OPENING THE BLACK BOX: AN EXPLAINABLE DEEP LEARNING APPROACH FOR ECG-BASED ARRHYTHMIA CLASSIFICATION

Deep learning techniques have demonstrated remarkable performance in electrocardiogram (ECG)–based arrhythmia classification; however, their adoption in clinical practice remains limited due to the lack of transparency and interpretability in model decision-making. Most existing approaches function as black boxes, offering minimal insight into how diagnostic conclusions are derived from ECG signals. To address this limitation, this study presents an explainable deep learning framework for ECG-based arrhythmia classification that combines reliable predictive performance with clinically meaningful interpretation. The proposed framework employs an end-to-end neural architecture integrating one-dimensional convolutional layers and recurrent modelling to automatically learn discriminative temporal–morphological representations from segmented ECG signals. A temporal attention mechanism is incorporated to explicitly quantify the contribution of individual heartbeats and signal regions to the final classification outcome. Furthermore, an explainability module maps model relevance back onto the ECG waveform, enabling intuitive visualization of diagnostically significant components such as the P-wave, QRS complex, and T-wave. The model is evaluated on multi-class arrhythmia classification tasks using standard performance metrics alongside detailed explainability analyses. Experimental results demonstrate that the proposed approach achieves robust classification performance while providing transparent, beat-level and segment-level explanations that are consistent with established cardiological knowledge. By opening the black box of deep ECG classification models, this work advances interpretable signal processing–driven artificial intelligence for cardiac diagnosis and supports the development of trustworthy clinical decision-support systems.