Research on ECG Classification Based on Graph Neural Networks

The rapid advancement of AI has made ECG-based arrhythmia classification crucial for cardiac diagnosis. Traditional methods face challenges in capturing complex spatiotemporal dependencies in ECG signals. We propose a novel GCN-Transformer hybrid model with Weighted Mutual Information (WMI) for enhanced graph construction, effectively modeling physiological relationships among ECG leads. Our approach employs GCN to capture spatial lead relationships and Transformer's self-attention to process spatiotemporal features. The WMI strategy significantly improves inter-lead correlation sensitivity. Extensive experiments on public ECG datasets show our model outperforms CNN and RNN baselines, achieving superior accuracy (3.7% improvement), sensitivity (4.2%), specificity (2.9%), and F1-score (3.5%). This work provides an effective automated ECG diagnosis solution with strong clinical potential.