Absolute Fracture Risk Assessment Using Lumbar Spine and Femoral Neck Bone Density Measurements: Derivation and Validation of a Hybrid System

ORIGINAL ARTICLE JBMR Absolute Fracture Risk Assessment Using Lumbar Spine and Femoral Neck Bone Density Measurements: Derivation and Validation of a Hybrid System William D Leslie 1,2 and Lisa M Lix 3 for the Manitoba Bone Density Program 1 Department of Medicine, University of Manitoba, Winnipeg, Canada 2 Department of Radiology, University of Manitoba, Winnipeg, Canada 3 School of Public Health, University of Saskatchewan, Saskatoon, Canada ABSTRACT The World Health Organization (WHO) Fracture Risk Assessment Tool (FRAX) computes 10-year probability of major osteoporotic fracture from multiple risk factors, including femoral neck (FN) T-scores. Lumbar spine (LS) measurements are not currently part of the FRAX formulation but are used widely in clinical practice, and this creates confusion when there is spine-hip discordance. Our objective was to develop a hybrid 10-year absolute fracture risk assessment system in which nonvertebral (NV) fracture risk was assessed from the FN and clinical vertebral (V) fracture risk was assessed from the LS. We identified 37,032 women age 45 years and older undergoing baseline FN and LS dual-energy X-ray absorptiometry (DXA; 1990 2005) from a population database that contains all clinical DXA results for the Province of Manitoba, Canada. Results were linked to longitudinal health service records for physician billings and hospitalizations to identify nontrauma vertebral and nonvertebral fracture codes after bone mineral density (BMD) testing. The population was randomly divided into equal-sized derivation and validation cohorts. Using the derivation cohort, three fracture risk prediction systems were created from Cox proportional hazards models (adjusted for age and multiple FRAX risk factors): FN to predict combined all fractures, FN to predict nonvertebral fractures, and LS to predict vertebral (without nonvertebral) fractures. The hybrid system was the sum of nonvertebral risk from the FN model and vertebral risk from the LS model. The FN and hybrid systems were both strongly predictive of overall fracture risk (p <.001). In the validation cohort, ROC analysis showed marginally better performance of the hybrid system versus the FN system for overall fracture prediction (p ¼.24) and significantly better performance for vertebral fracture prediction (p <.001). In a discordance subgroup with FN and LS T-score differences greater than 1 SD, there was a significant improvement in overall fracture prediction with the hybrid method (p ¼.025). Risk reclassification under the hybrid system showed better alignment with observed fracture risk, with 6.4% of the women reclassified to a different risk category. In conclusion, a hybrid 10-year absolute fracture risk assessment system based on combining FN and LS information is feasible. The improvement in fracture risk prediction is small but supports clinical interest in a system that integrates LS in fracture risk assessment. ß 2011 American Society for Bone and Mineral Research. KEY WORDS: BONE MINERAL DENSITY; DUAL-ENERGY X-RAY ABSORPTIOMETRY; OSTEOPOROSIS; FRACTURES; FRAX; ADMINISTRATIVE DATA; HISTORICAL COHORT STUDY Introduction Osteoporosis is a common condition affecting over one in five women over age 50 in Canada. (1) The ability to accurately gauge fracture risk is critical in identifying costeffective thresholds for intervention. (2,3) The World Health Organization (WHO) Collaborating Centre for Metabolic Bone Diseases at Sheffield recently has identified a set of seven clinical risk factors [ie, prior fragility fracture, a parental history of hip fracture, smoking, use of systemic corticosteroids, excess alcohol intake, body mass index (BMI), and rheumatoid arthritis] that, in addition to age and sex, contribute to fracture risk independent of bone mineral density (BMD). (3,4) The Fracture Risk Assessment Tool (FRAX), released in 2008, computes individualized 10-year probability of hip and osteoporotic fracture (ie, composite of hip, clinical spine, distal forearm, and proximal humerus) with and without BMD. (5) The total burden of osteoporotic fractures is much larger than from these four fracture sites alone. Received in original form June 30, 2010; revised form August 4, 2010; accepted August 31, 2010. Published online September 13, 2010. Address correspondence to: William D Leslie, MD, Department of Medicine (C5121), 409 Tache Avenue, Winnipeg, Manitoba R2H 2A6, Canada. E-mail: bleslie@sbgh.mb.ca For a Commentary on this article, please see Watts and Diab (J Bone Miner Res. 2011;458 459. DOI: 10.1002/jbmr.351). Journal of Bone and Mineral Research, Vol. 26, No. 3, March 2011, pp 460 467 DOI: 10.1002/jbmr.248 ß 2011 American Society for Bone and Mineral Research 460

