Improving arrhythmia classification using a residual split-attention pyramid network for class imbalance

Cardiovascular diseases are a leading cause of death worldwide, with arrhythmia being a key trigger for sudden cardiac death. It is characterized by its high concealment and sudden onset, severely threatening patients’ lives and increasing the medical burden. Although electrocardiogram (ECG) is the gold standard for arrhythmia diagnosis, its manual analysis is inefficient and prone to misdiagnosis, highlighting the urgent need for high-performance computer-aided diagnostic systems. Existing intelligent arrhythmia recognition models still face challenges such as inadequate feature extraction, class imbalance. To address these issues, this paper proposes an arrhythmia classification algorithm based on the Residual Split-Attention Pyramid Network (RSPNet). RSPNet takes ECG recurrence graphs as input and enhances feature expression by combining group convolution, depthwise separable convolution, dilated convolution, and feature pyramid networks. To address the class imbalance problem, a Tversky-Focal combined loss function (TFCL) is designed to dynamically adjust the weights of false positives and false negatives, focusing on hard-to-classify samples to improve the recognition rate of minority classes. Experimental results across multiple databases demonstrate that RSPNet combined with TFCL improves classification performance, achieving an Overall Accuracy of no less than 99.32% and an Overall F1−score of no less than 89.31%.