Suzuki K, Tatara K. Serum gamma-glutamyltransferase and risk of metabolic syndrome and type 2 diabetes in middle-aged Japanese men. Diabetes Care 2004;27(6): Wannamethee SG, Shaper AG, Lennon L, Whincup PH. Hepatic enzymes, the metabolic syndrome, and the risk of type 2 diabetes in older men. Diabetes Care 2005;28(12): Fraser A, Harris R, Sattar N, Ebrahim S, Davey Smith G, Lawlor DA. Alanine aminotransferase, gamma-glutamyltransferase, and incident diabetes: the British Women's Heart and Health Study and meta-analysis. Diabetes Care 2009;32(4): Schulze MB, Weikert C, Pischon T, et al. Use of multiple metabolic and genetic markers to improve the prediction of type 2 diabetes: the EPIC-Potsdam Study. Diabetes Care 2009;32(11): Balkau B, Lange C, Fezeu L, et al. Predicting diabetes: clinical, biological, and genetic approaches: data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR). Diabetes Care 2008;31(10): Wannamethee SG, Papacosta O, Whincup PH, et al. The potential for a two-stage diabetes risk algorithm combining non-laboratory-based scores with subsequent routine non-fasting blood tests: results from prospective studies in older men and women. Diabet Med 2011;28(1): Ghouri N, Preiss D, Sattar N. Liver enzymes, nonalcoholic fatty liver disease, and incident cardiovascular disease: a narrative review and clinical perspective of prospective data. Hepatology 2010;52(3): McGeechan K, Macaskill P, Irwig L, Liew G, Wong TY. Assessing new biomarkers and predictive models for use in clinical practice: a clinician's guide. Arch Intern Med 2008;168(21): Hlatky MA, Greenland P, Arnett DK, et al. Criteria for evaluation of novel markers of cardiovascular risk: a scientific statement from the American Heart Association. Circulation 2009;119(17): Abbasi A, Corpeleijn E, Peelen LM, et al. External validation of the KORA S4/F4 prediction models for the risk of developing type 2 diabetes in older adults: the PREVEND study. Eur J Epidemiol 2012; 27(1): Rathmann W, Kowall B, Heier M, et al. Prediction models for incident type 2 diabetes mellitusin the older population: KORA S4/F4 cohort study. Diabet Med 2010;27(10): Buijsse B, Simmons RK, Griffin SJ, Schulze MB. Risk assessment tools for identifying individuals at risk of developing type 2 diabetes. Epidemiol Rev 2011;33(1): Beulens JW, Monninkhof EM, Verschuren WM, et al. Cohort profile: the EPIC-NL study. Int J Epidemiol 2010;39(5): Lambers Heerspink HJ, Brantsma AH, de Zeeuw D, Bakker SJ, de Jong PE, Gansevoort RT. Albuminuria assessed from first-morning-void urine samples versus 24-hour urine collections as a predictor of cardiovascular morbidity and mortality. Am J Epidemiol 2008;168(8): Volovics A, Van Brandt PAD. Methods for the Analyses of Case-Cohort Studies. Biometrical Journal 1997;39(2): Sluijs I, van der AD, Beulens JW, et al. Ascertainment and verification of diabetes in the EPIC-NL study. Neth J Med 2010;68(1): Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin and risk of type 2 diabetes in the general population. Diabetologia 2011;54(9): Volovics A vdbp. Methods for the Analyses of Case-Cohort Studies. Biom J 1997;39: Abbasi A PL, Corpeleijn E, van der Schouw YT, Stolk RP, Spijkerman AM, van der A DL, Moons KGM, Navis G, Bakker SJ, Beulens JW Prediction models for risk of developing type 2 diabetes: systematic literature search and independent external validation study. BMJ 2012;345:e van Houwelingen HC. Validation, calibration, revision and combination of prognostic survival models. Stat Med 2000;19(24):

103 22. Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010;21(1): Cook NR, Ridker PM. Advances in measuring the effect of individual predictors of cardiovascular risk: the role of reclassification measures. Ann Intern Med 2009;150(11): Paynter NP, Mazer NA, Pradhan AD, Gaziano JM, Ridker PM, Cook NR. Cardiovascular risk prediction in diabetic men and women using hemoglobin A1c vs diabetes as a high-risk equivalent. Arch Intern Med 2011;171(19): Herder C, Baumert J, Zierer A, et al. Immunological and cardiometabolic risk factors in the prediction of type 2 diabetes and coronary events: MONICA/KORA Augsburg case-cohort study. PLoS One 2011;6(6):e Shafizadeh TB, Moler EJ, Kolberg JA, et al. Comparison of accuracy of diabetes risk score and components of the metabolic syndrome in assessing risk of incident type 2 diabetes in Inter99 cohort. PLoS One 2011;6(7):e Donders AR, van der Heijden GJ, Stijnen T, Moons KG. Review: a gentle introduction to imputation of missing values. J Clin Epidemiol 2006;59(10): Moons KG. Criteria for scientific evaluation of novel markers: a perspective. Clin Chem 2010;56(4): Chao C, Song Y, Cook N, et al. The lack of utility of circulating biomarkers of inflammation and endothelial dysfunction for type 2 diabetes risk prediction among postmenopausal women: the Women's Health Initiative Observational Study. Arch Intern Med 2010;170(17): Ding EL, Song Y, Malik VS, Liu S. Sex differences of endogenous sex hormones and risk of type 2 diabetes: a systematic review and meta-analysis. JAMA 2006;295(11): Association AD. Standards of medical care in diabetes Diabetes Care 2011;34 Suppl 1:S Herder C, Karakas M, Koenig W. Biomarkers for the prediction of type 2 diabetes and cardiovascular disease. Clin Pharmacol Ther 2011;90(1): Vergouwe Y, Moons KG, Steyerberg EW. External validity of risk models: Use of benchmark values to disentangle a case-mix effect from incorrect coefficients. Am J Epidemiol 2010;172(8): Abbasi A, Corpeleijn E, Meijer E, et al. Sex differences in the association between plasma copeptin and incident type 2 diabetes: the Prevention of Renal and Vascular Endstage Disease (PREVEND) study. Diabetologia 2012;55(7):

104 Appendix Observed Risk (7.5yr) A Ideal Nonparametric Predicted Risk (7.5yr) Observed Risk (7.5yr) B Ideal Nonparametric Predicted Risk (7.5yr) Figure S1. Calibration plots for comparison of the predicted 7.5-year risk of diabetes (according to the KORA basic model) against observed risk of developing type 2 diabetes. Panel A (the EPIC-NL case-cohort study), Panel B (the PREVEND cohort study). The ideal and nonparametric terms, the dashed line denotes the ideal calibration line (slope=1, intercept=0) and the dotted line denotes smooth calibration curve for each models. Hosmer-Lemeshow χ 2 statistic were 14.7 (P=0.10) and 7.8 (P=0.56) for the calibration performance of KORA basic model (after adjustment for the intercept and the slope) in the EPIC-NL and in the PREVEND studies, respectively. 101

105 Table S1. Missing data pattern in extrapolated EPIC-NL case-cohort study and PREVEND cohort study Variables Extrapolated EPIC-NL case-cohort study Missing Percent values PREVEND cohort study Missing values Percent Age Sex Incident type 2 diabetes Weight Height Body Mass Index Smoking Family history of diabetes Systolic blood pressure Diastolic blood pressure History of hypertension Antihypertensive medication Glucose HbA1c Uric acid GGT AST ALT Albumin EPIC-NL denotes European Prospective Investigation Into Cancer (the Netherlands), PREVEND Prevention of Renal and Vascular End-stage Disease, AST, aspartate aminotransferase, ALT, alanine aminotransferase, GGT, γ-glutamyl transpeptidasein, HbA1c glycated haemoglobin. 102

106 Table S2. Incremental predictive value of components of liver function tests for the risk of future type 2 diabetes Prediction Models C value (95% CI) Total sample KORABasic ( ) KORABasic + GGT ( ) KORABasic + ALT ) KORABasic + AST ( ) KORABasic + ALB ( ) KORABasic + ALT+ GGT ( ) KORABasic + LFTs ( ) EPIC-NL case-cohort study PREVEND cohort study P value IDI (P value) NRI (%) (P value) C value (95% CI) Ref. Ref. Ref ( ) < , < , < < , < , < , < ( ) ( ) ( ) ( ) ( ) ( ) P value IDI, P value NRI (%), P value Ref. Ref. Ref. < , < , < , < <0.001 < , < , , , <0.001 KORABasic model included data on age, BMI (kg m 2 ), ex-smoker (yes = 1, no = 0), current smoking (yes = 1, no = 0), parental diabetes (yes = 1, no = 0), hypertension (yes = 1, no = 0). EPIC-NL denotes European Prospective Investigation Into Cancer, PREVEND Prevention of Renal and Vascular End-stage Disease, CI confidence interval, IDI integrated discrimination improvement, NRI, net reclassification improvement, LFTs, liver function tests (including aspartate aminotransferase, alanine aminotransferase, γ-glutamyl transpeptidase and albumin), KORA Cooperative Health Research in the Region of Augsburg, HbA1c glycated haemoglobin, UAS serum uric acid. 103

107 104

108 Chapter 5 Maternal and paternal transmission of type 2 diabetes: influence of diet, lifestyle and adiposity Ali Abbasi 1,2,3 ; Eva Corpeleijn 1 ; Yvonne T. van der Schouw 3 ; Ronald P. Stolk 1 ; Annemieke M.W. Spijkerman 4 ; Daphne L. van der A 5 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 ; Joline W.J. Beulens 3 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands 4 Center for Prevention and Health Services Research, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands 5 Center for Nutrition and Health, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands J Intern Med. 2011;270:

109 Abstract Background Transmission of family history of type 2 diabetes to the next generation is stronger for maternal than paternal diabetes in some populations. The aim of the present study was to investigate whether this difference is explained by diet, lifestyle factors and or adiposity. Methods We analysed 35,174 participants from the Dutch contribution to the European Prospective Investigation into Cancer and Nutrition, a prospective population-based cohort (aged years) with a median follow-up of 10.2 years. Parental history of diabetes was self-reported. Occurrence of diabetes was mainly identified by self-report and verified by medical records. Results Amongst 35,174 participants, 799 incident cases of diabetes were observed. In age-and sex-adjusted analyses, hazard ratio (HR) and 95% confidence intervals (CIs) for diabetes by maternal and paternal diabetes were 2.66 ( ) and 2.40 ( ), respectively. Maternal transmission of risk of diabetes was explained by diet (9.4%), lifestyle factors including smoking, alcohol consumption, physical activity and educational level (7.8%) and by adiposity, i.e. body mass index and waist and hip circumference (23.5%). For paternal transmission, the corresponding values were 2.9%, 0.0% and 9.6%. After adjustment for diet, lifestyle factors and adiposity, the HRs for maternal (2.20; 95%CI, ) and paternal (2.23; 95% CI, ) transmission of diabetes were comparable. Conclusions Both maternal and paternal diabetes are associated with increased risk of type 2 diabetes, independently of diet, lifestyle and adiposity. The slightly higher risk conferred by maternal compared to paternal diabetes was explained by a larger contribution of diet, lifestyle factors and adiposity. 106

110 Introduction The prevalence of type 2 diabetes, a leading cause of cardiovascular morbidity and mortality, is increasing worldwide 1. Parental paternal and/or maternal history of diabetes is a major determinant of increased risk of diabetes 2-5. Family history may reflect complex relationships between genetic factors and environmental conditions that are important for developing diabetes 6. Thus, parental history of diabetes includes environmental risks (e.g. non-genetic familial behaviours, lifestyle and obesity) beyond the genetic risk factors for diabetes 4. A greater risk from maternal type 2 diabetes compared to paternal diabetes has been reported in some 7-9 but not all studies 2,4,5. A variety of explanations for this greater importance of maternal diabetes have included: genomic imprinting (ie the differential expression of inherited susceptibility genes in paternal or maternal generation 10 ; mutations in mitochondrial DNA, which are maternally inherited 11 ; and metabolic programming during intrauterine exposure 12. It is still not clear to what extent modifiable factors such as diet, lifestyle and obesity can explain the association between maternal or paternal diabetes and risk of diabetes. To our knowledge, only one prospective study among female nurses has investigated the contribution of excess adiposity and certain dietary habits 5. No such longitudinal data are available in men. The aim of this study was to prospectively investigate the association between parental history of diabetes maternal and/or paternal and risk of incident type 2 diabetes in a population-based cohort of male and female adults, ie the Dutch contribution to the European Prospective Investigation into Cancer and Nutrition (EPIC-NL). The EPIC-NL study was suitable for this purpose because detailed data on diabetes risk factors such as diet and lifestyle factors were collected in this cohort 13. Methods Study population and design The EPIC-NL cohort (n=40,011) includes the Monitoring Project on Risk Factors for Chronic Diseases (MORGEN) and Prospect cohorts, initiated between 1993 and Details of the EPIC-NL study design, recruitment and procedures have been described in more detail previously 13. Briefly, Prospect is a prospective populationbased cohort study of 17,357 women aged years who participated in a breast cancer screening programme. In the MORGEN study, 22,654 individuals aged years were recruited from Amsterdam, Doetinchem and Maastricht. A new random sample of about 5000 participants was examined each year. These rounds of enrolment add up to this number of individuals. The participation rates were 34.5% for Prospect and 45.0% for MORGEN. At baseline, a general questionnaire and a food frequency questionnaire (FFQ) were sent by post to all participants and these were returned after completion at the 107

111 medical examination. We excluded 1150 participants with missing data on baseline characteristics or extreme values for energy intake (<450 or >6000 kcal/day) and 2360 participants with unknown history of parental diabetes. Further subjects were excluded because of prevalent type 2 diabetes (n=507) or missing recordings of censoring time (n=820). Follow-up time was calculated from the date of enrolment to the date of diabetes diagnosis or death. All other participants were censored at the end of follow-up (January 2006). Finally, 35,174 participants were included in the cross-sectional and prospective analyses. All participants gave written informed consent prior to study inclusion. Both cohort studies complied with the Declaration of Helsinki. Prospect was approved by the Institutional Review Board of the University Medical Center Utrecht and MORGEN was approved by the Medical Ethics Committee of the Netherlands Organization for Applied Scientific Research. General measurements The general questionnaire contained questions on demographic characteristics and risk factors for the presence of chronic diseases. For both cohorts, coding of this information was standardized and merged into one uniform database. Body weight, height and waist and hip circumference were measured according to standard procedures. Smoking status was categorized into current, past and never smoker. Physical activity was assessed using a questionnaire validated in an elderly population and categorized as inactive, moderately inactive, moderately active and active, according to the Cambridge Physical Activity Score 14. Low education level was defined as primary education, lower vocational education or advanced elementary education. Blood pressure was measured twice on the left arm. The mean of the two blood pressure measurements was used in the analysis. In the Prospect study, systolic and diastolic blood pressures were measured with the participants in the supine position using a Boso Oscillomat (Bosch & Sohn, Jungingen, Germany), whereas a random-zero sphygmomanometer (Hawksley & Sons, Lancing, UK) with the participant in the sitting position was used in the MORGEN cohort. The comparability of these different measurement procedures has been described in more detail previously 15. The assessment of the Prospect cohort slightly overestimated blood pressure compared with the MORGEN cohort. Hypertension was defined based on self-report of diagnosis by a physician, measured hypertension ( 140 mmhg systolic blood pressure or 90 mmhg diastolic blood pressure) or the use of blood pressure-lowering medication. Hyperlipidaemia was defined based on selfreport of diagnosis by a physician or the use of lipid-lowering therapy. In both cohorts, daily food intake was determined using the same validated FFQ 16,17, which contains questions on the usual frequency of consumption of 79 main food groups during the year preceding enrolment. Overall, the questionnaire enables estimation of the average daily consumption of 178 foods. Intakes of different nutrients were adjusted for total energy intake using the regression residual method

112 Assessment of parental history of diabetes Parental history of diabetes was obtained by self-report. Participants were asked whether their biological mother and/or father had (whether alive or deceased) previously been diagnosed with diabetes. Parental history of diabetes was categorized as none, any parent(s) (mother and/or father), maternal only, paternal only or both. Assessment of type 2 diabetes Occurrence of diabetes during follow-up was self-reported via two follow-up questionnaires at 3- to 5-year intervals in the MORGEN and Prospect studies. In the Prospect study, incident cases of diabetes were also detected as glucosuria via a urinary glucose strip test, which was sent out with the first follow-up questionnaire. Diagnoses of diabetes were also obtained from the Dutch Center for Health Care Information, which holds a standardized computerized register of hospital discharge diagnoses. Follow-up was complete until 1 January Potential cases identified by these methods were verified against general practitioner or pharmacist records (Prospect only) via postal questionnaires 19. Diabetes was defined as being present when the diagnosis was confirmed by either of these methods. For 89% of participants with potential diabetes, verification information was available, and 72% were verified as having type 2 diabetes and were thus included in the analysis 19. Statistical analyses Baseline descriptive statistics of the continuous variables were reported as mean ± standard deviation (SD) and groups were compared using two-tailed Student s t-test or ANOVA. Categorical variables were presented as numbers and percentages and a χ 2 test was used to test the differences between participants without parental history of diabetes and those with each category of parental history of diabetes for these variables. Generalized linear models were used to assess the cross-sectional associations between parental diabetes and baseline parameters of obesity, including body mass index (BMI) and waist and hip circumference in participants. Multivariable models were adjusted for cohort (Prospect or MORGEN), age, sex, lifestyle factors (smoking, alcohol consumption, physical activity level and educational level) total energy intake and energy-adjusted dietary factors. The dietary factors included the amount of intake of fat, protein, carbohydrate, cholesterol, fibre, vitamin C and vitamin E. The estimated marginal means and 95% confidence intervals (CIs) were reported and linear regression β coefficients and 95% CIs were calculated for each category of parental history of diabetes compared to the reference group of participants without any parental history of diabetes. The association between parental history of diabetes and incident diabetes in participants was assessed by Cox proportional hazard regression. In the crude model (controlled for cohort), hazard ratios (HRs) and 95% CIs for diabetes were calculated for each category of parental diabetes against a reference group of participants 109

113 without parental history of diabetes. In Model 1, basic adjustments were made for age and sex. We assessed the effect of sex by including the interaction of sex with parental diabetes in this model. Moreover, the stratified analyses for sex were fitted in adjusted models for age and other covariates. Lifestyle factors were added in Model 2. Total energy intake and energy-adjusted dietary factors were added in Model 3. Parameters of obesity were subsequently included in the final model (Model 4). We then separately added each factor to Model 1 to determine its contribution to the association between parental history of diabetes and risk of diabetes. Inclusion of these factors in the model would be expected to attenuate the HR related to parental diabetes. We calculated the percentage attenuation of the HR for each category of parental diabetes. Percentage attenuation of HR was calculated as: (HR before addition HR after addition)/(hr before addition 1) 100. A P value of 0.05 or less from two-sided tests was considered statistically significant. All statistical analyses were carried out using Statistical Package for Social Sciences version 16 (SPSS Inc, Chicago, IL, USA) and STATA software version 10.0 (Stata-Corp LP, College Station, TX, USA). Results Baseline characteristics of the study population are summarized in Table 1 by parental diabetes status. When compared with participants without parental history of diabetes, those who reported paternal and/or maternal diabetes were older and more likely to be female, had a higher BMI, waist and hip circumference and blood pressure, a lower alcohol consumption and education level, were less physically active, and were more likely to experience cardiovascular morbidity. Parental diabetes was associated with a lower intake of total energy and carbohydrates, whereas the intake of protein, fat, fibre, vitamin C and vitamin E was higher in participants with parental diabetes. Cross-sectional analysis Table 2 shows the association between parental history of diabetes and parameters of obesity in participants. We calculated adjusted means of parameters of obesity in each category of parental diabetes accounting for cohort, age, sex, diet and lifestyle factors. Subjects with maternal and/or paternal diabetes had higher BMI (β coefficient, 0.65; 95% CI, ), waist circumference (β coefficient, 1.88; 95% CI, ) and hip circumference (β coefficient, 0.93; 95% CI, ) compared with participants without parental diabetes. This association was stronger for those with both maternal and paternal history of diabetes. Prospective analysis During a median follow-up of 10.2 years, we observed 799 incident cases of type 2 diabetes (rate of 2.2 per 1000 person-years). In the unadjusted analysis, participants with parental history of diabetes had an approximately 3-fold higher incidence rate of diabetes compared with those who had no parents with diabetes (Table 3; 1.7 vs

114 Table 1. Baseline characteristics by parental history of diabetes in the EPIC-NL study (n=35,174) a None Any parent(s) Only father Only mother Both parents No. (%) of participants 28,696 (81.6) 6478 (18.4) 2187 (6.2) 3941 (11.2) 350 (1.0) Age, y 48.4 (12.3) 1, * 51.6 (10.0) 49.7 (10.8) 2, * 52.6 (9.4) 52.0 (9.7) Female 21,026 (73.3) 1, * 5148 (79.5) 1682 (76.9) 3178 (80.6) 288 (82.3) Body mass index, kg/m (3.9) 1, * 26.4 (4.1) 26.0 (3.9) 2, * 26.6 (4.2) 27.2 (4.6) Waist circumference, cm 84.5 (11.2) 1, * 86.8 (11.5) 85.8 (11.3) 2, * 87.3 (11.6) 88.4 (12.0) Hip circumference, cm (7.9) 1, * (8.5) (7.9) 2, * (8.7) (9.7) Systolic blood pressure, mmhg 125 (19) 1, * 128 (19) 126 (19) 2, * 130 (20) 128 (19) Diastolic blood pressure, mmhg 77.4 (10.6) 1, * 78.8 (10.4) 77.9 (10.6) 2, * 79.4 (10.4) 78.2 (10.4) Alcohol consumption, g/week 11.4 (15.6) 1, * 9.9 (14.4) 10.6 (14.4) 9.7 (14.5) 7.9 (12.0) Current smoker 8560 (29.8) 1, * 1873 (28.9) 614 (28.1) 1171 (29.7) 88 (25.1) Low educational level b 15,606 (54.4) 4229 (65.3) 1250 (57.2) 2734 (69.4) 245 (70.0) Physical activity c Inactive Moderately inactive Moderately active Active 1,** 2450 (8.5) 8298 (28.9) 8045 (28.0) 9903 (34.5) 626 (9.7) 1837 (28.3) 1763 (27.2) 2252 (34.8) 185 (8.5) 649 (29.7) 604 (27.6) 749 (34.2) 402 (10.2) 1094 (27.8) 1049 (26.6) 1396 (35.4) 39 (11.1) 94 (26.9) 110 (31.4) 107 (30.6) Total energy intake, kcal/d (636.7) 1, * (601.0) (604.8) (602.4) (556.9) Nutrient intake, g/d d Protein 75.5 (10.9) 1, * 76.7 (10.9) 76.0 (10.7) 2, ** 76.9 (11.0) 78.5 (11.7) Fat 77.3 (11.3) 1, * 78.4 (11.5) 78.0 (10.8) 78.5 (11.9) 79.1 (11.2) Saturated fat 32.4 (5.8) 1, * 33.0 (5.9) 32.6 (5.6) 2, ** 33.1 (6.0) 33.1 (5.6) Monounsaturated fat 29.4 (5.1) 1, ** 29.6 (5.2) 29.6 (5.0) 29.6 (5.4) 29.9 (5.2) Polyunsaturated fat 14.9 (3.8) 1, * 15.1 (4.0) 15.1 (3.9) 15.1 (4.0) 15.5 (4.1) Cholesterol, mg/d (58.3) 1, * (62.6) (59.3) 2, *** (64.1) (63.9) Carbohydrates (30.7) (30.7) (30.0) (31.1) (31.3) Mono- and disaccharide (29.3) 1, *** (29.5) (28.5) (29.9) (30.9) Fibre 23.3 (4.8) 1, * 23.7 (4.8) 23.4 (4.7) 2, *** 23.8 (4.8) 23.7 (4.6) Vitamin C, mg/d (45.1) 1, * (46.1) (44.1) (47.0) (47.8) Vitamin E, mg/d 12.2 (3.2) 1, * 12.4 (3.3) 12.3 (3.3) 12.4 (3.3) 12.7 (3.5) Hyperlipidaemia e 2268 (7.9) 1, * 617 (9.5) 165 (7.5) 414 (10.5) 38 (10.9) Hypertension f 5889 (20.5) 1, * 1621 (25.0) 489 (22.4) 1023 (26.0) 109 (31.1) 111

115 Myocardial infarction 434 (1.5) 1, ** 121 (1.9) 28 (1.3) 82 (2.1) 11 (3.1) Stroke 309 (1.1) 1, * 101 (1.6) 26 (1.2) 65 (1.6) 10 (2.9) Cancer 1187 (4.1) 264 (4.1) 80 (3.7) 171 (4.3) 13 (3.7) a Data were given as mean (SD) for continuous variables, tested using two-tailed Student s t-test or ANOVA, and numbers (percentage) for categorical variables, tested using χ2 test. b Low education level was assigned for participants who had primary education, lower vocational education or advanced elementary education. c Physical activity level was defined based on the Cambridge Physical Activity Index. d Intake of nutrients was adjusted for total energy intake and given in g/d unless otherwise indicated. e Hyperlipidaemia was defined based on self-report of diagnosis by a physician or the use of lipid-lowering medications. f Hypertension was defined based on self-report of diagnosis by a physician,, measured hypertension (>140 systolic blood pressure or >90 diastolic blood pressure) or the use of blood pressure-lowering medications. 1 Comparisons between participants with any parental history of diabetes and none. 2 Comparisons between participants with only paternal and only maternal history of diabetes. *,P<0.001, **P<0.01 and ***P

116 Table 2. Cross-sectional association between parental history of diabetes and BMI and waist and hip circumference (n=35,174) Characteristics None Any parent(s) Only father Only mother Both parents Body mass index, kg/m 2 Estimated means (95% CI) a,b β coefficient (95% CI) Waist circumference, cm Estimated means (95% CI) a,b β coefficient (95% CI) Hip circumference, cm Estimated means (95% CI) a,b β coefficient (95% CI) 25.8 ( ) ( ) ( ) ( ) 0.65 ( ) 89.7 ( ) 1.88 ( ) ( ) 0.93 ( ) 26.2 ( ) 0.39 ( ) 89.1 ( ) 1.28 ( ) ( ) 0.50 ( ) 26.5 ( ) 0.75 ( ) 89.9 ( ) 2.12 ( ) ( ) 1.11 ( ) 27.0 ( ) 1.20 ( ) 90.9 ( ) 3.12 ( ) ( ) 1.55 ( ) a Estimated marginal means (95% CI) were presented in generalized linear models adjusted for cohort, age, sex, smoking, alcohol use, physical activity levels, low educational level, total energy intake and nutrients (fat, protein, carbohydrate, cholesterol, vitamin C, vitamin E and fibre). b P<0.001 for all comparisons between participants with any parental history of diabetes and none. 113

117 Table 3. Parental history of diabetes, contributing factors and risk of incident type 2 diabetes over 10 years (n=35,174) Risk of type 2 diabetes, HR (95% CI) Model adjustment None Any parent(s) Only father Only mother Both parents No. cases/participants 476/ / / / /350 Incidence rate, per 1000 person-year Crude model a 1 (ref) 2.90 ( ) 2.36 ( ) 2.91 ( ) 6.21 ( ) Model 1 1 (ref.) 2.75 ( ) 2.40 ( ) 2.66 ( ) 5.89 ( ) Model 2 1 (ref.) 2.64 ( ) 2.40 ( ) 2.53 ( ) 5.49 ( ) Model 3 1 (ref.) 2.58 ( ) 2.37 ( ) 2.46 ( ) 5.02 ( ) Model 4 1 (ref.) 2.32 ( ) 2.23 ( ) 2.20 ( ) 3.92 ( ) a Crude model was controlled for cohort. Model 1 was adjusted for age and sex. Model 2 was adjusted for factors in Model 1 plus smoking status, alcohol consumption, physical activity levels and low educational level. Model 3 was adjusted for factors in Model 2 plus total energy intake and energy-adjusted nutrients (fat, protein, carbohydrate, cholesterol, vitamin C, vitamin E and fibre). Model 4 was adjusted for factors in Model 3 plus body mass index and waist and hip circumference. 114