reference site for osteoporosis diagnosis and fracture risk assessment based on large epidemiologic studies that have documented its performance for fracture prediction and especially for assessment of hip fracture risk. (6) Although other BMD measurement sites also can be used for fracture risk assessment and for osteoporosis diagnosis, (7) they are not currently part of the FRAX formulation. Simulation studies have shown very little expected benefit from combining BMD measurement sites because results tend to be moderately correlated. (8) Large epidemiologic studies have failed to demonstrate a meaningful improvement in fracture risk stratification by using the lowest from among several measurements versus using the hip in isolation. (9,10) Notwithstanding these observations, lumbar spine BMD measurement with dual-energy X-ray absorptiometry (DXA) is used widely in clinical practice for initial assessment and monitoring, and there is uncertainty over how to interpret measurements that show discordance between the lumbar spine and femoral neck. Lumbar spine measurements are widely regarded as the best site for prediction of vertebral fractures, whereas hip measurements are the preferred assessment of hip fractures. (7) Completely ignoring a reduced lumbar spine measurement when the femoral neck is normal seems counterintuitive to many clinicians. We hypothesized that a hybrid risk assessment system might be constructed in which lumbar spine BMD was used to assess vertebral fracture risk and femoral neck BMD was used to assess nonvertebral fracture risk. Under this hybrid model, the relative advantages of the two measurement sites would be quantitatively integrated into a global measure of 10-year absolute fracture risk. We report the derivation and validation of this approach using a split-cohort design. Methods Patient population Fig. 1. Observed fracture risk at 10 years (Kaplan-Meier estimate) when moderate-risk cases under the femoral neck model were reclassified under the hybrid model. 95% CI bars are shown. Dashed lines indicate cutoffs between the low (<10%), moderate (10% to 20%), and high (>20%) fracture risk categories. FRAX is calibrated to use femoral neck BMD (or T-score from the Third National Health and Nutrition Examination Survey (NHANES III) white female reference population) in estimating fracture risk when BMD is included. The femoral neck was selected as the The study population for this retrospective cohort study consisted of all women age 45 years and older with valid DXA measurements from the lumbar spine and femoral neck. Women were required to have medical coverage from Manitoba Health during the observation period ending in March 2008. For women with more than one eligible set of measurements, only the first record was included. The final study population consisted of 37,032 women. The population was randomly divided into two equally sized subgroups, one for use in model derivation and the second for independent validation. The study was approved by the Research Ethics Board of the University of Manitoba and the Health Information Privacy Committee of Manitoba. Bone density measurements In the Province of Manitoba, Canada, health services are provided to virtually all residents through a single public health care system. Bone density testing with DXA has been managed as an integrated program since 1997; criteria and testing rates for this program have been published. (11) The program maintains a database of all DXA results that can be linked with other population-based computerized health databases through an FRACTURE PREDICTION WITH A HYBRID SYSTEM Journal of Bone and Mineral Research 461

