Abstract:

Objective We propose an efficient deep learning model to improve the classification accuracy in automatic classification tasks of 12-lead electrocardiogram (ECG) signals. Methods We designed a new ResLSTM-TemporalSE network architecture by incorporating a multi-layer Residual Long Short-Term Memory (ResLSTM) structure and introducing skip connections between LSTM layers to establish residual learning pathways for the temporal features. A temporal attention mechanism was integrated into the traditional Squeeze-and-Excitation (SE) module to enhance channel-wise feature representation while capturing long-term temporal dependencies within ECG signals, thereby an efficient hierarchical feature extraction framework was constructed. The model was validated using the public CPSC2018 dataset and a private clinical dataset from the Seventh Affiliated Hospital of Southern Medical University. Results The experimental results demonstrated that the model achieved a classification accuracy of 99.70% on the CPSC2018 test set, with precision, recall, and F1-score values of 0.9966, 0.9370, and 0.9653, respectively. On the private clinical dataset, it attained an accuracy of 82.77%, with precision, recall, and F1-score values of 0.6811, 0.8961, and 0.7723. Ablation studies confirmed the significant contributions of both the residual connections and the temporal attention module to model performance. Conclusion The ResLSTM-TemporalSE model effectively integrates spatiotemporal features of the ECG signals and demonstrates superior classification performance on the CPSC2018 benchmark while maintaining strong generalization capabilities in real-world clinical settings. This framework provides a robust solution for automated ECG analysis and holds significant promise for clinical applications.

Key words: electrocardiogram classification, deep learning, ResNet, Long Short-Term Memory network, Squeeze-and-Excitation module, ResLSTM-TemporalSE