118 per 1000 person-years, P<0.001). Despite the sex differences in each category of parental diabetes, there was no significant interaction of sex with parental diabetes (HR of interaction term, 0.86; 95% CI, ). In sex-stratified analyses, multivariable-adjusted HRs of diabetes for maternal and paternal history of diabetes were 2.50 (95% CI, ) and 2.14 (95% CI, ), respectively, in male participants. In females, these values were 2.15 (95% CI, ) and 2.32 (95%CI, ), respectively. In total, crude HRs of incident diabetes for maternal and paternal history of diabetes were 2.91 (95% CI, ) and 2.36 (95% CI, ), respectively, when compared with those who reported no parental diabetes. Model 1 in Table 3 shows that adjustment for age and sex modestly attenuated (13.1% reduction) the risk of diabetes by maternal diabetes (HR, 2.66; 95% CI, ]), whereas this did not contribute to the risk conferred by paternal diabetes (HR, 2.40; 95% CI, ). After multivariable adjustment (Model 4) for age, sex, diet, lifestyle factors and parameters of obesity, risk of diabetes was comparable for maternal (HR, 2.20; 95% CI, ) and paternal history of diabetes (HR, 2.23; 95% CI, ). It is interesting that age, sex, diet, lifestyle factors and parameters of obesity contributed more to the association between maternal diabetes and risk of diabetes (overall attenuation of 37.1%) than paternal diabetes (overall attenuation of 9.6%). Therefore, we separately added each factor to Model 1 to assess its contribution to the association between category of parental diabetes and risk of diabetes. Parameters of obesity explained 23.5% and 9.6%, respectively, of the risk estimation of diabetes by maternal and paternal diabetes. Risk estimation of maternal diabetes was partly explained (9.4%) by energy intake and dietary determinants, whereas this accounted for only 2.9% of the association between paternal diabetes and risk of diabetes. After adjustment for lifestyle factors, an attenuation of 7.8% was observed in the association between maternal diabetes and risk of diabetes, whereas risk estimation of paternal diabetes was not affected by lifestyle factors. Discussion In this prospective cohort with over 10 years of follow-up, we found that both maternal and paternal history of diabetes were associated with baseline diabetes risk factors and with an increased risk of incident type 2 diabetes in participants, independent of diet, lifestyle and adiposity. However, the association between maternal diabetes history and risk of diabetes was slightly stronger in the age- and sex-adjusted model compared with paternal history. More than one-third of the maternal transmission of diabetes was explained by age, sex, diet, lifestyle factors (smoking status, alcohol consumption, physical activity and educational level) and parameters of obesity. The association between paternal diabetes and incident diabetes, however, was explained only modestly (~10%) by diet and parameters of obesity. The main strengths of our study are its large sample size, prospective design, verification of incident diabetes and extensive information about participants diet 115

119 and lifestyle factors. Nevertheless, our study has some limitations. The EPIC-NL cohort almost exclusively comprised Caucasians from the Netherlands, and it is unclear whether our findings could be extended to other ethnic groups. Another limitation is that parental history of diabetes was obtained by self-report which is the usual method in single-generation cohorts. Furthermore, we excluded individuals with missing data or unknown parental history of diabetes. Having unknown family history has been shown to be more common for paternal than for maternal diabetes 20. However, the baseline characteristics of excluded individuals were similar to those who were included in our analysis. Therefore, it is unlikely that this would have led to recall bias or misclassification by category of parental diabetes of participants who did or did not develop diabetes. We relied on self-reported information about lifestyle and diet, which may be subject to misclassification. However, both the physical activity questionnaire and the FFQ have been validated previously 14,16,17. These studies showed that both questionnaires could be used to rank individuals according to their physical activity or diet. We therefore believe that this did not greatly influence our results. Finally, individuals with type 2 diabetes may remain undiagnosed for several months to years and diagnosis of diabetes is always challenging in observational studies. Some cases of type 2 diabetes may have been undetected, resulting in underestimation of the association between parental diabetes and risk of diabetes in participants. We first investigated the association between parental history of diabetes and baseline parameters of obesity in cross-sectional analyses. In multivariable-adjusted models, maternal or paternal diabetes was associated with higher BMI and waist and hip circumference. Having both maternal and paternal diabetes was associated with higher parameters of obesity. These findings are in agreement with those of other studies demonstrating that the presence of a maternal or paternal history of diabetes is associated with greater adiposity 21,22 and weight gain 23, thus suggesting that diabetes and obesity share some common heritable determinants 5,24. A slightly more important role of maternal, compared with paternal, transmission of diabetes was shown in the present study in the age- and sex-adjusted model. This difference (~25%) was explained by diet and adiposity as well as age, sex and lifestyle factors for maternal diabetes. Of note, multivariable models have not been used to assess these factors in previous studies investigating the increased importance of maternal transmission 8,9,21. Among these studies, adiposity substantially explained the risk of diabetes transmitted by maternal diabetes, whereas other factors contributed to a lesser extent. In the present study, both maternal and paternal transmission of diabetes were explained to some extent by obesity parameters, with a 2-fold higher contribution for maternal than for paternal diabetes. The contribution of dietary determinants was also larger in the association between maternal diabetes and risk of diabetes when compared with paternal diabetes. In addition, the influence of age, sex and lifestyle factors was confined to the association between maternal diabetes and risk of diabetes. This finding was not observed in a recent analysis from the Nurses Health 116

120 Study (NHS). In the NHS, BMI rather than waist and hip circumference largely explained the association between both maternal and paternal history of diabetes and risk of type 2 diabetes. In addition to BMI, it was suggested that a higher intake of red meat and sugar-sweetened beverages, and lack of alcohol consumption may explain part of the association between family history of diabetes and risk of diabetes 5. Of interest, in our study, parental history of diabetes was related to a lower intake of total energy and carbohydrates but a higher energy-adjusted intake of fat, protein, fibre, vitamin C and vitamin E. These differences modestly explained the association between maternal diabetes and risk of diabetes, whereas the effect was minimal for paternal transmission. The NHS included a sample of female nurses with limited variation in socioeconomic status and with relatively healthy lifestyle behaviours. This selected sample may have led to an underestimation of the extent to which lifestyle factors explain parental transmission of diabetes 5. The strong risk conferred by maternal or paternal diabetes was comparable after accounting for diet, lifestyle factors and adiposity. Of note, these factors differently explained maternal and paternal transmission of risk of diabetes. Our findings are consistent with previous evidence indicating a stronger magnitude of maternal transmission of obesity and its association with many lifestyle factors, compared with paternal transmission 25. A possible explanation for this is that the mother might have more influence on eating habits and other lifestyle behaviours while raising her children. Indeed, there may be more contact hours between mothers and children during childhood and in later life, and therefore the mother s lifestyle may be more of an example for her children than the father s. It has been shown that if the mother has a history of diabetes during pregnancy, her child is less likely to follow certain healthy dietary recommendations 26. Similarly, those with a maternal history of diabetes may be more prone to have an unhealthy diet and lifestyle throughout their lifetime. Finally, there is evidence to suggest an effect of maternal nutrition and weight maintenance during pregnancy on infant birth weight 27,28. Birth weight could in turn influence future risk of chronic diseases such as type 2 diabetes 29,30. Diabetes is a polygenic disease in which multiple genetic and environmental components play roles throughout all stages of the disease 6,31. Beyond the genetic heritability, parental history of diabetes also carries environmental risk factors. In other words, it seems that parents and children share common lifestyle behaviours which will be continued throughout the children s lifetime. These components, transmitted by maternal or paternal exposures, may explain the heterogeneity of diabetes transmission in different populations. We conclude that both maternal and paternal history of diabetes are associated with an increased risk of developing type 2 diabetes, independent of diet, lifestyle and adiposity. The slight excess risk conferred by maternal compared to paternal diabetes is explained by a larger contribution to this association of age, sex, diet, lifestyle factors and adiposity. 117

121 Acknowledgments This work was supported by the Netherlands Heart Foundation, the Dutch Diabetes Research Foundation and the Dutch Kidney Foundation. This research was performed within the framework of the Center for Translational Molecular Medicine (CTMM; project PREDICCt (grant 01C ). The EPIC-NL study was funded by Europe against Cancer Programme of the European Commission (SANCO), the Dutch Ministry of Health, the Dutch Cancer Society, the Netherlands Organisation for Health Research and Development (ZonMW), and World Cancer Research Fund (WCRF). We thank Statistics Netherlands and the PHARMO Institute for follow-up data on cancer, cardiovascular disease and vital status. None of the study sponsors had a role in the study design, data collection, analysis or interpretation, report writing, or the decision to submit the report for publication. No potential conflicts of interest relevant to this article were reported. 118

122 References 1. Wild S, Roglic G, Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the year 2000 and projections for Diabetes Care 2004; 27: Meigs JB, Cupples LA, Wilson PW. Parental transmission of type 2 diabetes: the Framingham Offspring Study. Diabetes 2000; 49: Wilson PW, Meigs JB, Sullivan L, Fox CS, NathanDM, D Agostino RB Sr. Prediction of incident diabetes mellitus in middle-aged adults: the Framingham Offspring Study. Arch Intern Med 2007; 167: Meigs JB, Shrader P, Sullivan LM, et al. Genotype Score in Addition to Common Risk Factors for Prediction of Type 2 Diabetes. N Engl J Med 2008; 359: van 't Riet E, Dekker JM, Sun Q, Nijpels G, Hu FB, van Dam RM. The role of adiposity and lifestyle in the relationship between family history of diabetes and 20-year incidence of type 2 diabetes in U.S. women. Diabetes Care 2010 doi: /dc Grarup N, Andersen G. (2007) Gene-environment interactions in the pathogenesis of type 2 diabetes and metabolism.curr Opin Clin Nutr Metab Care 2007; 10: Vaag A, Lehtovirta M, Thye-Rönn P, Groop L; European Group of Insulin Resistance. Metabolic impact of a family history of Type 2 diabetes. Results from a European multicentre study (EGIR). Diabet Med 2001; 18: Karter AJ, Rowell SE, Ackerson LM, et al. Excess maternal transmission of type 2 diabetes. The Northern California Kaiser Permanente Diabetes Registry. Diabetes Care 1999; 22: Groop L, Forsblom C, Lehtovirta M, et al. Metabolic consequences of a family history of NIDDM (The Botnia Study): evidence for sex-specific parental effects. Diabetes 1996; 45: Rampersaud E, Mitchell BD, Naj AC, Pollin TI. Investigating parent of origin effects in studies of type 2 diabetes and obesity. Curr Diabetes Rev 2008; 4: Maassen JA, Janssen GM, t Hart LM. Molecular mechanisms of mitochondrial diabetes (MIDD). Ann Med 2005; 37: Fetita LS, Sobngwi E, Serradas P, Calvo F, Gautier JF. Consequences of fetal exposure to maternal diabetes in offspring, J Clin Endocrinol Metab 2006; 91: Beulens JW, Monninkhof EM, Verschuren WM. Cohort profile: The EPIC-NL study. Int J Epidemiol 2009 doi: /ije/dyp Voorrips LE, Ravelli AC, Dongelmans PC, Deurenberg P, Van Staveren WA. A physical activity questionnaire for the elderly. Med Sci Sports Exerc 1991; 23: Schulze MB, Kroke A, Saracci R, Boeing H. The effect of differences in measurement procedure on the comparability of blood pressure estimates in multi-centre studies. Blood Press Monit 2002; 7: Ocké MC, Bueno-de-Mesquita HB, Goddijn HE, et al. The Dutch EPIC Food Frequency Questionnaire. I. Description of the questionnaire, and relative validity and reproducibility for food groups. Int J Epidemiol 1997; 26: S37-S Ocké MC, Bueno-de-Mesquita HB, Pols MA, Smit HA, van Staveren WA, Kromhout D. The Dutch EPIC Food Frequency Questionnaire. II. Relative validity and reproducibility for nutrients. Int J Epidemiol 1997; 26: S49-S Schulze MB, Liu S, Rimm EB, Manson JE, Willett WC, Hu FB. Glycemic index, glycemic load, and dietary fiber intake and incidence of type 2 diabetes in younger and middle-aged women. Am J Clin Nutr 2004; 80: Sluijs I, van der A DL, Beulens JW, et al Ascertainment and verification of diabetes in the EPIC- NL study. Neth J Med 2010; 68: Thorand B, Liese AD, Metzger MH, Reitmeir P, Schneider A, Löwel H. Can inaccuracy of reported parental history of diabetes explain the maternal transmission hypothesis for diabetes? Int J Epidemiol 2001; 30:

123 21. Lee SC, Pu YB, Chow CC, et al. Diabetes in Hong Kong Chinese: evidence for familial clustering and parental effects, Diabetes Care 2000; 23: Shaw JT, Purdie DM, Neil HA, Levy JC, Turner RC. The relative risks of hyperglycaemia, obesity and dyslipidaemia in the relatives of patients with Type II diabetes mellitus. Diabetologia 1999; 42: Samocha-Bonet D, Campbell LV, Viardot A, et al. A family history of type 2 diabetes increases risk factors associated with overfeeding. Diabetologia 2010; 53: Rice T, Bouchard C, Perusse L, Rao DC. Familial clustering of multiple measures of adiposity and fat distribution in the Quebec Family Study: a trivariate analysis of percent body fat, body mass index, and trunk-to-extremity skin fold ratio. Int J Obes Relat Metab Disord 1995;19: Cooper R, Hyppönen E, Berry D, Power C. Associations between parental and offspring adiposity up to midlife: the contribution of adult lifestyle factors in the 1958 British Birth Cohort Study. Am J Clin Nutr 2010 doi: /ajcn Kvehaugen AS, Andersen LF, Staff AC. Dietary intake and physical activity in women and offspring after pregnancies complicated by preeclampsia or diabetes mellitus. Acta Obstet Gynecol Scand 2010;89: Pirkola J, Pouta A, Bloigu A, et al. Risks of overweight and abdominal obesity at age 16 years associated with prenatal exposures to maternal prepregnancy overweight and gestational diabetes mellitus. Diabetes Care 2010; 33: Steyn NP, Mann J, Bennett PH, et al. Diet, nutrition and the prevention of type 2 diabetes. Public Health Nutr 2004;7: Eriksson JG, Osmond C, Kajantie E, Forsén TJ, Barker DJ. Patterns of growth among children who later develop type 2 diabetes or its risk factors. Diabetologia 2006;49: Whincup PH, Kaye SJ, Owen CG, et al. Birth weight and risk of type 2 diabetes: a systematic review. JAMA 2008;300: Bruce DG, Van Minnen K, Davis WA, et al. Maternal family history of diabetes is associated with a reduced risk of cardiovascular disease in women with type 2 diabetes: the Fremantle Diabetes Study. Diabetes Care 2010; 33:

124 Chapter 5a Commentary: Both multiplicative and additive components may contribute to parental transmission of type 2 diabetes a response to K. Hemminki and X. Li and J. Sundquist and K. Sundquist Ali Abbasi 1,2,3 ; Joline W.J. Beulens 3 ; Harold Snieder 4 ; Stephan J.L. Bakker 2 1 Department of Epidemiology, University Medical Center Groningen, University of Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University Medical Center Groningen, University of Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Centre Utrecht, Utrecht, the Netherlands 4 Unit of Genetic Epidemiology and Bioinformatics, Department of Epidemiology, University Medical Center Groningen, Groningen, the Netherlands J Intern Med. 2011;270:291-2

125 Dear Sir, We thank Hemminki and coworkers for their interest in our recent findings concerning parental transmission of type 2 diabetes 1. Based on their and our data 1,2, they concluded that paternal and maternal history of diabetes simply add rather than act multiplicatively to increase risk for diabetes in offspring. They do so because in the Swedish model a fully multiplicative model was rejected on statistical grounds. We agree that based on these data it is very unlikely that paternal and maternal transmission of diabetes act in an entirely multiplicative manner. However, both in their and our study, the relative risk predicted by an additive model (3.7 for their study and 3.4 for our study) is lower than the actually observed risk (4.3 for their study and 3.9 for our study). Earlier findings were also consistent with these results, with relative risks of 3.6 in Pima Indians and 5.9 Framingham Offspring predicted by an additive model, while actually observed risks were higher, with values of 3.9 and 6.1 respectively 3,4 Thus, data of all 4 studies to date are in the same direction, with a higher actually observed risk in subjects with two affected parents than predicted by an additive model. Therefore, given that a multiplicative model does not fit the data, acceptance of an additive model as sole valid alternative seems an oversimplification, with the truth lying in between both of these extremes. It is important to acknowledge the possibility of non-additive genetic effects (epistasis or gene-gene interactions) of individual loci playing a role in transmission of risk of type 2 diabetes, because this is considered a source of missing heritability in complex diseases, including type 2 diabetes 5. As many low-penetrance loci may contribute to genetic susceptibility for type 2 diabetes 6, such a scenario does not seem unrealistic. References 1. Abbasi A, Corpeleijn E, van der Schouw YT, et al. Maternal and paternal transmission of type 2 diabetes: influence of diet, lifestyle and adiposity. J Intern Med 2011; doi: /j x. 2. Hemminki K, Li X, Sundquist K, Sundquist J. Familial risks for type 2 diabetes in Sweden. Diabetes Care 2010;33: Knowler WC, Pettitt DJ, Savage PJ, Bennett PH. Diabetes incidence in Pima Indians: contributions of obesity and parental diabetes. Am J Epidemiol 1981;113: Meigs JB, Cupples LA, Wilson PW. Parental transmission of type 2 diabetes: the Framingham Offspring Study. Diabetes 2000;49: Manolio TA, Collins FS, Cox NJ, et al. Finding the missing heritability of complex diseases. Nature 2009;461: Salanti G, Southam L, Altshuler D, et al. Underlying genetic models of inheritance in established type 2 diabetes associations. Am J Epidemiol 2009;170:

126 Chapter 6 Parental history of type 2 diabetes and cardiometabolic biomarkers in offspring Ali Abbasi 1,2,3 ; Eva Corpeleijn 1 ; Yvonne T. van der Schouw 3 ; Ronald P. Stolk 1 ; Annemieke M.W. Spijkerman 4 ; Daphne L. van der A 5 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 ; Joline W.J. Beulens 3 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands 4 Center for Prevention and Health Services Research, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands 5 Center for Nutrition and Health, National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands Eur J Clin Invest. 2012;42:974-82

127 Abstract Background Parental history of type 2 diabetes (T2D) is associated with cardiometabolic risk. We aimed to investigate the associations of parental history of T2D with cardiometabolic biomarkers and to subsequently investigate to what extent these putative associations were explained by modifiable factors. Methods Cross-sectionally, we analysed a random sample of 2,001 participants without T2D (20 70 yr) from the European Prospective Investigation into Cancer and Nutrition Netherlands (EPIC-NL). Plasma levels of 12 biomarkers total, HDL and LDL-cholesterol, triglycerides, HbA1c, gamma-glutamyltransferase (GGT), alanine aminotransferase (ALT), asparate aminotransferase (AST), albumin, uric acid, creatinine and high-sensitivity CRP (hs-crp) were assessed according to categories of parental history of T2D. Results In age and sex adjusted analyses, offspring with parental history of T2D had significantly higher ALT (β=0.074, 95% confidence interval [95%CI], 0.023;0.126) and AST levels (β=0.033, 95%CI,0.001;0.066) and a trend towards higher HbA1c (β=0.011, 95%CI,-0.001;0.024) and GGT (β=0.049, 95%CI,-0.015;0.112) levels. Adjustment for diet, smoking, alcohol intake, physical activity and educational level modestly attenuated the magnitude of these associations, but they remained significant for ALT and borderline significant for AST. After further adjustment for adiposity, additional attenuation was observed, but the association remained significant for ALT. Only maternal history of T2D was associated with higher ALT levels. T2D in both parents was associated with increased levels of all liver enzymes, but the association remained significant for GGT after adjustment for adiposity. Overall, the modifiable factors explained 21.2%-45.4% of these associations. The contribution of adiposity was 18.2%-38.9%. Conclusions We conclude that parental history of T2D was associated with higher non-fasting levels of liver enzymes in a general population without T2D. Adiposity substantially contributed to these associations. The contribution of diet and lifestyle factors was modest. 124

128 Introduction Type 2 diabetes (T2D) is associated with a broad range of metabolic components. There is evidence suggesting familial transmission of the metabolic components to offspring 1. Family history of T2D contains both genetic and environmental information 2. We and others recently demonstrated that diet, lifestyle factors and adiposity contribute to T2D risk exerted by parental history of T2D 3, 4. Diet and lifestyle intervention has been shown to modify cardiometabolic risk factors 5. A limited number of studies evaluated the associations of parental history of T2D with levels of some diabetes-related metabolic biomarkers These studies found no associations of parental history of T2D with the inflammatory marker high sensitivity C-reactive protein (hs-crp), glycaemia index HbA1c and blood lipid profile components, including total, and HDL-cholesterol and triglycerides 6, 7, 10. They, however, found a positive association of parental history of T2D with the liver enzymes such as gamma-glutamyltranspeptidase (GGT) in non-diabetic offspring 6. Evidence suggests that elevated liver enzymes as surrogate marker of non-alcoholic fatty liver disease (NAFLD) comprises a new component of the metabolic syndrome 11, 12. NAFLD may be linked to uric acid 13, while uric acid in turn is linked to both the metabolic syndrome and renal dysfunction 14. Of note, the associations of parental history of T2D with uric acid and renal profile components like serum creatinine have not been examined yet. Moreover, it is still unclear whether modifiable factors including diet, lifestyle and adiposity have any effects on the associations of parental history of T2D with aforementioned cardiometabolic biomarkers. We investigated whether parental history of T2D is associated with a broad panel of cardiometabolic biomarkers. We did this in a random sample of adults without T2D from the European Prospective Investigation into Cancer and Nutrition (EPIC) study in the Netherlands (EPIC-NL). We hypothesize that modifiable factors affect these associations; we therefore examined to what extent adiposity, diet and lifestyle factors contribute to the associations of parental history of T2D with levels of biomarkers. Methods Study population and design This study was performed in a random sample (6.5%, n=2,604) of the baseline cohort of the EPIC-NL study in the Netherlands (n=40,011). The EPIC-NL cohort comprises the Monitoring Project on Risk Factors for Chronic Diseases (MORGEN) and Prospect cohorts, set up simultaneously between 1993 and Details of the study design, recruitment and study procedures (including the random sample) were described in detail elsewhere 15. From the random sample, we excluded 43 participants who had T2D (mainly defined by self-report and verified by medical records) and 560 with missing data on 125

129 baseline characteristics, extreme values for energy intake (<450 or>6,000 kcal/day) or unknown parental history of T2D, leaving 2,001 participants for this cross-sectional analysis. The baseline characteristics of excluded participants were similar to those who were included in our analysis (Table S1). We asked participants whether their biological mother and/or father had been diagnosed with T2D. Parental history of T2D was categorized as none, any parent(s) (mother and or father), maternal only, paternal only or both parental. All participants gave written informed consent prior to study inclusion. The EPIC-NL cohort complies with the Declaration of Helsinki and was approved by the relevant local Medical Ethics Committees. Reporting of the study conforms to STROBE along with references to STROBE and the broader EQUATOR guidelines 16. Biomarker measurements We collected non-fasting blood samples were collected from participants at baseline. Blood samples were fractionated into aliquots and stored at -196 C for future use. HbA1c was measured in erythrocytes using an immunoturbidimetric latex test. Biomarkers were assessed in EDTA (in MORGEN cohort) or citrate (in Prospect cohort) plasma. We compared EDTA and citrate measurements and validated these against serum in a sample of 50 participants, observing very good to excellent correlations 15. Albumin and creatinine (Jaffé method) were measured using a colorimetric method. Alanine aminotransferase (ALT), asparate aminotransferase (AST), GGT, total cholesterol, triglycerides and uric acid were measured using enzymatic methods, whereas hs-crp was measured with a turbidimetric method. We measured HDL and LDL-cholesterol using a homogeneous assay with enzymatic endpoint. These assays were all performed on an autoanalyser (LX20, Beckman Coulter, Mijdrecht, the Netherlands). Our technicians were blinded to the participants characteristics. Statistical analyses We used generalized linear models to assess the association between parental history of T2D and cardiometabolic biomarkers in model 1, adjusted for cohort, age and sex. Model 2 was further adjusted for smoking, alcohol use, physical activity level, educational level, total energy intake, and energy-adjusted intakes of fat, protein, carbohydrate, fibre, vitamin C, and vitamin E 3, 15, while model 3 also included body mass index (BMI) and waist circumference. The β regression coefficients (95%CI) for biomarker levels (indicating the increase or decrease in log-transformed level of biomarker) in each category of parental history of T2D were calculated. Inclusion of these factors in the model would be expected to attenuate the β coefficients 3. We calculated the percentage attenuation of the β coefficients for each category of parental history of T2D. Percentage attenuation of β coefficient was calculated as: (β coefficient before addition β coefficient after addition) (β coefficient before addition) 100. To account for the use of non-fasting blood samples, we performed sensitivity analyses after exclusion of those with high blood glucose ( 7.8 mmol/l; 126

130 n=34) and we adjusted the associations for time since last meal or drink (postprandial time). We considered a P value of 0.05 or less from two-sided tests statistically significant. All the statistical analyses were carried out using Statistical Package for Social Sciences version 17 (SPSS Inc, Chicago, Illinois, USA). Results We summarized baseline characteristics of the study population according to parental history of T2D in Table 1. When compared with offspring without parental history of T2D, those who reported any parental history of T2D had higher levels of HbA1c, LDL-cholesterol, ALT and GGT. Moreover, a trend towards higher AST levels in offspring with parental history of T2D than those without parental history of T2D was found. We observed no difference for levels of other biomarkers between each category of parental history of T2D. In our dataset, there were 560 individuals with missing data on baseline characteristics. We presented the comparison between included and excluded participants and found no differences (Table S1). In regression model adjusted for cohort, age and sex, BMI was associated with levels of liver enzymes (β ranged from to 0.316; p<0.001) in both offspring with and without parental history of T2D. The R 2 change after adding BMI to this model showed to what extent BMI could explain the variation of levels of liver enzymes. This contribution of BMI in the variation of AST (6.9% vs. 0.9%) and GGT (9.9% vs. 3.8%) was stronger for those with parental history of T2D. Parental history of diabetes and cardiometabolic biomarkers The association between parental history of T2D and biomarker levels in offspring adjusted for covariates is shown in table 2. In model 1, offspring with any parental history of T2D had significantly higher levels of ALT (β=0.074, 95%CI[0.023;0.126]) and AST (β=0.033, 95%CI[0.001;0.066]). Moreover, levels of HbA1c (β=0.011, 95%CI[ ;0.024]) and GGT (β=0.049, 95%CI[-0.015;0.112]) tended to be higher in offspring with any parental history of T2D. The association between any parental history of T2D and level of these biomarker levels was modestly attenuated in model 2. Adjustment for adiposity (model 3) further attenuated the associations, but they remained significant for ALT (β=0.052, 95% CI [0.002;0.101]) and borderline significant for AST (β= % CI [-0.006;0.059]). The overall attenuations for HbA1c, ALT and AST were 36.4%, 29.7% and 21.2%, respectively. To examine to what extent adiposity itself contribute to the association of parental history of T2D with HbA1c and liver enzymes, we separately added parameters of obesity and obesity (BMI 30 kg/m 2 ) to model 1. Parameters of obesity explained 18.2%, 28.4% and 21.2%, respectively, of the association of parental history of T2D with HbA1c ALT and AST, while presence of obesity explained 9.1%, 10.8% and 9.1%, respectively of these associations. 127

131 Table 1. Baseline characteristics of the random sample according to parental history of type 2 diabetes (n= 2,001) Category of parental history of type of diabetes Variables None (n= 1,611) Any parents (n= 390) Only father (n= 142) Only mother (n= 227) Both parents (n= 21) p value a p value b Age, year 49.5 (12.2) 52.0 (9.3) 52.3 (9.8) 52.2 (8.8) 47.6 (9.4) < Female 1220 (75.7) 300 (76.9) 107 (75.7) 175 (77.1) 18 (85.7) Body mass index, kg/m (4.0) 26.7 (3.9) 26.4 (3.8) 26.6 (3.8) 29.1 (5.0) <0.001 <0.001 Waist circumference, cm 85.3 (11.4) 87.2 (11.1) 86.6 (10.9) 87.2 (10.9) 90.9 (14.0) Alcohol consumption, g/week 11.5 (15.0) 9.7 (12.8) 10.4 (13.6) 9.3 (11.9) 9.3 (1.7) Current smoker 489 (30.4) 107 (27.4) 36 (25.4) 66 (29.1) 5 (23.8) Low educational level c 916 (56.9) 244 (62.6) 77 (54.2) 156 (68.7) 11 (52.4) Physically active d 520 (32.3) 141 (36.2) 46 (32.4) 84 (37.0) 11 (52.4) Total energy intake, kcal/d (599.1) (582.5) (551.6) (601.4) (563.2) Nutrient intake e, g/d Protein Fat Carbohydrates Fiber Vitamin C, mg/d Vitamin E, mg/d 75.8 (10.9) 77.3 (11.2) (31.2) 23.3 (4.6) (46.2) 12.1 (3.2) 77.7 (10.9) 78.4 (10.5) (23.4) 24.0 (4.7) (44.1) 12.3 (3.3) 78.1 (9.7) 77.1 (9.8) (27.9) 24.1 (4.6) (39.4) 11.9 (2.8) 77.0 (11.4) 79.1 (10.7) (28.1) 23.9 (4.8) (46.6) 12.5 (3.6) 82.1 (11.7) 78.5 (12.2) (34.2) 24.6 (3.8) (47.9) 12.0 (3.0) HbA1c, % 5.36 ( ) 5.46 ( ) 5.52 ( ) 5.44 ( ) 5.45 ( ) < Total cholesterol, mmol/l 5.2 ( ) 5.3 ( ) 5.3 ( ) 5.3 ( ) 5.1 ( ) HDL-C, mmol/l 1.21 ( ) 1.21 ( ) 1.22 ( ) 1.22 ( ) 1.16 ( ) LDL-C, mmol/l 3.07 ( ) 3.16 ( ) 3.25 ( ) 3.16 ( ) 3.00 ( ) Triglycerides, mmol/l 1.34 ( ) 1.28 ( ) 1.39 ( ) 1.23 ( ) 1.48 ( ) ALT, IU/l 14.5 ( ) 14.9 ( ) 14.5 ( ) 15.1 ( ) 15.4 ( ) AST, IU/l 19.9 ( ) 20.2 ( ) 20.4 ( ) 20.1 ( ) ( ) GGT, IU/l 20.3 ( ) 21.3 ( ) 20.3 ( ) 21.8 ( ) 32.2 ( ) Albumin, g/l 38.7 ( ) 37.5 ( ) 38.1 ( ) 37.1 ( ) 39.6 ( ) Uric acid, µmol/l ( ) ( ) 252.8( ) ( ) ( ) Creatinine, µmol /l 61.4 ( ) 60.8 ( ) 61.3 ( ) 60.0 ( ) 60.9 ( ) hs-crp, mg/l 1.24 ( ) 1.26 ( ) 1.27 ( ) 1.24 ( ) 1.15 ( )