anonymous personal identifier. (12) The DXA database has been described previously with completeness and accuracy in excess of 99%. DXA scans were performed and analyzed in accordance with manufacturer recommendations. Lumbar spine T-scores (ie, number of SDs above or below young-adult mean BMD) and Z-scores (ie, number of SDs above or below age-matched mean BMD) were calculated using the manufacturer s US white female reference values. Hip T-scores and Z-scores were calculated from the revised NHANES III white female reference values (Prodigy Version 8.8, GE Lunar Corp., Madison, WI, USA). (13,14) Vertebral levels affected by artifact were excluded by experienced physicians using conventional criteria. (15) Prior to 2000, DXA measurements were performed with a pencil-beam instrument (Lunar DPX, GE Lunar Corp.), and after this date, a fan-beam instrument was used (Lunar Prodigy, GE Lunar Corp.). Instruments were cross-calibrated using anthropomorphic phantoms and 59 volunteers. No clinically significant differences were identified (T-score differences < 0.2). Densitometers showed stable long-term performance [coefficient of variation (CV) < 0.5%] and satisfactory in vivo precision (CV ¼ 1.7% for L 1 L 4 and 1.1% for the total hip). (16) Definitions of fractures and other clinical risk factors Fractures and other medical diagnoses was assessed through a combination of hospital discharge abstracts (ie, diagnoses and procedures coded using the ICD-9-CM prior to 2004 and ICD-10- CA thereafter) and physician billing claims (ie, coded using ICD-9- CM). (17) Use of systemic corticosteroids and other medications was obtained by linkage to the provincial Drug Program Information Network (DPIN) database, with drugs classified according to the Anatomical Therapeutic Chemical (ATC) System of the WHO. (18) Each prescription record contains the date of dispensation; an exact identification of the dispensed drug, including substance, strength, route, and dosage form; the number of doses provided; the anticipated duration of the prescription in days; and a code for prescribing physician and dispensing pharmacy. The pharmacy database is accurate both for capture of drug dispensations and the prescription details. (19) Longitudinal health service records were assessed for the presence of hip, clinical vertebral, forearm, and humerus fracture codes (collectively designated as osteoporotic ) before and after BMD testing that were not associated with trauma codes. (20) Incident fractures were defined as fractures that occurred after the index BMD measurement and generated two or more sitespecific fracture codes in any diagnosis field (hospitalization or physician visit). We required that hip fractures and forearm fractures be accompanied by a site-specific fracture reduction, fixation, or casting code because this enhances the diagnostic and temporal specificity for an acute fracture event. To minimize potential misclassification of prior fractures as incident fractures, we required that there be no hospitalization or physician visit(s) with the same fracture type in the 6 months preceding an incident fracture diagnosis. For purposes of the FRAX calculation, prior fragility fracture was taken to be an osteoporotic fracture prior to BMD testing. A diagnosis of rheumatoid arthritis was taken from physician office visits and/or hospitalizations with a compatible ICD-9-CM/ICD-10-CA code in a 3-year period prior to BMD testing. Proxies were used for smoking [chronic obstructive pulmonary disease (COPD) diagnosis] and high alcohol intake (alcohol or substance abuse diagnosis). Prolonged corticosteroid use (over 90 days dispensed in the year prior to DXA testing at a mean prednisone-equivalent dose of 7.5 mg/day or greater) was obtained from the provincial pharmacy system. Weight and height were recorded at the time of the DXA examination (prior to 2000, this was by self-report, and starting in 2000, height was assessed with a wall-mounted stadiometer and weight was assessed without shoes using a standard floor scale). BMI (in kg/ m 2 ) was calculated as weight (in kg) divided by height squared (in m). Secondary causes of osteoporosis, other than rheumatoid arthritis, do not contribute to the probability of fracture as calculated by FRAX when information on BMD is present and therefore was not considered in this analysis. Model construction Three fracture-prediction models were developed. The reference model was femoral neck BMD to predict all osteoporotic fractures (nonvertebral and vertebral). The hybrid comparator model reflected a simple sum of two mutually exclusive component models: femoral neck BMD to predict nonvertebral fractures and lumbar spine BMD to predict isolated clinical vertebral fractures. The requirement for additive independence in the nonvertebral and vertebral fracture models required that individuals satisfying both definitions be assigned to one or the other to avoid overestimation in the summed fracture risk. Therefore, women who sustained both nonvertebral and vertebral fractures during follow-up were included with the former and not the latter. The fracture-prediction models included age and BMD, with parameters derived from Cox proportional hazards regression. The hazard ratios (HRs) were adjusted