Publication Details
Abstract
Cardiovascular disease (CVD) is one of the main causes of death across the globe, the provision of a rapid diagnostic service for arrhythmia is an essential component in the process of enhancing the quality of life of cardiac patients. The automatic diagnosis of arrhythmia from electrocardiogram (ECG) data has garnered a significant amount of scientific interest. However, the automatic extraction of both the morphological and temporal aspects of ECG signals remains a tough task currently. For the purpose of multi-class arrhythmia classification, we propose a novel hybrid deep learning architecture that integrates Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and an attention mechanism. A complete data preprocessing pipeline was incorporated into the suggested model, which was further evaluated with the help of the MIT-BIH Arrhythmia database. The preprocessing steps include discrete wavelet transform (DWT) based denoising, R-peak extraction, heartbeat-based segmentation, normalization, and class balance through under-sampling and synthetic minority over-sampling technique (SMOTE). Spatial feature extraction of ECG signals were extracted by CNN module, while the temporal relationship between successive heartbeats were modeled by LSTM module. The attention mechanism allows the model to allocate variable significance to the most pertinent segments of the ECG signal that correlate to clinically relevant areas. Class-performance analysis demonstrated that the proposed model exhibits high levels of sensitivity and specificity across all arrhythmia classes with nearly perfect classification rates for normal and ventricular beat classes. Some minor misclassification errors occurred primarily due to morphological similarity between some classes and the large disparity in class sizes. Our proposed model provides a good compromise between the trade-offs between accuracy, computational cost, and interpretability thus, demonstrating superiority over many of the other state-of-the-art approaches. The integration of the attention mechanism into our proposed model enhances both its performance and explainability, rendering it appropriate for application in real-world clinical environments. This study shows great potential for the development of hybrid deep learning architectures for reliable and efficient automatic ECG-based diagnosis.