132 Data were given as mean (SD) or median (IQR) for continuous variables and numbers (percentage) for categorical variables. a For the comparison between participants with any parental history of type 2 diabetes and none using χ2 test (categorical data), and t test or Mann-Whitney U test (continuous data). b For the comparison between participants with paternal, maternal or both history of type 2 diabetes and none (as referent) using χ2 test (categorical data), and ANOVA or Kruskal-Wallis (continuous data). c Low education level was assigned for participants who had primary education up to completing intermediate vocational education. d Physical activity level was defined based on Cambridge Physical Activity Index. e Intakes of nutrients were adjusted for total energy intake and given as g/d unless otherwise indicated. HbA1c denotes glycosylated haemoglobin, HDL-C high-density cholesterol, LDL-C low-density cholesterol, ALT alanine transaminase, AST aspartate transaminase, GGT gamma glutamyl transpeptidase and hs-crp high sensitivity C-reactive protein. Table 2. Association between parental history of type 2 diabetes and offspring biomarker levels (n=2,001) β regression coefficients (95%CI) by category of parental history of type 2 diabetes a Log-transformed biomarker Any parents (n= 390) HbA1c, (%) Model 1 p value Model 2 p value Model 3 p value Total cholesterol, mmol/l Model 1 p value Model 2 p value Model 3 p value HDL-C, mmol/l Model 1 p value Model ( to 0.024) ( to 0.022) ( to 0.020) ( to 0.041) ( to 0.040) ( to 0.039) ( to 0.056) ( to 0.061) Only father (n= 142) ( to 0.035) ( to 0.035) ( to 0.034) ( to 0.053) ( to 0.054) ( to 0.054) ( to 0.083) ( to 0.078) Only mother (n= 227) ( to 0.024) ( to 0.020) ( to 0.019) ( to 0.051) ( to 0.050) ( to 0.048) ( to 0.067) ( to 0.079) Both parents (n= 21) ( to 0.067) ( to 0.068) ( to 0.057) ( to 0.090) ( to 0.088) ( to 0.080) ( to 0.089) ( to 0.070) 129

133 p value Model 3 p value LDL-C, mmol/l Model 1 p value Model 2 p value Model 3 p value Triglycerides, mmol/l Model 1 p value Model 2 p value Model 3 p value ALT, IU/l Model 1 p value Model 2 p value Model 3 p value AST, IU/l Model 1 p value Model 2 p value Model 3 p value GGT, IU/l Model 1 p value ( to 0.062) ( to 0.075) ( to 0.072) ( to 0.067) ( to 0.075) ( to 0.074) ( to 0.049) (0.023 to 0.126) (0.017 to 0.120) (0.002 to 0.101) (0.001 to 0.066) ( to 0.065) ( to 0.059) ( to 0.112) ( to 0.081) ( to 0.101) ( to 0.100) ( to 0.097) ( to 0.147) ( to 0.160) ( to 0.146) ( to 0.110) ( to 0.106) ( to 0.094) ( to 0.080) ( to 0.077) ( to 0.075) ( to 0.071) ( to 0.080) ( to 0.089) ( to 0.084) ( to 0.079) ( to 0.076) ( to 0.064) ( to 0.038) (0.020 to 0.149) (0.016 to 0.145) (0.003 to 0.127) ( to 0.068) ( to 0.069) ( to 0.063) ( to 0.140) ( to 0.128) ( to 0.150) ( to 0.144) ( to 0.117) ( to 0.263) ( to 0.289) ( to 0.159) (0.062 to 0.462) (0.032 to 0.431) ( to 0.335) (0.000 to 0.257) ( to 0.240) ( to 0.211) (0.178 to 0.668)

134 Model 2 p value Model 3 p value Albumin, g/l Model 1 p value Model 2 p value Model 3 p value Uric acid, µmol/l Model 1 p value Model 2 p value Model 3 p value Creatinine, mg/l Model 1 p value Model 2 p value Model 3 p value hs-crp, mg/l Model 1 p value Model 2 p value Model 3 p value ( to 0.121) ( to 0.100) ( to 0.016) ( to 0.016) ( to 0.016) ( to 0.036) ( to 0.037) ( to 0.023) ( to 0.012) ( to 0.015) ( to 0.013) ( to 0.182) ( to 0.188) ( to 0.130) ( to 0.082) ( to 0.068) (0.002 to 0.032) (0.000 to 0.030) (0.001 to 0.031) ( to 0.044) ( to 0.047) ( to 0.036) ( to 0.033) ( to 0.033) ( to 0.037) ( to 0.229) ( to 0.264) ( to 0.214) ( to 0.148) ( to 0.129) ( to 0.011) ( to 0.011) ( to 0.012) ( to 0.044) ( to 0.044) (0.040 to 0.032) ( to 0.012) (0.048 to 0.016) (0.049 to 0.015) ( to 0.244) ( to 0.231) ( to 0.181) (0.191 to 0.675) < (0.095 to 0.566) ( to 0.059) ( to 0.052) ( to 0.055) ( to 0.177) ( to 0.169) ( to 0.096) ( to 0.112) ( to 0.113) ( to 0.105) ( to 0.647) ( to 0.696) ( to 0.386)

135 HbA1c denotes glycosylated haemoglobin, HDL-C high-density cholesterol, LDL-C low-density cholesterol, ALT Alanine transaminase, AST Aspartate transaminase, GGT Gamma glutamyl transpeptidase and hs-crp high sensitivity C-reactive protein. a Model 1 was adjusted for cohort, age and sex. Model 2 was adjusted for model 1 plus smoking, alcohol use, physical activity level, educational level, total energy intake, and energy-adjusted dietary factors, including the amount of intake of fat, protein, carbohydrate, fiber, vitamin C, and vitamin E. Model 3 was adjusted for model 2 plus body mass index, waist circumference. Significant P values were indicated in bold type. 132

136 Maternal versus paternal history of diabetes In subsequent analyses, we examined whether biomarkers levels differed by maternal or paternal history of T2D. In model 1, offspring with only maternal history of T2D had higher ALT levels (β=0.084, 95%CI[0.020;0.149]). Offspring with only paternal history of T2D had higher albumin levels (β=0.017, 95%CI[0.002;0.032]). Those with both maternal and paternal history of T2D had higher GGT levels (β=0.423, 95%CI [0.178;0.668]), higher ALT levels (β=0.262, 95%CI[0.062;0.462]) and higher AST levels (β=0.129, 95%CI[(0.000; 0.257]). In model 3, statistically significant higher levels of GGT (β=0.331, 95%CI[0.095;0.566]) were found for offspring with both diabetic parents, and higher levels of ALT (β=0.065, 95%CI[0.003;0.127]) for offspring with only maternal history of T2D (Table 2). The overall attenuation of the association of maternal history of T2D with ALT was 22.6%. The overall attenuations of the association of both parental history of T2D with AST, ALT and GGT were 34.8%, 45.4% and 21.7%, respectively. To examine to what extent adiposity itself contribute to the associations of maternal and both parental history of T2D with most-related biomarkers, we separately added parameters of adiposity and obesity to model 1 in each category. Parameters of obesity explained 23.8% of the association of maternal history of T2D with ALT, but presence of obesity explained 4.8% of this association. Parameters of obesity explained 38.9% and 26.7%, respectively, of the association of both parental history of T2D with ALT and GGT. Obesity explained 26.3% and 18.4%, respectively, of these associations. In sensitivity analyses, we excluded those with high blood glucose ( 7.8 mmol/l; n=34) and further adjusted for time since last meal or drink (postprandial time) and observed similar results (data not shown). Discussion In this cross-sectional analysis, we found that any parental history of T2D was associated with higher levels of liver enzymes (ALT and AST) in adults without T2D. This was particularly true for maternal history of T2D. Offspring with both diabetic parents had higher GGT levels as well, whereas those who reported only paternal diabetes had higher albumin levels. The associations between parental history of diabetes and liver enzymes were partly explained by diet and lifestyle factors (smoking status, alcohol consumption, physical activity and educational level). The parameters of adiposity contributed substantially to these associations. Given the extensive information about modifiable factors of the participants, we were able to show the contribution of these factors in the associations between parental history of diabetes T2D and a panel of biomarkers in adults without T2D. Nevertheless, our study has some limitations. The study is a cross-sectional investigation and causal relationships cannot be inferred. Another limitation is that parental history of T2D was obtained by self-report. Furthermore, we excluded the individuals with missing data or unknown parental history of T2D. However, most of 133

137 the baseline characteristics of excluded participants were similar to those who were included in our analysis. Therefore, it is unlikely that this may have led to selection bias in some categories of parental history of T2D. We used data on non-fasting blood samples as fasting samples were not available, but, further adjustment for postprandial time did not change the results. Although non-fasting conditions could be argued as a severe limitation of our study, there is evidence suggesting that overnight fasting status might have minimal effects on concentrations of most biomarkers which we studied 17, 18, also in our cohort 19. Of note, data on non-fasting lipids such as triglycerides and inflammation-related biomarkers such as CRP and adiponectin have been shown to improve the risk prediction of T2D and cardiovascular events independent of traditional cardiometabolic risk profile Finally, we had no data on more specific factors to T2D, such as HOMA/insulin and adiponectin to examine their associations with parental history of T2D, but the metabolic syndrome was not a primary focus of this study. Among different cardiometabolic biomarkers, only liver enzymes were associated with parental history of T2D. Our findings are in line with the limited prior information suggesting that female workers with family history of T2D had elevated levels of ALT, AST and GGT 6. However, this study also observed an association with triglycerides, which we could not confirm 6. Similar to our results, the recent studies also showed no association of family history of T2D with lipid profile components and proinflammatory markers in adults without T2D 7, 8, 10. Our results were similar by excluding those with high non-fasting glucose concentrations and were therefore more likely to be a direct consequence of parental history of T2D rather than secondary to pre-diabetes state. Adjusting for diet and lifestyle attenuated the association between parental history of T2D and liver function profile. This support the suggestion of lifestyle modification has positive effect on cardiometabolic risk 5. Further adjustment for adiposity led to a more attenuation in the associations. In fact, this indicated that adiposity contributed to parental transmission of liver function profile independent from diet and lifestyle factors. Furthermore, increased adiposity, in itself, could substantially explain the associations of parental history of T2D with HbA1c and liver enzymes. These findings are generally in line with those of other epidemiologic studies showing that adiposity play an important role in transmission of cardiometablic risk 10, 21. Mild elevated levels of liver enzymes may have occurred in early or subclinical metabolic abnormality underlying NAFLD. Of note, the lack of association of parental history of T2D with markers of glycaemia, lipids and low grade inflammation may show that these subclinical changes in liver function are important for further progression to pre-diabetes state and T2D 22, 23. This extends the available information for a possible link between NAFLD, adiposity and T2D. Of note, these traits partly share common genetic and environmental components 24. Further prospective investigations are warranted showing whether the link between liver function and adiposity contributes to transmission of risk of T2D and its complications to the next generation

138 In our study, there was a stronger association of maternal history of T2D with liver enzymes. This difference might be partly explained by genetic and environmental conditions such as X-linked traits, gene imprinting intra-uterine programming and a more prominent maternal role in raising children. A recent study among Japanese men showed that maternal but not paternal adiposity was associated with a higher level of ALT 25. Similar to our data, another study on offspring of diabetic parents has shown that defects in insulin sensitivity and β-cell glucose sensitivity were accentuated along the maternal line of inheritance 10. The specific relation of maternal history of T2D and liver enzymes observed in our study could also be explained by the fact that the mother might have a more prominent role in raising her children from early life, i.e. during pregnancy, to later life. For example, it has been shown that maternal pre-pregnancy overweight increased risk of offspring overweight and abdominal obesity 26. Adiposity in turn associates with elevated liver enzymes and NAFLD In conclusion, we found that parental history of T2D was associated with higher non-fasting levels of ALT, AST and GGT in a general population without T2D. Adiposity, substantially, contributed to the associations between parental history of T2D and liver function profile. The contribution of diet and lifestyle factors was modest. 135

139 Acknowledgments This work was supported by the Netherlands Heart Foundation, the Dutch Diabetes Research Foundation and the Dutch Kidney Foundation. This research was performed within the framework of the Center for Translational Molecular Medicine (CTMM; project PREDICCt (grant 01C ). The EPIC-NL study was funded by Europe against Cancer Programme of the European Commission (SANCO), the Dutch Ministry of Health, the Dutch Cancer Society, the Netherlands Organisation for Health Research and Development (ZonMW), and World Cancer Research Fund (WCRF). We thank Statistics Netherlands and the PHARMO Institute for follow-up data on cancer, cardiovascular disease and vital status. None of the study sponsors had a role in the study design, data collection, analysis or interpretation, report writing, or the decision to submit the report for publication. No potential conflicts of interest relevant to this article were reported. 136

140 References 1. Liese AD, Mayer-Davis EJ, Tyroler HA, et al. Familial components of the multiple metabolic syndrome: the ARIC study. Diabetologia 1997;40(8): Meigs JB, Shrader P, Sullivan LM, et al. Genotype score in addition to common risk factors for prediction of type 2 diabetes. N Engl J Med 2008;359(21): Abbasi A, Corpeleijn E, van der Schouw YT, et al. Maternal and paternal transmission of type 2 diabetes: influence of diet, lifestyle and adiposity. J Intern Med 2011;270(4): van 't Riet E, Dekker JM, Sun Q, Nijpels G, Hu FB, van Dam RM. Role of adiposity and lifestyle in the relationship between family history of diabetes and 20-year incidence of type 2 diabetes in U.S. women. Diabetes Care 2010;33(4): Goodpaster BH, Delany JP, Otto AD, et al. Effects of diet and physical activity interventions on weight loss and cardiometabolic risk factors in severely obese adults: a randomized trial. JAMA 2010;304(16): Inoue K, Matsumoto M, Miyoshi Y, Kobayashi Y. Elevated liver enzymes in women with a family history of diabetes. Diabetes Res Clin Pract 2008;79(3):e Goldfine AB, Beckman JA, Betensky RA, et al. Family history of diabetes is a major determinant of endothelial function. J Am Coll Cardiol 2006;47(12): Ford ES, Giles WH, Mokdad AH. Family history of diabetes or cardiovascular disease and C- reactive protein concentration: findings from the National Health and Nutrition Examination Survey, Am J Prev Med 2005;29(5 Suppl 1): Anjana RM, Lakshminarayanan S, Deepa M, Farooq S, Pradeepa R, Mohan V. Parental history of type 2 diabetes mellitus, metabolic syndrome, and cardiometabolic risk factors in Asian Indian adolescents. Metabolism 2009;58(3): Natali A, Muscelli E, Mari A, et al. Insulin sensitivity and beta-cell function in the offspring of type 2 diabetic patients: impact of line of inheritance. J Clin Endocrinol Metab 2010;95(10): Haffner SM. Relationship of metabolic risk factors and development of cardiovascular disease and diabetes. Obesity (Silver Spring) 2006;14 Suppl 3:121S-7S. 12. Musso G, Gambino R, Bo S, et al. Should nonalcoholic fatty liver disease be included in the definition of metabolic syndrome? A cross-sectional comparison with Adult Treatment Panel III criteria in nonobese nondiabetic subjects. Diabetes Care 2008;31(3): Samuel VT. Fructose induced lipogenesis: from sugar to fat to insulin resistance. Trends Endocrinol Metab 2011;22(2): Cirillo P, Sato W, Reungjui S, et al. Uric acid, the metabolic syndrome, and renal disease. J Am Soc Nephrol 2006;17(12 Suppl 3):S Beulens JW, Monninkhof EM, Verschuren WM, et al. Cohort profile: the EPIC-NL study. Int J Epidemiol 2010;39(5): Simera I, Moher D, Hoey J, Schulz KF, Altman DG. A catalogue of reporting guidelines for health research. Eur J Clin Invest 2010;40(1): Langsted A, Freiberg JJ, Nordestgaard BG. Fasting and nonfasting lipid levels: influence of normal food intake on lipids, lipoproteins, apolipoproteins, and cardiovascular risk prediction. Circulation 2008;118(20): Herder C, Baumert J, Zierer A, et al. Immunological and cardiometabolic risk factors in the prediction of type 2 diabetes and coronary events: MONICA/KORA Augsburg case-cohort study. PLoS One 2011;6(6):e van Dieren S, Nothlings U, van der Schouw YT, et al. Non-fasting lipids and risk of cardiovascular disease in patients with diabetes mellitus. Diabetologia 2011;54(1): Bansal S, Buring JE, Rifai N, Mora S, Sacks FM, Ridker PM. Fasting compared with nonfasting triglycerides and risk of cardiovascular events in women. JAMA 2007;298(3):

141 21. Srinivasan SR, Frontini MG, Berenson GS. Longitudinal changes in risk variables of insulin resistance syndrome from childhood to young adulthood in offspring of parents with type 2 diabetes: the Bogalusa Heart Study. Metabolism 2003;52(4):443-50; discussion Fraser A, Harris R, Sattar N, Ebrahim S, Davey Smith G, Lawlor DA. Alanine aminotransferase, gamma-glutamyltransferase, and incident diabetes: the British Women's Heart and Health Study and meta-analysis. Diabetes Care 2009;32(4): Forbes S, Taylor-Robinson SD, Patel N, Allan P, Walker BR, Johnston DG. Increased prevalence of non-alcoholic fatty liver disease in European women with a history of gestational diabetes. Diabetologia 2011;54(3): Goessling W, Massaro JM, Vasan RS, D'Agostino RB, Sr., Ellison RC, Fox CS. Aminotransferase levels and 20-year risk of metabolic syndrome, diabetes, and cardiovascular disease. Gastroenterology 2008;135(6): , 44 e Kazumi T, Kawaguchi A, Yoshino G. Associations of middle-aged mother's but not father's body mass index with 18-year-old son's waist circumferences, birth weight, and serum hepatic enzyme levels. Metabolism 2005;54(4): Pirkola J, Pouta A, Bloigu A, et al. Risks of overweight and abdominal obesity at age 16 years associated with prenatal exposures to maternal prepregnancy overweight and gestational diabetes mellitus. Diabetes Care 2010;33(5): Zelber-Sagi S, Nitzan-Kaluski D, Halpern Z, Oren R. Prevalence of primary non-alcoholic fatty liver disease in a population-based study and its association with biochemical and anthropometric measures. Liver Int 2006;26(7): Nakao K, Nakata K, Ohtsubo N, et al. Association between nonalcoholic fatty liver, markers of obesity, and serum leptin level in young adults. Am J Gastroenterol 2002;97(7): Church TS, Kuk JL, Ross R, Priest EL, Biltoft E, Blair SN. Association of cardiorespiratory fitness, body mass index, and waist circumference to nonalcoholic fatty liver disease. Gastroenterology 2006;130(7):

142 Appendix Table S1. Baseline characteristics of the random sample according to missing data Variables Total (n= 2,604) a With missing data (n= 603) b With complete data (n= 2,001) Age, year 49.4 (11.8) 47.5 (12.1) 50.0 (11.7) Female 1945 (74.7) 425 (70.5) 1520 (76.0) Body mass index, kg/m (4.1) 25.4 (4.5) 25.9 (4.0) Waist circumference, cm 85.5 (11.7) 84.9 (12.7) 85.7 (11.7) Alcohol consumption, g/week 11.2 (15.8) 11.2 (19.3) 11.2 (14.6) Current smoker 793 (30.5) 197 (32.9) 596 (29.8) Low educational level c 1491 (57.8) 331 (57.1) 1160 (58.0) Physically active d 846 (32.5) 185 (30.7) 661 (33.0) Maternal history of diabetes 317 (12.7) 69 (13.7) 248 (12.4) Paternal history of diabetes 207 (8.4) 44 (9.3) 163 (8.1) Total energy intake, kcal/d (613.2) (666.3) (596.3) Nutrient intake e, g/d Protein Fat Carbohydrates Fiber Vitamin C, mg/d Vitamin E, mg/d HbA1c, % mmol/mol 76.0 (11.1) 77.5 (11.3) (31.3) 23.3 (4.8) (46.7) 12.2 (3.3) 75.5 (11.5) 77.4 (11.9) (33.1) 23.3 (5.3) (49.4) 12.3 (3.5) 76.2 (10.9) 77.5 (11.1) (30.7) 23.5 (4.7) (45.8) 12.1 (3.2) 5.38 ( ) 35 (31-39) 5.38 ( ) 35 (31-40) 5.38 ( ) 35 (32-39) Total cholesterol, mmol/l 5.2 ( ) 5.1 ( ) 5.2 ( ) LDL-C, mmol/l 3.06 ( ) 2.89 ( ) 3.08 ( ) Triglycerides, mmol/l 1.31 ( ) 1.25 ( ) 1.33 ( ) ALT, IU/l 14.6 ( ) 14.3 ( ) 14.6 ( ) AST, IU/l 20.0 ( ) 19.9 ( ) 20.0 ( ) GGT, IU/l 20.4 ( ) 19.7 ( ) 20.5 ( ) Albumin, g/l 38.9 ( ) 40.4 ( ) 38.6 ( ) Uric acid, µmol/l ( ) ( ) ( ) Creatinine, µmol /l 61.4 ( ) 61.8 ( ) 61.3 ( ) hs-crp, mg/l 1.29 ( ) 1.72 ( ) 1.24 ( ) Data were given as mean (SD) or median (IQR) for continuous variables and numbers (percentage) for categorical variables. a Data were available for 2269 to 2604 individuals. b Data were available for 268 to 603 individuals. 139

143 140

144 Chapter 7 Plasma procalcitonin is associated with obesity, insulin resistance and the metabolic syndrome Ali Abbasi 1,2 ; Eva Corpeleijn 1 ; Douwe Postmus 1 ; Ron T. Gansevoort 2 ; Paul E. de Jong 2 ; Rijk O.B. Gans 2 ; Joachim Struck 3 ; Hans Hillege 1 ; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Department of Research, BRAHMS AG (Part of Thermo Fisher Scientific), Hennigsdorf Germany J Clin Endocrinol Metab. 2010;95:E26-E31

145 Abstract Background Procalcitonin, a well-known biomarker of sepsis and bacterial infections, is produced by adipose tissue and has potential as a marker for chronic low-grade inflammation. The objective of this study was to investigate whether plasma procalcitonin levels in the normal range are associated with obesity, insulin resistance and metabolic syndrome (MS) in the general population. Methods Plasma procalcitonin ( ng/ml) was measured in 3197 men and 3638 women (aged 28 to 75 years) of the Prevention of Renal and Vascular End-stage Disease (PREVEND) study using an ultra-sensitive immunoluminometric assay. MS was defined according to Adult Treatment Panel III criteria. Results Median (interquartile range) plasma procalcitonin was ( ) ng/ml in men and ( ) ng/ml in women (P <0.001). Plasma procalcitonin was positively associated with BMI and waist circumference. In both sexes, cross-sectional associations of plasma procalcitonin with insulin resistance and components of the MS remained significant after adjustment for age, BMI, waist circumference and other covariates. The age-adjusted odds ratio (OR) for MS was 3.2 (95% confidence interval [CI]= ) in men and 4.1 (95%CI, ) in women, when comparing the highest to the lowest quartile of plasma procalcitonin. The multivariate-adjusted OR for MS was 1.9 (95%CI= ) in men and 2.3 (95%CI= ) in women. The multivariate-adjusted OR for insulin resistance was 3.3 (95%CI= ) in men and 2.5 (95%CI= ) in women. Conclusions Elevated plasma procalcitonin levels in the normal range are associated with measures of obesity, insulin resistance and metabolic syndrome in the general population. 142

146 Introduction There are strong links between obesity, insulin resistance and components of the metabolic syndrome. Chronic low grade inflammation has been implicated in the pathophysiology of these three intertwined entities 1, 2. Procalcitonin, a 116-aminoacid polypeptide, is the precursor of calcitonin hormone produced by neuroendocrine C-cells of the thyroid and K-cells of the lung, encoded from the calcitonin I (CALC I) gene on chromosome Procalcitonin is best known as a biomarker of infection and severe systemic inflammation 6, 7. Recent studies show that adipose tissue is capable of expressing and secreting procalcitonin This makes procalcitonin a potential biomarker for obesity-related low gradeinflammation. There are no data addressing the significance of variation in plasma procalcitonin levels in the general population. So far, procalcitonin level in the normal population has been studied only in a small sample, and only an association of procalcitonin with sex was acknowledged 11. We hypothesize that plasma procalcitonin may be associated with measures of obesity, insulin resistance and metabolic risk factors. Methods This cross-sectional analysis was conducted on the participants from the Prevention of Renal and Vascular Endstage Disease (PREVEND) study in the general population (age ranged between 28 and 75 years) of the city of Groningen, the Netherlands. Details of the study design, recruitment, and procedures have been published elsewhere 12. Plasma procalcitonin was measured in 7,690 participants from the samples of the baseline screening. At first, we excluded 25 participants with procalcitonin level > 0.1 ng/ml. Further exclusion was for 385 individuals who had no documented fasting blood samples or missing data for other variables, leaving 3,137 men and 3,638 women (total, n=6,835) for the present analysis. The PREVEND study was approved by the local medical ethics committee, University Medical Center Groningen, and conformed to the principles outlined in the Declaration of Helsinki. All participants gave written informed consent. Blood pressure was measured in supine position with an automatic device (Dinamap XL Model 9300, Johnson-Johnson Medical, Tampa, FL). Smoking and alcohol use were based on self-reports. Metabolic syndrome was defined according to the National Cholesterol Education Program s Adult Treatment Panel III report (ATP III) criteria 13, as participants having at least 3 of the following: 1) Waist circumference > 35 inches (> 88 cm) in women or > 40 inches (> 102 cm) in men, 2) blood pressure 130/ 85 mmhg or treatment for hypertension, 3) fasting triglycerides 150 mg/dl ( 1.7 mmol/l), 4) HDL cholesterol 40mg/dL ( 1.0 mmol/l) in men or 50 mg/dl 143

147 ( 1.3 mmol/l) in women, and 5) fasting blood glucose 110 mg/dl ( 6.1 mmol/l) or treatment for type 2 diabetes. Insulin resistance was assessed based on the homeostasis model assessment for insulin resisatnce (HOMA-IR) that is calculated using the following formula: [glucose (mmol/l) insulin (mu/ml)]/22.5] 14. We defined insulin resistance as a HOMA-IR score in upper sex-specific quartiles. In baseline samples, serum and urinary creatinine, total cholesterol, and plasma glucose were measured by dry chemistry (Eastman Kodak, Rochester, New York). High density lipoprotein (HDL) cholesterol was measured with a homogeneous method (direct HDL, Aeroset TM System, Abbott Laboratories, Abbott Park, Illinois). Triglycerides were measured enzymatically. High-sensitivity C- reactive protein (hs-crp) was determined by nephelometry (BN II, Dade Behring, Marburg, Germany). Urinary albumin excretion (UAE) was measured as the mean of two 24-hour urine collections by nephelometry with a threshold of 2.3 mg/l (Dade Behring Diagnostic, Marburg, Germany). Insulin was measured with an AxSym auto-analyzer (Abbott Diagnostics, Amstelveen, The Netherlands). Procalcitonin was measured by a novel commercially available immuno-luminometric assays (B.R.A.H.M.S PCT sensitive LIA, Hennigsdorf, Germany). Assays were performed in EDTA-plasma aliquots that had been stored frozen at -80 C, without prior thawing and re-freezing. The intraassay CV at 0.1 ng/ml was 6% and at 0.03 ng/ml it was 8%. The Functional assay sensitivity, defined as the lowest concentration to be determined with an interassay CV of 20% was ng/ml. The lowest detection limit was ng/ml. The assay technique has been described previously 11. All technicians were blinded to the participants characteristics. Continuous variables were compared by using one-way ANOVA or a Kruskal- Wallis test and a χ 2 test was used for the categorical variables to test for differences across quartiles of procalcitonin. We evaluated the association of log2 procalcitonin level with the components (continuous) of the metabolic syndrome using univariate and multivariate-adjusted linear regression models in sex-stratified analyses. Regression coefficients with 95%confidence intervals (CIs) were determined. We performed univariate and multivariate-adjusted logistic regression models to test the association between plasma procalcitonin level and presence of the metabolic syndrome and insulin resistance. The models were adjusted for age, measures of obesity, hs-crp, tobacco smoking, alcohol use, history of cardiovascular disease and hormone replacement therapy (for women). A P value of 0.05 or less from two-sided tests was considered statistically significant. The statistical analyses were performed using SPSS 16.0 statistical software (SPSS Inc., Chicago, IL). Results Of 3,197 men and 3,638 women, 631 (19.7%) and 616 (16.9%) had metabolic syndrome, respectively. Median (IQR) procalcitonin levels were ( ) ng/ml in men and ( ) ng/ml in women (P <0.001). Anthropometric and clinical characteristics of the study population are summarized in Table 1 for men 144