for multiple risk factors that are considered in the FRAX tool [BMI, prior osteoporotic fracture, diagnosed rheumatoid arthritis, COPD (as a proxy for smoking), substance diagnosis (as a proxy for high alcohol intake), and prolonged corticosteroid use]. These proxy measures for smoking and high alcohol intake have been validated to predict major osteoporotic fractures in our data set independent of other FRAX variables. (21) Use of osteoprotective medication in the year after BMD measurement also was included in the adjustments. The prediction models did not attempt to explicitly incorporate all the FRAX risk factors because the details of the FRAX model are not sufficiently well described for their inclusion, and it is uncertain whether they would apply equally for nonvertebral and vertebral fractures. This does not affect the current analysis, which isolates the variables of interest, namely, lumbar spine and femoral neck BMD, thereby providing a direct evaluation of the feasibility and utility of a combined risk assessment system. Age as a continuous variable was included in the models given its consistent association with fracture risk (independent of BMD) and age-related declines in BMD. There was no additional benefit to inclusion of age as a quadratic term. Death was considered a competing hazard in accordance with the approach adopted by the WHO Collaborating Centre. 462 Journal of Bone and Mineral Research LESLIE AND LIX

Statistics All results are reported as mean SD unless otherwise stated. Group comparisons for continuous data were with the Student s t test and for categorical data used a chi-square test. Cox proportional hazards models were used to estimate hazard ratios (HRs) for age and BMD adjusted for FRAX risk factors (ie, BMI, prior osteoporotic fracture, rheumatoid arthritis, COPD, substance abuse, and recent corticosteroid use) as well as osteoprotective medication use in the year after BMD testing. Parental history of hip fracture was not available. The proportional hazard assumption was confirmed graphically from log[ log(survival)] versus log(time) plots. Statistical analyses were performed with SPSS for Windows, Version 16.0 (SPSS, Inc., Chicago, IL, USA). Overall performance of the prediction system was assessed from the area under the curve (AUC) for receiver operating characteristic (ROC) plots. AUCs were compared using the nonparametric method of DeLong and colleagues, (22) which allows for comparison of correlated curves originating from a common population (AccuROC 2.5, Accumetric Corp., Montreal, Quebec, Canada). Absolute 10-year fracture probabilities under the femoral neck and hybrid prediction systems were categorized as low risk (< 10%), moderate risk (10% to 20%), and high risk (> 20%) in accordance with Canadian reporting guidelines. (23) The number of individuals in whom the hybrid system reclassified risk to a different category was determined according to the method of Janes and colleagues. (24) The number of fractures under the two prediction systems were cross-tabulated, and the linear trend in fracture rates when a risk category under the femoral neck system was reclassified under the hybrid system was assessed using the Cochran-Armitage test. Within each subgroup, fracture outcomes to 10 years were estimated using the Kaplan-Meier method. Results Population The baseline characteristics of the 18,447 women in the derivation subgroup and the 18,585 women in the validation subgroup are summarized in Table 1. No significant differences (all p >.05) were seen in any of the baseline covariates or in subsequent outcomes. Derivation cohort During a mean 5.6 years of follow-up (range up to 10 years), 1748 women (4.7%) experienced one or more incident fractures satisfying one of the primary definitions for vertebral, hip, forearm, and humerus fracture. Of these, 1484 (4.0%) involved a nonvertebral fracture, and 264 (0.7%) were an isolated clinical vertebral fracture. Forty-four women had both vertebral and nonvertebral fractures and were counted with the nonvertebral fracture group. Preliminary analyses looked at the ability of lumbar spine and femoral neck BMD to predict nonvertebral and vertebral fractures both alone and in combination (Table 2). There was moderate correlation between lumbar spine and femoral neck T- scores (Pearson r 2 ¼ 0.50). Each BMD measurement site was able to predict vertebral and nonvertebral fractures (all p <.001). When both the lumbar spine and femoral neck were included in the prediction model, the lumbar spine (p <.001) but not the femoral neck (p ¼.275) was predictive of