Interpretable Deep Learning Framework for Real-Time ECG Classification and Diagnosis

Automatic arrhythmia detection requires effective and clear models because cardiovascular disease is increasingly common. Convolutional neural networks (CNNs) have proven to be the most effective at ECG classification, but their black box nature limits their practical applications. This paper presents an explainable deep learning framework with Grad-CAM for auxiliary supervision to retrace the decision paths in ECG classification using CNNs. A 1D CNN was trained with the MIT-BIH Arrhythmia Database, where ECG signals were classified as normal sinus rhythm (NSR), atrial fibrillation (AFib), premature ventricular contractions (PVC), left bundle branch block (LBBB), and right bundle branch block (RBBB). The model’s classification accuracy was 98.2%, which indicates high reliability. With Grad-CAM, to the model’s focus and relevant clinical features of the ECG, important parts of the ECG were steered by QRS complex and intervals between beats, thus increasing trust in the deep learning systems. These findings assist in understanding the impact of XAI in the automation of medical decision-making processes. Attention and hybrid approaches will be the focus of other studies to improve the explainability and diagnostic accuracy of the systems.