Hepatocellular carcinoma (HCC) is a leading. Development of Risk Scoring System for Stratifying Population for Hepatocellular Carcinoma Screening

Transcription

1 Development of Risk Scoring System for Stratifying Population for Hepatocellular Carcinoma Screening Yi-Chun Hung, 1 Chih-Lin Lin, 2 Chun-Jen Liu, 3 Hung Hung, 1 Shi-Ming Lin, 4 Shou-Dong Lee, 5,6 Pei-Jer Chen, 3 Shu-Chun Chuang, 7 and Ming-Whei Yu 1 The age and risk level that warrants hepatocellular carcinoma (HCC) screening remains to be defined. To develop risk scores for stratifying average-risk population for mass HCC screening, we conducted a pooled analysis using data from three cohorts involving 12,377 Taiwanese adults years of age. During 191,240.3 person-years of follow-up, 387 HCCs occurred. We derived risk scores from Cox s model in two thirds of participants and used another one third for model validation. Besides assessing discrimination and calibration, we performed decision curve analysis to translate findings into public health policy. A risk score according to age, sex, alanine aminotransferase, previous chronic liver disease, family history of HCC, and cumulative smoking had good discriminatory accuracy in both model derivation and validation sets (c-statistics for 3-, 5-, and 10-year risk prediction: ). It also performed well across cohorts and diverse subgroups. Decision curve analyses revealed that use of the score in selecting persons for screening improved benefit at threshold probabilities of >2% 10-year risk, compared with current guidelines and a strategy of screening all hepatitis B carriers. Using 10-year risk 2% as a threshold for initiating screening, the screening age ranged from 20 to 60 years, depending on the tertile of risk scores and status of hepatitis B/C virus infection. Combining risk-score tertile levels and hepatitis virus status to stratify participants was more sensitive than current guidelines for HCC detection within 10 years (89.4% vs. 76.8%), especially for young-onset HCCs <50 years (79.4% vs. 40.6%), under slightly lower specificity (67.8% vs. 71.8%). Conclusion: A simple HCC prediction algorithm was developed using accessible variables combined with hepatitis virus status, which allows selection of asymptomatic persons for priority of HCC screening. (HEPATOLOGY 2015;61: ) See Editorial on Page 1783 Hepatocellular carcinoma (HCC) is a leading cause of cancer-related deaths worldwide. Chronic infection with hepatitis B virus (HBV) accounts for 50%-80% of global HCC cases, whereas 10%-25% of cases are attributable to hepatitis C virus (HCV). 1 There is evidence showing that community-wide ultrasonography (USG)-based mass screening for HCC reduces mortality, 2,3 but less than 20% of patients with diagnosed HCC reported receiving prediagnosis regular screening. 4 Current expert guidelines recommend performing HCC screening in all patients with cirrhosis and for chronic HBV carriers, including Asian men 40 years Abbreviations: AASLD, American Association for the Study of Liver Diseases; ALT, alanine aminotransferase; anti-hcv, antibodies against hepatitis C virus; CGMH, Chang Gung Memorial Hospital; CI, confidence interval; CLD, chronic liver disease; GECC, Government Employees Central Clinics; HBsAg, hepatitis B surface antigen; HBV, hepatitis B virus; HCC, hepatocellular carcinoma; HCCFAM, hepatocellular carcinoma family; HCV, hepatitis C virus; HR, hazard ratio; USG, ultrasonography. From the 1 1nstitute of Epidemiology and Preventive Medicine, College of Public Health, National Taiwan University, Taipei, Taiwan; 2 Department of Gastroenterology, Ren-Ai Branch, Taipei City Hospital, Taipei, Taiwan; 3 Division of Gastroenterology, Department of Internal Medicine, National Taiwan University Hospital and Graduate Institute of Clinical Medicine, National Taiwan University College of Medicine, Taipei, Taiwan; 4 Liver Research Unit, Chang Gung Memorial Hospital, Chang Gung University College of Medicine, Taipei, Taiwan; 5 Faculty of Medicine, National Yang-Ming University School of Medicine, Taipei, Taiwan; 6 Division of Gastroenterology, Department of Medicine, Cheng Hsin General Hospital, Taipei, Taiwan; and 7 Division of Biostatistics and Bioinformatics, Institute of Population Health Sciences, National Health Research Institutes, Zhunan, Taiwan. Received September 5, 2014; accepted November 14, Additional Supporting Information may be found at onlinelibrary.wiley.com/doi/ /hep.27610/suppinfo. This study was funded by the Taiwan Ministry of Science and Technology National Research Program for Biopharmaceuticals (grants NSC B , NSC B , NSC B , and MOST B ). 1934