148 and women separately. Participants with high procalcitonin levels were older, more obese and more likely to fulfill criteria for the metabolic syndrome. They also had lower insulin sensitivity, lower creatinine clearance, higher hs-crp and higher UAE. Men with high procalcitonin were more likely to be smoker or ex-smoker, whereas women with high procalcitonin were less likely to use alcohol. In men, across quartiles of body mass index (BMI) median (IQR) procalcitonin levels gradually increased from ( ) ng/ml in the first to ( ) ng/ml in the fourth quartile (P<0.001) (Figure S1). In women, this was from ( ) ng/ml in the first to ( ) ng/ml in the fourth quartile (P<0.001). In men, across quartiles of waist circumference, median (IQR) procalcitonin increased from ( ) ng/ml in the first to ( ) ng/ml in the fourth quartile (P<0.001). In women, this was from ( ) ng/ml in the first to ( ) ng/ml in the fourth quartile (P<0.001). Associations of components of the metabolic syndrome waist circumference, systolic blood pressure, diastolic blood pressure, triglycerides, HDL cholesterol, glucose and insulin resistance fasting insulin, HOMA-IR with procalcitonin were independent of age in linear regression analyses. Further adjustment for BMI attenuated these associations, but they remained statistically significant. Subsequent further adjustments for hs-crp, smoking status, alcohol intake, history of cardiovascular disease and hormone replacement therapy (for women) did not materially change these associations except for blood pressure, which lost significance in women (Table S1). Logistic regression analyses (Table 2) show that risk for the metabolic syndrome and insulin resistance increased across procalcitonin quartiles. In multivariate-adjusted models, the odds ratio s (ORs) for metabolic syndrome and insulin resistance in the highest quartile compared with the lowest were 1.9 (95% CI, ) and 3.3 (95% CI, ) in men, and 2.3 (95% CI, ) and 2.5 (95%CI, ) in women, respectively. Discussion To the best of our knowledge, this study explored for the first time the association of plasma procalcitonin with measures of obesity and metabolic and cardiovascular risk factors in a large sample of the general population. An important finding is that variation in plasma procalcitonin within the normal range is associated with insulin resistance and the metabolic syndrome in apparently healthy men and women. The association of plasma procalcitonin with insulin resistance and metabolic syndrome was independent of age, measures of obesity, hs-crp, history of cardiovascular disease and health behaviors. The current results are in line with experimental and observational data that suggest that plasma procalcitonin can be an inflammatory biomarker even in the absence of signs of systemic infection or sepsis 8-10, 15, 16. Human adipose tissue depots have been identified as major non-neuroendocrine calcitonin mrna expression sites 8, 9, and in vitro secretion of procalcitonin by adipocytes was stimulated by activated 145

149 Table 1. Anthropometric and clinical characteristics of participants according to sex-specific quartiles of plasma procalcitonin (n=6,835) Procalcitonin Quartiles, ng/ml P value* ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) M F Characteristic No. of participants M F Age, yr M 46.0± ± ± ±12.9 <0.001 F 42.5± ± ± ±12.3 <0.001 History of Cardiovascular disease, no. (%) M 60 (6.8) 58 (7.2) 75 (9.8) 81 (11.0) F 16 (1.5) 20 (2.2) 35 (4.0) 38 (4.6) <0.001 BMI, kg/m 2 M 25.0± ± ± ±3.8 <0.001 F 24.4± ± ± ±5.2 <0.001 < (26.9) 374 (50.9) 163 (22.2) 247 (32.1) 390 (50.7) 132 (17.2) 314 (39.0) 402 (49.9) 90 (11.2) M 465 (52.4) 365 (41.1) 57 (6.4) Normal weight, no. (%) Overweight, no. (%) < (31.5) 319 (38.9) 243 (29.6) 391 (45.0) 316 (36.4) 161 (18.5) 498 (55.0) 294 (32.5) 114 (12.6) F 668 (64.0) 293 (28.1) 82 (7.9) Obese, no. (%) Normal weight, no. (%) Overweight, no. (%) Obese, no. (%) 146 Waist circumference, cm M 89.6± ± ± ±11.5 <0.001 F 80.0± ± ± ±12.3 <0.001 Systolic blood pressure, mmhg M 124.9± ± ± ±18.6 <0.001 F 113.6± ± ± ±21.4 <0.001 Diastolic blood pressure, mmhg M 72.3± ± ± ±9.9 <0.001 F 66.7± ± ± ±8.5 <0.001 Total Cholestrol, mmol/l M 5.4± ± ± ±1.1 <0.001 F 5.3± ± ± ±1.2 <0.001 HDL cholesterol, mmol/l M 1.2± ± ± ±0.3 <0.001 F 1.6± ± ± ±0.4 <0.001 Triglyceride, mmol/l M 1.3± ± ± ±1.5 <0.001 F 1.0± ± ± ±1.0 <0.001 Glucose, mmol/l M 4.8± ± ± ±1.6 <0.001 F 4.5± ± ± ±1.6 <0.001 Insulin, mu/l M 6.8 ( ) 7.9 ( ) 9.2 ( ) 10.5 ( ) <0.001

150 F 6.5 ( ) 7.1 ( ) 7.8 ( ) 9.9 ( ) <0.001 HOMA-IR M 1.4 ( ) 1.7 ( ) 2.0 ( ) 2.4 ( ) <0.001 F 1.3 ( ) 1.4 ( ) 1.6 ( ) 2.2 ( ) <0.001 M < (22.7) 348 (47.3) 220 (29.9) 175 (22.8) 310 (40.3) 284 (36.9) 216 (26.8) 301 (37.3) 289 (35.9) 266 (30.0) 318 (35.9) 303 (34.2) Tobacco smoking, no. (%) Never Quitted Current smoker F (34.0) 271 (33.0) 271 (33.0) 313 (36.1) 285 (32.8) 270 (31.1) 294 (32.5) 285 (31.5) 327 (36.1) 356 (34.1) 344 (33.0) 343 (32.9) Never Quitted Current smoker M (18.5) 97 (13.2) 253 (34.4) 181 (24.6) 68 (9.3) 143 (18.6) 88 (11.4) 263 (34.2) 211 (27.4) 64 (8.3) 143 (17.7) 94 (11.7) 307 (38.1) 201 (24.9) 61 (7.6) 125 (14.1) 90 (10.1) 370 (41.7) 223 (25.1) 79 (8.9) Alcohol use, no. (%) Never 1 to 4 drinks per month 2 to 7 drinks per week 1 to 3 drinks per day 4 drinks per day 147 F < (38.1) 167 (20.3) 216 (26.3) 105 (12.8) 20 (2.4) 308 (35.5) 155 (17.9) 245 (28.2) 147 (16.9) 13 (1.5) 265 (29.2) 183 (20.2) 305 (33.7) 130 (14.3) 23 (2.5) 280 (26.8) 200 (19.2) 375 (36.0) 173 (16.6) 15 (1.4) Never 1 to 4 drinks per month 2 to 7 drinks per week 1 to 3 drinks per day 4 drinks per day Creatinine clearance, ml/min M 117.4± ± ± ±27.4 <0.001 F 98.1± ± ± ±24.5 <0.001 Urine albumin excretion, mg/24h M 9.1 ( ) 9.9 ( ) 10.7 ( ) 13.0 ( ) <0.001 F 7.8 ( ) 8.1 ( ) 8.5 ( ) 10.0 ( ) <0.001 hs-crp, mg/l (n=6617) M 0.7 ( ) 1.0 ( ) 1.3 ( ) 2.0 ( ) <0.001 F 0.9 ( ) 1.1 ( ) 1.4 ( ) 2.3 ( ) <0.001 Abbreviations: M, male; F, female; BMI, body mass index; HDL, high-density lipoprotein; hs-crp, high sensitivity C-reactive protein; HOMA-IR, homeostasis model assessment for insulin resistance. *P values are based on χ 2 test for categorical data, spearman rank correlation for ordinal data and ANOVA or Kruskal-Wallis for continuous data; depending on the normality of the data, which were presented by mean±s.d. or median (interquartile range).

151 Table 2. Odds ratios for metabolic syndrome and insulin resistance according to sex-specific quartiles of plasma procalcitonin Procalcitonin Quartiles, ng/ml P value for trend Men No. of person Metabolic syndrome, No. of cases Unadjusted OR (95% CI) Age-adjusted OR (95% CI) Multivariate OR (95% CI) ( ) 1.6 ( ) 1.2 ( ) 2.4 ( ) 2.1 ( ) 1.3 ( ) 3.9 ( ) 3.2 ( ) 1.9 ( ) Insulin resistance, * No of cases Unadjusted OR (95% CI) Age-adjusted OR (95% CI) Multivariate OR (95% CI) ( ) 1.9 ( ) 1.6 ( ) 3.1 ( ) 2.8 ( ) 2.1 ( ) 5.3 ( ) 4.7 ( ) 3.3 ( ) Women No. of person Metabolic syndrome, No. of cases Unadjusted OR (95% CI) Age-adjusted OR (95% CI) ( ) 1.5 ( ) 1.3 ( ) 3.3 ( ) 2.2 ( ) 1.3 ( ) 8.0 ( ) 4.1 ( ) 2.3 ( ) Multivariate OR (95% CI) Insulin resistance, * No of cases Unadjusted OR (95% CI) Age-adjusted OR (95% CI) Multivariate OR (95% CI) ( ) 1.4 ( ) 1.2 ( ) 2.7 ( ) 2.1 ( ) 1.4 ( ) 5.6 ( ) 3.9 ( ) 2.5 ( ) Abbreviation: OR, odds ratio; CI, confidence interval; hs-crp, high sensitivity C-reactive protein. * Insulin resistance defined as cases with homeostasis model assessment score in upper quartile, namely above 2.8 and 2.3 for men and women, respectively. odds ratios with corresponding 95% confidence intervals (95% CIs) has been adjusted for age, BMI, hs-crp, tobacco smoking, alcohol use, history of cardiovascular diseases and hormone replacement therapy (for women) in 6,617 participants with hs-crp availabe. <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <

152 macrophages 9. Since obesity is associated with increased macrophage infiltration into adipose tissue, a similar scenario may play a role in vivo. In a recent publication, an association of plasma procalcitonin with central body fat distribution was found in women with polycystic ovary syndrome 15. In line with this, we found a significant independent association of waist circumference with procalcitonin in both sexes. The associations of plasma procalcitonin levels with insulin resistance and components of the metabolic syndrome were attenuated after adjustment for BMI and therefore dependent on BMI. This supports our view that circulating levels of procalcitonin are partly dependent on adipose tissue mass. However, another part of the associations was independent of BMI. Possible explanations are that circulating levels of procalcitonin are related to adipocyte function rather than mass, or that other factors that link inflammation to the metabolic syndrome play a role, e.g. nonassessed atherosclerosis 17 or periodontitis 18. This study extends the available information for procalcitonin to a role as a biomarker of non-infectious conditions, namely the metabolic and cardiovascular arena. Moreover, since plasma procalcitonin can now be measured within the normal range, it warrants further research into its potential to identify individuals at risk of cardiovascular and chronic metabolic disease. There are several limitations of this study. The study is a cross-sectional investigation and causal relationships of procalcitonin as a novel biomarker of the metabolic syndrome and insulin resistance can not be inferred. Another limitation is the use of insulin resistance based on HOMA-IR instead of the gold standard hyperinsulinemic euglycemic clamp technique. While our study included apparently healthy adults, mostly recruited from Caucasians in the Netherlands, it is unclear if our findings would be replicable in other regions and among un-healthy individuals with cardiovascular or other comorbidities. In conclusion, our findings based on community-based data show that higher plasma procalcitonin levels in the normal range are associated with increased measures of obesity, components of the metabolic syndrome, and greater risk of having metabolic syndrome and insulin resistance. Because associations only partly depend on BMI, plasma procalcitonin may serve as a new marker for adipocyte dysfunction, chronic low-grade inflammation or both. 149

153 Acknowledgments This work was supported by the Netherlands Heart Foundation, Dutch Diabetes Research Foundation and Dutch Kidney Foundation. This research was performed within the framework of CTMM, the Center for Translational Molecular Medicine ( project PREDICCt (grant 01C ). Disclosures: Dr. J. Struck is an employee of B.R.A.H.M.S AG, which is the manufacturer of and holds patent rights on the procalcitonin assay. The present study was not financed by BRAHMS AG. No other author has anything to declare. 150

154 References 1. Eckel RH, Grundy SM, Zimmet PZ. The metabolic syndrome. Lancet 2005;365(9468): Fernandez-Real JM, Ricart W. Insulin resistance and chronic cardiovascular inflammatory syndrome. Endocr Rev 2003;24(3): Jacobs JW, Lund PK, Potts JT, Jr., Bell NH, Habener JF. Procalcitonin is a glycoprotein. J Biol Chem 1981;256(6): Rosenfeld MG, Mermod JJ, Amara SG, et al. Production of a novel neuropeptide encoded by the calcitonin gene via tissue-specific RNA processing. Nature 1983;304(5922): Becker KL, Nylen ES, White JC, Muller B, Snider RH, Jr. Clinical review 167: Procalcitonin and the calcitonin gene family of peptides in inflammation, infection, and sepsis: a journey from calcitonin back to its precursors. J Clin Endocrinol Metab 2004;89(4): Assicot M, Gendrel D, Carsin H, Raymond J, Guilbaud J, Bohuon C. High serum procalcitonin concentrations in patients with sepsis and infection. Lancet 1993;341(8844): Briel M, Schuetz P, Mueller B, et al. Procalcitonin-guided antibiotic use vs a standard approach for acute respiratory tract infections in primary care. Arch Intern Med 2008;168(18):2000-7; discussion Linscheid P, Seboek D, Nylen ES, et al. In vitro and in vivo calcitonin I gene expression in parenchymal cells: a novel product of human adipose tissue. Endocrinology 2003;144(12): Linscheid P, Seboek D, Schaer DJ, Zulewski H, Keller U, Muller B. Expression and secretion of procalcitonin and calcitonin gene-related peptide by adherent monocytes and by macrophageactivated adipocytes. Crit Care Med 2004;32(8): Linscheid P, Seboek D, Zulewski H, Keller U, Muller B. Autocrine/paracrine role of inflammation-mediated calcitonin gene-related peptide and adrenomedullin expression in human adipose tissue. Endocrinology 2005;146(6): Morgenthaler NG, Struck J, Fischer-Schulz C, Seidel-Mueller E, Beier W, Bergmann A. Detection of procalcitonin (PCT) in healthy controls and patients with local infection by a sensitive ILMA. Clin Lab 2002;48(5-6): Lambers Heerspink HJ, Brantsma AH, de Zeeuw D, Bakker SJ, de Jong PE, Gansevoort RT. Albuminuria assessed from first-morning-void urine samples versus 24-hour urine collections as a predictor of cardiovascular morbidity and mortality. Am J Epidemiol 2008;168(8): National Cholesterol Education Program. Third Report of the National Cholesterol Education Program (NCEP) Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults (Adult Treatment Panel III) final report. Circulation 2002;106(25): Matthews DR, Hosker JP, Rudenski AS, Naylor BA, Treacher DF, Turner RC. Homeostasis model assessment: insulin resistance and beta-cell function from fasting plasma glucose and insulin concentrations in man. Diabetologia 1985;28(7): Puder JJ, Varga S, Kraenzlin M, De Geyter C, Keller U, Muller B. Central fat excess in polycystic ovary syndrome: relation to low-grade inflammation and insulin resistance. J Clin Endocrinol Metab 2005;90(11): van Ree RM, de Vries AP, Oterdoom LH, et al. Plasma procalcitonin is an independent predictor of graft failure late after renal transplantation. Transplantation 2009;88(2): Holewijn S, den Heijer M, Swinkels DW, Stalenhoef AF, de Graaf J. The metabolic syndrome and its traits as risk factors for subclinical atherosclerosis. J Clin Endocrinol Metab 2009;94(8): D'Aiuto F, Sabbah W, Netuveli G, et al. Association of the metabolic syndrome with severe periodontitis in a large U.S. population-based survey. J Clin Endocrinol Metab 2008;93(10):

155 Appendix Figure S1. Association of plasma procalcitonin levels with body mass index (A) and waist circumference (B) in men and women 152

156 Table S1. Regression coefficients for Log2 procalcitonin with components of metabolic syndrome, fasting insulin and HOMA-IR* Glucose, mmol/l Model 1 Model 2 Model 3 Model 4 Model 5 Insulin, mu/l Model 1 Model 2 Model 3 Model 4 Model 5 HOMA-IR Model 1 Model 2 Model 3 Model 4 Model 5 Waist circumference, cm Model 1 Model 2 Model 3 Model 4 Model 5 Systolic blood pressure, mmhg Model 1 Model 2 Model 3 Unstandardized β (95%CI) 0.45 ( ) 0.34 ( ) 0.24 ( ) 0.23 ( ) 0.22 ( ) 3.15 ( ) 2.88 ( ) 1.53 ( ) 1.43 ( ) 1.42 ( ) 1.04 ( ) 0.92 ( ) 0.55 ( ) 0.52 ( ) 0.52 ( ) 5.97 ( ) 4.50 ( ) ( ) 6.18 ( ) 3.60 ( ) 1.60 ( ) Men Women Standardized β P value Unstandardized β (95%CI) <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 < <0.001 <0.001 < ( ) 0.42 ( ) 0.31 ( ) 0.29 ( ) 0.24 ( ) 4.21 ( ) 3.65 ( ) 2.21 ( ) 2.05 ( ) 1.85 ( ) 1.34 ( ) 1.11 ( ) 0.73 ( ) 0.69 ( ) 0.60 ( ) 9.15 ( ) 5.76 ( ) ( ) ( ) 3.90 ( ) 2.17 ( ) Standardized β P value <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 < <0.001 <0.001 <

157 Model 4 Model 5 Diastolic blood pressure, mmhg Model 1 Model 2 Model 3 Model 4 Model 5 HDL cholesterol, mmol/l Model 1 Model 2 Model 3 Model 4 Model 5 Triglyceride, mmol/l Model 1 Model 2 Model 3 Model 4 Model ( ) 1.65 ( ) 3.37 ( ) 1.87 ( ) 0.95 ( ) 0.88 ( ) 0.84 ( ) ( ) ( ) ( ) ( ) ( ) 0.46 ( ) 0.45 ( ) 0.32 ( ) 0.30 ( ) 0.31 ( ) <0.001 < <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 < ( ) 1.31 ( ) 3.87 ( ) 1.43 ( ) 0.92 ( ) 0.80 ( ) 0.73 ( ) ( ) ( ) ( ) ( ) ( ) 0.54 ( ) 0.42 ( ) 0.34 ( ) 0.32 ( ) 0.29 ( ) Abbreviations: HOMA-IR, homeostasis model assessment for insulin resistance; HDL, high density cholesterol; hs-crp, high sensitivity C-reactive protein. *Per doubling increase of procalcitonin level, unstandardized regression coefficients, β, represents the every unit change of metabolic risk factors. Model 1 is the unadjusted β for Log2 procalcitonin level with the variables. Model 2 presents the age-adjusted β. In model 3, BMI is added to model 2 (not for waist circumference). In model 4, waist circumference is added to model 3 (not for waist circumference). In model 5, hs-crp, tobacco smoking, alcohol use, history of cardiovascular disease and hormone replacement therapy (for women) are added to model 4 in 6617 participants with hs-crp availabe <0.001 < <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <

158 Chapter 8 Sex differences in the association between plasma copeptin and incident type 2 diabetes: the PREVEND study Ali Abbasi 1,2 ; Eva Corpeleijn 1 ; Esther Meijer 2 ; Douwe Postmus 1 ; Ron T. Gansevoort 2 ; Rijk O.B. Gans 2 ; Joachim Struck 3 ; Hans Hillege 1 ; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Department of Research, BRAHMS AG (Part of Thermo Fisher Scientific), Hennigsdorf Germany Diabetologia. 2012;55:

159 Abstract Background Vasopressin plays a role in osmoregulation, glucose homeostasis and inflammation. Therefore, plasma copeptin, the stable C-terminal portion of the precursor of vasopressin, has strong potential as a biomarker for the cardio-metabolic syndrome and diabetes. Previous results were contradictory, which may be explained by differences between men and women in responsiveness to the vasopressin system. The aim of this study was to evaluate the value of copeptin for prediction of future type 2 diabetes in men and women separately. Methods From the PREVEND study, 4,063 women and 3,909 men without diabetes at baseline were included. A total of 208 women and 288 men developed diabetes during a median follow-up of 7.7 years. Results In multivariable-adjusted models, we observed a stronger association of copeptin with future risk of diabetes in women (OR:1.49;95%CI:1.24,1.79) than in men (OR:1.01;95%CI:0.85,1.19) (pinteraction<0.01). Addition of copeptin to the DESIR (the Data from the Epidemiological Study on the Insulin Resistance Syndrome) clinical model improved the discriminative value of copeptin (C-statistic,+0.007, p=0.02) and reclassification (IDI=0.004, p<0.01) only in women. We, however, observed no improvement in men. The additive value of copeptin in women was maintained when other independent predictors for glucose homeostasis, renal function, and inflammation, such as glucose, hs-crp and 24-hour urinary albumin excretion (UAE), were included in the model. Conclusions The association of plasma copeptin with the risk of developing diabetes was stronger in women than in men. Plasma copeptin alone and along with existing biomarkers (glucose, hs-crp and UAE) significantly improved the risk prediction of diabetes in women. 156

160 Introduction The arginine vasopressin (AVP) stress-adaptation system has been shown to play a role in glucose homeostasis in both experimental and human studies 1 2. Epidemiological studies investigating the prospective association between plasma AVP levels and risk of type 2 diabetes are scarce. The main reason may be that reliable measurements of AVP are difficult in large scale sample sizes. AVP in blood is mainly bound to platelets in circulation and is unstable in isolated plasma 3, 4. In addition, most AVP measurements have relatively limited sensitivity. Recently, an assay for copeptin, the C-terminal portion of the precursor of AVP has been developed 5. Copeptin can be a reliable marker of AVP secretion and a surrogate for circulating AVP concentration 3, 5. One recent study found that high baseline levels of copeptin are associated with increased risk for development of type 2 diabetes 6. The link between the AVP stress-adaptation system and type 2 diabetes may lie in stimulatory effects of AVP on hepatic glucose production 7, effects on insulin release from the pancreas 8, stimulation of endogenous cortisol secretion 9, and adverse effects on whole-body insulin resistance 10. Of note, there are marked differences in responsiveness of the AVP stress- adaptation system between men and women 11, 12. We hypothesized that there could be a difference in the association of copeptin with type 2 diabetes between men and women. In a previous study in the population in which we would like to test this hypothesis, we found independent associations of high sensitivity C-reactive protein (hs-crp) and 24-hour urinary albumin excretion (UAE) with the risk of type 2 diabetes in the general population 13. The associations of hs-crp and UAE with the risk of type 2 diabetes have also been found in several other studies We aimed to test whether the association of copeptin with type 2 diabetes was independent of other covariates including clinical variables and more established biomarkers like glucose, hs-crp and 24-hour UAE. In addition, the present study evaluates the predictive ability of copeptin for the risk of developing type 2 diabetes in men and women separately. The predictive ability is evaluated with addition to an existing sex specific prediction model. Next, we performed a comparison of the potential additive value of copeptin to glucose, hs-crp and 24- hour UAE. Methods Study population and design The study population was obtained from the Prevention of Renal and Vascular Endstage Disease (PREVEND) study, a Dutch cohort drawn from the general population (age ranged between 28 and 75 years) of the city of Groningen, the Netherlands. Details on study design, recruitment, and procedures have been published elsewhere 18. Of 8,592 participants in the baseline cohort, we excluded

161 individuals with diabetes at baseline (self-reported physician diagnosis and screendetected prevalent cases) and 289 with missing data on follow-up, leaving 4,063 nondiabetic women and 3,909 men for the our post-hoc analysis. The PREVEND study was approved by the local medical ethics committee, University Medical Center Groningen, and was performed according to the principles outlined in the Declaration of Helsinki. All participants gave written informed consent. Clinical and biomarker measurements The first screening took place in 1997 to 1998; the second in 2001 to 2003 and the third in 2003 to In each screening, the participants underwent two outpatient visits to assess medical history, anthropometry and cardiovascular and metabolic risk factors, and they had to collect two 24-hour urine samples. Information on use of medication was completed and confirmed by using data from pharmacy registries of all community pharmacies in the city of Groningen 19. In 89.9% of all participants, blood samples for measurement of copeptin were taken after overnight fasting. Total cholesterol and plasma glucose were measured by dry chemistry (Eastman Kodak, Rochester, New York). High density lipoprotein (HDL) cholesterol was measured with a homogeneous method (direct HDL, Aeroset TM System, Abbott Laboratories, Abbott Park, Illinois). Hypertension was defined by self-reported physician diagnosis, use of antihypertensive medication, or blood pressure 140/90mmHg. Triacylglycerol was measured enzymatically. hs-crp was determined by nephelometry (BN II, Dade Behring, Marburg, Germany). UAE - given as the mean of the two 24-h urine excretions - was determined by nephelometry with a threshold of 2.3 mg/l and intra- and inter-assay coefficients of variation of less than 2.2% and less than 2.6%, respectively (Dade Behring Diagnostic, Marburg, Germany). Plasma copeptin level was measured by a new sandwich immunoassay (B.R.A.H.M.S GmbH/Thermo Fisher Scientific, Hennigsdorf/Berlin, Germany), which was described previously 5, 20. The lower detection limit was 0.4 pmol/l and the functional assay sensitivity (20% inter-assay coefficient of variation) was less than 1 pmol/l 5. All the technicians were blinded to the participants characteristics. Outcome definition Incident cases of type 2 diabetes were ascertained if one or more of the following criteria were met: 1) fasting plasma glucose 7.0 mmol/l (126 mg/dl); 2) random sample plasma glucose 11.1 mmol/l (200 mg/dl); 3) self-report of a physician diagnosis; 4) use of antidiabetic medications based on a central pharmacy registration 21. We included cases from 3 months after the baseline screening visits ( ) until January

162 Statistical Analyses Continuous data were compared by using one-way ANOVA or a Kruskal-Wallis test, as, when applicable. A χ2 test was used for the comparison of categorical variables to test differences across sex specific quartiles of copeptin. Because of significant sex differences, we investigated the associations between baseline characteristics and plasma copeptin levels separately for women and men. We applied logistic regression analyses to examine the hypothesis that plasma copeptin is associated with the risk of developing type 2 diabetes in women and men. Odds ratios (ORs) with 95% confidence intervals (CIs) for type 2 diabetes were calculated according to base-two logarithmically transformed copeptin. In further analyses, these associations were tested across sex specific quartiles of copeptin while the lowest quartile considered as the reference. In model 1, basic adjustment was for age. In model 2, we further adjusted for alcohol use, smoking status, and family history of diabetes as covariates which can be confounding of the association between copeptin and risk of diabetes. In model 3, we further adjusted for covariates included in the metabolic syndrome, i.e., waist circumference, hypertension, HDL-cholesterol, triacylglycerol and fasting glucose. In model 4, we further adjusted for hs-crp and 24-hour UAE. We examined the added value of copeptin for the risk prediction of developing diabetes on top of the existing DESIR clinical models. The DESIR models were chosen because it has prediction rules separately for women and men 22. The DESIR models included data on family history of diabetes, waist circumference, hypertension, in women; and data on smoking status, waist circumference and hypertension in men 22. To evaluate the added value of copeptin, we compared the prediction of the DESIR models, as the reference, to that of the models including log2 copeptin. Next, we added log2 hs-crp and log2 UAE to the DESIR models and examined if these two conventional cardiometablic biomarkers could improve the risk prediction of diabetes. Thereafter, to evaluate the value of copeptin over existing biomarkers, we added log2 copeptin along with log2 hs-crp and log2 UAE to the DESIR models. Finally, we added glucose, a strong predictor for diabetes, to the DESIR models and examined whether copeptin, hs-crp and 24-h UAE could improve prediction above the models incorporating glucose. We examined improvement of diabetes prediction in terms of discrimination and integrated discrimination improvement (IDI) 23, 24. The Discrimination performance denotes to what extent the model distinguishes between individuals with and without incident diabetes; a value of 1 implies a perfect discrimination and a value of 0.5 implies performance no better than chance. Discrimination was examined by calculating the C-statistic with 95%CI. IDI, a continuous measure of reclassification, was calculated by subtracting the mean difference of predicted risk between the DESIR models to that of the models including biomarkers for those who developed diabetes from the corresponding risks for those who did not develop diabetes. A significant p value of IDI represents an improved prediction 23, 24. For most baseline variables, <1% was missing, whereas this was up to 8% for self-reported variables. A single imputation and predictive mean matching was 159

