Abstract
PURPOSE: Electrocardiogram (ECG) is the most commonly used non-invasive diagnostic tool for cardiovascular diseases. However, its interpretation requires significant expertise, which may not be available across all medical specializations. Automated detection of heart abnormalities can assist clinicians in making accurate and efficient diagnoses. This study aims to develop an explainable machine-learning framework for classifying myocardial infarction (MI) and other heart abnormalities using ECG images. METHODS: Publicly available ECG images from health centers in South Asia were used for this study. The dataset consists of four classes: Normal, Abnormal, Myocardial Infarction (MI), and Previous History of MI. The ECG images were preprocessed and transformed into time-series signals for efficient computation. These signals were then used to train an ensemble model, XGBoost, and a deep learning model, Convolutional Neural Network (CNN). To enhance learning of both spatial and temporal features, we proposed CNNBoost, a Multilevel Explainable Stacked Ensemble model where CNN-extracted spatial features were concatenated with time-series data and processed by XGBoost. SHapley Additive Explanations (SHAP) were used to identify diagnostically significant ECG leads, validated by a senior cardiologist. RESULTS: The proposed framework achieved 99% accuracy, 98.58% AUC, and 96.47% AUPRC, demonstrating effective MI classification. The explainability component identified critical ECG leads, reinforcing model trustworthiness. CONCLUSION: By integrating deep learning and ensemble learning, our approach enhances ECG classification while ensuring clinical relevance. CNNBoost's ability to learn spatial and temporal features improves interpretability and decision support. Cardiologist validation confirms its potential for real-world healthcare applications, reducing misdiagnoses and aiding clinical decision-making.