Multimodal Machine Learning for Automated Angina Pectoris Recognition Using MIMIC-IV Clinical Data and ECG Signals

Abstract

This study presents a methodologically rigorous machine learning pipeline for automated recognition of angina pectoris using the MIMIC-IV electronic health record database. Three systematic errors common in prior work — incorrect ICD-10 cohort selection, data leakage through population-level imputation, and cardiac patient contamination of the control group — were identified and corrected, yielding a balanced cohort of 58,486 hospital admissions. Six supervised learning algorithms were evaluated on 34 clinical features. Gradient Boosting achieved the best performance on tabular data (ROC-AUC = 0.797, Recall = 0.818). A multimodal architecture integrating 12-lead ECG signals with tabular biomarkers improved Recall to 0.846. SHAP analysis identified age, serum creatinine, and Troponin T as the primary clinical predictors