Comparable Predictive Performance of Non-VCTE–Based Machine Learning Model for HBV-Infected Hepatocellular Carcinoma Risk

Abstract

Background

Accurate risk stratification for hepatocellular carcinoma (HCC) among chronic hepatitis B (CHB) patients remains challenging. Vibration-controlled transient elastography (VCTE) is useful for HCC prediction; however, its nationwide use might be limited in CHB-endemic or resource-limited regions. We developed and validated non-VCTE–based machine learning (ML) models for HCC prediction, integrating routine clinical and laboratory parameters.

Methods

We analyzed a multicenter cohort of CHB patients receiving entecavir or tenofovir. ML-based survival models were developed and validated, with their performance compared both with and without VCTE-derived parameters (liver stiffness measurement/controlled attenuation parameter) to analyze VCTE's role amidst routine variables. Non-VCTE–based ML models were then compared against the modified PAGE-B (mPAGE-B) score. Model performance was assessed by Uno's C-index, area under the ROC curve (AUC), and Brier score at 5-, 6-, and 7-year follow-up using cross-validation and external validation.

Results

Among 2736 patients (training n = 1954; validation n = 782), no statistically significant difference in predictive performance was observed between ML-based models with and without VCTE-derived parameters. Non-VCTE–based ML models consistently demonstrated superior overall performance and a statistically significant difference compared to the mPAGE-B score. The non-VCTE–based RSF model demonstrated the highest discrimination (5-year C-index: 0.821; Brier score: 0.0592) in external validation. Statistically significant improvements over the mPAGE-B score were observed across all metrics except for rpCox's AUC up to 5 years.

Conclusions

ML-based models provide superior long-term HCC risk prediction in CHB patients compared to the mPAGE-B score. Importantly, their performance remains robust even without VCTE-derived parameters, suggesting that effective risk stratification is achievable across diverse clinical settings, especially CHB-endemic or resource-limited regions.