Interpretable machine learning enables early and accurate detection of drug‐induced liver injury: A multicenter study with real‐world clinical translation

Abstract Drug‐induced liver injury (DILI) remains a major clinical challenge due to the absence of specific biomarkers and dependence on subjective diagnostic criteria. This study presents an interpretable gradient‐boosting decision tree model (XGB‐D) that uses routine laboratory data to enable early and accurate DILI detection. Developed through a multicenter cohort of 36,199 patients, XGB‐D shows superior diagnostic performance (area under the curve [AUC] = 0.971) compared with conventional methods and demonstrates robust generalizability across three independent validation cohorts (AUC = 0.881–0.935). SHAP (SHapley Additive exPlanations) analysis identifies alanine aminotransferase and C‐reactive protein as key contributors, revealing mechanistic links between hepatocellular damage and inflammatory responses. In prospective real‐world monitoring, XGB‐D detected DILI signals 2–4 weeks earlier than expert assessment in 29.4% of cases, supporting timely clinical intervention. By integrating interpretable machine learning with clinical hepatology, this work establishes a scalable and transparent framework for precision toxicology with significant implications for drug safety evaluation and personalized medicine.