The Fusion Model of ResNet and GRU Based on Simplified Self-Attention for ECG Classification on PTB-XL Dataset

The electrocardiogram (ECG) depicts the orderly depolarization process of the pacemaker, atria and ventricles during each cardiac cycle. Waveforms captured by sophisticated cardiographic devices will be rendered into graphical forms. It is extensively used in clinical diagnoses and monitoring of cardiac pathologies such as arrhythmias, hypertrophy, myocardial infarction, ischemia, and other diseases affecting heart function. Our team has recently devised a ground-breaking model which synergistically integrates the formidable ResNet and GRU models, accompanied with a simplified self-attention mechanism. The model is capable of meticulously categorizing electrocardiograms into various diagnostic classsifications including 5 super-diagnostic categories, 23 sub-diagnostic categories, and 44 diagnostic categories, respectively. The model utilised in our study is notable for its incorporation of a versatile sliding window algorithm. This approach greatly enhances the system’s ability to learn by allowing for the adjustment of window size, which in turn expands the range of feature scope and facilitates the acquisition of a wider range of features during the training process. The aforementioned algorithm enables the system to attain an unparalleled degree of adaptability, thereby making it an ideal tool for the high-precision categorization of intricate physiological data. We use AUC as the metric to evaluate our best-performing model, and in the 5 super-diagnostic categories, our model outperformed the best baseline LSTM_BiDir model by 0.3, 0.2, 0.1, 0.9, and 0.5% in accuracy, AUC, recall, precision, and F1-score, respectively. In the 23 sub-diagnostic categories, our model outperformed the best baseline ResNet-Wang model by 0.1%, 7.8%, and 4.6% in AUC, recall, and F1-score, respectively. Additionally, in the 44 diagnostic categories, our model outperformed the best baseline LSTM model by 0.4, 9.8, 0.5, and 8.3% in AUC, recall, precision, and F1-score, respectively.