2 HEPATOLOGY, Vol. 61, No. 6, 2015 HUNG ET AL and women 50 years. 5 The guideline is intended to provide cost-effective screening strategies, but substantial numbers of at-risk persons below these age limits still develop HCC. 6 Furthermore, some data suggest considering earlier screening for persons with firstdegree family history of HCC, but the age to start screening has not been established. 5,7 There is a wide heterogeneity in rates of progression from healthy HBV or HCV carrier state to HCC. 7,8 For population-based screening, an accurate HCC riskprediction model for stratifying population into risk categories is needed. At the population level, an ideal model is easy to use and thus could enhance awareness of risk and encourage at-risk people to be screened. Until now, model developments for predicting HCC in this way have been limited. The only published HCC model that attempted to assess risk toward the thought of use in the general public was constructed based on a single Taiwanese cohort from a private health care center after an average follow-up of 8.5 years. 9 This study ignored previous liver disease and family history in risk prediction, which are important concerns in HCC screening. 5,7 As with most other studies focusing on the development of models to predict disease risk, clinical utility of different HCC risk models was not evaluated. We set out to develop and validate a score predictive of risk for HCC screening. The scoring system is based on widely accessible information on 12,377 adults from three cohort studies. In particular, we included a cohort of first-degree relatives of patients with HCC who are at high risk for young-onset HCC. 7 To gauge the future potential of applying a prediction model in population-based screening, we performed decision curve analysis and investigated whether the risk score in combination with main risk factors (i.e., HBV/HCV) could reliably stratify risks for HCC at selected ages. Patients and Methods Study Population. We included three cohorts of different sources with some results published in the pooled cohort. 7,10-12 In general, the study participants in the pooled cohort had a higher proportion of highrisk groups characterized by male sex, hepatitis B surface antigen (HBsAg) positivity, and HCC family history. The cohort enrolled from the Government Employees Central Clinics (GECC) consisted of civil servants who attended the clinic for an annual free physical examination. Participants of the Chang Gung Memorial Hospital (CGMH) cohort are HBsAgpositive adults who visited the Hepatitis B Carrier Clinic of CGMH for routine checkup. This cohort consisted of study populations of two waves recruited in and , respectively. Participants of the HCC family (HCCFAM) cohort are the relatives of patients with HCC, consisting of first- and seconddegree relatives and relatives in law by marriage, who belong to a multicenter study designed to explore genetic and environmental determinants of HCC. 11,12 At recruitment, in-person interviews were conducted by trained research assistants according to structured questionnaires. Questionnaires may differ across calendar time, but all contain core elements, including demographics, habits of tobacco and alcohol use, and personal and family medical histories. All participants were tested for HBsAg and serum alanine aminotransferase (ALT). Except for the CGMH cohort recruited in , antibodies against hepatitis C virus (anti- HCV) were assayed when blood samples were available. In total, 12,377 individuals were included. Table 1 gives an overview of each of the three cohorts. There are differences in the cohort compositions of sex, HBsAg status, and HCC family history because of the cohort recruitment policy for special interest. This study was approved by the research ethics committee at the National Taiwan University, and all participants provided informed consent. Cohort Follow-up. All three cohorts were followed for HCC incidence and mortality through linking the profile of personal identification number to the national cancer and death registry databases. Follow-up of participants who were seropositive for HBsAg or anti-hcv were also achieved through medical examinations, including USG and blood testing. 7,10,12 At the end of 2010, there were 387 (91 of whom were detected under age 50 refer to young-onset HCC) incident cases of HCC in 191,240.3 person-years of follow-up as well as 1,001 deaths (256 deaths were Address reprint requests to: Ming-Whei Yu, Ph.D., Institute of Epidemiology and Preventive Medicine, College of Public Health, National Taiwan University, Room 522, 17 Xuzhou Road, Zhongzheng District, Taipei City 10055, Taiwan. yumw@ntu.edu.tw; fax: Copyright VC 2015 by the American Association for the Study of Liver Diseases. View this article online at wileyonlinelibrary.com. DOI /hep Potential conflicts of interest: Nothing to report.

