A Novel Noise-Resilient and Explainable Machine Learning Framework for Accurate and Robust ECG-Based Heart Disease Diagnosis

An electrocardiogram (ECG) is essential for diagnosing cardiac abnormalities. Automated heartbeat classification enables continuous heart monitoring and early diagnosis. Current machine learning-based methods for heart disease diagnosis from ECG face challenges such as sensitivity to noise, limited adaptivity across different patients, and computational complexity, hindering their adoption. This study introduces a statistical shape model-based method for classifying different heartbeat types: normal, myocardial infarction, premature ventricular contraction, and right bundle branch block. The innovation of this approach lies in its adaptability to shape variability of ECG signals across different patients, as well as its noise-resilience and computational efficiency, allowing for real-time, precise diagnosis of cardiac conditions across diverse ECG morphologies. The model performance was validated on a dataset of 270 patients. It exhibited a strong performance with recall (97.49%), precision (97.73%), F1 Score (97.58%), and accuracy (98.63%) using a support vector machine classifier. When tested with varying levels of signal-to-noise ratio between 12 and -6 dB, the model maintained robust performance with recall (95.49%), precision (95.57%), F1 Score (95.49%), and accuracy (97.44%), demonstrating noise-resilience. Given its high accuracy, robustness, and computa-tional efficiency, this method is well-suited for real-time ECG-based diagnosis and continuous heart monitoring and can be implemented on battery-operated wearables.