An Enhanced Neural Network Algorithm for ECG Diagnosis Based on Dynamic Weighted Attention and Hierarchical Contrastive Learning

This study proposes an enhanced neural network algorithm for electrocardiogram (ECG) diagnosis, addressing limitations in existing methods regarding accuracy, noise resistance, and rare case identification. The algorithm incorporates three innovative components: a multi-scale feature extraction network, a dynamic weighted attention mechanism, and a hierarchical contrastive learning strategy. The multi-scale feature extraction network, based on an improved 1D-ResNet architecture, effectively captures ECG features at different temporal scales. The dynamic weighted attention mechanism enhances the model's ability to identify key signal regions by adaptively adjusting attention weights in both channel and temporal dimensions. The hierarchical contrastive learning strategy implements feature learning at both signal and category levels, significantly improving the model's capability to identify rare cases. Experimental results on two public datasets, PTB-XL and MIT -BIB, demonstrate significant improvements in overall accuracy (88.2%), sensitivity (87.0%), and specificity (89.4%), while exhibiting excellent robustness under noisy conditions. Notably, for rare cases with occurrence rates below 1 %, the recognition accuracy improved by 15.3 % compared to baseline models. These results validate the algorithm's potential for practical clinical applications.