3 1936 HUNG ET AL. HEPATOLOGY, June 2015 Sampling strategy Table 1. Design and Characteristics of the Three Cohorts GECC CGMH HCCFAM All HBsAg-positive men were invited; HBsAg-negative men were one-to-one matched to HBsAg-positive men on age and ethnic background HBsAg-positive individuals who visited the Hepatitis B Carrier Clinic of CGMH for routine checkup Relatives of HCC patients who entered inpatient care in CGMH, Taipei Veterans General Hospital, and National Taiwan University Hospital Total no. of participants 5,364 3,517 3,496 Eligibility criteria Men ages years HBsAg-positive men and women ages Men and women ages years years Exclusion criteria Previous HCC Previous HCC Previous HCC Enrollment period August 1989 June 1992 Two waves of recruitment: First wave: January 1998 December 2006 August 1988 July 1991; Second wave: August 2003 April 2004 Female, % HBsAg positive, % With anti-hcv data, % (data are available for the second 99.6 wave of recruitment) Anti-HCV positive, % Years of follow-up, % < Median of follow-up 19.3 ( ) 21.4 ( ) 9.6 ( ) time, years (IQR) No. of incident HCC Person-years 99, , ,662.2 Incidence of HCC (1 in 100,000 per year) Abbreviation: IQR, interquartile range. HCC related) in 192,871.8 person-years (detailed information on HCC detection and antiviral treatments is provided in the Supporting Information). Statistical Analysis. Before building the models, we screened all potential predictor variables by Cox s regression to examine their associations with HCC. Hazard ratios (HRs) and 95% confidence intervals (CIs) were calculated as estimates of relative risk. Variables that were statistically significantly associated with HCC in univariate analysis were chosen for use in multivariable regression analysis to construct the final models. Initially, each data set from the three cohorts were analyzed separately to assess the associations between risk factors and HCC to guide the building of the risk model. Because different cohorts shared almost a similar set of risk factors (Supporting Table 1) and the goal was to develop a unified risk-scoring system that would be the best for population risk stratification, we then built final models from the pooled cohort to calculate moreprecise overall relative risk estimates. Two thirds of the pooled cohort of 12,377 subjects were randomly assigned to a model derivation data (training) set (n 5 8,252) and one third (n 5 4,125) were reserved as an independent validation set. These two sets were similar in baseline characteristics and 5- and 10-year risk of HCC (Supporting Table 2). Considering comparing performance between models and the availability of the anti-hcv information, four risk models and corresponding risk score were developed. We started with a simple model (model 1) that included age, sex, and ALT (model 1), followed by incorporation of health history into the model (model 2). Finally, model 2 was extended by adding HBsAg (model 3) or both HBsAg and anti-hcv (model 4). In models 1 and 2, the stratified Cox s proportional hazards model was used to account for differences in the baseline hazard function by strata of HBsAg, which also allows adjustment for its strong effect. Discriminatory accuracy for predicting risk of HCC was assessed by calculating the c-statistic, analogous to the area under the receiver operator characteristic curve. A c-statistic of 0.5 represents no discrimination power, and an index of 1 indicates a perfect discrimination. In addition to using a split-sample technique to evaluate the model performance by focusing on aspects of internal validation, bootstrapping and calibration were also used. Bootstrap-corrected Harrell s c-statistics were quantified after fitting to a Cox s model based on 200 resamples. 13 Coefficients of the Cox s model were converted into integer scores by

4 HEPATOLOGY, Vol. 61, No. 6, 2015 HUNG ET AL dividing each regression coefficient by the coefficient for a 5-year increment in age as the reported procedure, 14 and the 10-year predicted risk of HCC of a given individual was calculated from the Cox s model (computational formula for predicted 10-year risk is provided in the Supporting Information). Calibration of the model was assessed by plotting the predicted versus observed probabilities within 10 years and by Hosmer-Lemeshow s goodness-of-fit test. Scoring metrics developed in the training set were tested in the validation set to evaluate model performance. We used Kaplan-Meier s analysis to determine cumulative risks of HCC by tertiles of risk scores to investigate the impact of changes in risk scores on the population burden of HCC. To assess the clinical usefulness of our prediction models, we also compared estimated net benefits for risk-model score versus other clinical alternatives across all possible risk thresholds by using a decision curve plot. 15 Sensitivity Analysis. To further examine how might the enrichment of study population with a high proportion of HBV-infected people impact on the discriminatory accuracy of the HCC-related risk score, we conducted sensitivity analysis by generating random samples in which we randomly selected