163 applied for missing data. In the current analysis, a weighted method was performed to compensate for baseline enrichment of the PREVEND participants with high urinary albumin concentration (>10 mg/l). Given the strong predictive value of glucose, we performed a sensitivity analysis with exclusion of individuals (women, n= 305; men, n= 538) with impaired fasting glucose (IFG) at baseline. IFG was defined by the ADA criteria of fasting glucose of mmol/l 25. Next, we repeated analyses after excluding those who used antihypertensive medications (women, n= 569; men, n= 617). In addition, we assessed whether the different components of the DESIR models might have affected the predictive value of copeptin. To do this, we fitted the model for women and examined the effect of adding copeptin in men. A p-value of 0.05 or less, two-sided, was considered statistically significant. All the statistical analyses were carried out using IBM SPSS Statistics 19 and R for Windows ( Results Baseline clinical characteristics The associations between baseline clinical characteristics and plasma copeptin, stratified by sex, are summarized in Table 1. Median (interquartile range [IQR]) copeptin levels were higher in men, i.e ( ) pmol/l in men and 3.59 ( ) pmol/l in women (p<0.001). For both men and women, across sex-specific quartiles of copeptin, higher copeptin was positively related to age, body mass index (BMI), waist circumference, high blood pressure, fasting blood glucose, and total cholesterol, but was not related to HDL-cholesterol. Also hs-crp and UAE increased with higher copeptin levels. Women with higher copeptin were more likely to be smoker, whereas men with higher copeptin had higher triglycerides. Plasma copeptin and type 2 diabetes During median (IQR) follow-up for 7.7 ( ) years, 208 (5.1%) women and 288 (7.4%) men developed type 2 diabetes. Table 2 depicts the association of copeptin and the risk of type 2 diabetes, calculated per doubling (per log2-unit increase) of copeptin levels and over sex specific quartiles separately for women and men. In women, the crude OR (95%CI) for the risk of type 2 diabetes was 1.60 (1.37,1.85) per doubling of copeptin levels. After adjustment for age (model 1), smoking, alcohol use, and family history of diabetes (model 2), and waist circumference, hypertension, fasting glucose, HDL-cholesterol and triglycerides (model 3) this association (OR:1.50; 95%CI: 1.25,1.80) remained statistically significant. In model 4, adjustment for hs-crp and 24-hour UAE did not further change the association of copeptin with type 2 diabetes (OR:1.49; 95%CI:1.24,1.79). In men, crude OR (95%CI) for the risk of developing type 2 diabetes was 1.19 (1.03,1.37) per doubling of copeptin levels. Adjustment for the variables in model 2 did not materially change this association (OR:1.18; 95%CI:1.01,1.37). Further 160

164 adjustments for waist circumference, hypertension, fasting glucose, HDL-cholesterol and triglycerides attenuated the association to non-significance (p=0.74). Higher copeptin levels were a significantly stronger predictor of type 2 diabetes in women than in men (p<0.01 for interaction) in both crude and multivariable-adjusted models. We also repeated the main analyses to examine the association between copeptin and the risk of diabetes in the entire sample. The age and sex- adjusted OR for the risk of type 2 diabetes were 1.31 (1.18,1.46) per doubling of copeptin levels. After multivariable-adjustment, including age, sex and the other variables in model 4, the risk was attenuated to 1.21 (1.07,1.37). Prognostic value of plasma copeptin We examined the predictive value of copeptin for the risk of developing type 2 diabetes when added with the DESIR clinical models (Table 3) 22. In women, the DESIR model, including data on family history of diabetes, waist circumference, hypertension as predictors, showed a C-statistic of ( ). Addition of copeptin (log2) significantly improved the C-statistic (a change of ; p=0.02) of the model, and led to an IDI of (p<0.01). Addition of hs-crp (log2) and 24-hour UAE (log2) improved the C-statistic (a change of ; p=0.09) and led to IDI of (p<0.02). After addition of copeptin along with hs-crp and 24-hour UAE, we observed a change of for the C-statistic and IDI of (p=0.01). The DESIR model and glucose showed a C-statistic of (0.860,0.911). Addition of copeptin significantly improved the C-statistic (a change of ; p=0.01) of the DESIR model and glucose. When hs-crp and 24-hour UAE were included along with the DESIR model and glucose, we observed non-significant improvements (a change of C- statistic: ; p=0.11). In men, the DESIR model including data on smoking status, waist circumference and hypertension as predictors, showed a C-statistic of (0.681,0.745), which was considerably lower than in the women. Addition of copeptin (log2) alone did not improve the prediction in terms of discrimination (p=0.40) and reclassification (IDI of ; p=0.15). Addition of hs-crp (log2) and 24-hour UAE (log2) significantly improved the C-statistic (a change of 0.011; p=0.01) and led to IDI of (p=0.003). The DESIR model and glucose showed a C-statistic of ( ). When hs-crp and 24-hour UAE were included along with the DESIR model and glucose, we observed borderline improvements (a change of C-statistic: ; p=0.07). Sensitivity analyses First, when we excluded the individuals with IFG at baseline, the crude OR (95%CI) for the risk of diabetes was 1.93 (1.58,2.36) per doubling of copeptin levels in women. The adjusted OR for model 4 was 1.81 (1.42,2.30). In men, the crude and adjusted ORs were 1.21 (1.00,1.47) and 1.03 ( ), respectively (p<0.01 for interaction by sex in both the crude analyses and in model 4). The DESIR model combined with glucose showed a C-statistic of (0.756,0.845) in women. Addition of copeptin 161

165 Table 1. Baseline clinical characteristics of participants in total and according to quartiles of plasma copeptin Sex specific quartiles p value * Women Total No. of participants 4,063 1,001 1,025 1,019 1,018 - Copeptin level- pg/ml 3.6 ( ) 1.8 ( ) 2.9 ( ) 4.4 ( ) 7.6 ( ) - Age- yr 47.7± ± ± ± ±12.3 <0.001 Family history of diabetes- % 830 (20.4) 202 (20.2) 190 (18.5) 208 (20.4) 230 (22.6) 0.16 Current smoker - % 1374 (33.8) 275 (27.5) 313 (30.5) 366 (35.9) 420 (41.3) <0.001 Ever alcohol use- % 2718 (67.2) 641 (64.2) 707 (69.2) 684 (67.5) 686 (67.9) 0.10 BMI- kg/m ± ± ± ± ±5.2 <0.001 Waist circumference- cm 82.9± ± ± ± ±13.6 <0.001 Systolic blood pressure- mmhg 119.4± ± ± ± ± Diastolic blood pressure- mmhg 68.7± ± ± ± ± Hypertension- % 959 (23.6) 222 (22.2) 216 (21.1) 235 (23.1) 286 (28.1) Glucose- mmol/l 4.62± ± ± ± ± Total cholesterol- mmol/l 5.60± ± ± ± ± HDL cholesterol- mmol/l 1.49± ± ± ± ± Triacylglycerol - mmol/l 1.05 ( ) 1.02 ( ) 1.04 ( ) 1.06 ( ) 1.06 ( ) 0.13 hs-crp- mg/l 1.31 ( ) 1.20 ( ) 1.26 ( ) 1.35 ( ) 1.46 ( ) 0.03 UAE- mg/24hour 10.4 ( ) 7.3 ( ) 8.2 ( ) 8.3 ( ) 9.8 ( ) <0.001 Men No. of participants 3, Copeptin level- pg/ml 6.2 ( ) 3.5 ( ) 4.6 ( ) 7.6 ( ) 12.5 ( ) - Age- yr 50.2± ± ± ± ±12.8 <0.001 Family history of diabetes- % 753 (19.3) 187 (19.2) 169 (17.2) 206 (21.1) 191 (19.5) 0.20 Current smoker - % 1369 (35.0) 319 (32.8) 347 (35.4) 361 (36.9) 342 (35.0) 0.29 Ever alcohol use- % 3210 (82.5) 794 (81.8) 821 (84.3) 798 (81.9) 797 (81.9) 0.40 BMI- kg/m ± ± ± ± ±3.8 <0.001 Waist circumference- cm 93.6± ± ± ± ±11.4 <0.001 Systolic blood pressure- mmhg 128.7± ± ± ± ±18.8 <0.001 Diastolic blood pressure- mmhg 74.7± ± ± ± ±10.0 <0.001 Hypertension- % 1278 (32.7) 273 (28.1) 317 (32.3) 316 (32.3) 372 (38.0) <

166 Glucose- mmol/l 4.86± ± ± ± ±0.70 <0.001 Total cholesterol- mmol/l 5.68± ± ± ± ±1.17 <0.001 HDL cholesterol- mmol/l 1.16± ± ± ± ± Triacylglycerol - mmol/l 1.28 ( ) 1.24 ( ) 1.23 ( ) 1.30 ( ) 1.35 ( ) <0.001 hs-crp- mg/l 1.20 ( ) 1.03 ( ) 1.08 ( ) 1.22 ( ) 1.45 ( ) <0.001 UAE - mg/24hour 8.3 ( ) 9.3 ( ) 9.7 ( ) 10.7 ( ) 11.9 ( ) <0.001 BMI denotes body mass index, HDL high density lipoprotein, hs-crp high sensitivity C-reactive protein and UAE urine albumin excretion. Data were given as mean (SD) for continuous variables, tested using ANOVA or Kruskal-Wallis, and numbers (percentage) for categorical variables, tested using χ2 test. * Univariate analyses were for comparison across sex-specific quartiles of plasma copeptin. 163

167 Table 2. ORs (95% CI) for incident type 2 diabetes according to quartiles of plasma copeptin Sex specific quartiles OR (95% CI) per Log2-unit increase * Women (n=4,063) No. of cases (%) 31 (3.1) 42 (4.1) 59 (5.8) 76 (7.5) Crude analysis (0.79,2.26) 2.54 (1.58,4.06) 3.82 (2.42,6.03) 1.60 (1.37,1.85) <0.001 Model (1.83,2.41) 2.29 (1.42,3.69) 3.24 (2.03,5.16) 1.47 (1.26,1.71) <0.001 Model (0.86,2.52) 2.33 (1.43,3.77) 3.22 (2.01,5.15) 1.45 (1.24,1.69) <0.001 Model (0.89,3.11) 3.46 (1.95,6.13) 3.57 (2.03,6.27) 1.50 (1.25,1.80) <0.001 Model (0.88,3. 10) 3.37 (1.90,6.00) 3.54 (2.01,6.24) 1.49 (1.24,1.79) <0.001 p value * Men (n=3,909) No. of cases (%) 58 (6.0) 62 (6.3) 88 (9.0) 80 (8.2) Crude analysis b (0.65,1.49) 1.30 (0.88,1.93) 1.57 (1.07,2.32) 1.19 (1.03,1.37) 0.02 Model (0.66,1.52) 1.31 (0.88,1.95) 1.46 (0.99,2.16) 1.17 (1.01,1.35) 0.04 Model (0.67,1.55) 1.28 (0.86,1.91) 1.48 (0.99,2.19) 1.18 (1.01,1.37) 0.03 Model (0.76,1.84) 1.28 (0.83,1.95) 1.03 (0.66,1.59) 1.03 (0.87,1.21) 0.74 Model (0.74,1.80) 1.25 (0.81,1.91) 1.00 (0.63,1.52) 1.01 (0.85,1.19) 0.95 * OR (95% CI; p value) expressed per unit increase in log2-transformed levels of plasma copeptin Model 1 is adjusted for age; model 2 is adjusted for age plus alcohol use, smoking status, and family history of diabetes; model 3 is adjusted for variables in model 3 plus waist circumference, hypertension, fasting glucose, HDL-cholesterol and triacylglycerol; model 4 is adjusted for variables in model 3 plus high sensitivity C-reactive protein and 24-hour urine albumin excretion. 164

168 Table 3. Added value of plasma copetin above the DESIR model for the prediction risk of developing type 2 diabetes* Women Men C,statistic for the DESIR model (95%CI) * (0.795,0.850) (0.686,0.745) C,statistic for the DESIR model plus copeptin (95%CI) (0.803,0.855) (0.685,0.744) p value for change of C,statistic IDI (p value) (<0.01) (0.15) C,statistic for the DESIR model plus hs,crp and UAE (95%CI) (0.805,0.857) (0.700,0.757) p value for change of C,statistic IDI (p value) (0.01) (0.003) C,statistic for the DESIR model plus copeptin, hs,crp and UAE (95%CI) (0.810,0.860) (0.700,0.757) p value for change of C,statistic IDI (p value) (0.01) (0.002) DESIR denotes Data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR), IDI, integrated discrimination improvement, hs-crp, high sensitivity C-reactive protein, and UAE, 24-hour urine albumin excretion. * The DESIR models includes data on family history of diabetes, waist circumference, hypertension, in women, and data on smoking status, waist circumference and hypertension in men. The DESIR model was considered as reference. 165

169 significantly improved the C-statistic (a change of ; p=0.05) of the DESIR model combined with glucose. The DESIR model combined with glucose showed a C- statistic of (0.721,0.801) in men. Addition of copeptin did not improve the C- statistic (p=0.63). Second, we calculated the risk of diabetes per doubling of copeptin in individuals who did not use antihypertensive medication. The crude OR and adjusted OR for model 4 in women were 1.75 (1.43,2.14) and 1.65 (1.27,2.13), respectively. The crude and adjusted ORs in men were 1.24 (1.05,1.47) and 1.02 (0.84,1.25), respectively. Third, we fitted the model for women and calculated the C-statistic for predicting the risk of diabetes in men. In men, the model for prediction of diabetes risk in women, based on family history of diabetes, waist circumference, hypertension as predictors, showed a C-statistic of (0.711,0.769). Addition of copeptin did not improve the C-statistic (p=0.94) in men similar to our finding for applying the DESIR model for men. Discussion In this population-based cohort, we demonstrated that plasma copeptin, as a reliable surrogate marker for AVP, is of additive value to predict future type 2 diabetes. Furthermore, we show that the association between copeptin and the risk of developing type 2 diabetes is modified by sex. In women, addition of copeptin to the DESIR model significantly improved the risk prediction of diabetes in terms of discrimination and reclassification. It is true that fasting glucose was a very good predictor of incident type 2 diabetes, because it is part of the diagnosis. Despite of this, women in the fourth quartile of copeptin had a 3.5-times higher risk for developing type 2 diabetes compared with those in the first quartile of copeptin when we adjusted for fasting glucose and other clinical variables. Of note, along with glucose and existing biomarkers for inflammation i.e. hs-crp, and renal function i.e. 24-h UAE, the addition of copeptin to the model further improved the risk prediction of diabetes. In men, we observed that addition of copeptin did not improve the risk prediction of diabetes in terms of discrimination and reclassification. In addition, the association of copeptin with the risk of diabetes particularly strengthened in women when we further excluded the individuals with IFG at baseline. Thus, it is particularly important for the assessment of the value of novel biomarkers to take into account the possible effect of sex on the risk prediction of diabetes. There are several prediction models including data on demographics, anthropometric measures and lifestyle factors which have been developed for the risk of diabetes in the general population 26, 27. In these models, sex has been incorporated as one of the most commonly used predictors for the risk of diabetes 27. In our study, we used the DESIR clinical model, because the DESIR models were developed for men and women separately 22. Another aspect regarding risk prediction is the clinical utility of novel biomarkers like copeptin. The change of C-statistics is interpreted as whether 166

170 addition of biomarkers may improve the ability of model to assign a higher probability of risk to cases compared with non-cases 24. The C-statistic is considered as one of main commonly reported measures. However, it may be insensitive for small improvements of prediction 24, 28. Alternatively, the IDI can be calculated as a measure of continuous reclassification. A significant IDI is interpreted as that addition of biomarkers to the model increases the difference in average predicted risk between cases and non-cases 23, 24. It is difficult to judge whether the statistically significant improvements in the risk prediction of diabetes may be clinically relevant. To answer this question, one should first define the clinical relevant categories for the risk of diabetes and assign correct movements of cases and non-cases into risk strata 24, 28. Currently, widely accepted cut points and the number of categories are lacking for the risk prediction of diabetes. Thus, further studies will need to replicate current findings in other settings and subsequently assess the clinical utility of novel biomarkers like copeptin. With regard to the differences in copeptin level in men and women, the higher plasma copeptin level in men was consistently observed in our and previous studies 4, Sex is one of the major determinants of plasma levels of copeptin. The range in copeptin levels is comparable between men and women, and the difference in absolute copeptin level is not likely the explanation for the difference in predictive ability for men and women. More in general, it is worthy to note that most prediction models including data on common risk factors have shown a better performance in women when compared with men 26. Various known biomarkers, like hs-crp, insulin and endogenous sex hormone, improved the risk prediction of type 2 diabetes differently in women and in men 16, 32, 33. We and others have shown before that higher plasma copeptin levels were positively associated with the metabolic syndrome, insulin resistance, inflammatory marker of hs-crp and higher UAE in cross-sectional studies 4, 29, 30. Likewise, all these conditions are known as predictors for the risk of type 2 diabetes 13. In extension of these studies, two previous studies investigated the association of copeptin with the risk of type 2 diabetes 6, 32. Malmö Diet and Cancer (MDC) study showed that copeptin, independently of a wide range of clinical risk factors, predicts the risk of type 2 diabetes in the general population 6. However, the FINRISK97 study could not find an independent association 32. The fact that we found a stronger association of copeptin with the risk of type 2 diabetes in women than in men might partly be explained by differences in population characteristics compared to other studies. For example, the MDC study from a Swedish population-based cohort included 4,472 participants with 174 incident cases 6 who were older mean age of 58 years and contained around 60% women. In the MDC study, including a higher numbers of women who had comparable copeptin levels to that of our study, a potential sexrelated effect on the association of copeptin with the diabetes risk was not addressed. The FINRISK97 study from a cohort of 7,827 participants with 417 incident cases included similar numbers of women and men 32. In the FINRISK97 study, a higher but non-significant risk of type 2 diabetes per one SD increase of copeptin was found 167

171 in total and sex-stratified population 32. In this latter study, the range of copeptin levels was smaller than the MDC study and our study for both women and men. Theoretically, this smaller range may also lead to overlap of copeptin levels between individuals with and without type 2 diabetes in the latter study which limits the predictive value of copeptin above clinical risk factors 34. The finding that the AVP system may provide promising biomarkers for the prediction of type 2 diabetes is in line with experimental data showing that AVP system has various actions on underlying pathways involved in the pathogenesis of type 2 diabetes. AVP stimulates glycogenolysis and gluconeogenesis through the V1a receptors in the liver 7. In addition, AVP have been shown to induce glucagon and insulin release from pancreas which is mediated via V1b receptors of islet cells 10. Furthermore, AVP, via the same receptor (V1b), exerts stimulatory effects in maintaining basal secretion of ACTH and corticosterone, and in modulating HPA activity under stress conditions 9. In another aspect, insulin sensitivity signalling was oppositely modulated by AVP effects via both V1a and V1b in adipose tissue of mice 35. Previous experimental and clinical data show differences between men and women in responsiveness to the vasopressin system. Both AVP V1a and V1b receptors have been shown to be more sensitive to some effects of AVP in women than in men 11, 36. In another aspect, women have markedly lower AVP expression and lower AVP levels due to modulatory actions of estrogen on the nuclear receptors in cells of the paraventricular nucleus 37. One may assume that a lower tolerance to changes in AVP levels has a stronger effect in women than in men. In conclusion, it is particularly important for the assessment of the value of novel biomarkers to take into account the possible sex differences 38, 39. We found a stronger association of plasma copeptin with the risk of type 2 diabetes in women than in men. In women, copeptin was an independent predictor for type 2 diabetes with the added predictive value on top of the existing prediction model along with glucose and existing biomarkers for inflammation (hs-crp) and renal function (UAE), whereas in men, copeptin showed no added predictive value. 168

172 Acknowledgments This work was supported by the Netherlands Heart Foundation, Dutch Diabetes Research Foundation and Dutch Kidney Foundation. This research was performed within the framework of CTMM, the Center for Translational Molecular Medicine ( project PREDICCt (grant 01C ). We thank Prof. dr. L.T.W. de Jong-van den Berg and dr. S.T. Visser from the Department of Social Pharmacy, Pharmacoepidemiology and Pharmacotherapy, Groningen University Institute for Drug Exploration, University Medical Center Groningen and University of Groningen, The Netherlands for providing the data on pharmacy-registered use of antidiabetic medication. Duality of interest: Dr Struck is an employee of B.R.A.H.M.S GmbH/Thermo Fisher Scientific, a company, which holds patent rights on and manufactures the copeptin assay. The present study was not financed by B.R.A.H.M.S GmbH. No other author has anything to potential conflicts of interest relevant to this article. None of the study sponsors had a role in study design; in data collection, analysis, or interpretation; in writing the report; or in the decision to submit for publication. 169

173 References 1. Aoyagi T, Birumachi J, Hiroyama M, et al. Alteration of glucose homeostasis in V1a vasopressin receptor-deficient mice. Endocrinology 2007;148(5): Zerbe RL, Vinicor F, Robertson GL. Plasma vasopressin in uncontrolled diabetes mellitus. Diabetes 1979;28(5): Struck J, Morgenthaler NG, Bergmann A. Copeptin, a stable peptide derived from the vasopressin precursor, is elevated in serum of sepsis patients. Peptides 2005;26(12): Meijer E, Bakker SJ, Halbesma N, de Jong PE, Struck J, Gansevoort RT. Copeptin, a surrogate marker of vasopressin, is associated with microalbuminuria in a large population cohort. Kidney Int 2010;77(1): Morgenthaler NG, Struck J, Alonso C, Bergmann A. Assay for the measurement of copeptin, a stable peptide derived from the precursor of vasopressin. Clin Chem 2006;52(1): Enhorning S, Wang TJ, Nilsson PM, et al. Plasma copeptin and the risk of diabetes mellitus. Circulation (19): Kirk CJ, Rodrigues LM, Hems DA. The influence of vasopressin and related peptides on glycogen phosphorylase activity and phosphatidylinositol metabolism in hepatocytes. Biochem J 1979;178(2): Oshikawa S, Tanoue A, Koshimizu TA, Kitagawa Y, Tsujimoto G. Vasopressin stimulates insulin release from islet cells through V1b receptors: a combined pharmacological/knockout approach. Mol Pharmacol 2004;65(3): Tanoue A, Ito S, Honda K, et al. The vasopressin V1b receptor critically regulates hypothalamic-pituitary-adrenal axis activity under both stress and resting conditions. J Clin Invest 2004;113(2): Fujiwara Y, Hiroyama M, Sanbe A, et al. Insulin hypersensitivity in mice lacking the V1b vasopressin receptor. J Physiol 2007;584(Pt 1): Stachenfeld NS, Splenser AE, Calzone WL, Taylor MP, Keefe DL. Sex differences in osmotic regulation of AVP and renal sodium handling. J Appl Physiol 2001;91(4): Simmler LD, Hysek CM, Liechti ME. Sex differences in the effects of MDMA (ecstasy) on plasma copeptin in healthy subjects. J Clin Endocrinol Metab 2011;96(9): Brantsma AH, Bakker SJ, Hillege HL, de Zeeuw D, de Jong PE, Gansevoort RT. Urinary albumin excretion and its relation with C-reactive protein and the metabolic syndrome in the prediction of type 2 diabetes. Diabetes Care 2005;28(10): Halimi JM, Bonnet F, Lange C, Balkau B, Tichet J, Marre M. Urinary albumin excretion is a risk factor for diabetes mellitus in men, independently of initial metabolic profile and development of insulin resistance. The DESIR Study. J Hypertens 2008;26(11): Thorand B, Lowel H, Schneider A, et al. C-reactive protein as a predictor for incident diabetes mellitus among middle-aged men: results from the MONICA Augsburg cohort study, Arch Intern Med 2003;163(1): Thorand B, Baumert J, Kolb H, et al. Sex differences in the prediction of type 2 diabetes by inflammatory markers: results from the MONICA/KORA Augsburg case-cohort study, Diabetes Care 2007;30(4): Pradhan AD, Manson JE, Rifai N, Buring JE, Ridker PM. C-reactive protein, interleukin 6, and risk of developing type 2 diabetes mellitus. JAMA 2001;286(3): Lambers Heerspink HJ, Brantsma AH, de Zeeuw D, Bakker SJ, de Jong PE, Gansevoort RT. Albuminuria assessed from first-morning-void urine samples versus 24-hour urine collections as a predictor of cardiovascular morbidity and mortality. Am J Epidemiol 2008;168(8): Monster TB, Janssen WM, de Jong PE, de Jong-van den Berg LT. Pharmacy data in epidemiological studies: an easy to obtain and reliable tool. Pharmacoepidemiol Drug Saf 2002;11(5):

174 20. Fenske W, Stork S, Blechschmidt A, Maier SG, Morgenthaler NG, Allolio B. Copeptin in the differential diagnosis of hyponatremia. J Clin Endocrinol Metab 2009;94(1): Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin and risk of type 2 diabetes in the general population. Diabetologia 2011;54(9): Balkau B, Lange C, Fezeu L, et al. Predicting diabetes: clinical, biological, and genetic approaches: data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR). Diabetes Care 2008;31(10): Pencina MJ, D'Agostino RB, Sr., Steyerberg EW. Extensions of net reclassification improvement calculations to measure usefulness of new biomarkers. Stat Med 2011;30(1): Cook NR, Ridker PM. Advances in measuring the effect of individual predictors of cardiovascular risk: the role of reclassification measures. Ann Intern Med 2009;150(11): American Diabetes Association. Standards of medical care in diabetes Diabetes Care 2007;30 Suppl 1:S4-S Buijsse B, Simmons RK, Griffin SJ, Schulze MB. Risk assessment tools for identifying individuals at risk of developing type 2 diabetes. Epidemiol Rev 2011;33(1): Collins GS, Mallett S, Omar O, Yu LM. Developing risk prediction models for type 2 diabetes: a systematic review of methodology and reporting. BMC Med 2011;9: McGeechan K, Macaskill P, Irwig L, Liew G, Wong TY. Assessing new biomarkers and predictive models for use in clinical practice: a clinician's guide. Arch Intern Med 2008;168(21): Enhorning S, Struck J, Wirfalt E, Hedblad B, Morgenthaler NG, Melander O. Plasma copeptin, a unifying factor behind the metabolic syndrome. J Clin Endocrinol Metab (7):E Saleem U, Khaleghi M, Morgenthaler NG, et al. Plasma carboxy-terminal provasopressin (copeptin): a novel marker of insulin resistance and metabolic syndrome. J Clin Endocrinol Metab 2009;94(7): Bhandari SS, Loke I, Davies JE, Squire IB, Struck J, Ng LL. Gender and renal function influence plasma levels of copeptin in healthy individuals. Clin Sci (Lond) 2009;116(3): Salomaa V, Havulinna A, Saarela O, et al. Thirty-one novel biomarkers as predictors for clinically incident diabetes. PLoS One 2010;5(4):e Ding EL, Song Y, Malik VS, Liu S. Sex differences of endogenous sex hormones and risk of type 2 diabetes: a systematic review and meta-analysis. JAMA 2006;295(11): Herder C, Karakas M, Koenig W. Biomarkers for the prediction of type 2 diabetes and cardiovascular disease. Clin Pharmacol Ther 2011;90(1): Nakamura K, Aoyagi T, Hiroyama M, et al. Both V(1A) and V(1B) vasopressin receptors deficiency result in impaired glucose tolerance. Eur J Pharmacol 2009;613(1-3): Kajantie E, Phillips DI. The effects of sex and hormonal status on the physiological response to acute psychosocial stress. Psychoneuroendocrinology 2006;31(2): Rhodes ME, Rubin RT. Functional sex differences ('sexual diergism') of central nervous system cholinergic systems, vasopressin, and hypothalamic-pituitary-adrenal axis activity in mammals: a selective review. Brain Res Brain Res Rev 1999;30(2): Taking sex into account in medicine. Lancet 2011;378(9806): Oertelt-Prigione S, Parol R, Krohn S, Preissner R, Regitz-Zagrosek V. Analysis of sex and gender-specific research reveals a common increase in publications and marked differences between disciplines. BMC Med 2010;8:

175 172

176 Chapter 9 HDL cholesterol, apolipoprotein A-I and HDL particle composition independently predict incident type 2 diabetes mellitus in the general population: The PREVEND Study Ali Abbasi 1,2 ; Eva Corpeleijn 1 ; Ron T. Gansevoort 2 ; Rijk O.B. Gans 2 ; Hans Hillege 1 ; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 ; Robin P.F. Dullaart 2 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands Submitted to Diabetologia 2013

177 Abstract Background High density lipoproteins (HDL) may directly stimulate β-cell function and glucose metabolism. We determined relationships of fasting HDL-cholesterol, plasma apolipoprotein (apo) A-I, apoa-ii and of HDL-cholesterol/apoA-I and HDLcholesterol/apoA-II ratios, as estimates of HDL particle composition, with incident type 2 diabetes mellitus. Methods A prospective study was carried out in the PREVEND cohort after excluding subjects with diabetes at baseline (n=6,820; age years). Type 2 diabetes was defined by ADA criteria. The association of HDL-related variables with incident type 2 diabetes was determined by multivariate logistic regression analyses. Results After a median follow-up of 7.7 years, 394 incident cases of type 2 diabetes were ascertained (5.8 %). After adjustment for age, sex, family history of diabetes, body mass index, hypertension, alcohol and smoking, odd ratios (ORs) for diabetes were 0.55 (95 % CI: ; P<0.001), 0.81 ( ; P=0.002), 0.02 ( ; P<0.001) and 0.03 ( ; P<0.001) per 1-SD increase in HDL-cholesterol, apoa-i and in the HDL-cholesterol/apoA-I and the HDL-cholesterol/apoA-II ratios, respectively. In contrast, apoa-ii did not predict incident diabetes (OR: 1.02 ( ; P=0.71). The relationships of HDL-cholesterol and the ratios of HDLcholesterol/apoA-I and HDL-cholesterol/apoA-II remained significant after further adjustment for baseline glucose and triglycerides (P 0.001). Conclusions Higher HDL cholesterol and apoa-i levels, as well as higher HDLcholesterol/apoA-I and HDL-cholesterol/apoA-II ratios are strong and independent predictors of lower risk of future type 2 diabetes. We suggest that larger HDL particles may specifically confer lower risk of diabetes development. 174

178 Introduction The epidemic of type 2 diabetes mellitus poses a major public health concern worldwide 1, 2. Much attention is currently being paid to the development and validation of diabetes prediction models in order to optimize guidelines for diabetes prevention 3-5. Many of the proposed diabetes risk scores take account of metabolic syndrome components, including high density lipoprotein (HDL)-cholesterol 4, 5. Indeed, lower levels of HDL-cholesterol confer a higher incidence of type 2 diabetes in various ethnic populations and age groups 6-12, analogous to the cardiovascular protection attributed to HDL 13, 14. Such an inverse relationship of HDL-cholesterol with diabetes development is not surprising in the context of co-existing obesity and disturbances in lipoprotein metabolism in subjects at high risk for diabetes Importantly, evidence has accumulated recently that HDL may also be directly involved in the pathogenesis of type 2 diabetes mellitus by virtue of its ability to enhance pancreatic β-cell function and glucose uptake in skeletal muscle In line, defects in functional properties of HDL have been shown to result in increased susceptibility of pancreatic β-cells to oxidative stress, apoptosis, islet inflammation and cholesterol accumulation 20. HDL is able to restore oxidized low density lipoprotein (LDL)-induced impairment of insulin processing in vitro, whereas free apolipoprotein (apo) A-I, being the most abundant protein constituent of HDL, and also reconstituted HDL particles and native HDL have been shown to stimulate insulin secretion by enhancing cholesterol efflux out of β-cells In agreement with a contributory role of HDL functionality on the maintenance of insulin secretion, both the anti-oxidative capacity of HDL and the ability of plasma to stimulate cholesterol efflux from cultured fibroblasts have been found to be independent determinants of β-cell function in well-controlled type 2 diabetes mellitus 24. Of further relevance, apoa-i stimulates the AMP-activated protein kinase pathway in myocytes in vitro 25. Reconstituted HDL infusion also stimulates this pathway in skeletal muscle from subjects with type 2 diabetes in vivo 21. HDL could, therefore, not only lower plasma glucose by stimulating insulin secretion, but also by stimulation of glucose uptake via an insulin independent mechanism 18, 21. Despite current focus on the allegedly beneficial effects of HDL on glucose homeostasis, it is still unknown whether the major apolipoproteins of HDL, apoa-i and apoa-ii, predict incident type 2 diabetes mellitus in the general population. The same is true for HDL particle characteristics. Importantly, apoa-i and apoa-ii exert specific effects on HDL functional properties 16, 26, are protein constituents of distinct HDL subfractions, i.e. LpA-I and LpA-I:A-II particles 27, 28, and may have dissimilar potential in predicting cardiovascular disease 29, 30, and possibly also in predicting diabetes. For these reasons, it is clinically relevant to discern whether diabetes development is not only predicted by HDL-cholesterol, but also by plasma levels of apoai and apoa-ii. 175

179 The present study was initiated to determine the strength of associations of incident type 2 diabetes mellitus with HDL-cholesterol, plasma apoa-i, apoa-ii, as well as with HDL particle composition, as estimated by the ratios of HDLcholesterol/apoA-I and HDL-cholesterol/apoA-II. To this end we carried out a prospective study in the population-based prevention of renal and cardiovascular end-stage disease (PREVEND) cohort. Methods Study population and design The PREVEND study was approved by the local medical ethics committee, University Medical Center Groningen, and was performed according to the principles outlined in the Declaration of Helsinki. All participants gave written informed consent. Details on study design, recruitment, and procedures have been reported elsewhere 31 The study population is based on the Prevention of Renal and Vascular Endstage Disease (PREVEND) study, a Dutch cohort drawn from the general population (age ranged between 28 and 75 years) of the city of Groningen, the Netherlands. From the baseline cohort (n= 8,592; consisting of 6,000 individuals with a urinary albumin concentration (UAC) of 10 mg/l and 2,592 individuals with a UAC <10 mg/l), we first excluded 336 individuals with prevalent cases of diabetes. These cases were defined by either a self-report of physician diagnosis or screening at first visits ( ). Other exclusions were for 285 subjects with no follow-up data or who could not be linked to pharmacy registry and 807 individuals with nonfasting blood sampling or those using lipid-lowering agents (n=344), leaving 6820 participants who were free of baseline diabetes for our cohort analysis. Clinical and laboratory measurements During three rounds of screening from until , the participants underwent two outpatient visits to assess medical history, anthropometry and cardiovascular and metabolic risk factors, and they had to collect two 24-hour urine samples. We collected information on use of medications via data from pharmacy registries of all community pharmacies in the city of Groningen 32. Hypertension was defined by self-reported physician diagnosis, use of antihypertensive medication, or blood pressure 140/90mmHg. Total cholesterol and plasma glucose were measured by dry chemistry (Eastman Kodak, Rochester, New York). HDL-cholesterol (HDL-C) was measured with a homogeneous method (direct HDL, Aeroset TM System, Abbott Laboratories, Abbott Park, Illinois). Serum apoa-i and apoa-ii were determined by nephelometry applying commercially available reagents for Dade Behring nephelometer systems (BN II; Dade Behring, Marburg, Germany; apoa-i test kit, code no. OUED; apoa-ii test kit, code no. OQBA). Fasting insulin was measured with an AxSym autoanalyzer (Abbott Diagnostics, Amstelveen, The Netherlands). Insulin 176

180 resistance was assessed based on the homeostasis model assessment for insulin resistance (HOMA-IR) that is calculated by the following formula: [glucose (mmol/l) insulin (mu/l]/ Triglycerides were measured enzymatically. UAC was determined by nephelometry with a threshold of 2.3 mg/l and intra- and inter-assay coefficients of variation of less than 2.2% and less than 2.6%, respectively (Dade Behring Diagnostic, Marburg, Germany). Urinary albumin excretion (UAE) is given as the mean of the two 24-h urine collections. All the technicians were blinded to the participants characteristics. Outcome definition Incident cases of type 2 diabetes mellitus were ascertained as described 4, 34. Briefly, type 2 diabetes was ascertained if one or more of the following criteria were met: 1) fasting plasma glucose 7.0 mmol/l (126 mg/dl); 2) random sample plasma glucose 11.1 mmol/l (200 mg/dl); 3) self-report of a physician diagnosis and 4) use of glucose-lowering medication retrieved from a central pharmacy registry 33, 34. We included cases from 3 months after the baseline screening visits until January Statistical Analyses Continuous data were compared by using Student t tests or Mann Whitney U tests, where applicable. We used χ 2 tests for the comparison of categorical variables between individuals with and without incident type 2 diabetes. Logistic regression analysis was used to examine the associations of HDL-variables, i.e. HDL-C, apoa-i, apoa-ii, HDL-cholesterol/apoA-I and HDL-cholesterol/apoA-II ratios with the risk of developing type 2 diabetes. Odds ratios (ORs) for type 2 diabetes were calculated per SD change for each HDL variable with 95% confidence intervals (CIs). In model 1, basic adjustment was for age and sex. In model 2, we further adjusted for BMI or waist circumference. In model 3, we additionally adjusted for family history of diabetes, hypertension, alcohol use and smoking. In models 4 and 5, we further adjusted for fasting glucose, and triglycerides plus fasting glucose, respectively. Additionally, we also carried out analysis in which we adjusted for HOMA-IR, as covariate. In model A, we adjusted for those variables in model 3 and for HOMA-IR. In model B, we further adjusted for triglycerides. For HOMA-IR and triglycerides, logarithmic transformation with base 2 (log2) was used. Given the expected strong negative relationship of HDL-C with triglycerides, a potential confounder in the associations of interest, we also calculated ORs (95%CI) of HDL variables across each tertile of triglycerides. To this end the associations of diabetes incidence with HDL-C, the HDL-C/apoA-I ratio and the HDL-C/apoA-II ratio were ascertained with the lowest HDL variable being used as the reference category in each triglyceride tertile. Finally, we calculated interaction terms for HDL- C sex in each model, and carried out sex-specific analysis. In view of the enrichment of the PREVEND participants with microalbuminuric subjects 31, we accounted for 24-hour UAE at baseline as another 177

181 potential confounding factor in secondary analysis. Next, we repeated regression models by using a weighted method to compensate for baseline enrichment of the PREVEND participants with high UAC (i.e., 10 mg/l or greater). For most baseline variables, <1% was missing, whereas this was up to 8% for self-reported variables such as family history of diabetes mellitus. A single imputation and predictive mean matching method was applied for missing data. Two-sided P-values < 0.05 were considered statistically significant. All the statistical analyses were carried out using IBM SPSS Statistics 19 and R for Windows ( Results During median (interquartile range [IQR]) follow-up for 7.7 ( ) years, 394 individuals (5.8%) developed new-onset type 2 diabetes mellitus. Baseline clinical and laboratory characteristics of the total cohort, and a comparison of individuals who developed new-onset type 2 diabetes vs. individuals who remained free of diabetes are shown in Table 1. Individuals with incident type 2 diabetes were older, more likely to be male, more obese, more likely to have a family history of diabetes and more likely to have hypertension than those who did not develop diabetes (P<0.001). Levels of fasting glucose, insulin, HOMA-IR, total cholesterol, triglycerides and 24-h UAE were significantly higher in individuals who developed new-onset type 2 diabetes vs. those without incident diabetes. Concentrations of HDL-C, apoa- I, apoa-ii, and ratios of HDL-C/apoA-I and HDL-C/apoA-II were significantly lower in individuals who developed new-onset type 2 diabetes mellitus than in subjects who did not develop diabetes (P<0.001 for all). HDL variables and risk of type 2 diabetes ORs (95% CI) for incident type 2 diabetes per 1-SD increase of HDL-C, apoa-i, apoa- II, and the ratios of HDL-C/apoA-I and HDL-C/apoA-II are shown in Table 2. In ageand sex-adjusted analysis, all HDL-related variables were significantly associated with risk of incident type 2 diabetes (P<0.001), except for apoa-ii (P=0.77) (model 1). Further adjustment for BMI, family history of diabetes, hypertension, alcohol use and smoking did not materially change these associations (models 2 and 3). After further adjustment for baseline fasting glucose and triglycerides, HDL-C, as well as the HDL- C/apoA-I and the HDL-C/apoA-II ratio remained independently associated with risk of incident diabetes (models 4 and 5). In the fully adjusted models, the strongest effect size was observed for the HDL-C/apoA-I and HDL-C/apoA-II ratios (ORs of 0.14 and 0.12, respectively; P<0.001 for each). Furthermore, the directions and the strengths of the relationships were similar in analysis in which we adjusted for waist circumference instead of BMI (Table S1). As shown in Figure 1, we also observed that the risk of diabetes of HDL-C and the ratios of HDL-C/apoA-I and HDL-C/apoA-II did not differ across triglyceride tertiles (P>0.10 for interactions). 178

182 Table 1. Baseline clinical and laboratory characteristics of participants Total Non-cases Incident cases of type 2 diabetes P-value No. of participants no. (%) 6,820 (100) 6,426 (94.2) 394 (5.8) - Male no. (%) 3247 (47.6) 3021 (47.0) 226 (57.4) <0.001 Age- yr 48.8 ± ± ± 10.8 <0.001 Family history of diabetes no. (%) 1339 (19.6) 1196 (18.6) 143 (36.3) <0.001 Smoking no. (%) Current 2296 (33.7) 2162 (33.6) 134 (34.0) Former 2463 (36.1) 2308 (35.9) 155 (39.3) 0.22 Never 2061 (30.2) 1956 (30.4) 105 (26.6) Alcohol use no. (%) >= 4 drinks per day 347 (5.1) 326 (5.1) 21 (5.3) 1-3 drinks per day 1334 (19.6) 1259 (19.6) 75 (19.0) 2-7 drinks per week 2315 (33.9) 2205 (34.3) 110 (27.9) drinks per month 1105 (16.2) 1034 (16.1) 71 (18.0) Almost never 1719 (25.2) 1602 (24.9) 117 (29.7) Systolic blood pressure- mmhg ± ± ± 20.5 <0.001 Diastolic blood pressure- mmhg 71.5 ± ± ± 9.5 <

183 Hypertension no. (%) 1790 (26.2) 1593 (24.8) 197 (50.0) <0.001 BMI- kg/m ± ± ± 4.8 <0.001 Waist circumference- cm 87.8 ± ± ± 12.3 <0.001 Glucose- mmol/l 4.7 ± ± ± 0.7 <0.001 IFG no. (%) 689 (89.9) 467 (7.3) 222 (56.3) <0.001 Insulin - mu/l 7.7 ( ) 7.5 ( ) 8.6 ( ) <0.001 HOMA-IR 1.60 ( ) 1.55 ( ) 3.17 ( ) <0.001 Total cholesterol- mmol/l 5.65 ± ± ± 1.12 <0.001 HDL cholesterol- mmol/l 1.34 ± ± ± 0.29 <0.001 Triglycerides- mmol/l 1.12 ( ) 1.10 ( ) 1.63 ( ) <0.001 Apo A-I, g/l 1.40 ± ± ± 0.24 <0.001 Apo A-II, g/l 0.34 ± ± ±0.06 <0.001 HDL-C/ApoA-I ratio 0.96± ± ±0.16 <0.001 HDL-C/ApoA-II ratio 3.96 ± ± ±0.83 <0.001 UAE - mg/24hour 9.2 ( ) 9.0 ( ) 15.1 ( ) <0.001 Data as mean (±SD) and as median (interquartile ranges). Apo, apolipoprotein, BMI, body mass index; IFG, impaired fasting glucose (fasting glucose of 6.1 to 6.9 mmol/l (110 mg/dl to 125 mg/dl); HOMA-IR, homeostasis model assessment for insulin resistance: HDL, high-density lipoprotein; UAE, urinary albumin excretion. 180

184 Table 2. Relationships of high density lipoprotein (HDL) variables with the risk of developing type 2 diabetes mellitus OR (95% CI) per SD- increase HDL variables Model 1 P- Model 2 P- value value Model 3 P- value Model 4 P- value Model 5 P- value HDL-C 0.44 ( ) < ( ) Apo A-I 0.72 ( ) < ( ) Apo A-II 0.98 ( ) ( ) HDL-C/ApoA-I ratio ( ) HDL-C/ApoA-II ratio ( ) < ( ) < ( ) < ( ) < ( ) < ( ) ( ) ( ) ( ) ( ) ( ) ( ) 0.83 < ( ) < ( ) < ( ) <0.001 < ( ) < ( ) < ( ) <0.001 Apo, apolipoprotein; HDL-C, high density lipoprotein cholesterol Model 1: adjusted for age and sex Model 2: model 1+ body mass index Model 3: model 2+, family history of diabetes, hypertension, alcohol use and smoking Model 4: model 3+ glucose Model 5: model 4 + triglycerides 181

185 ORs for incident type 2 diabetes mellitus per 1-SD increase of HDL-C, apoa-i, apoa-ii, the HDL-C/apoA-I ratio and the HDL-C/apoA-II ratio with HOMA-IR incorporated in model 4 instead of plasma glucose are shown in Table 3. In multivariable adjusted analysis now controlling for insulin resistance, triglycerides and other clinical diabetes risk factors, all results remained essentially similar. 182

186 Figure 1. Risk of developing type 2 diabetes mellitus according to tertiles of high density lipoprotein cholesterol (HDL-C) (panel A), the HDL-C/apolipoprotein (apo) A-I ratio (panel B) and the HDL-C/apoA-II ratio (panel C) stratified by TG tertiles. The OR (95%CI) were calculated by logistic regression models adjusted for age and sex. The individuals in the 1 st. HDL tertile were considered as reference category (P> 0.10 for interactions). Secondary analyses In secondary analyses, we additionally accounted for 24-h UAE as an additional covariate (other included confounders as in model 5, Table 2). We observed that HDL-C and ratios of HDL-C to apolipoproteins were strongly associated with risk of incident type 2 diabetes after multivariable adjustment for 24-h UAE, fasting glucose and triglycerides plus other clinical diabetes risk factors (Table S2). To account for baseline enrichment of the PREVEND cohort with microalbuminuric subjects, we subsequently repeated the analysis in model 5 when we weighted for individuals with a mean UAC above 10 mg/l. The complex design analysis hardly affected the ORs in model 5 (Table S3). Finally, the associations of HDL-C and the ratios of HDL- C to apolipoproteins with the risk of incident type 2 diabetes were not modified by sex (P>0.10 for interactions). In sex-stratified analyses, the multivariable-adjusted ORs (adjusted for baseline fasting glucose, triglycerides and other clinical risk factors) per 1-SD increase of HDL-C were 0.70 ( , P=0.007) and 0.74 ( , P=0.025) in men and women, respectively. 183

187 Table 3. Relationships of high density lipoprotein (HDL) variables with the risk of developing type 2 diabetes mellitus after adjustment for homeostasis model assessment (HOMA-IR) and other clinical factors OR (95% CI) per SD- increase HDL variables Model A P- value Model B P- value HDL-C 0.66 ( ) < ( ) Apo A-I 0.88 ( ) ( ) 0.33 Apo A-II 1.02 ( ) ( ) 0.91 HDL-C/ApoA-I ratio 0.07 ( ) < ( ) HDL-C/ApoA-II ratio 0.07 ( ) < ( ) <0.001 Apo, apolipoprotein, HDL-C, high density lipoprotein cholesterol Model A: Adjustments were for age, sex, BMI, smoking, alcohol use, family history of diabetes, hypertension and HOMA-IR Model B: model A + triglycerides Discussion This prospective study in a predominantly Caucasian population shows that the ageand sex-adjusted relationship of incident type 2 diabetes mellitus with HDLcholesterol is at least in part independent of other metabolic syndrome components, including (central) obesity, hypertension, fasting plasma glucose and triglycerides, as well as of a positive family history of diabetes. Higher HDL-cholesterol levels also predicted a lower risk of diabetes after further adjustment of alcohol consumption and smoking, important environmental factors that govern HDL-cholesterol. More strikingly, incident diabetes was predicted by plasma apoa-i levels as well, in marked contrast with the lack of any independent relationship of diabetes development with plasma apoa-ii. As judged from the respective ORs, the relationship of incident diabetes with apoa-i was weaker than that with HDLcholesterol. As a consequence, we observed robust inverse relationships of incident diabetes with the HDL-cholesterol/apoA-I and the HDL-cholesterol/apoA-II ratio. Of note, the particularly low ORs are partly a consequence of the fact that small changes in ratios can reflect relatively large changes in HDL-cholesterol or apolipoprotein differences. The present data, therefore, suggests that apoa-i rather than apoa-ii may influence diabetes development, and agree with the hypothesis that HDL compositional characteristics may be involved in the pathogenesis of type 2 diabetes mellitus. Besides increasing age and a positive family history of diabetes, (central) obesity, hypertension and higher initial plasma glucose are relevant predictors of incident type 2 diabetes mellitus 3-5. This makes it necessary to take account of these cardio-metabolic risk factors in order to determine the strength of the relationships of incident diabetes with HDL-related variables. Previous reports have already demonstrated that lower HDL-cholesterol relates to increased diabetes risk after adjustment for several metabolic syndrome components in predominantly middleaged Caucasian men 9, 10, Korean subjects 12, and even in native American children and adolescents 11. In some early surveys, the impact of HDL- cholesterol on diabetes 184

188 incidence was found to be confined to women 7, 8. In the PREVEND cohort, we did not find an interaction between sex and HDL cholesterol on incident diabetes, and the point estimates of HDL cholesterol for diabetes risk were similar for men and women separately. In view of the strong negative relationships of HDL-cholesterol and HDL particle size with plasma triglycerides 16, 27, 28, 35, 36, it is also relevant that the impact of the HDL-cholesterol/apoA-I and the HDL-cholesterol/apoA-II ratio on diabetes development was not appreciably modified by plasma triglyceride levels. Although not unequivocally established, the impact of apoa-ii on HDL remodelling, as well as on HDL s functional properties, such as the ability to promote cell-derived cholesterol efflux, is likely to be distinct and less pronounced compared to the multifaceted actions of apoa-i, which are generally considered to be anti-atherogenic 26-28, 37. In this context, it is plausible to propose that the lack of effect of apoa-ii compared to apoa-i on incident diabetes would translate into a differentiated role of these apolipoproteins in their ability to protect against diabetes development. Of further interest, the lowest odds ratios for incident diabetes were found for the HDL cholesterol/apoa-i and the HDL cholesterol/apoa-ii ratio. Higher ratios of HDLcholesterol compared to its major apolipoproteins provide estimates of more cholesterol loaded and, hence, larger HDL particles 16, Therefore, the strong negative relationship of these ratios with diabetes development would raise the possibility that larger HDL particles may particularly predict lower diabetes incidence. In keeping with this hypothesis and in line with the lack of relationship of diabetes development with apoa-ii, HDL particles that contain both apoa-i and apoa-ii, i,e, LpA-I:A-II particles, are smaller than LpA-I particles, that contain apoa-i only 28. Obviously, the possible diabetes-protective impact of various HDL subfractions including pre β-hdl and the contribution of specific HDL-associated proteins, such as the anti-oxidative enzymes, paraoxonase-i and lipoproteinassociated phospholipase A-2 26, 27, 38, needs to be more precisely defined in future human studies. Several methodological aspects and limitations of our study need to be considered. The PREVEND cohort is enriched with microalbuminuric subjects 31. In this cohort, microalbuminuria was shown to predict incident diabetes independently of metabolic syndrome components 39. Nonetheless, the incidence of type 2 diabetes, among PREVEND participants during a median follow-up of 7.7 years with January first 2009 as census date, ascertained according to 1997 ADA criteria, as well as by self-reported physician diagnosis of type 2 diabetes or the use of oral glucose lowering drugs, roughly agrees with current projections of diabetes prevalence in other established market economies 1, 2, 4, 5. Clearly, it is relevant to take account of microalbuminuria upon establishing the relationship of diabetes development with HDL. Additional adjustment for 24-h UAE did not materially change the ORs of any HDL-related variables for incident diabetes. Furthermore, a secondary analysis weighted for the enrichment of the PREVEND cohort with higher urine albumin concentrations 31 revealed comparable diabetes risk estimates for the HDL cholesterol/apoa-i and the HDL cholesterol/apoa-ii ratio. Hence, clinically 185

189 important bias in the interpretation of the current results attributable to overrepresentation of microalbuminuric subjects consequent to the focus of PREVEND on renal disease is very unlikely. For these reasons, we interpret our data to be representative for the general population. In conclusion, our study demonstrates for the first time that new onset type 2 diabetes mellitus is not only predicted by lower HDL cholesterol, but even more strongly by lower HDL-cholesterol/apoA-I and HDL-cholesterol/apoA-II ratios. We suggest that the composition of HDL particles may specifically impact on diabetes development with larger HDL particles conferring lower risk. 186

190 Acknowledgments We thank Prof. dr. L.T.W. de Jong-van den Berg and dr. S.T. Visser from the Department of Social Pharmacy, Pharmacoepidemiology and Pharmacotherapy, Groningen University Institute for Drug Exploration, University of Groningen, University Medical Center Groningen for providing the data on pharmacy-registered use of glucose lowering medication. The technical assistance of J.J. Duker is highly appreciated. We acknowledge Dade Behring (Marburg, Germany) for supplying equipment (Behring Nephelometer II) and analytes for the determination of apolipoproteins and other metabolites. This study is supported by the Netherlands Heart Foundation, the Dutch Diabetes Research Foundation and the Dutch Kidney Foundation. This research was performed within the framework of CTMM, the Center for Translational Molecular Medicine ( project PREDICCt (grant 01C ), and by grants from the Netherlands Heart Foundation (grant ) and the Jan Kornelis de Cock Foundation Groningen, the Netherlands. The authors do not have any conflicts of interest to disclose. 187

191 References 1. Shaw JE, Sicree RA, Zimmet PZ. Global estimates of the prevalence of diabetes for 2010 and Diabetes Res Clin Pract 2010;87(1): Wild S, Roglic G, Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the year 2000 and projections for Diabetes Care 2004;27(5): Abbasi A, Corpeleijn E, Peelen LM, et al. External validation of the KORA S4/F4 prediction models for the risk of developing type 2 diabetes in older adults: the PREVEND study. Eur J Epidemiol 2012;27(1): Abbasi A, Peelen LM, Corpeleijn E, et al. Prediction models for risk of developing type 2 diabetes: systematic literature search and independent external validation study. Bmj 2012;345:e Noble D, Mathur R, Dent T, Meads C, Greenhalgh T. Risk models and scores for type 2 diabetes: systematic review. BMJ 2011;343:d Haffner SM, Stern MP, Hazuda HP, Mitchell BD, Patterson JK. Cardiovascular risk factors in confirmed prediabetic individuals. Does the clock for coronary heart disease start ticking before the onset of clinical diabetes? JAMA 1990;263(21): Njolstad I, Arnesen E, Lund-Larsen PG. Sex differences in risk factors for clinical diabetes mellitus in a general population: a 12-year follow-up of the Finnmark Study. Am J Epidemiol 1998;147(1): Fagot-Campagna A, Knowler WC, Narayan KM, Hanson RL, Saaddine J, Howard BV. HDL cholesterol subfractions and risk of developing type 2 diabetes among Pima Indians. Diabetes Care 1999;22(2): von Eckardstein A, Schulte H, Assmann G. Risk for diabetes mellitus in middle-aged Caucasian male participants of the PROCAM study: implications for the definition of impaired fasting glucose by the American Diabetes Association. Prospective Cardiovascular Munster. J Clin Endocrinol Metab 2000;85(9): Wilson PW, Meigs JB, Sullivan L, Fox CS, Nathan DM, D'Agostino RB, Sr. Prediction of incident diabetes mellitus in middle-aged adults: the Framingham Offspring Study. Arch Intern Med 2007;167(10): Franks PW, Hanson RL, Knowler WC, et al. Childhood predictors of young-onset type 2 diabetes. Diabetes 2007;56(12): Seo MH, Bae JC, Park SE, et al. Association of lipid and lipoprotein profiles with future development of type 2 diabetes in nondiabetic Korean subjects: a 4-year retrospective, longitudinal study. J Clin Endocrinol Metab 2011;96(12):E Lewington S, Whitlock G, Clarke R, et al. Blood cholesterol and vascular mortality by age, sex, and blood pressure: a meta-analysis of individual data from 61 prospective studies with 55,000 vascular deaths. Lancet 2007;370(9602): Di Angelantonio E, Sarwar N, Perry P, et al. Major lipids, apolipoproteins, and risk of vascular disease. JAMA 2009;302(18): Haffner SM. The metabolic syndrome: inflammation, diabetes mellitus, and cardiovascular disease. Am J Cardiol 2006;97(2A):3A-11A. 16. Borggreve SE, De Vries R, Dullaart RP. Alterations in high-density lipoprotein metabolism and reverse cholesterol transport in insulin resistance and type 2 diabetes mellitus: role of lipolytic enzymes, lecithin:cholesterol acyltransferase and lipid transfer proteins. Eur J Clin Invest 2003;33(12): Schmidt MI, Duncan BB. Diabesity: an inflammatory metabolic condition. Clin Chem Lab Med 2003;41(9): Drew BG, Rye KA, Duffy SJ, Barter P, Kingwell BA. The emerging role of HDL in glucose metabolism. Nat Rev Endocrinol 2012;8(4): von Eckardstein A, Sibler RA. Possible contributions of lipoproteins and cholesterol to the pathogenesis of diabetes mellitus type 2. Curr Opin Lipidol 2011;22(1):