vertebral fracture, whereas the femoral neck (p <.001) but not the lumbar spine (p ¼.757) was predictive of nonvertebral fractures. The independence of the lumbar spine and femoral neck for prediction of vertebral and nonvertebral fractures, respectively, provides justification for the two-compartment model described in this article. After adjustment for FRAX risk factors and treatment, the lumbar spine was strongly predictive of isolated vertebral fractures (HR ¼ 1.84 per SD, HR per T-score score ¼ 1.49), and the femoral neck was strongly predictive of nonvertebral fractures (HR ¼ 1.66 per SD, HR per T-score score ¼ 1.68). Gradients of risk for fracture prediction are similar to those from a previous metaanalysis. (7) Validation cohort ROC AUC measures for fracture discrimination for the validation subgroup are summarized in Table 3. There was a small Table 1. Study Population Baseline Characteristics (n ¼ 37,032) Derivation (n ¼ 18,447) Validation (n ¼ 18,585) p Value Age (years) 63.6 10.4 63.6 10.3.504 BMI (kg/cm 2 ) 26.6 5.2 26.6 5.3.673 Rheumatoid arthritis (%) 641 (3.5%) 703 (3.8%).113 COPD (%) 1346 (7.3%) 1393 (7.5%).465 Substance abuse diagnosis (%) 486 (2.6%) 472 (2.5%).565 Corticosteroid use before (%) 812 (4.4%) 832 (4.5%).726 Lumbar spine T-score 1.3 1.5 1.3 1.5.708 Lumbar spine Z-score 0.1 1.4 0.1 1.4.717 Femoral neck T-score 1.4 1.0 1.4 1.0.634 Femoral neck Z-score 0.1 1.0 0.1 1.0.612 Minimum T-score osteoporotic (%) 5451 (29.5%) 5485 (29.5%).939 Prior osteoporotic fracture 2354 (12.8%) 2357 (12.7%).820 Incident osteoporotic fracture 840 (4.6%) 908 (4.9%).132 Observation period (years) 5.5 2.8 5.6 2.8.103 Data expressed as mean SD or n (%). BMI ¼ body mass index; COPD ¼ chronic obstructive pulmonary disease. FRACTURE PREDICTION WITH A HYBRID SYSTEM Journal of Bone and Mineral Research 463

Table 2. Hazard Ratios (HRs) for Fracture per T-Score Unit Used in Constructing the Prediction Models and Gradient of Risk (HR per SD) HR per T-score unit a (95% CI) HR per SD a (95% CI) Construction of FN model FN for any fracture 1.65 (1.5 1.81) 1.63 (1.48 1.78) Construction of hybrid model FN for nonvertebral fracture 1.68 (1.51 1.86) 1.66 (1.50 1.83) LS for vertebral fracture 1.49 (1.28 1.73) 1.84 (1.46 2.32) FN ¼ femoral neck; LS ¼ lumbar spine. a Values are from adjusted Cox proportional hazards models using the derivation cohort (n ¼ 18,447) and are adjusted for BMI, prior osteoporotic fracture, diagnosed rheumatoid arthritis, COPD (as a proxy for smoking), substance diagnosis (as a proxy for high alcohol intake), prolonged corticosteroid use, and use of osteoprotective medication in the year after BMD measurement. improvement in vertebral fracture prediction under the hybrid model [0.731 (0.212)] compared with the femoral neck model [0.7235 (0.217), p <.001]. No significant difference was seen between these prediction models for nonvertebral fractures or all fractures combined. When analysis was restricted to 6434 women from the validation subgroup with a T-score difference between the lumbar spine and femoral neck exceeding 1 SD, there was a small improvement in overall fracture prediction with the hybrid model versus the femoral neck model (p ¼.025). The effect of including lumbar spine in the risk assessment algorithm was assessed in terms of reclassification and fracture risk stratification (Table 4). Based on predefined 10-year fracture risk categories (low < 10%, moderate 10% to 20%, high > 20%), there was a low level (6.4%) of categorical reclassification that was greatest for those at moderate risk under the femoral neck system (8.0%), followed by the low-risk group (6.1%) and least for the high-risk group (<0.1%). Among women in the moderaterisk category under the femoral neck system, those who were reclassified to low risk had mean lumbar spine T-scores of þ0.2, those who remained in the moderate-risk category had mean lumbar spine T-scores of 2.2, and those who moved into the high-risk category had mean lumbar spine T-scores of 3.4. Tenyear fracture risk for women in the moderate-risk category under the femoral neck system estimated by the Kaplan-Meier method was 18.2% for the validation data set. After reclassification under the hybrid system, 10-year fracture estimates were 6.3% for those reclassified as low risk, 18.6% for those reclassified as moderate risk, and 25.5% for those reclassified as high risk ( p <.001 for trend). Similar results were seen for the overall data set and for the discordance subgroup. Table 4 shows that 675 (39%) of the major osteoporotic fractures occurred in women designated as low risk under