<scp>HBV</scp> ‐Specific T Cell Responses for Short‐Term Hepatitis Progression Prediction: A Comparative, Interpretable Machine Learning Study

BACKGROUND & AIMS: As central determinants of viral control and immunopathology, hepatitis B virus (HBV)-specific T-cell responses provide more clinically meaningful information than routine biochemical and virological markers. Nevertheless, most existing prediction models rely primarily on clinical parameters and rarely incorporate HBV-specific T-cell responses, potentially limiting their predictive performance and biological interpretability. To develop an accurate and interpretable model for predicting hepatitis progression, we evaluated diverse machine learning (ML) methods integrating multisource data. METHODS: We enrolled a cohort of 479 patients and divided them into training and testing cohorts based on admission time. Clinical data, treatment regimens, and HBV-specific T-cell immune responses were collected. A comprehensive benchmarking of 10 ML models was conducted through 5-fold cross-validation on the training cohort across varying feature combinations and algorithmic performances. The ML models were independently evaluated on the testing cohort, and significant predictive factors were identified. RESULTS: Based on alanine aminotransferase (ALT) levels at 6- and 12-month follow-ups, the patients were stratified into hepatitis (ALT > 40 U/L) and non-hepatitis (ALT ≤ 40 U/L) groups. The predictive performance was significantly improved after integrating clinical indicator features (CIF) and HBV-specific T-cell features (STCF). An XGBoost model based on six selected features (three CIF and three STCF) demonstrated especially robust performance, achieving AUCs of 0.874 (validation) and 0.880 (testing) at 6 months follow up and 0.851 (validation) and 0.845 (testing) at 12 months. To facilitate clinical application, a web-based tool was developed for personalized risk assessment. CONCLUSIONS: Incorporating HBV-specific T-cell responses into the predictive framework improves the prediction of disease progression among HBV-infected individuals. The developed interpretable model is a potentially valuable tool for early risk stratification, facilitating proactive monitoring.