A lightweight progressive residual convolutional network for interpretable arrhythmia classification

Arrhythmia poses a serious threat to human health, necessitating research into automated diagnostic methods. Although advances have been made in arrhythmia classification using deep learning, existing models often suffer from high computational cost and poor interpretability, limiting their clinical applicability. To this end, this paper proposes a lightweight progressive residual convolutional attention network (DResCA). Through the synergistic design of stage-adaptive downsampling and channel attention, this network reduces parameter redundancy while enhancing deep ECG feature extraction. To enhance clinical interpretability, template matching and attention visualization are employed to reveal the key ECG regions underlying the model's diagnoses. Five-class classification experiments on the MIT-BIH Arrhythmia Database demonstrate that the model achieves 98.35% accuracy and 97.99% macro-F1 score while maintaining fewer parameters than mainstream deep learning models. These characteristics provide a practical technical pathway for real-time arrhythmia detection on wearable edge computing devices.