FRAX, of whom 67 (10%) were appropriately reclassified to a higher-risk category under the hybrid system. Discussion This analysis provides proof of concept that incorporation of lumbar spine BMD in fracture risk assessment provides a small but demonstrable improvement in fracture risk stratification compared with femoral neck BMD alone. The improvement was more evident among those with discordance between the lumbar spine and femoral neck T-scores exceeding 1 SD. When analyzed according to fracture site, all the gain in fracture prediction under the hybrid system was enhanced risk prediction Table 3. Area Under the Curve (AUC) for Receiver Operator Characteristic Curve Analysis of Fracture Prediction FN model Hybrid model p Value, a FN versus hybrid model Overall data set, n ¼ 37,032 Nonvertebral fracture 0.687 (.007) 0.687 (.007).18 Vertebral fracture 0.719 (.016) 0.727 (.016) <.001 Any fracture 0.695 (.007) 0.695 (.007).42 Validation data set, n ¼ 18,585 Nonvertebral fracture 0.693 (.010) 0.692 (.010).24 Vertebral fracture 0.723 (.022) 0.731 (.021) <.001 Any fracture 0.701 (.009) 0.701 (.009).75 Discordance subgroup, n ¼ 6434 b Nonvertebral fracture 0.671 (.018) 0.672 (.018).46 Vertebral fracture 0.729 (.040) 0.750 (.038) <.001 Any fracture 0.682 (.016) 0.687 (.016).025 Data expressed as mean (SE). a p Values are from the nonparametric method of DeLong and colleagues. (22) b Validation cohort subjects with femoral neck and lumbar spine T-score differences > 1 SD. 464 Journal of Bone and Mineral Research LESLIE AND LIX

Table 4. Observed Fracture Risk at 10 Years (Kaplan-Meier Estimate) According to Risk Reclassification Under the Femoral Neck (FN) and Hybrid Models for All Patients Femoral neck model Overall Low risk (< 10%) Hybrid model Moderate risk (10% to 20%) High risk (> 20%) p Trend Low risk (< 10%) n 24,660 23,167 1493 0 Fractures 675 608 67 0 <.001 % Fractures at 10 years (SE) 6.9 (0.3) 6.4 (0.3) 14.9 (1.9) N/A % Reclassified 6.1% 6.1% N/A Moderate risk (10% to 20%) n 10,830 50 9968 812 Fractures 803 3 712 88 <.001 % Fractures at 10 years (SE) 17.5 (0.7) 11.0 (6.0) 16.9 (0.7) 25.5 (2.4) % Reclassified 8.0% 0.5% 7.5% High risk (> 20%) n 1542 0 1 1541 Fractures 270 0 0 270 N/A % Fractures at 10 years (SE) 40.3 (2.7) N/A N/A 46.4 (4.9) % Reclassified 0.1% N/A 0.1% Overall n 37,032 23,217 11,462 2353 Fractures 1748 611 779 358 <.001 % Fractures at 10 years (SE) 11.2 (0.3) 6.4 (0.3) 16.8 (0.7) 35.2 (2.0) for vertebral fractures, whereas nonvertebral fracture prediction was unchanged. At first glance, the minimal changes in AUC could be used to argue that the more complex two-site risk assessment system offers little advantage over the single-site femoral neck system and is not worth the trouble. Although there is truth to this statement, it must be acknowledged that lumbar spine BMD measurements are performed routinely in clinical practice both for baseline risk assessment and for monitoring purposes. When confronted with highly discordant measurements (lumbar spine worse than femoral neck), clinicians are in a quandary about how this should be integrated into the decision-making process. Under some guidelines, such as those from the National Osteoporosis Foundation, treatment would be recommended for a lumbar spine T-score in the osteoporotic range regardless of the estimated risk. (25) Other national guidelines, such as those from the United Kingdom, do not have specific treatment recommendations for individuals with osteoporotic lumbar spine BMD when 10-year fracture risk prediction from the femoral neck does not achieve the intervention threshold. (26) Previous 2005 Canadian guidelines attempted to address the issue of site discordance by substituting the minimum T-score for the femoral neck T-score, (23) but this systematically overestimates fracture risk (27) and does not consider site-specific differences in fractures or the way BMD declines with age. (7,28) Where lumbar spine and hip measurements are both performed for clinical purposes, using a procedure that accurately reflects the contribution of each measurement site to fracture risk clearly is preferred. FRAX is a work in progress, and there is active discussion about changes and enhancements that will improve its clinical performance. Incorporating the lumbar spine in the FRAX algorithm is one such option, whereas inclusion of additional clinical risk factors (eg, falls) also has been proposed. Although not the primary focus of this analysis, it is worth noting that the improvement in fracture prediction from including the FRAX clinical risk factors in fracture risk