Interpretable Machine Learning Framework for Predicting Major Adverse Cardiovascular Events in Rheumatoid Arthritis Using Electronic Health Records: Multicenter Cohort Study.

BACKGROUND: Patients with rheumatoid arthritis (RA) face higher risks of major adverse cardiovascular events than the general population. While machine learning offers powerful predictive capabilities, its clinical adoption is hindered by the "black-box" nature of complex algorithms. OBJECTIVE: This study aimed to develop interpretable survival models to predict the risk of major adverse cardiovascular events in patients with RA, providing transparent and actionable insights for personalized clinical prognosis management. METHODS: Using data from the Taipei Medical University Clinical Research Database (2011-2022) for 2461 patients with RA, machine learning survival models, including random survival forest (RSF), DeepSurv, and Cox-Time, were compared with the traditional Cox proportional hazards model. Performance was evaluated using the C-index and integrated Brier score. Permutation importance and Shapley additive explanations (SHAP) analyses were integrated to provide explainability for individual-level risk predictions. RESULTS: RSF demonstrated superior performance, achieving a C-index of 0.8771 and an integrated Brier score of 0.0775. Permutation importance identified key features, including creatinine, conventional synthetic disease-modifying antirheumatic drugs, C-reactive protein, alanine aminotransferase, and age at RA diagnosis. SHAP analysis further quantified feature-specific effects, revealing both protective and risk-increasing associations between medications and laboratory indicators. CONCLUSIONS: RSF outperformed traditional methods, and integrating SHAP enabled transparent, personalized risk interpretation, translating complex models into actionable insights for clinicians. This approach empowers clinicians to identify high-risk individuals and advances precision medicine in rheumatology. Future work should include temporal validation using data from later years and external validation using datasets from other health care systems to further assess model generalizability.