Explainable AI for automatic heart disease diagnosis using 3DFMMecg features: A novel ECG-based approach

The electrocardiogram (ECG), a gold standard in cardiac diagnostics, is increasingly combined with Artificial Intelligence (AI) methods to enhance its clinical utility. However, many recent studies have prioritised performance over clinical interpretability by focusing on Deep Learning (DL) techniques, which offer limited explainability since they do not directly correlate with clinical features. This lack of transparency causes mistrust among physicians and hinders adoption in daily clinical practice. We developed novel, self-explainable features from the 3DFMMecg model parametrisation, enabling highly accurate and clinically interpretable ML classifiers for cardiovascular pathology diagnosis from 12-lead ECG signals. We evaluated our approach on PTB-XL+, a widely used dataset of annotated ECG recordings. Our framework outperforms existing feature-based methods in four out of six classification tasks, achieving macro-AUC between 0.88 and 0.95 and weighted macro-AUC between 0.90 and 0.95, comparable to and in some tasks surpassing DL approaches. We further show that the model maintains high diagnostic accuracy when using only three of the standard twelve ECG leads, with less than 6% loss in performance, enabling deployment in mobile, wearable, or resource-constrained environments. Feature importance analyses using SHapley Additive exPlanations (SHAP) confirm strong alignment between model predictions and established clinical markers, such as QRS width and T-wave amplitude parameters. These results underscore the potential of 3DFMMecg-based pipelines toward reliable, transparent, and accessible ECG-based diagnostic systems.