Modeling 5-Year Hepatocellular Carcinoma Risk in Alaska Native Peoples With Hepatitis B Virus Infection.

Background and Aims: Modeling hepatocellular carcinoma (HCC) risk in Alaska Native (AN) peoples with chronic hepatitis B virus (HBV) infection is important for risk stratification and surveillance. Existing HCC risk prediction models use baseline characteristics ascertained at the time of HBV diagnosis, rather than predicting HCC risk within 5 years of a relevant time point (such as a clinic visit), and do not include the HBV genotype (GT). We aimed to develop an HCC risk prediction model that addresses these limitations. Methods: We used longitudinal data from a cohort of 1163 AN peoples with HBV. We considered age, sex, GT, serum alpha fetoprotein (AFP), along with serum alanine transaminase, albumin, aspartate aminotransferase, bilirubin, hepatitis B-e-antigen, platelet count, and fibrosis 4 score. To build a 5-year risk model, we structured the longitudinal data into multiple 5-year segments, using AFP as the landmark biomarker. We used the generalized estimation equation approach as well as the Random Forest approach to build risk prediction models. Results: Among the 11 predictors included in our final models, AFP was the most important followed by platelet count and GT. Based on cross-validation, the generalized estimation equation model had an area under the receiver operating characteristic curve of 0.81, with 46.5% sensitivity at 90% specificity for 5-year HCC risk prediction. The Random Forest model was superior with an area under the receiver operating characteristic curve of 0.88 and 70% sensitivity at 90% specificity, outperforming the PAGE-B, mPAGE-B, REACH-B and REAL-B models. Conclusion: We developed an HCC risk prediction model using rich information from different time points in a patient's disease trajectory. Our model can accurately estimate HCC risk at different time points during follow-up for risk stratification and risk-based surveillance.