An explainable hybrid deep learning model for ECG classification of cardiac conditions

Detecting cardiac dysfunction through electrocardiogram (ECG) traces remains a critical challenge in medical diagnostics. Although state-of-the-art classification models have demonstrated promising results, further reducing misclassification rates is essential for reliable clinical decision-making. This study presents a deep learning framework that integrates autoencoder-based dimensionality reduction, employing Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), followed by a multi-head attention mechanism with four attention heads for classification tasks. An extensive ablation study was performed to assess the effectiveness of the autoencoder and transformer components. To address the inherent opacity of deep learning models, the proposed approach incorporates explainable artificial intelligence (XAI) to enhance model transparency. Specifically, a weight tracking mechanism was developed to identify critical data segments influencing classification decisions, thereby ensuring the model's suitability for clinical applications where interpretability is essential. The proposed model demonstrated its effectiveness by accurately classifying five distinct heart conditions in the MIT-BIH dataset and two conditions in the PTBDB dataset. It achieved superior performance compared to existing methods, attaining an F1-score of 0.99 on both datasets.