Interpretability-driven feature engineering enables compact and accurate multi-outcome risk stratification at hospitalization for chronic hepatitis B

Clinical artificial intelligence (AI) often trades predictive performance against interpretability. This study develops an interpretability-driven feature engineering framework for multi-outcome risk stratification at hospitalization for chronic hepatitis B. A total of 36,024 hospitalization records from 22,001 patients were included. The framework integrating SHapley Additive exPlanations (SHAP)-based main and interaction effects was used to construct compact machine learning models for classifying disease non-progression, liver cirrhosis (LC), and hepatocellular carcinoma (HCC). SHAP main effects selected the top 15 features per outcome; TreeSHAP interactions selected four pairwise terms. Performance was evaluated using patient-level data splitting, temporal validation, and independent external validation, with calibration and decision-curve analyses. The full-feature models achieved superior discrimination (AUCs: 0.927, 0.897, and 0.899) for classifying hospitalization-recorded disease non-progression, LC, and HCC, respectively. The corresponding compact models maintained strong discrimination (AUCs: 0.911, 0.882, and 0.890) with 80.8% fewer features. Interaction models yielded subtle AUC shifts (AUCs: 0.912, 0.881, and 0.891) while characterizing potential clinical joint-effect patterns. Temporal validation yielded AUCs of 0.863, 0.827, and 0.847, while external validation yielded AUCs of 0.829, 0.865, and 0.827, respectively. The framework produced compact and transparent models for contemporaneous risk stratification in clinical practice, seamlessly uniting predictive accuracy with multi-level interpretability.