Comparative Performance of 14 HCC Prediction Models in CHB: A Dynamic Validation at Serial On-Treatment Timepoints.

Abstract

To assess comparative performance of 14 hepatocellular carcinoma (HCC) prediction models in chronic hepatitis B (CHB) patients using on-treatment values at different timepoints. Based on a nationwide prospective cohort of 986 treatment-naive CHB patients undergoing entecavir therapy with every 26-week follow-up, 14 HCC risk scores were calculated using on-treatment values at week 26, 52, 78, and 104, respectively. Model performance predicting 3-year HCC was assessed using time-dependent area under the receiver operating characteristic curve (AUC) and calibration index. Model cutoffs were validated through common diagnostic accuracy measures. During median 4.7-year follow-up, 56 (7.5%) developed HCC. Discrimination using on-treatment values within first 2 years was generally acceptable for most models (AUCs ranging from 0.68 to 0.81), except for REACH-B, NGM-HCC, and PAGE-B, although AUCs slightly decreased from week 26 to 104. Of these, REAL-B, CAMD, GAG-HCC, AASL-HCC, LSM-HCC, mPAGE-B, and mREACH-BII showed highest discrimination with AUCs ranging from 0.76 to 0.81, 0.72 to 0.76, 0.70 to 0.76, and 0.71 to 0.74 when reassessment at week 26, 52, 78, and 104, respectively. With reassessment within first 2 years, both REAL-B and CAMD calibrated well (Brier score ranging from 0.037 to 0.052). Of 9 models reporting cutoffs, REAL-B, AASL-HCC, and mPAGE-B using on-treatment values could identify 30%-40% of patients as low risk with minimal HCC incidence in the low-risk group (0.40% [REAL-B]-1.56% [mPAGE-B]). In this undergoing antiviral treatment CHB cohort, most HCC prediction models performed well even using on-treatment values during first 2 years, particularly REAL-B, AASL-HCC, CAMD, and mPAGE-B model.