HBsAgpositive subjects of variable sizes from all the HBsAgpositive subjects to create populations of 10%, 15%, and 20% HBsAg-positive rate, respectively. At each level of HBsAg-positive rate, we selected 100 random samples and illustrated the dispersion of c-statistics. Results Relative Risk Comparisons. As a result of the cohort structure, there were some differences in riskfactor profiles or magnitude of HRs between the three cohorts. Chronic hepatitis virus infection appears to be the most significant predictor. Shared risk factors for different cohorts were age, chronic liver disease (CLD), high-normal ALT (25 U/L), and smoking 18 pack-years (cut-off point of top tertile). Male sex and first-degree family history of HCC were also associated with increased risk, but no association was observed between high alcohol intake (males, 40 g/day; females, 20 g/day) and HCC (Supporting Table 1). Risk Prediction Models and Discrimination. Table 2 shows HRs and risk scores for each variable derived from the four models. The c-statistic for model 1 was 0.78 (95% CI ; bootstrap-corrected c-statistic ), which did not improve significantly with the addition of health history (c-statistic ; 95% CI ; bootstrap-corrected c-statistic ; P for model 1 vs. model 2). The addition of HBsAg in model 3 significantly increased the c-statistic to 0.85 (95% CI ; bootstrapcorrected c-statistic ; P < for model 2 vs. model 3), but there was no appreciable improvement in the c-statistic when taking into account anti-hcv (P for model 3 vs. model 4). We obtained similar results in all four models whether using the entire or training cohort (n 5 8,252). The model estimates were robust as shown through 1,000 bootstrap replications (Supporting Table 3). Evaluation of Predicted Risk Scores. Using risk scores (calculation in detail is provided in Supporting Table 4), we next measured model fit in the validation cohort. In general, the calibration plots suggest excellent calibration agreement of observed probabilities with predicted probabilities of developing HCC in 10 years in all models (Fig. 1). In addition, the c-statistics for predicting 3-, 5-, and 10-year risk were , , , and , respectively, for models 1-4 in the training cohort; the corresponding figures were , , , and 0.87 in the validation data set. The models were then refit to the pooled cohort database, showing a c-statistic of The discriminative ability of the risk scores from the four models (model 1: ; model 2: ; model 3: ; model 4: ) maintained when the analysis was restricted to each of the three cohorts or subgroups stratified by age, sex, or HBV/HCV status in the pooled cohort (Fig. 2). Sensitivity Analysis. The sensitivity analysis revealed extensive overlap of the values for c-statistics between samples with varying HBsAg-positive rates from 10% to 20%, and reduction in HBsAg-positive proportion from 62.3% (whole cohort) to 10% had little impact on the estimation of c-statistics (Fig. 3). Assessment of Potential for Public Health Application. In models 3 and 4, risk scores were heavily weighted on the effect of HBV/HCV and produced a striking disparity between hepatitis virus infected and uninfected participants, even though a high heterogeneity in risk of HCC is known for hepatitis virus infected individuals. 7,8,10-12 For model 3, a dramatic difference in risk scores also appears between HBV- and HCV-infected participants, even though both groups are at high risk (Supporting Fig. 1). Thus, we restricted following decision curve analysis to models 1 and 2. Figure 4 compares the net benefits of using risk models for stratifying people for screening with those of other clinical alternatives, including the AASLD (American Association of the Study of the Liver Diseases) guidelines for Asian populations with hepatitis

5 1938 HUNG ET AL. HEPATOLOGY, June 2015 Table 2. Characteristics of Participants in the Pooled Cohort and Variables Associated With HCC in Cox s Models Characteristic Model 1 Model 2 Model 3 Model 4 Total No.* No. of HCC* Risk Risk Risk (n 5 12,377) (n 5 387) HR (95% CI) Score HR (95% CI) Score HR (95% CI) Score HR (95% CI) Risk Score Sex Female 3, Male 9, ( ) ( ) ( ) ( ) 3 Age, years Per year 1.08 ( ) 1.08 ( ) 1.08 ( ) 1.07 ( ) , , , , ALT (U/L) <25 8, , ( ) ( ) ( ) ( ) 3 Previous CLD No 11, Yes ( ) ( ) ( ) 3 First-degree family history of HCC No 8, Yes 3, ( ) ( ) ( ) 2 Smoking (pack-year) <18 11, , ( ) ( ) ( ) 1 HBsAg ( ) 4, (1) 7, ( ) 7 HBsAg(1) or Anti-HCV(1) No 4, Yes 7, ( ) 9 Model performance statistics c-statistic (95% CI) 0.78 ( ) 0.79 ( ) 0.85 ( ) 0.85 ( ) Bootstrap-corrected Harrell s c-statistic Discriminative ability was assessed in terms of c-statistic based on the risk equation from the corresponding Cox s model. *Excluding participants with missing values. Stratified Cox s model was used to account for difference in the baseline hazard function for strata of HBsAg. Risk scores were generated from variables regression coefficients of corresponding Cox s models from the training cohort (calculation in detail is provided in Supporting Table 4). Calculated based on 200 data sets using bootstrapping procedures.

