Multimodal ECG-PPG Clinical Fusion for Myocardial Infarction Classification Using Ensemble Learning

Abstract

This study presents a comparative analysis of multimodal ECG, PPG, and clinical feature fusion for myocardial infarction (MI) classification using four ensemble learning algorithms: Random Forest, XGBoost, LightGBM, and CatBoost. The experiments were conducted in two classification scenarios: binary classification for normal vs. MI and multi-class classification for normal, STEMI, NSTEMI, and old MI. Five feature scenarios were evaluated, including clinical-only, ECG-only, PPG-only, ECG + PPG, and ECG + PPG + clinical. The results show that ECG features were the most dominant modality for MI classification. In binary classification, XGBoost with ECG-only features achieved perfect performance with accuracy, macro F1-score, macro recall, and MCC of 1.0000. For multi-class classification, the best result was obtained by CatBoost using ECG + PPG + clinical features, achieving an accuracy of 0.9000, a macro F1-score of 0.5394, and an MCC of 0.6912. These findings indicate that multimodal fusion is more beneficial for MI subtype classification, while ECG-only features are highly effective for binary MI detection