assessment over BMD alone has been an area of uncertainty. FRAX with BMD outperformed FRAX without BMD or BMD alone in a pooled analysis from Kanis and colleagues. (5) Attempts to confirm this in additional cohorts have met with mixed results. Trémollieres and colleagues (29) did not demonstrate value in using FRAX over BMD alone, but a recent report from Manitoba did find that FRAX with BMD provided better stratification for major osteoporotic and hip fractures than BMD or FRAX clinical risk factors alone. (21) Limitations to this analysis are acknowledged. Although a split-cohort derivation/validation procedure was used, completely independent validation in different populations and contexts would be preferred. A major challenge in the study of the epidemiology of osteoporotic fractures concerns the difficulty in ascertaining vertebral fractures. The protean definitions that have been applied clinically, morphometrically, and administratively speak to this challenge. Notwithstanding these legitimate difficulties, administrative definitions for vertebral fractures have been developed and tested with acceptable results. (30) Since the predominant difficulty is underrecognition of vertebral fractures, this actually would introduce a bias against the hybrid system, as noted earlier. Furthermore, it does not seriously undermine the conceptual framework of this analysis. The actual improvement in risk stratification under the hybrid system is likely underestimated owing to the intentional use of a highly restrictive definition of clinical vertebral fracture. Larger gains would be expected with FRACTURE PREDICTION WITH A HYBRID SYSTEM Journal of Bone and Mineral Research 465

better vertebral fracture recognition. Whether the existing FRAX formulation would admit to inclusion of multiple BMD measurement sites is also unclear. FRAX already considers a multitude of interactions, and inclusion of additional interactions with lumbar spine BMD and/or higher-order interactions may not be feasible. The calculations described are also unsuitable for clinical practice because the risk calculations require computerization. A simple, clinician-friendly approach that has face validity to clinicians probably would meet with greater acceptance. Referral bias in this clinical population is likely responsible for the higher than expected prevalence of rheumatoid arthritis. Despite this, mean Z-scores for the spine and hip were very close to the manufacturer s age-matched mean reference values (ie, Z-score close to zero). In summary, we have shown that it is feasible to create a hybrid 10-year absolute fracture risk assessment system that combines lumbar spine and femoral neck BMD information. Risk reclassification under this hybrid system is responsive to lumbar spine measurements and provides a small improvement in risk stratification that is most evident in those at moderate risk under the femoral neck system and where there is discordance between the two BMD measurement sites. This supports clinical interest in a system that quantitatively integrates lumbar spine BMD in the fracture risk assess paradigm. Although it cannot be quantified, clinicians are likely to have greater confidence in a risk assessment system that has face validity by adjusting risk in the setting of discordant lumbar spine measurements. Disclosures WL: Speaker bureau and unrestricted research grants: Merck Frosst Canada; research honoraria and unrestricted educational grants: Sanofi-Aventis, Procter & Gamble Pharmaceuticals Canada; unrestricted research grants: Novartis, Amgen Pharmaceuticals Canada; unrestricted educational grants: Genzyme Canada; Advisory boards: Genzyme Canada, Novartis, Amgen Pharmaceuticals Canada. LL: Unrestricted research grant: Amgen. Both the authors state that they have no conflicts of interest. Acknowledgments We are indebted to Manitoba Health for providing data (HIPC File No. 2007/2008 35). The results and conclusions are those of the authors, and no official endorsement by Manitoba Health is intended or should be inferred. This article has been reviewed and approved by the members of the Manitoba Bone Density Program Committee. References 1. Berger C, Goltzman D, Langsetmo L, et al. Peak bone mass from longitudinal data: implications for the prevalence, pathophysiology, and diagnosis of osteoporosis. J Bone Miner Res. 2010;25:1948 1957. 2. Kanis JA, Oden A, Johnell O, Jonsson B, De Laet C, Dawson A. The burden of osteoporotic fractures: a method for setting intervention thresholds. Osteoporos Int. 2001;12:417 427. 