Explainable Parallel Spatio-Temporal Hybrid Architecture for Clinically Robust ECG Arrhythmia Detection

Abstract

Cardiovascular morbidity especially cardiac arrhythmias require accurate and early diagnosis. Although deep learning models can perform with the level of the expert in arrhythmia recognition, their opaque, black-box nature usually prevents their practical usage. To solve this, the present work will offer a new Explainable Parallel Hybrid Deep Learning Framework. The architecture simultaneously combines the use of 1D-Convolutional Neural Networks (CNNs) to extract morphological features and Bidirectional Long Short-Term Memory (BiLSTM) units to predict temporal contexts. The model is further improved with a Multi-Head Self-Attention mechanism and a Gradient-weighted Class Activation Mapping (Grad-CAM) module to guarantee performance and interpretability respectively. Analyzed with a stratified part of MIT-BIH Arrhythmia Database, the structure showed the state-of-the-art results, having a classification accuracy of 99.74, a precision of 99.48 and clinically critical recall of 100% with Premature Ventricular Contractions (PVCs). Moreover, as a rigorous test of its generalizability and clinical strength, cross-domain testing was performed on an unseen section of the INCART 12-lead Arrhythmia Database, with an overall accuracy of 98.17 and a recall of 91.70 with no fine-tuning. Lastly, the Grad-CAM visualization confirms the medical relevance of the model by indicating the anomalies like the QRS complex, thus helping to close the gap between the efficacy of the algorithm and medical trust.