DMST-AGN: dual-branch multi-scale spatiotemporal agent attention with graph convolutional network for ECG arrhythmia classification

Arrhythmias classification plays an important role for the diagnosis and treatment of cardiovascular diseases. Deep learning based on 12-lead ECG arrhythmia classification methods primarily focus on temporal feature extraction. However multi-scale features and inter-lead spatial correlations are often neglected, which are essential for detecting both localized waveform anomalies and long-term rhythm disorders. To overcome these limitations, this study proposed a novel dual-branch neural network architecture that simultaneously captures multi-scale temporal patterns and inter-lead spatial relationships. The temporal branch combines multi-scale convolutional modules with feature fusion mechanisms to automatically extract and integrate ECG features at different scales, moreover an improved attention mechanism is proposed to capture key temporal patterns. The spatial branch constructs cosine similarity graphs by computing inter-lead cosine similarities and extracts spatial features using a multi-layer graph convolutional network. Finally, the fused spatiotemporal features are utilized for arrhythmia classification. Comprehensive experiments on three benchmark datasets (CPSC-2018, PTB-XL, and Chapman) demonstrate the superiority of proposed approach, achieving state-of-the-art F1-scores of 85.51 %, 36.50 %, and 97.04 %, respectively, thus validating its effectiveness in multi-class arrhythmia classification.