6 HEPATOLOGY, Vol. 61, No. 6, 2015 HUNG ET AL Fig. 1. Calibration plots for 10-year prediction for HCC events in the entire cohort (A) and validation cohort (B). The risk score-predicted probabilities of HCC in 10 years were stratified in five equally sized subgroups. For each subgroup, the average predicted probability (x-axis) was plotted against the Kaplan-Meier s probability observed in this series of subjects (y-axis). The 95% CIs of the Kaplan-Meier s estimates are indicated with vertical lines. The dashed line indicates the reference line for ideal model. P values, calculated according to Hosmer-Lemeshow s goodnessof-fit test, indicate that the four models fit well. B (i.e., males 40 years and females 50 years) and the extreme strategies of screening everyone or no one, in overall and in HBsAg-positive participants. The predicted net benefit is the total true-positive classifications minus the total false-positive classifications, weighted by the odds of the risk threshold. 15 This analysis showed that the ability of both models to predict HCC provided a better benefit than all clinical alternatives across 10-year risk thresholds from 2% to 15%. In this range, model 2 offers the highest net benefit, especially for the HBV-infected participants. Cumulative HCC Risk by Risk Score. As expected, there was a step-wise increase in risk score with advancing age, because advancing age confers increased risk for HCC and because of a greater burden of risk factors with aging. Figure 5A displays cumulative HCC risks by tertile of risk scores from model 2 at each index age. In the absence of hepatitis virus, the risks appear to be very low at all ages. For hepatitis virus infected participants, at age under 50, only those in the highest tertile had a 10-year risk of >2%, the lower bound of the threshold probabilities that the model can confer benefits Fig. 2. Indices of discriminative ability for risk scores according to different models for predicting 3- (A), 5- (B), and 10-year (C) HCC risk. Risk scores were generated from variables coefficients of Cox s models in the training cohort and supplied as inputs to analysis of c-statistics in logistic regression to test discrimination in different populations.

7 1940 HUNG ET AL. HEPATOLOGY, June 2015 Fig. 3. Sensitivity analysis of potential effects of different HBsAg-positive rates on indices of discriminative ability for risk scores according to models 1-4 for predicting 5- and 10-year HCC risk. We generated 100 samples at each level of HBsAg-positive proportion, in which subsets of HBsAgpositive subjects were randomly selected from 7,708 HBsAg-positive subjects to combine with all the 4,656 HBsAg-negative subjects when generating populations with HBsAg-positive proportion from 10% to 20%. for screening. At ages 50-59, hepatitis virus infected participants in the middle tertile still had a substantial risk for HCC. At older ages, all three categories of risk score had a 10-year risk of >2% (Fig. 5B). Selection for HCC Screening With the Risk Model Versus AASLD Guideline. In the comparison of using model 2 and threshold of 2% 10-year risk (Fig. 5B) with AASLD criteria for selection of persons for screening, the sensitivities were 89.4% versus 76.8% and the specificities were 67.8% versus 71.8% for HCC detection within 10 years. Notably, model 2 (79.4%) identified 38.8% more of the young-onset HCC cases diagnosed under age 50 years than did the AASLD criteria (40.6%). Similar results were noted across time (Table 3). Discussion Using a pooled cohort, in which 75% of the participants were between the ages of 20 and 49 years at baseline and have been followed over years across the

8 HEPATOLOGY, Vol. 61, No. 6, 2015 HUNG ET AL Fig. 4. Decision curve analysis of the effect of risk scores on the detection of HCC in 10 years in the entire cohort (A) and in the HBsAg(1) participants (B). Risk scores were generated from models 1 and 2. AASLD criteria: Asian male HBV carriers 40 years; female HBV carriers 50 years. predominant age period of HCC onset, we have developed a risk-scoring system based on multivariable analysis in combination with main risk factors (i.e., HBV/ HCV) to identify persons at high risk for HCC and assessed the utility of the scoring metrics at identifying persons for HCC screening in practice, including considerations of actual risks and potential benefits and harms. The models account for important risk factors, which have strong, biologically etiological roles in HCC, and also confirmed previously reported risk factors as independent predictors for the development of HCC, including high-normal ALT, previous CLD, family history of HCC, and cigarette smoking. 