192 20. Kruit JK, Brunham LR, Verchere CB, Hayden MR. HDL and LDL cholesterol significantly influence beta-cell function in type 2 diabetes mellitus. Curr Opin Lipidol 2010;21(3): Drew BG, Duffy SJ, Formosa MF, et al. High-density lipoprotein modulates glucose metabolism in patients with type 2 diabetes mellitus. Circulation 2009;119(15): Fryirs MA, Barter PJ, Appavoo M, et al. Effects of high-density lipoproteins on pancreatic betacell insulin secretion. Arterioscler Thromb Vasc Biol 2010;30(8): Abderrahmani A, Niederhauser G, Favre D, et al. Human high-density lipoprotein particles prevent activation of the JNK pathway induced by human oxidised low-density lipoprotein particles in pancreatic beta cells. Diabetologia 2007;50(6): Dullaart RP, Annema W, de Boer JF, Tietge UJ. Pancreatic beta-cell function relates positively to HDL functionality in well-controlled type 2 diabetes mellitus. Atherosclerosis 2012;222(2): Han R, Lai R, Ding Q, et al. Apolipoprotein A-I stimulates AMP-activated protein kinase and improves glucose metabolism. Diabetologia 2007;50(9): Kontush A, Chapman MJ. Functionally defective high-density lipoprotein: a new therapeutic target at the crossroads of dyslipidemia, inflammation, and atherosclerosis. Pharmacol Rev 2006;58(3): Camont L, Chapman MJ, Kontush A. Biological activities of HDL subpopulations and their relevance to cardiovascular disease. Trends Mol Med 2011;17(10): Asztalos BF, Tani M, Schaefer EJ. Metabolic and functional relevance of HDL subspecies. Curr Opin Lipidol 2011;22(3): Corsetti JP, Bakker SJ, Sparks CE, Dullaart RP. Apolipoprotein A-II influences apolipoprotein E-linked cardiovascular disease risk in women with high levels of HDL cholesterol and C-reactive protein. PLoS One 2012;7(6):e Birjmohun RS, Dallinga-Thie GM, Kuivenhoven JA, et al. Apolipoprotein A-II is inversely associated with risk of future coronary artery disease. Circulation 2007;116(18): Lambers Heerspink HJ, Brantsma AH, de Zeeuw D, Bakker SJ, de Jong PE, Gansevoort RT. Albuminuria assessed from first-morning-void urine samples versus 24-hour urine collections as a predictor of cardiovascular morbidity and mortality. Am J Epidemiol 2008;168(8): Monster TB, Janssen WM, de Jong PE, de Jong-van den Berg LT. Pharmacy data in epidemiological studies: an easy to obtain and reliable tool. Pharmacoepidemiol Drug Saf 2002;11(5): Matthews DR, Hosker JP, Rudenski AS, Naylor BA, Treacher DF, Turner RC. Homeostasis model assessment: insulin resistance and beta-cell function from fasting plasma glucose and insulin concentrations in man. Diabetologia 1985;28(7): Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin and risk of type 2 diabetes in the general population. Diabetologia 2011;54(9): Borggreve SE, Hillege HL, Wolffenbuttel BH, et al. The effect of cholesteryl ester transfer protein -629C->A promoter polymorphism on high-density lipoprotein cholesterol is dependent on serum triglycerides. J Clin Endocrinol Metab 2005;90(7): El Harchaoui K, Arsenault BJ, Franssen R, et al. High-density lipoprotein particle size and concentration and coronary risk. Ann Intern Med 2009;150(2): Tailleux A, Duriez P, Fruchart JC, Clavey V. Apolipoprotein A-II, HDL metabolism and atherosclerosis. Atherosclerosis 2002;164(1): Koren-Gluzer M, Aviram M, Meilin E, Hayek T. The antioxidant HDL-associated paraoxonase- 1 (PON1) attenuates diabetes development and stimulates beta-cell insulin release. Atherosclerosis 2011;219(2): Brantsma AH, Bakker SJ, Hillege HL, de Zeeuw D, de Jong PE, Gansevoort RT. Urinary albumin excretion and its relation with C-reactive protein and the metabolic syndrome in the prediction of type 2 diabetes. Diabetes Care 2005;28(10):

193 Table S1. Relationships of high density lipoprotein (HDL) variables with the risk of developing type 2 diabetes mellitus after adjustment for waist circumference and other clinical factors OR (95% CI) per SD- increase HDL variables Model 1 P- value Model 2 P- value Model 3 P- value Model 4 P- value Model 5 P- value HDL-C 0.44 ( ) < ( ) Apo A-I 0.72 ( ) < ( ) Apo A-II 0.98 ( ) ( ) HDL-C/ApoA-I ratio ( ) HDL-C/ApoA-II ratio ( ) < ( ) < ( ) < ( ) ( ) ( ) < ( ) < ( ) Apo, apolipoprotein; HDL-C, high density lipoprotein cholesterol Model 1: adjusted for age and sex Model 2: model 1+ waist circumference Model 3: model 2+, family history of diabetes, hypertension, alcohol use and smoking Model 4: model 3+ glucose Model 5: model 4 + triglycerides < ( ) ( ) ( ) < ( ) < ( ) < ( ) ( ) ( ) 0.78 < ( ) <0.001 < ( ) <

194 Table S2: Association of high density lipoprotein (HDL) variables with the risk of developing type 2 diabetes mellitus after adjustment for 24-h urinary albumin excretion and other clinical factors HDL variables OR (95% CI) P-value HDL-C 0.74 ( ) Apo A-I 0.95 ( ) 0.52 Apo A-II 1.02 ( ) 0.73 HDL-C/ApoA-I ratio 0.15 ( ) HDL-C/ApoA-II ratio 0.12 ( ) <0.001 Apo, apolipoprotein; HDL-C, high density lipoprotein cholesterol. Adjustments were for age, sex, body mass index, family history of diabetes, hypertension, alcohol use, smoking, fasting glucose and 24-h urine albumin excretion Table S3: Association of high density lipoprotein (HDL) variables with the risk of developing type 2 diabetes mellitus by complex design analysis* HDL variables OR (95%CI) P-value HDL-C 0.78 ( ) 0.01 Apo A-I 1.00 ( ) 0.98 Apo A-II 1.15 ( ) 0.06 HDL-C/ApoA-I ratio 0.18 ( ) HDL-C/ApoA-II ratio 0.13 ( ) <0.001 * A complex design analysis was performed to account for the baseline enrichment of the PREVEND population with individuals with a urine albumin concentration of 10 mg/l or greater. Adjustments were for age, sex, body mass index, family history of diabetes, hypertension, alcohol use, smoking and fasting glucose 191

195 192

196 Chapter 10 Plasma procalcitonin and risk of type 2 diabetes in the general population Ali Abbasi 1,2 ; Eva Corpeleijn 1 ; Douwe Postmus 1 ; Ron T. Gansevoort 2 ; Paul E. de Jong 2 ; Rijk O.B. Gans 2 ; Joachim Struck 3 ; Hans Hillege 1 ; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Department of Research, BRAHMS AG (Part of Thermo Fisher Scientific), Hennigsdorf Germany Diabetologia. 2011;54:2463-5

197 To the Editor Chronic low-grade inflammation is a key feature of the pathophysiology of obesity, insulin resistance and type 2 diabetes (T2D) 1. In vitro studies have found that parenchymal cells of tissues, including adipocytes secrete procalcitonin in response to stimulation by activated macrophages 2. Since obesity is associated with increased presence of activated macrophages in adipose tissue, a similar scenario may play a role in vivo 2, 3. In an initial study, we found that procalcitonin is associated with insulin resistance and all components of the metabolic syndrome 3. Our current aim is to test the prospective association of plasma procalcitonin with incident T2D in the general population, and to compare its predictive value with that of high sensitivity C-reactive protein (hs-crp). The study population was obtained from the Prevention of Renal and Vascular End-stage Disease (PREVEND) study, Groningen, the Netherlands. Details of the study design, recruitment, and measurements have been published elsewhere 3, 4. Of 8,592 participants in this cohort, procalcitonin was missing in 677 individuals. Further exclusion was for, 295 individuals with diabetes at baseline and 1,002 individuals with missing data on covariates or follow-up data on development of diabetes. The current analyses were performed on 6,618 non-diabetic participants with complete data. In baseline samples, plasma procalcitonin was measured by a novel commercially available immuno-luminometric assay (B.R.A.H.M.S PCT sensitive LIA, BRAHMS GmbH, Hennigsdorf,Germany). Assays were performed in EDTA-plasma aliquots that had been stored frozen at -80 C, without prior thawing and re-freezing. The intraassay CV was 6% at 0.1 ng/ml and 8% at 0.03 ng/ml. The functional assay sensitivity, defined as the lowest concentration to be determined with an interassay CV of 20% was ng/ml and lowest detection limit ng/ml. The assay technique for procalcitonin and other biomarkers has been described previously 3, 5. Incident T2D was ascertained if 1 or more available conditions were met: selfreport of physician diagnosis; fasting plasma glucose of 7.0 mmol/l; random sample plasma glucose 11.1 mmol/l; use of antidiabetic agents according to a central pharmacy-registration. We applied logistic regression models to test the associations of procalcitonin and hs-crp with incident diabetes. To assess added value of procalcitonin and hs- CRP, we examined improvement of diabetes prediction in terms of C statistic, a measure of discrimination, and integrated discrimination improvement (IDI), a measure of reclassification. A p value of 0.05 or less from two-sided tests was considered statistically significant. The statistical analyses were performed using SPSS 18.0 (SPSS Inc., Chicago, IL) and R version (Vienna,Austria) ( The median (interquartile range[iqr]) of plasma procalcitonin was ( ) ng/ml; men had a higher level than women. Anthropometric and clinical characteristics are summarized in Table S1. Participants with highern 194

198 procalcitonin were older, more obese and had higher hs-crp (correlation coefficient = 0.24; p<0.001) (Figure S1). During mean (SD) follow-up for 7.6 (0.8) years, 385 participants developed T2D. Median (IQR) procalcitonin levels were ( ) and ( ) ng/ml in incident cases and non-cases respectively (p<0.001). Odds ratios (ORs) for incident T2D by quartiles of procalcitonin and hs-crp are presented in Table 1. A graded increase in risk was observed in crude analysis, after adjustment for age, sex, smoking and alcohol intake (model 2), and additional adjustment for hypertension and parental history of diabetes (model 3) (p<0.001). In model 4, with further adjustment for BMI and waist circumference, the OR for incident T2D in the fourth quartile of procalcitonin was 1.74 (95%CI, ) compared with the lowest quartile (p=0.008). The association of hs-crp lost significance with this further adjustment (Table 1). In a secondary analysis, the association of procalcitonin with incident T2D remained significant after adjustment for HDL-cholesterol (OR 1.25; 95%CI, ; p=0.04). C-statistic improved from 0.78 ( ) to 0.79 ( ) (p=0.05) and IDI was (p=0.02) after adding procalcitonin to a clinical prediction model including sex, smoking, waist circumference, hypertension and family history of diabetes 6. We observed no improvement in c statistic (p=0.23) or IDI (p=0.35) when hs-crp was added instead of procalcitonin. We found plasma procalcitonin to be an independent predictor of incident T2D in the general population. Particularly, plasma procalcitonin was more strongly associated with incident T2D than hs-crp after accounting for adiposity. The link between obesity and diabetes is mediated through both low-grade inflammation and non-inflammatory processes 1, 7. In our data, adjustment for adiposity attenuated the association of procalcitonin with T2D risk. This supports obesity partly contributes to this association. Another part might be explained by adipocyte dysfunction, other inflammatory conditions or lipid markers rather than adipose tissue mass 3. In line with prior evidence 7, the association between hs-crp and T2D materially lost significance after accounting for adiposity. This suggests that procalcitonin as a pro-inflammatory predictor of T2D may be more independent of obesity than hs-crp. Our findings suggest that the calcitonin-related system may play a role in the pathophysiology of diabetes. Experimental studies have demonstrated biological activity of procalcitonin on calcitonin receptor family complexes, affecting vascular tone, insulin sensitivity and insulin secretion by the pancreatic beta-cells 2, 8. Some limitations of this study should be noted. While our study only recruited Caucasians in the Netherlands, it is unclear if our findings would be replicable in others. Another limitation was for the excluded individuals with missing data or without confirmed fasting blood sampling. However, only numerically small differences in baseline characteristics were found between those who included in the study and the excluded individuals. Moreover, we had no data on other inflammatory markers such as interleukin-6 or new diabetes risk factors such as 195

199 Table 1. Odds ratios for incident type 2 diabetes by quartiles of plasma procalcitonin and hs-crp (n= 6,618) Incidence rate, per 1000 person-years Crude analysis Model 1 Model 2 Model 3 Model 4 1 ( 0.012) Quartiles of plasma procalcitonin, ng/ml, a OR (95% CI) Per Unit Increase in Log b 2 ( ) 3 ( ) 4 ( 0.020) ( ) 1.70 ( ) 1.68 ( ) 1.64 ( ) 1.45 ( ) 2.41 ( ) 1.67 ( ) 1.64 ( ) 1.58 ( ) 1.18 ( ) 4.80 ( ) 2.91 ( ) ( ) 2.59 ( ) 1.74 ( ) P value for Trend <0.001 <0.001 <0.001 < ( ) 1.50 ( ) 1.50 ( ) 1.45 ( ) 1.32 ( ) P value <0.001 <0.001 <0.001 < Incidence rate, per 1000 person-years Crude analysis c Model 1 Model 2 Model 3 Model 4 1 ( 0.54) Quartiles of hs-crp, mg/l, OR (95% CI) 2 ( ) 3 ( ) 4 ( 2.77) ( ) 1.83 ( ) 1.79 ( ) 1.66 ( ) 1.21 ( ) 3.79 ( ) 2.79 ( ) 2.68 ( ) 2.38 ( ) 1.44 ( ) 4.83 ( ) 3.55 ( ) 3.33 ( ) 2.91 ( ) 1.42 ( ) Abbreviations: hs-crp, high-sensitivity C-reactive protein; OR, odds ratio; CI, confidence interval. a Plasma procalcitonin levels were measured within low range of <0.1 ng/ml. b Odds ratios expressed per unit increase in log2-transformed level of procalcitonin and hs-crp. c Analyses based on a sample of participants with data on hs-crp (n= 6,393[ 361 incident cases]). Model 1 is adjusted for age and sex. Model 2 is adjusted for variables in model 1 plus alcohol use and smoking status. Model 3 is adjusted for variables in model 2 plus hypertension and parental history of diabetes. Model 4 is adjusted for variables in model 3 plus body mass index and waist circumference <0.001 <0.001 <0.001 < ( ) 1.30 ( ) 1.28 ( ) 1.25 ( ) 1.06 ( ) <0.001 <0.001 <0.001 <

200 gamma glutamyl transferase to compare their predictive value for T2D with that of procalcitonin. In conclusion, plasma procalcitonin levels are associated with incident T2D independent of common diabetes risk factors. Our findings may be considered an opening for further studies on a potential role of the calcitonin-related system in the pathophysiology of diabetes. 197

201 Acknowledgments This work was supported by the Netherlands Heart Foundation, Dutch Diabetes Research Foundation and Dutch Kidney Foundation. This research was performed within the framework of CTMM, the Center for Translational Molecular Medicine ( project PREDICCt (grant 01C ). Duality of interest statement: Dr Struck is an employee of BRAHMS GmbH, a company which manufactures the procalcitonin assay and holds patent rights on procalcitonin. The present study was not financed by BRAHMS GmbH. No other author has anything to declare. None of the study sponsors had a role in study design; in data collection, analysis, or interpretation; in writing the report; or in the decision to submit for publication. 198

202 References 1. Wellen KE, Hotamisligil GS. Inflammation, stress, and diabetes. J Clin Invest 2005;115(5): Becker KL, Nylen ES, White JC, Muller B, Snider RH, Jr. Clinical review 167: Procalcitonin and the calcitonin gene family of peptides in inflammation, infection, and sepsis: a journey from calcitonin back to its precursors. J Clin Endocrinol Metab 2004;89(4): Abbasi A, Corpeleijn E, Postmus D, et al. Plasma procalcitonin is associated with obesity, insulin resistance, and the metabolic syndrome. J Clin Endocrinol Metab 2010;95(9):E Lambers Heerspink HJ, Brantsma AH, de Zeeuw D, Bakker SJ, de Jong PE, Gansevoort RT. Albuminuria assessed from first-morning-void urine samples versus 24-hour urine collections as a predictor of cardiovascular morbidity and mortality. Am J Epidemiol 2008;168(8): Morgenthaler NG, Struck J, Fischer-Schulz C, Seidel-Mueller E, Beier W, Bergmann A. Detection of procalcitonin (PCT) in healthy controls and patients with local infection by a sensitive ILMA. Clin Lab 2002;48(5-6): Balkau B, Lange C, Fezeu L, et al. Predicting diabetes: clinical, biological, and genetic approaches: data from the Epidemiological Study on the Insulin Resistance Syndrome (DESIR). Diabetes Care 2008;31(10): Lee CC, Adler AI, Sandhu MS, et al. Association of C-reactive protein with type 2 diabetes: prospective analysis and meta-analysis. Diabetologia 2009;52(6): Martinez A, Kapas S, Miller MJ, Ward Y, Cuttitta F. Coexpression of receptors for adrenomedullin, calcitonin gene-related peptide, and amylin in pancreatic beta-cells. Endocrinology 2000;141(1):

203 Appendix Table S1. Baseline characteristics of participants for the whole population and according to quartiles of plasma procalcitonin Characteristic Whole population Procalcitonin Quartiles, ng/ml a,b ( 0.012) ( ) ( ) ( 0.020) No. of participants Age, y 48.3± ± ± ± ±12.8 Men, no. (%) 3154 (47.8) 284 (22.2) 687 (36.1) 1006 (56.6) 1195 (72.1) Weight, kg 77.7± ± ± ± ±14.1 Height, cm 173.1± ± ± ± ±9.3 BMI, kg/m ± ± ± ± ±4.2 Waist circumference, cm 87.6± ± ± ± ±12.1 Systolic blood pressure, mmhg 123.4± ± ± ± ±19.5 Diastolic blood pressure, mmhg 71.4± ± ± ± ±9.7 Total Cholesterol, mmol/l 5.6± ± ± ± ±1.1 HDL cholesterol, mmol/l 1.3± ± ± ± ±0.3 Triglyceride, mmol/l 1.1 ( ) 0.9 ( ) 1.0 ( ) 1.2 ( ) 1.4 ( ) Glucose, mmol/l 4.7± ± ± ± ±0.7 Insulin, pmol/l 46.8 ( ) 39.0 ( ) 43.8 ( ) 49.8 ( ) 60.0 ( ) HOMA-IR 1.6 ( ) 1.3 ( ) 1.5 ( ) 1.7 ( ) 2.2 ( ) Tobacco smoking, no. (%) Never Ex-smoker Current smoker Alcohol use, no. (%) Never 1 to 4 drinks per month 2 to 7 drinks per week 1 to 3 drinks per day 4 drinks per day 1979 (30.0) 2384 (36.1) 2239 (33.9) 1611 (24.4) 1024 (15.5) 2306 (35.0) 1326 (20.1) 330 (5.0) 419 (32.7) 439 (34.2) 424 (33.1) 298 (23.2) 229 (17.9) 467 (36.4) 247 (19.3) 41 (3.2) 600 (31.5) 626 (32.9) 676 (35.5) 453 (23.8) 299 (15.7) 704 (37.4) 358 (18.8) 88 (4.6) 521 (29.3) 646 (36.4) 609 (34.3) 461 (26.0) 253 (14.2) 609 (34.3) 364 (20.5) 89 (5.0) 449 (27.0) 679 (41.0) 530 (32.0) 404 (24.4) 245 (14.8) 532 (32.1) 361 (21.8) 116 (7.0) Urine albumin excretion, mg/24h 9.2 ( ) 8.1 ( ) 8.6 ( ) 9.2 ( ) 11.2 ( ) hs-crp, mg/l 1.2 ( ) 0.8 ( ) 1.0 ( ) 1.2 ( ) 1.8 ( ) Abbreviations: BMI, body mass index; HDL, high-density lipoprotein; HOMA-IR, homeostasis model assessment for insulin resistance, hs-crp, high sensitivity C-reactive protein. Data are percent, mean±sd or median (interquartile range). 200

204 a All P values were <0.001 for the comparison across procalcitonin quartiles using χ 2 test (categorical data), spearman rank correlation (ordinal data), and ANOVA or Kruskal-Wallis (continuous data). b From included sample, 6,523 to 6,606 participants had complete data on other biomarkers. Figure S1. Association between log2 high-sensitivity C-reactive protein and log2 procalcitonin, (β= 0.24, P <0.001) 201

205 202

206 Chapter 11 Peroxiredoxin 4, a novel circulating biomarker for oxidative stress and the risk of incident cardiovascular disease and all-cause mortality Ali Abbasi 1,2,3 ; Eva Corpeleijn 1 ; Douwe Postmus 1 ; Ron T. Gansevoort 2 ; Paul E. de Jong 2 ; Rijk O.B. Gans 2 ; Joachim Struck 3 ; Janin Schulte 4 ; Hans Hillege 1 ; Pim van der Harst 4 ; Linda M. Peelen 3 ; Joline W.J. Beulens 3 ; Ronald P. Stolk 1 ; Gerjan Navis 2 ; Stephan J.L. Bakker 2 1 Department of Epidemiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 2 Department of Internal Medicine, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands 3 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands 4 Department of Research, Thermo Fisher Scientific / BRAHMS GmbH, Hennigsdorf Germany 5 Department of Cardiology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands J Am Heart Assoc. 2012;1(5):e002956

207 Abstract Background Oxidative stress has been suggested to play a key role in the development of cardiovascular disease (CVD). The aim of our study was to investigate the associations of serum peroxiredoxin 4 (Prx4), a hydrogen peroxide degrading peroxidase, with incident CVD and all-cause mortality. We subsequently examined the incremental value of Prx4 for the risk prediction of CVD compared with the Framingham risk score (FRS). Methods We performed Cox regression analyses in 8,141 participants without history of CVD (aged years; women 52.6%) from the Prevention of Renal and Vascular End-stage Disease (PREVEND) study in Groningen, the Netherlands. Serum Prx4 was measured by an immunoluminometric assay in baseline samples. Main outcomes were: 1) incident CVD events or CVD mortality; and 2) all-cause mortality during a median follow-up of 10.5 years. Results In total, 708 (7.8%) participants developed CVD events or CVD mortality, and 517 (6.3%) participants died. Baseline serum Prx4 levels were significantly higher in participants with incident CVD events or CVD mortality and in those who died than in participants who remained free of outcomes (both P<0.001). In multivariable models with adjustment for Framingham risk factors, hazard ratios were 1.16 (95%CI, , P<0.001) for incident CVD events or CVD mortality, and 1.17 (95%CI, , P=0.003) for all-cause mortality per doubling of Prx4 levels. After addition of Prx4 to the FRS, the net reclassification improvement was 2.7% (P=0.01) using 10-year risk categories of CVD. Conclusions Elevated serum Prx4 levels are associated with significantly higher risk of incident CVD events or CVD mortality, and all-cause mortality after adjustment for clinical risk factors. Addition of Prx4 to the FRS marginally improved risk prediction of future CVD. 204

208 Introduction Experimental and clinical studies suggest that oxidative stress plays a key role in the pathogenesis of cardiovascular disease (CVD) 1-4. Oxidative stress status is usually defined as overproduction of reactive oxygen/nitrogen species in imbalance with endogenous antioxidant defences, which in turn result in increased oxidative damage 5. Several biomarkers, including target oxidation products and antioxidants have been proposed for assessment of the level of oxidative stress, but clinical data examining association between marker(s) and CVD independent of common risk factors are limited 5-7. Recently, peroxiredoxin 4 (Prx4), which is a secretable and stable isoform of the Prx family of antioxidant peroxidases 8, has been found in the circulation of humans. Prx4 can be precisely measured by a validated immunoassay 9. Previous evidence showed an abundant cellular antioxidant activity of Prx4 and other Prx isoforms in all mammals protecting against oxidative stress So far, a limited number of smallscale studies have evaluated association of the serum Prx4 with clinical data 9, 14. In these studies, serum levels of Prx4 were increased in septic patients when compared to that of healthy individuals and were positively associated with well-established inflammatory markers like procalcitonin, C-reactive protein (CRP) and interleukin 6 (IL-6) 9, 14. Recently, a study in patients presenting to emergency departments showed the incremental prognostic value of Prxr4 to predict 30 day survival beyond usual risk predictors 15. We aimed to investigate whether serum Prx4 is a predictor of CVD and allcause mortality. For this study, we used data of a large scale, observational cohort of general population and examined the association of Prx4 with incident CVD events or CVD mortality, and all-cause mortality. Since a major clinical application of a biomarker lies within risk stratification and guided preventive strategies 16-19, we also evaluated the incremental predictive value of Prx4 above the Framingham risk score for the 10-year risk of CVD. Methods Study population and design The study population was obtained from the Prevention of Renal and Vascular Endstage Disease (PREVEND) study, a Dutch cohort drawn from the general population (age ranged between 28 and 75 years) of the city of Groningen, the Netherlands between 1997 and We have reported details of the study design and recruitment of participants elsewhere 20, 21. Briefly, 40,856 individuals (47.8%) completed a questionnaire on demographics, history of cardiovascular and metabolic outcomes, medication use and pregnancy prior to their first visit, collecting early morning urine sample in a vial to measure urinary albumin concentration. Those who were unable or unwilling to participate, individuals using insulin and pregnant women were 205

209 excluded. The baseline PREVEND were recruited from a total of 6,000 individuals with a urinary albumin concentration of 10 mg/l or greater and a random control sample of individuals with a urinary albumin concentration of less than 10 mg/l (n=2,592). In the baseline cohort, serum Prx4 assay was missing for 370 participants, leaving 8,222 for the baseline cross-sectional analyses. The PREVEND study was approved by the local medical ethics committee, University Medical Center Groningen, and conformed to the principles outlined in the Declaration of Helsinki. All participants gave written informed consent. Clinical and biomarker measurements In the baseline screening, study participants underwent two outpatient visits to assess demographics, anthropometric measurements, cardiovascular and metabolic risk factors, health behaviours, and medical family history and to collect two 24-hour urine samples on 2 consecutive days. Furthermore, information on medication use was substantiated with use of pharmacy-based data from all community pharmacies in the city of Groningen 22. Smoking and alcohol use were based on self-reports. Hypertension was defined based on self-report of diagnosis by a physician, measured hypertension ( 140/90 mmhg systolic/diastolic blood pressure) or the use of blood pressure-lowering agents. Prevalent cases of type 2 diabetes were ascertained if one or more of the following criteria were met: 1) a fasting plasma glucose of 7.0 mmol/l (126 mg/dl) or random sample plasma glucose of 11.1 mmol/l (200 mg/dl); or 2) self-report of diagnosis by a physician; or 3) use of glucose-lowering agents according to a central pharmacy registration 23. Prevalent CVD was defined based on self-reported physician diagnosis of cardiac, cerebral and peripheral events by a physician. Kidney disease was defined based on prior history of kidney disease requiring dialysis or estimated glomerular filtration rate (egfr) below 60 ml/min/1.73m 2. We used the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation to calculate egfr 24. In all participants, blood sample measurements for biomarkers were taken after an overnight fast. Serum Prx4 level was measured retrospectively in analogously stored baseline serum samples by a novel immunoluminometric assay, which was described previously 9. The functional assay sensitivity (interassay coefficient of variation <20%) was 0.51 U/L. The intraassay coefficient of variation was <8% throughout range of Prx4 levels 9, 14. Insulin was measured with an AxSym autoanalyzer (Abbott Diagnostics, Amstelveen, The Netherlands). Details on assays for total cholesterol, HDL-cholesterol, triglycerides, hs-crp and procalcitonin have been described previously 25. These baseline assays were performed in EDTA-plasma aliquots that had been stored frozen at -80 C without previous thawing and refreezing. 24-hour urinary albumin excretion (UAE) - given as the mean of the two 24-hour urine excretions - was measured by nephelometry with a threshold of 2.3 mg/l and intra- and inter-assay coefficients of variation of less than 2.2% and less 206

