A Multimodal and Explainable Machine Learning Approach to Diagnosing Multi-Class Ejection Fraction from Electrocardiograms

Researchers from Hartford HealthCare and MIT have unveiled a multimodal machine learning framework that diagnoses left ventricular ejection fraction (LVEF) from standard 12-lead electrocardiograms (ECGs) and electronic health record (EHR) data. The model, based on XGBoost, classifies LVEF into four clinically relevant strata: normal (>50%), mildly reduced (40-50%), moderately reduced (30-40%), and severely reduced (<30%). It was trained on 36,784 ECG-echocardiogram pairs from 30,952 outpatients and validated on 19,966 ECGs from a subsequent period, demonstrating robust temporal generalizability. The multimodal approach achieved one-vs-rest AUROCs of 0.95 for severe reduction, 0.92 for moderate, 0.82 for mild, and 0.91 for normal, outperforming both ECG-only and EHR-only baselines.

Explainability is a key feature: using SHAP (SHapley Additive exPlanations) attributions, the model identifies the most influential ECG waveform features and EHR variables driving each prediction. This addresses a critical barrier to clinical adoption of AI in cardiology. The findings suggest that ECG-based LVEF stratification could serve as a practical, low-cost screening tool in primary care and resource-constrained settings, triaging patients for confirmatory echocardiography. The study, published on arXiv (arXiv:2604.25942), represents a significant step toward democratizing heart failure diagnosis without specialized imaging equipment.