7,9,12,16 In addition to model 2 variables, studies have reported increased HCC risk associated with heavy alcohol consumption, but no agreement exists on the dose-effect relation between alcohol intake and HCC. 16,17 We defined high alcohol intake as consumption of ethanol 40 g/day in men and 20 g/day in women, which has been used previously, 16 and observed no association between alcohol and HCC. The relative importance of alcohol consumption in the etiology of HCC may differ appreciably by ethnicity and geographical areas, 18 but a recent largescale Danish cohort study indicated that the 5-year cumulative risk for HCC, even in patients with diagnosed alcoholic cirrhosis, was 1%, and only 1.8% of their deaths were HCC related. 19 The data provide evidence against HCC screening of patients with alcoholic cirrhosis. Although ingestion of aflatoxins may be a cofactor for HBV/HCV-induced carcinogenesis in our study area, no effort in this study was made to assess aflatoxins, mainly because there is no cheap and fast method to conduct large-scale investigation of biomarkers of aflatoxins. Compared to model 1, the addition of family history in model 2 is important because it is one of the main concerns for early screening. The c-statistics attained by risk score according to model 2 range from 0.70 to 0.87 across a range of subgroups (such as HBV 1,HBV, and HCV 1 ), suggesting high and consistent predictive discrimination. The discriminatory power of the model compared well with, or superior to, those reported for existing risk models for other types of cancer or chronic diseases, such as the Gail model for breast cancer, the PLCO (prostate, lung, colorectal, and ovarian) model for lung cancer, and the Framingham Risk Score for cardiovascular disease We used the model because the risk factor data are available for general public, and this makes the model able to be widely applied as an interactive risk assessment tool for effective risk communication. More-complex models might further improve discriminatory ability. However, the c-statistics of model 2 are also within a similar range of the values derived from other published HCC risk models limited to either hepatitis B or hepatitis C, including those with almost all possible viral and clinical variables included We found that the simplest model with age, sex, and ALT showing that an overall c- statistic of >0.7 for predicting HCC across time was remarkable, which is similar to what has been observed in a previous cohort study involving nearly 400,000

9 1942 HUNG ET AL. HEPATOLOGY, June 2015 Fig. 5. Kaplan-Meier s estimates of cumulative HCC risks of participants at ages of 20-39, 40-49, 50-59, and 60 years by tertile of risk scores according to model 2 (A) and subgroups eligible for screening at the threshold of 2% 10-year risk (B). Ten-year risk estimates are shown at the right-hand side of the dashed vertical line of each graph. Abbreviations: NA, not applicable because 2 participants were included; T, tertile of risk score (T1 5 first tertile; T2 5 second tertile; T3 5 third tertile). Taiwanese people from a private health management institution. 9 Thus, the use of factors for predicting HCC in the present study has probably approached a limit. Risk-predictive models with good discriminatory accuracy are potentially valuable for identifying a highrisk population for screening, but high discrimination is easier to attain in data that are heterogeneous with regard to risk, irrespective of risk level. Actually, c-statistic does not address the potential value of a predictive model where the main aim is stratification into meaningful risk categories linking clinical decision making. We showed that our developed risk score, which reflect aggregate levels of risk-factor burden, helped to stratify the population into diverse risk

10 HEPATOLOGY, Vol. 61, No. 6, 2015 HUNG ET AL Table 3. Predictive Accuracy of HCC Classification 5-Year Follow-up 10-Year Follow-up Whole Follow-up Period Selection Criteria for Screening Criteria Positive Criteria Negative Criteria Positive Criteria Negative Criteria Positive Criteria Negative Risk-scoring system* Young-onset HCC 76.9% TPR 23.1% FNR 79.4% TPR 20.6% FNR 78.9% TPR 21.1% FNR Late-onset HCC 98.4% TPR 1.6% FNR 93.9% TPR 6.1% FNR 91.2% TPR 8.8% FNR Sensitivity 90.2% 89.4% 88.2% Specificity 67.3% 67.8% 68.5% AASLD criteria Young-onset HCC 48.7% TPR 51.3% FNR 40.6% TPR 59.4% FNR 30.8% TPR 69.2% FNR Late-onset HCC 92.4% TPR 7.6% FNR 92.9% TPR 7.1% FNR 87.2% TPR 12.8% FNR Sensitivity 76.2% 76.8% 73.9% Specificity 71.4% 71.8% 72.4% There were 91 young-onset (<50 years) and 296 late-onset cases with HCC identified during follow-up. *Calculating risk scores was based on model 2. Participants who were criteria positive were those who had a 10-year risk of >2%, as defined according to the threshold probability for initiating screening. AASLD selection criteria for screening for Asian hepatitis B carriers: males 40 years; females 50 years. Abbreviations: TPR, true-positive rate; FNR, false-negative rate. categories. Nevertheless, the actual risks were very low in the absence of hepatitis virus infection. It was also found that participants with HBV/HCV and risk scores in the highest tertile had a rate for HCC that was substantially greater than the group with HBV/HCV, but