Development and prospective validation of a machine learning model for risk stratification of drug-induced liver injury using real-world clinical data.

Background: Drug-induced liver injury (DILI) is difficult to diagnose and manage in routine care because it lacks pathognomonic biomarkers and is often recognized only after clinically meaningful injury has occurred. Existing computational approaches are largely drug-centric and do not routinely incorporate patient-level clinical data available in electronic health records (EHRs). Objectives: To develop and temporally validate a machine learning model for episode-level risk stratification of Roussel Uclaf Causality Assessment Method (RUCAM)-defined DILI using routinely available baseline clinical data. Design: This was an observational cohort study conducted at Hai Phong International Hospital using linked EHR, laboratory, and pharmacy data. The final labeled cohort was partitioned chronologically at the patient level into a retrospective development cohort (2019-2023) and a temporally subsequent prospective validation cohort (2024-2025). Methods: Eligible drug-exposure episodes with complete baseline liver biochemistry and key exposure covariates were included. Analysis-ready episodes were monitored for biochemical liver injury triggers, and trigger-positive episodes underwent clinical review and RUCAM adjudication. DILI was defined as RUCAM ⩾6. Predictors were limited to baseline demographics, comorbidities, laboratory values, drug-exposure features, and FDA DILIrank 2.0 metadata. Candidate models included logistic regression, elastic-net logistic regression, random forest, ExtraTrees, XGBoost, and LightGBM. Results: Among 5095 eligible episodes from 3579 patients, 2786 episodes from 2712 patients were analysis-ready after exclusions. The final labeled cohort comprised 274 DILI-positive and 2512 non-DILI episodes. The prospective validation cohort included 828 episodes, of which 108 (13.0%) were DILI-positive. Tree-based ensemble models outperformed regression-based models. Logistic regression achieved an area under the receiver operating characteristic curve (AUROC) of 0.777 and an area under the precision-recall curve (PR-AUC) of 0.349, whereas the final LightGBM model achieved an AUROC of 0.965 (95% CI 0.942-0.983), a PR-AUC of 0.903 (95% CI 0.856-0.943), and a Brier score of 0.034. Conclusion: A prospectively validated machine learning model using routinely collected baseline clinical data showed excellent performance for DILI risk stratification and may strengthen hospital pharmacovigilance.