Omar Cheikhrouhou

Accurate classification of atrial arrhythmias from single-lead ECG recordings presents significant clinical challenges, particularly for atrial flutter (AFL) subtypes characterized by varying atrioventricular (AV) conduction ratios. These distinctions are crucial for therapeutic decision-making; however, traditional manual ECG interpretation remains labor-intensive and requires expert analysis. This study introduces two key contributions. First, we propose a novel labeling methodology that automatically identifies AFL subtypes based on AV conduction patterns (from 1:1 to 4:1 ratios). Second, we design a lightweight deep learning framework optimized for single-lead (lead II) ECG classification across multiple rhythm types, including AFL subtypes, atrial fibrillation, atrial tachycardia, and sinus rhythm. The proposed hybrid architecture combines one-dimensional depthwise separable convolutions, squeeze-and-excitation attention mechanisms, residual connections, and bidirectional GRU layers to enhance temporal feature representation while maintaining computational efficiency. The model achieves strong performance with 98.23% accuracy, 98.27% F1-score, 98.25% precision, 99.70% specificity, and 98.29% recall on the Chapman ECG dataset. We validated the model using stratified 10-fold cross-validation, with the best-performing fold reaching an accuracy of 98.48%. Additionally, external validation was performed on selected atrial rhythm records from the MIT-BIH Arrhythmia Database, achieving an average F1-score of 0.85 across targeted atrial arrhythmia classes and perfect classification F1-score of 1.0 for AFL samples. To enhance clinical interpretability, we integrated Explainable AI through Gradient-weighted Class Activation Mapping (Grad-CAM), providing visual explanations that highlight the temporal ECG regions most influential for each classification decision. The lightweight design makes this approach highly suitable for deployment in resource-constrained environments, particularly wearable ECG devices for continuous cardiac monitoring and real-time clinical decision support.