risk scores in the lowest tertile (10-year risk: 6.5% vs. 0.2%). These results reinforce the influence of main risk factors on the risks for HCC and also strengthen the importance of main risk factors in combination of risk scores for morefocused management recommendations. Accordingly, we sought to combine main risk factors and risk scores into a single clinically relevant method of risk stratification, which also takes into account risk heterogeneity in hepatitis virus infected people in delivering screening programs. Risk assessment can be performed by a two-step procedure: one is main risk-factor assessment and the other is risk-score calculation. Model 2 is valuable for effective risk communication, and calculating risk score can be Web- or paper-based. We propose that individuals with a high risk score should be properly informed in a timely manner for assessment of main risk factors (i.e., HBV/HCV); and for individuals known to have a main risk factor, including those in areas where hepatitis B is endemic and there is a surveillance system of HBV/ HCV, the next step is to carry out 10-year risk assessment, as measured by our age-specific scoring system (Fig. 5), to identify individuals whose 10-year risk warrants consideration of screening. In practice, the framework of screening also involves deciding what level of risk should be considered for screening. Unlike CVD, cardiovascular disease, for which a 10-year risk of 20% has been recommended for preventive interventions, no consensus on the threshold risk is available in cancer screening, although, in general, a 10-year risk of 2% is considered to be at high risk, 21,26 andthisisalsothecaseforhcc. 5 It was suggested that screening became cost-effective once the incidence of HCC exceeds 0.2% per year (or 10-year risk: 2%). 5 Therefore, routine HCC screening was recommended for HBV carriers to start at ages 40 and 50 in Asian men and women, respectively. 5 However, if we assume that the risk threshold at which the screening program might have benefits, it makes sense to offer screening to all individuals with that level of risk, irrespective of age. Our findings showing that high-risk scores in young hepatitis virus infected adults <40 years of age produced a relatively high cumulative risk, especially those with an affected first-degree relative, and argue against the practice based on the current age criterion. According to decision curve analysis, we chose the 10- year risk of 2% as the plausible threshold probabilities for initiating HCC screening. Detailed calculation of sensitivity and specificity indicate that our scoring system selected 13% more persons who received a diagnosis of HCC in 10 years for screening than did the AASLD criteria 5 under slight reduction in specificity. Most important, our model would include 80% of HCC patients diagnosed under age 50 for screening, whereas the corresponding figure would be only 40% for the AASLD criteria. The derivation and assessment of risk prediction models using multiple cohorts with varying levels and distributions of risk factors might reduce overfitting and increase generalizability, in comparison with analysis of single cohorts. 27 The age range of our study population was 75% between ages 20 and 49 at recruitment, which might be younger than most other cohorts that have been reported on. 23,24 However, despite it being developed in a younger population, our model was equally accurate to predict HCC while analysis was limited to participants 40 years. Our study subjects were selected to enrich for HBV-infected participants, resulting in a

11 1944 HUNG ET AL. HEPATOLOGY, June 2015 higher average incidence of HCC than many other populations. 1 To address this concern, we have attempted to construct models using the stratified Cox s regression accounting for difference in the baseline hazard function between those with and without HBV. Indeed, the discriminative ability of model 2 was reasonably well in HBsAg-negative participants, in which the major determinant of HCC is HCV. In this sense, our approach might be applicable to external populations with variable prevalence of HBV infection. Furthermore, we have conducted sensitivity analysis by generating random samples with varying HBsAg-positive rates to clarify the potential impacts of our enrichment strategy with a high proportion of HBV-infected people, and the results revealed that the discriminative abilities of different models are not sensitive to change in HBsAg-positive rates. In addition to older age groups, we have provided data about how to stratify risk for HCC among persons <40 years in order to target screening to those at high risk. Our risk score, which was based on widely accessible variables, combined with HBV/HCV status and family history, was more sensitive at identifying persons for HCC screening, as compared with current expert guidelines. The predictor variables we used for construction of the risk-scoring system are common risk factors underlying