210 than 2.6%, respectively (Dade Behring Diagnostic, Marburg, Germany). All technicians were blinded to the participants characteristics. Outcome definition In prospective data, we ascertained the main outcomes as following: 1) incident CVD events, 2) incident CVD events or CVD mortality; and 3) all-cause mortality (up to January ). Information (on hospitalization) for incident CVD events was obtained from PRISMANT, the Dutch national registry of hospital discharge diagnoses. The validity of this database has been shown to be good, with 84% of primary diagnoses and 87% of secondary diagnoses matching the diagnoses recorded in patients charts 26. Data were coded according to the International Classification of Diseases (ICD), 9th revision and the classification of interventions. The incident CVD events were classified as acute myocardial infarction (ICD-code 410), acute and subacute ischemic heart disease (411), occlusion or stenosis of the precerebral (433) or cerebral arteries (434) and procedures including coronary artery bypass grafting or percutaneous transluminal coronary angioplasty, and other vascular interventions namely percutaneous transluminal angioplasty or bypass grafting of aorta and peripheral vessels. Data on mortality were obtained through the municipal registration. Cause of mortality was ascertained by linking the number of the mortality certificate to the primary cause of mortality as coded by the Dutch Central Bureau of Statistics. Survival time was defined as the period from the baseline to the date of first incident CVD events, CVD mortality, date of death or 1 January In case a person had moved to an unknown destination, the date on which the person was removed from the municipal registry was used as censor date 27. Statistical Analyses Data awere shown as mean± standard deviation (SD) or median (quartiles 1 and 3 [Q1-Q3]) for continuous variables which were compared by one-way ANOVA or Kruskal-Wallis tests as appropriate. Frequency was used for categorical variables which were compared by χ 2 test across tertiles of Prx4. We calculated Spearman correlation coefficients of Prx4 with age, systolic blood pressure, body mass index (BMI), waist circumference, glucose, total Cholesterol, HDL cholesterol, triglyceride, hs-crp, procalcitonin and 24-hour UAE. We used backward-elimination regression models to examine which of the clinical variables were independently associated with Prx4 as a dependent variable. The distribution of Prx4 was highly skewed. To normalise the distribution, we performed logarithmic transformation of the values of Prx4 prior to analyses. We used the logarithm base 2 (log2) to allow for interpretation of results per doubling of Prx4. Interpretation of results expressed per doubling of Prx4 seem more meaningful than interpretation per factor 10 change or per factor e change, which would have been the case respectively if transformation according to base 10 or transformation according to the natural logarithm would have been applied. We used 207

211 Cox proportional-hazards regression in crude and multivariable-adjusted models to examine the associations of Prx4 with incident CVD and all-cause mortality. We adjusted for age and sex in model 1. In model 2, we adjusted for the Framingham risk factors including age, sex, smoking, systolic blood pressure, use of antihypertensive therapy, diabetes at baseline, total cholesterol, and high-density lipoprotein (HDL)- cholesterol 28. We tested the assumptions of proportional-hazards for Cox regression models by Shoenfeld s global tests. Finally, in stepwise adjustments, we included alcohol use, triglycerides, high sensitivity (hs)-crp and 24-hour urine albumin excretion (UAE) to model 2. To assess incremental value of Prx4 for the risk prediction of CVD, we examined improvement of prediction of CVD as compared to the Framingham risk score. To do so we calculated 10-year general CVD risk based on the Framingham Risk Score 28 and based on a model with the Framingham Risk Score and log2 Prx4. Subsequently the models were compared in terms of the following measures, taking into account the time-to-event nature of the data 18, 19, 29, 30 : 1) Harrell s C-statistic for the Cox proportional-hazards regression to quantify the discrimination performance of the models (ability to distinguish between individuals with and without outcome); 2) net reclassification improvement (NRI) to examine if individuals with and without outcome were correctly reclassified, (using the threshold values of <6%, 6 to 20% and 20% for categories of low, medium and high risk, respectively 28, 31 ); and 3) integrated discrimination improvement (IDI), a continuous measure of reclassification. Of the baseline sample of 8,592 participants, 451 had prior history of CVD. To do the prospective analyses, we first excluded these prevalent cases of CVD, leaving 8,141 participants. For most baseline variables, <1% was missing; however, this was up to 8% for self-reported variables. We performed a single imputation with predictive mean matching for missing data. This method can be used for skewed data with less than 10% missingness, because it produce less biased estimates for nonlinear model and imputations are in the metric of the observed data 32, 33. Moreover, a weighted method was performed to compensate the baseline enrichment for the PREVEND participants with urinary albumin concentration (UAC) 10 mg/l. Given the frequency of individuals with UAC 10 mg/l (24.4%) in our general population 20, 21,27, we calculated the weight by sampling fractions. Those with UAC 10 mg/l had weight equal to 0.35 and those with UAC <10 mg/l weight equal to Subsequently, we performed secondary analyses to take into account residual confounding. To do this, we incorporated other covariates which might be confounding of the association between Prx4 and the risk of incident CVD in combination with the Framingham risk factors (model 2). First, we further adjusted for BMI or waist circumference in separate models. Second, we adjusted for family history of CVD and examined the effects of kidney disease on this association. Third, we calculated the metabolic syndrome which was defined according to the National Cholesterol Education Program s Adult Treatment Panel III report criteria 25. And then we adjusted for the metabolic syndrome or insulin in combination with variables 208

212 in model 2. Next, we examined the association of Prx4 with each component of CVD events or CVD mortality including myocardial infarction, cerebrovascular disease and cardiovascular mortality. In addition, we performed another analysis including those who had prior history of CVD. In total population, we further adjusted history of CVD in combination with variables in model 2. And then, we performed a similar analysis in participants with prior history of prior CVD. We used Cox proportionalhazards regression with fractional polynomials to search for the best fitting functional form of Prx4 in the model for incident CVD (model 2). Given the number of each event, we had 80% power at a 0.05 significance level to detect a HR equal to 1.29 for myocardial infarction, 1.25 for cerebrovascular disease and 1.23 for CVD mortality. All the statistical analyses were carried out using IBM SPSS 19.0 and R version (Vienna, Austria) ( cran.r-project.org/). Results Baseline characteristics Baseline clinical characteristics of total population and corresponding tertiles of serum Prx4 are summarized in Table 1. Median (Q1-Q3) Prx4 levels were 0.71 ( ) U/L in men and 0.66 ( ) U/L in women (P<0.001). Across tertiles of Prx4, those who had higher Prx4 levels were older, more obese, less frequent alcohol drinker and more likely to have a history of CVD, hypertension and prevalent diabetes. Prx4 levels were positively correlated with age, BMI, waist circumference, systolic blood pressure, glucose, triglyceride, insulin, 24-hour UAE and inflammatory biomarkers of hs-crp and procalcitonin, and inversely correlated with HDLcholesterol (p<0.001 for all correlations, Table 2). In backward-elimination regression model, log2 Prx4 were positively associated with age (β=0.006, P<0.001), prior history of CVD (β=0.203, P=0.001), triglyceride (β=0.048, P<0.001), log2 hs-crp (β=0.102, P<0.001), log2 UAE (β=0.050, P<0.001) and log2 insulin (β=0.102, P<0.001) and inversely associated with female sex (β= , P=0.03), alcohol use (β= , P<0.001) and total cholesterol (β= , P<0.001 ) (Table 3). Incident CVD and all-cause mortality During median (Q1-Q3) follow-up of 10.5 ( ) years, 608 participants (7.5%) developed incident CVD events, 708 participants (8.7%) developed incident CVD events or CVD mortality, and 517 (6.3%) participants died (of which 135 from cardiovascular causes). Median (Q1-Q3) Prx4 levels were 0.88 ( ) U/L and 0.85 ( ) U/L in participants who developed incident CVD events or CVD mortality and all-cause mortality, respectively. In Figure 1, the cumulative survival to incident CVD events and to incident CVD events or CVD mortality is shown based on tertile groups. The crude cumulative incidence rates (per 1,000 person-year), and hazard ratios (HR) (95% confidence interval [CI]) in crude and multivariable-adjusted 209

213 Table 1. Baseline characteristics of study participants for total population's corresponding tertiles of serum peroxiredoxin 4 * Total Tertiles Characteristic 8,222 2,730 2,671 2,821 Peroxiredoxin 4 (U/l) 0.69 ( ) 0.37 ( ) 0.68 ( ) 1.38 ( ) Age yr 49.2± ± ± ±13.2 Male sex no.(%) 4107 (50.0) 1276 (46.7) 1325 (49.6) 1506 (53.4) Prior history of CVD no.(%) 431 (5.4) 80 (3.0) 116 (4.5) 235 (8.7) Family history of CVD no.(%) 3,817 (50.2) 1,292 (50.7) 1,249 (50.3) 1,276 (49.7) Current smoker no.(%) 2787 (34.0) 1071 (39.4) 897 (33.7) 819 (29.1) Ex-smoker no.(%) 2984 (36.4) 886 (32.6) 954 (35.9) 1144 (40.7) Alcohol drinker no.(%) 6110 (74.7) 2121 (78.0) 2021 (75.9) 1968 (70.2) BMI kg/m ± ± ± ±4.5 Waist circumference cm 88.5± ± ± ±13.8 Systolic blood pressure mmhg 124.5± ± ± ±21.1 Diastolic blood pressure mmhg 71.8± ± ± ±10.2 Antihypertensive therapy no.(%) 1282 (15.6) 294 (10.8) 396 (14.8) 592 (21.0) Prevalent diabetes no.(%) 315 (4.0) 49 (1.9) 96 (3.7) 170 (6.2) Fasting glucose mmol/l 4.9± ± ± ±1.4 Total cholesterol mmol/l 5.64± ± ± ±1.16 HDL cholesterol mmol/l 1.32± ± ± ±0.40 Triglyceride mmol/l 1.16 ( ) 1.11 ( ) 1.13 ( ) 1.26 ( ) Metabolic syndrome no.(%) 1,378 (18.2) 314 (12.3) 444 (17.9) 629 (24.5) Fasting insulin mu/l 8.0 ( ) 7.2 ( ) 8.0 ( ) 9.1 ( ) hs-crp mg/l 1.27 ( ) 0.93 ( ) 1.22 ( ) 1.85 ( ) Procalcitonin ng/ml ( ) ( ) ( ) ( ) UAE mg/24h 9.45 ( ) 8.76 ( ) 9.11 ( ) ( ) * Data are mean (±SD) and median (quartiles 1 and 3) for continuous variables and percentage for categorical variables in complete baseline data set. For clinical variables, up to 1.2% was missing. For self-reported data, % was missing. For biomarkers, % was missing. BMI denotes body mass index which is the weight in kilogram divided by the square of the height in meters. CVD denotes cardiovascular disease, BMI body mass index which is the weight in kilogram divided by the square of the height in meters, HDL high-density lipoprotein, UAE urine albumin excretion and hs-crp high sensitivity C-reactive protein. The metabolic syndrome was defined according to the National Cholesterol Education Program s Adult Treatment Panel III report criteria. P <0.001for the comparison among all peroxiredoxin 4 tertile group, except for total cholesterol (P=0.005) and family history of CVD (P=0.77). 210

214 models for the risk of developing incident CVD events, incident CVD events or CVD mortality, and all-cause mortality are shown in Table 4. Age-and sex-adjusted HR (95% CI) were ranging from 1.32 ( ) for incident CVD events to 1.40 ( ) for all-cause mortality when compared the highest tertile to the first tertile of Prx4 (p for trend <0.001). In a model adjusted for the Framingham risk factors, Prx4 was significantly associated with the increased risk of incident CVD and all-cause mortality. The proportional-hazards assumptions were met for all models. Table 2. Spearman correlation coefficients of serum peroxiredoxin 4 with baseline variables* Variables correlation coefficients (95%CI) Age (0.131 to 0.177) Systolic blood pressure (0.117 to 0.164) Body mass index (0.128 to 0.170) Waist circumference (0.146 to 0.187) Glucose (0.088 to 0.135) Total cholesterol 0.020(0.000 to 0.044) HDL-cholesterol ( to ) Triglyceride (0.080 to 0.133) Insulin ( ) hs-crp (0.205 to 0.249) Procalcitonin (0.107 to 0.153) UAE (0.104 to 0.150) BMI denotes body mass index which is the weight in kilogram divided by the square of the height in meters, HDL high-density lipoprotein, hs-crp high sensitivity C-reactive protein and UAE urine albumin excretion. * Data were available for 7,638 to 8,222 participants. We used bootstrapping method to calculate 95% confidence interval (CI). P values were <0.001 for all correlations, except for total cholesterol (p=0.096). In further models, stepwise adjustment for alcohol use, triglyceride, hs-crp and 24-hour UAE minimally attenuated the associations of Prx4 with incident CVD events or CVD mortality. This was comparable to calculation of HR per doubling Prx4 levels for each outcome. We observed a 15% increased risk of incident CVD events independent of the Framingham risk factors per doubling Prx4 levels (HR, 1.15; 95% CI, 1.05 to 1.26). This was a 16% and a 17% increase for incident CVD events or CVD mortality (HR, 1.16; 95% CI, 1.06 to 1.27), and all-cause mortality (HR, 1.17; 95% CI, 1.06 to 1.29), respectively (Table 4). In subsequent analyses, the associations of Prx4 with the risk of either incident CVD or all-cause mortality were similar for men and women (data not shown). To assess the incremental predictive value of Prx4 for the risk of CVD, we added Prx4 to the Framingham Risk Score as continuous variable 28. In our data set, the Framingham risk score had a C-statistic (95% CI) of 0.80 ( ) for the 10-year risk of CVD. Addition of Prx4 improved modestly C-statistic to 0.81 ( ) (P=0.02), and led to IDI of (P<0.001) and NRI of 2.7% (95%CI, 0.7% to 4.7%; P=0.01) (Table 5). In patients without incident CVD events or CVD mortality, use of Prx4 reclassified 1% and 4% of participants in lower and higher risk categories (<6% 211

215 and above 20%), respectively. In patients with incident CVD events or CVD mortality, use of Prx4 reclassified 2% and 4% of participants in lower and higher risk categories, respectively. Cumulative survival to incident CVD events or CVD mortality A 1st tertile 2nd tertile 3rd tertile B Years Cumulative survival to incident CVD events st tertile 2nd tertile 3rd tertile Years Figure 1. The cumulative probability of incident CVD events (Panel A) and incident CVD events or CVD mortality (Panel B) is shown by tertiles of Prx4. In overall, the log-rank tests were significant for all outcomes according to the tertiles of serum peroxiredoxin 4 (P<0.001). 212

216 Table 3. Association of baseline variables with serum peroxiredoxin 4 as dependent variable* Unadjusted Age-and sex-adjusted Multivariable-adjusted β coefficients (SE) P value β coefficients (SE) P value β coefficients (SE) P value Age, per increase of 1 year (0.001) < (0.001) < (0.001) <0.001 Sex, female vs. male (0.022) < (0.022) (0.027) 0.03 Prior history of CVD, yes vs. no (0.049) < (0.050) < (0.052) Smoking, yes vs. no (0.013) (0.013) 0.67 Alcohol use, yes vs. no (0.025) < (0.025) < (0.026) BMI, per increase of 1 kg/m (0.003) < (0.003) <0.001 Waist circumference, per increase of 1 cm (0.001) < (0.001) < (0.001) Systolic blood pressure, per increase of 1 mmhg (0.001) < (0.001) < (0.001) Antihypertensive therapy, yes vs. no (0.030) < (0.023) < (0.035) 0.10 Prevalent diabetes, yes vs. no (0.057) < (0.058) <0.001 Total cholesterol, per increase of 1 mmol/l (0.035) (0.010) (0.011) <0.001 HDL cholesterol, per increase of 1 mmol/l (0.027) < (0.029) <0.001 Triglyceride, per increase of 1 mmol/l (0.011) < (0.011) < (0.013) <0.001 Insulin, per increase of log2-unit (0.013) < (0.013) < (0.017) <0.001 hs-crp, per increase of log2-unit (0.023) < (0.007) < (0.008) <0.001 Procalcitonin, per increase of log2-unit (0.007) < (0.024) < (0.026) hour UAE, per increase of log2-unit (0.009) < (0.009) < (0.010) <0.001 BMI denotes body mass index which is the weight in kilogram divided by the square of the height in meters, SE standard error, HDL high-density lipoprotein, hs-crp high sensitivity C-reactive protein and UAE urine albumin excretion. * Base-two logarithmically-transformed serum level of peroxiredoxin 4 was considered as dependent variable. Backward selection was used when dropped non-significant variables (probability for removal was 0.10 by F test). 213

217 Table 4. Association of serum peroxiredoxin 4 with incident CVD events, incident CVD events or CVD mortality, and all-cause mortality (n=8,141)* HR (95% CI) or No. of Cases (Incidence Rate) According to Tertiles of Prx HR (95% CI) Per Log2 Unit Increase* Incident CVD events No. of cases, per 1000 person-years 151 (5.8) 204 (7.8) 308 (12.3) Unadjusted analysis (0.95 to 1.48) 1.73 (1.40 to 2.13) 1.32 (1.21 to 1.44) <0.00 Model (0.84 to 1.30) 1.32 (1.07 to 1.63) 1.21 (1.10 to 1.33) <0.00 Model (0.84 to 1.32) 1.26 (1.01 to 1.57) 1.15 (1.05 to 1.26) Model 2+alcohol use (0.84 to 1.32) 1.25 (1.01 to 1.55) 1.14 (1.04 to 1.25) Model 2+TG (0.85 to 1.33) 1.28 (1.03 to 1.59) 1.16 (1.06 to 1.27) Model 2+CRP (0.84 to 1.31) 1.22 (0.98 to 1.52) 1.11 (1.01 to 1.22) 0.02 Model 2+UAE (0.84 to 1.32) 1.25 (1.01 to 1.55) 1.15 (1.05 to 1.25) Model 2+alcohol use, TG, CRP, and UAE (0.84 to 1.32) 1.23 (1.00 to 1.53) 1.12 (1.02 to 1.23) 0.02 Incident CVD events or CVD mortality No. of cases, per 1000 person-years 160 (6.1) 215 (8.2) 333 (13.3) Unadjusted analysis (0.94 to 1.45) 1.78 (1.46 to 2.19) 1.33 (1.23 to 1.45) <0.00 Model (0.82 to 1.27) 1.35 (1.10 to 1.66) 1.22 (1.12 to 1.33) <0.00 Model (0.83 to 1.29) 1.29 (1.05 to 1.59) 1.16 (1.06 to 1.27) <0.00 Model 2+alcohol use (0.83 to 1.29) 1.29 (1.04 to 1.58) 1.16 (1.06 to 1.26) Model 2+TG (0.84 to 1.30) 1.32 (1.07 to 1.62) 1.18 (1.08 to 1.28) <0.00 Model 2+CRP (0.82 to 1.28) 1.26 (1.02 to 1.56) 1.13 (1.03 to 1.23) 0.01 Model 2+UAE (0.83 to 1.29) 1.29 (1.05 to 1.59) 1.16 (1.06 to 1.27) Model 2+alcohol use, TG, CRP, and UAE (0.83 to 1.29) 1.27 (1.03 to 1.57) 1.13 (1.03 to 1.24) Incident all-cause mortality No. of cases, per 1000 person-years 117 (4.4) 160 (5.9) 240 (9.1) Unadjusted analysis (1.10 to 1.86) 2.00 (1.56 to 2.57) 1.36 (1.23 to 1.50) <0.00 Model (0.96 to 1.62) 1.40 (1.08 to 1.79) 1.20 (1.08 to 1.33) <0.00 P 214

218 Model (0.98 to 1.66) 1.41 (1.09 to 1.82) 1.17 (1.06 to 1.29) Model 2+alcohol use (0.98 to 1.66) 1.41 (1.09 to 1.82) 1.17 (1.06 to 1.30) Model 2+TG (0.97 to 1.65) 1.39 (1.07 to 1.79) 1.16 (1.05 to 1.29) Model 2+CRP (0.95 to 1.62) 1.26 (0.97 to 1.64) 1.09 (0.98 to 1.21) 0.11 Model 2+UAE (0.98 to 1.66) 1.40 (1.08 to 1.81) 1.15 (1.04 to 1.28) Model 2+alcohol use, TG, CRP, and UAE (0.94 to 1.60) 1.24 (0.96 to 1.61) 1.08 (0.97 to 1.20) 0.18 Cardiovascular disease (CVD) events were defined as a composite of incident cardiac, cerebral and peripheral vascular events. Participants with history of CVD were excluded. These associations did not differ by sex. Hazard ratios (HR) (95% confidence interval [CI]) have been adjusted for: Model 1: age and sex Model 2: the Framingham risk factors including age, sex, smoking, systolic blood pressure, use of anti-hypertensive therapy, diabetes at baseline, total cholesterol, HDL-cholesterol. * Base-two logarithmically-transformed Prx4, CRP, and 24-hour UAE were analyzed as continuous variable. Individuals with prevalent CVD were excluded. 215

219 Table 5. Reclassification of participants for the 10-year risk prediction of cardiovascular disease corresponding to the Framingham risk score and after adding serum peroxiredoxin 4* Framingham risk score with Prx4 Framingham risk category Low risk Intermediate risk High risk Reclassification (%) In participants without outcome Low risk Intermediate risk High risk In participants with outcome Low risk Intermediate risk High risk Total sample Low risk Intermediate risk High risk * Corresponding the Framingham risk score and after adding serum peroxiredoxin 4, low risk denotes less than 6%, intermediate risk 6 to 20% and high risk more than 20% for the 10-year of cardiovascular disease. Net reclassification index (95%CI) was 2.7 (0.7 to 4.7); P= 0.01, and integrated discrimination improvement (95% CI) of ( to ); P value<

220 Secondary analyses Tables 6 and 7 show the results of secondary analyses. Separately, adjustment for BMI or waist circumference in combination with the Framingham risk factors (i.e., model 2) did not materially change the association of Prx4 with risk of incident CVD events or CVD mortality. Moreover, our results adjusted for both kidney disease and family history of CVD were similar to that of model 2. Additionally, further adjustment for the metabolic syndrome or insulin in model 2 did not affect on the association. We also investigated whether Prx4 was associated with the components of incident CVD events or CVD mortality including MI, cerebrovascular disease and CVD mortality. After adjustment for the variables in model 2, the HRs (95%CI) in the highest tertile compared to the first tertile of Prx4 were 1.03 ( ), 1.28 ( ) and 1.22 ( ) for myocardial infarction, cerebrovascular disease and CVD mortality, respectively. Table 6. Association of serum peroxiredoxin 4 with incident cardiovascular events or mortality (n=8,141) HR (95%CI) by tertiles of peroxiredoxin 4 (U/L) Model Adjusted for model ( ) 1.29 ( ) Adjusted for model 2 + BMI ( ) 1.29 ( ) Adjusted for model 2 + waist circumference ( ) 1.31 ( ) Adjusted for model 2 + family history of CVD ( ) 1.30 ( ) Adjusted for sex, age, smoking + metabolic syndrome ( ) 1.38 ( ) Adjusted for model 2 + insulin ( ) 1.31 ( ) Adjusted for model 2 + kidney disease ( ) 1.30 ( ) Hazard ratios (HR) (95% confidence interval [CI]) have been adjusted for model 2, in which the Framingham risk factors age, sex, smoking, systolic blood pressure, use of anti-hypertensive therapy, diabetes at baseline, total cholesterol, HDL-cholesterol were included. Kidney disease was defined based on prior history of kidney disease requiring dialysis or estimated glomerular filtration rate (egfr) below 60 ml/min/1.73m2. We used the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation to calculate egfr. In another analysis, we examined the association of Prx4 with risk of incident CVD after adjustment for the variables in model 2 and prior history of CVD in total population. The adjusted HR (95%CI) in the highest tertile compared to the first tertile of Prx4 was 1.32 ( ) for incident CVD events or CVD mortality. In participants with prior history of CVD, the adjusted HR (95%CI), for the variables in model 2, in the highest tertile compared to the first tertile of Prx4 was 1.07 ( ) for incident CVD events or CVD mortality (n=181). Figure 2 depicts the relationship of continuous Prx4 with incident CVD events or CVD mortality. We plotted for Framingham risk factors adjusted (model 2) HRs and their 95%CIs as a function of Prx4. The optimal transformation was one in which the terms (Prx4) 1/2 and (Prx4) 1/2 ln(prx4) were incorporated. The solid line demonstrates that after a slight decrease in risk with levels slightly higher than the 217

221 lowest ones that can be detected, the risk associated with increasing levels of Prx4 steeply increases, until a plateau is reached with high levels of Prx4. Incremental value for risk prediction with addition of (Prx4) 1/2 and (Prx4) 1/2 ln(prx4) to the model rather than log2-linear transformed Prx4 (C-statistic=0.81, 95%CI, ; IDI=0.004, P<0.001; NRI=3.6% 95%CI, 1.5% to 5.7%, P<0.001) was slightly higher, but very similar to that with log2-linear transformed Prx4. Table 7. Association of serum peroxiredoxin 4 with incident myocardial infarction, cerebrovascular events and cardiovascular mortality (n=8,141) Tertiles of peroxiredoxin 4 (U/L) Incident myocardial infarction No. of Cases (%) 51 (5.8) 63 (7.8) 94 (10.1) Unadjusted HR (95%CI) ( ) 1.43 ( ) Mutlivariate-adjusted HR (95%CI)* ( ) 1.06 ( ) Incident cerebrovascular disease No. of Cases (%) 39 (6.1) 58 (8.2) 88 (13.3) Unadjusted HR (95%CI) ( ) 1.73 ( ) Mutlivariate-adjusted HR (95%CI)* ( ) 1.28 ( ) Incident cardiovascular mortality No. of Cases (%) 28 (1.0) 44(5.9) 63 (9.1) Unadjusted HR (95%CI) ( ) 2.10 ( ) Multivariate-adjusted HR (95%CI)* ( ) 1.38 ( ) * Hazard ratios (HR) (95% confidence interval [CI]) have been adjusted for model 2 in which the Framingham risk factors included, age, sex, smoking, systolic blood pressure, use of anti-hypertensive therapy, diabetes at baseline, total cholesterol, HDL-cholesterol. Discussion In this study, we demonstrated that serum Prx4, a circulating biomarker with antioxidant properties, was associated with the most common risk factors of CVD in a general population cohort enriched with individuals with microalbuminuria. We found it to have an statistically significant positive association with age, history of CVD, systolic blood pressure, antihypertensive therapy, triglycerides, hs-crp, UAE and procalcitonin, while there was an inverse association with alcohol use and total cholesterol. Moreover, higher serum Prx4 levels were associated with significantly higher risk of incident CVD and all-cause mortality. For potential clinical application, we examined the incremental predictive value of Prx4 compared with the Framingham risk score. Particularly, Prx4 modestly improved prediction for the10- year CVD risk when added to the Framingham risk score in terms of discriminative ability and net reclassification. 218

222 Figure 2. The relationship of peroxiredoxin 4 with incident CVD events or CVD mortality. Data are shown for 8,141 participants without CVD at baseline. The plotted hazard ratio (95%CI) was adjusted for the Framingham risk factors and centered on the Prx4 median value. The optimal transformation of Prx4 consisted of one in which the terms (Prx4) 1/2 and (Prx4) 1/2 ln(prx4) were incorporated. Our findings showing the associations of serum Prx4 levels with cardiovascular risk factors and events support previous clinical studies suggesting a role of oxidative stress in the pathogenesis of CVD 1, 2, 4, 34. In an earlier study, a higher urinary excretion of oxidative stress indices was reported in individuals with renovascular hypertension which was correlated with endothelium-dependent vasodilatation 4. Another study has shown the independent association of oxidized low-density lipoprotein with the incidence of metabolic syndrome 2. In line with this, genotypes and serum activity of paraoxonase 1, an HDL-related antioxidant, has been shown to be associated with systemic oxidative stress and cardiovascular outcomes in humans 1. We now extend accumulating information obtained from animal and human studies on Prx4 of the family of thiol-dependent antioxidants in this era. An animal model of type 1 diabetes has indicated that Prx4 may have a pivotal role in the suppression of apoptosis and the proliferation of progenitor cells to protect against oxidative stress-induced β-cell dysfunction 12. In line with this, a higher gene expression of Prx4 has been found in the islets of high-fat diet model of β-cell dysfunction 35 and down-regulated expression of Prx4 in islet in diabetic mice with chronic hyperglycaemia 36. Moreover, recent studies suggested that Prx4 might be involved in the protection against celiac disease and cancer in pancreas and lung with an increased expression and production of Prx4 in the related human tissues. 12, 37,