3. Kanis JA, Borgstrom F, De Laet C, et al. Assessment of fracture risk. Osteoporos Int. 2005;16:581 589. 4. Kanis JA, Johnell O, Oden A, Johansson H, McCloskey E. FRAX and the assessment of fracture probability in men and women from the UK. Osteoporos Int. 2008;19:385 397. 5. Kanis JA, Oden A, Johnell O, et al. The use of clinical risk factors enhances the performance of BMD in the prediction of hip and osteoporotic fractures in men and women. Osteoporos Int. 2007;18: 1033 1046. 6. Johnell O, Kanis JA, Oden A, et al. Predictive value of BMD for hip and other fractures. J Bone Miner Res. 2005;20:1185 1194. 7. Marshall D, Johnell O, Wedel H. Meta-analysis of how well measures of bone mineral density predict occurrence of osteoporotic fractures. BMJ. 1996;312:1254 1259. 8. Blake GM, Patel R, Knapp KM, Fogelman I. Does the combination of two BMD measurements improve fracture discrimination? J Bone Miner Res. 2003;18:1955 1963. 9. Kanis JA, Johnell O, Oden A, et al. The use of multiple sites for the diagnosis of osteoporosis. Osteoporos Int. 2006;17:527 534. 10. Leslie WD, Lix LM, Tsang JF, Caetano PA. Single-site vs multisite bone density measurement for fracture prediction. Arch Intern Med. 2007;167:1641 1647. 11. Leslie WD, MacWilliam L, Lix L, Caetano P, Finlayson GS. A populationbased study of osteoporosis testing and treatment following introduction of a new bone densitometry service. Osteoporos Int. 2005; 16:773 782. 12. Leslie WD, Caetano PA, MacWilliam LR, Finlayson GS. Construction and validation of a population-based bone densitometry database. J Clin Densitom. 2005;8:25 30. 13. Binkley N, Kiebzak GM, Lewiecki EM, et al. Recalculation of the NHANES Database SD Improves T-Score Agreement and Reduces Osteoporosis Prevalence. J Bone Miner Res. 2005;20:195 201. 14. Kanis JA, McCloskey EV, Johansson H, Oden A, Melton LJ III, Khaltaev N. A reference standard for the description of osteoporosis. Bone. 2008;42:467 475. 15. Hansen KE, Binkley N, Christian R, et al. Interobserver reproducibility of criteria for vertebral body exclusion. J Bone Miner Res. 2005;20: 501 508. 16. Leslie WD. The importance of spectrum bias on bone density monitoring in clinical practice. Bone. 2006;39:361 368. 17. Roos NP, Shapiro E. Revisiting the Manitoba Centre for Health Policy and Evaluation and its population-based health information system. Med Care. 1999;37:JS10 JS14. 18. WHO Collaborating Centre for Drug Statistics Methodology. Guidelines for ATC classification and DDD assignment. Oslo, 2005. 19. Kozyrskyj AL, Mustard CA. Validation of an electronic, populationbased prescription database. Ann Pharmacother. 1998;32:1152 1157. 20. Leslie WD, Tsang JF, Caetano PA, Lix LM. Effectiveness of bone density measurement for predicting osteoporotic fractures in clinical practice. J Clin Endocrinol Metab. 2007;92:77 81. 21. Leslie WD, Lix LM, Johansson H, Oden A, McCloskey E, Kanis JA. Independent clinical validation of a Canadian FRAX tool: Fracture prediction and model calibration. J Bone Miner Res. 2010;25:2350 2358. 22. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44:837 845. 23. Siminoski K, Leslie WD, Frame H, et al. Recommendations for bone mineral density reporting in Canada. Can Assoc Radiol J. 2005;56: 178 188. 24. Janes H, Pepe MS, Gu W. Assessing the value of risk predictions by using risk stratification tables. Ann Intern Med. 2008;149:751 760. 25. Dawson-Hughes B, Looker AC, Tosteson AN, Johansson H, Kanis JA, Melton LJ III. The potential impact of new National Osteoporosis 466 Journal of Bone and Mineral Research LESLIE AND LIX

Foundation guidance on treatment patterns. Osteoporos Int. 2010;21:41 52. 26. Kanis JA, McCloskey EV, Johansson H, Strom O, Borgstrom F, Oden A. Case finding for the management of osteoporosis with FRAX((R))- assessment and intervention thresholds for the UK. Osteoporos Int. 2008;19:1395 1408. 27. Leslie WD, Lix LM. Simplified 10-year absolute fracture risk assessment: a comparison of men and women. J Clin Densitom. 2010;13: 141 146. 28. Faulkner KG, von SE, Miller P. Discordance in patient classification using T-scores. J Clin Densitom. 1999;2:343 350. 29. Tremollieres FA, Pouilles JM, Drewniak N, Laparra J, Ribot CA, rgent- Molina P. Fracture risk prediction using BMD and clinical risk factors in early postmenopausal women: sensitivity of the WHO FRAX tool. J Bone Miner Res. 2010;25:1002 1009. 30. Curtis JR, Mudano AS, Solomon DH, Xi J, Melton ME, Saag KG. Identification and validation of vertebral compression fractures using administrative claims data. Med Care. 2009;47:69 72. FRACTURE PREDICTION WITH A HYBRID SYSTEM Journal of Bone and Mineral Research 467