most HCCs in the world. Future studies should test this scoring system in other populations. References 1. Venook AP, Papandreou C, Furuse J, de Guevara LL. The incidence and epidemiology of hepatocellular carcinoma: a global and regional perspective. Oncologist 2010;15(Suppl 4): Zhang BH, Yang BH, Tang ZY. Randomized controlled trial of screening for hepatocellular carcinoma. J Cancer Res Clin Oncol 2004;130: Chen TH, Chen CJ, Yen MF, Lu SN, Sun CA, Huang GT, et al. Ultrasound screening and risk factors for death from hepatocellular carcinoma in a high risk group in Taiwan. Int J Cancer 2002;98: Davila JA, Morgan RO, Richardson PA, Du XL, McGlynn KA, El- Serag HB. Use of surveillance for hepatocellular carcinoma among patients with cirrhosis in the United States. HEPATOLOGY 2010;52: Bruix J, Sherman M. American Association for the Study of Liver Diseases. Management of hepatocellular carcinoma: an update. HEPATO- LOGY 2011;53: Shih WL, Chang HC, Liaw YF, Lin SM, Lee SD, Chen PJ, et al. Influences of tobacco and alcohol use on hepatocellular carcinoma survival. Int J Cancer 2012;131: Yu MW, Chang HC, Liaw YF, Lin SM, Lee SD, Liu CJ, et al. Familial risk of hepatocellular carcinoma among chronic hepatitis B carriers and their relatives. J Natl Cancer Inst 2000;92: Freeman AJ, Dore GJ, Law MG, Thorpe M, Von Overbeck J, Lloyd AR, et al. Estimating progression to cirrhosis in chronic hepatitis C virus infection. HEPATOLOGY 2001;34: Wen CP, Lin J, Yang YC, Tsai MK, Tsao CK, Etzel C, et al. Hepatocellular carcinoma risk prediction model for the general population: the predictive power of transaminases. J Natl Cancer Inst 2012;104: Yu MW, Shih WL, Lin CL, Liu CJ, Jian JW, Tsai KS, et al. Body-mass index and progression of hepatitis B: a population-based cohort study in men. J Clin Oncol 2008;26: Huang HH, Shih WL, Li YH, Wu CF, Chen PJ, Lin CL, et al. Hepatitis B viraemia: its heritability and association with common genetic variation in the interferon gamma signalling pathway. Gut 2011;60: Yu MW, Chang HC, Chen PJ, Liu CJ, Liaw YF, Lin SM, et al. Increased risk for hepatitis B-related liver cirrhosis in relatives of patients with hepatocellular carcinoma in northern Taiwan. Int J Epidemiol 2002;31: 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: Sullivan LM, Massaro JM, D Agostino RB, Sr. Presentation of multivariate data for clinical use: The Framingham Study risk score functions. Stat Med 2004;23: Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making 2006;26: Trichopoulos D, Bamia C, Lagiou P, Fedirko V, Trepo E, Jenab M, et al. Hepatocellular carcinoma risk factors and disease burden in a European cohort: a nested case-control study. J Natl Cancer Inst 2011;103: Donato F, Tagger A, Gelatti U, Parrinello G, Boffetta P, Albertini A, et al. Alcohol and hepatocellular carcinoma: the effect of lifetime intake and hepatitis virus infections in men and women. Am J Epidemiol 2002;155: Welzel TM, Graubard BI, Quraishi S, Zeuzem S, Davila JA, El-Serag HB, et al. Population-attributable fractions of risk factors for hepatocellular carcinoma in the United States. Am J Gastroenterol 2013;108: Jepsen P, Ott P, Andersen PK, Sørensen HT, Vilstrup H. Risk for hepatocellular carcinoma in patients with alcoholic cirrhosis: a Danish nationwide cohort study. Ann Intern Med 2012;156: Mealiffe ME, Stokowski RP, Rhees BK, Prentice RL, Pettinger M, Hinds DA. Assessment of clinical validity of a breast cancer risk model combining genetic and clinical information. J Natl Cancer Inst 2010;102: Tammem agi MC, Katki HA, Hocking WG, Church TR, Caporaso N, Kvale PA, et al. Selection criteria for lung-cancer screening. N Engl J Med 2013;368: D Agostino RB, Sr., Grundy S, Sullivan LM, Wilson P; CHD Risk Prediction Group. Validation of the Framingham coronary heart disease prediction scores: results of a multiple ethnic groups investigation. JAMA 2001;286: Lee MH, Yang HI, Liu J, Batrla-Utermann R, Jen CL, Iloeje UH, et al. Prediction models of long-term cirrhosis and hepatocellular carcinoma risk in chronic hepatitis B patients: risk scores integrating host and virus profiles. HEPATOLOGY 2013;58: Wong VW, Chan SL, Mo F, Chan TC, Loong HH, Wong GL, et al. Clinical scoring system to predict hepatocellular carcinoma in chronic hepatitis B carriers. J Clin Oncol 2010;28: Hoshida Y, Villanueva A, Sangiovanni A, Sole M, Hur C, Andersson KL, et al. Prognostic gene expression signature for patients with hepatitis C-related early-stage cirrhosis. Gastroenterology 2013;144: Pharoah PD, Antoniou AC, Easton DF, Ponder BA. Polygenes, risk prediction, and targeted prevention of breast cancer. N Engl J Med 2008;358: Royston P, Parmar MK, Sylvester R. Construction and validation of a prognostic model across several studies, with an application in superficial bladder cancer. Stat Med 2004;23: Author names in bold designate shared co-first authorship. Supporting Information Additional Supporting Information may be found at onlinelibrary.wiley.com/doi/ /hep.27610/suppinfo.