A machine learning model for early warning of treatment failure in autoimmune hepatitis: a 10-year real-world study

This is the authors' abstract. We don't add key points for this paper.

Autoimmune hepatitis (AIH) can progress rapidly to end-stage liver disease without timely remission. Interpatient heterogeneity and severe presentations often preclude standard immunosuppression, increasing early clinical deterioration in hospitalized patients even with standard care. We developed and independently validated a machine-learning model to predict treatment failure risk upon admission. We conducted a dual-cohort study comprising a retrospective discovery cohort (n = 314, 2017–2025) and a consecutive temporal validation cohort (n = 66, 2025–2026). Following ensemble-based feature selection, 13 machine-learning algorithms were evaluated. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) and benchmarked against the conventional Model for End-Stage Liver Disease (MELD) and Child-Pugh score. Seven predictors were identified: alanine aminotransferase (ALT), total bile acid (TBA), Child-Pugh score, immunoglobulin M (IgM), total bilirubin (TBil), alkaline phosphatase (ALP), and gamma-glutamyl transferase (GGT). The Gradient Boosting Machine (GBM) achieved superior discrimination (AUC = 0.762), significantly outperforming MELD (AUC = 0.555) and Child-Pugh (AUC = 0.611) scores, with satisfactory calibration and clinical utility. Predictive efficacy was maintained in the temporal validation cohort (AUC = 0.755). SHAP analysis revealed a dual biological trade-off: pronounced inflammatory activity (elevated ALT) inversely correlated with treatment failure risk, whereas severe cholestasis and impaired hepatic reserve drove adverse outcomes. This non-invasive GBM model quantifies the complex clinical interplay among active inflammation, cholestatic severity, and depleted hepatic reserve. Overcoming limitations of static scoring systems, it provides a reliable, actionable tool for personalized risk stratification and preemptive therapeutic decision-making in heterogeneous AIH patients upon index admission. When patients are hospitalized for autoimmune hepatitis, it is often difficult for doctors to predict who will worsen despite receiving standard treatments. We developed an artificial intelligence tool that uses routine blood tests to calculate a patient’s risk of treatment failure the moment they are admitted to the hospital. This early warning system allows doctors to quickly identify high-risk patients and personalize their care with more effective or intensive therapies right from the start.