Online Co-Learning Strategy Based on Multibranch Multiscale Network for 12-Lead ECG Classification

The coexistence of multiple cardiovascular diseases (CVDs) will increase the burden on cardiac function and increase the patient’s risk of death. The 12-lead electrocardiogram (ECG) offers a more comprehensive assessment of cardiac electrical activity and serves as a vital tool for multilabel ECG classification. However, 12-lead ECG is high-dimensional data, and designing the model structure capable of capturing the complex relationships between different leads is important. In addition, deep learning tends to overfit small samples, so improving the model’s generalization ability is crucial. In response, we have developed an online co-learning (OCL) strategy based on multibranch multiscale residual fusion network (2MRFNet) for 12-lead ECG multilabel classification. The 2MRFNet primarily comprises two key modules: the multibranch multiscale residual (MBSR) module and the scale heterogeneous attention fusion (SHAF) module. These modules are used to extract and fuse features on multiscale from three groups of leads: standard, limb, and chest leads. The OCL strategy effectively leverages the transfer of feature information between models to enhance model generalization capability. The PTB-XL and CPSC-2018 datasets were used as validation sets. The proposed method achieved an average AUC score of 93.38% and an average $F1$ score of 48.51% for five tasks on the PTB-XL dataset. On the CPSC-2018 dataset, the corresponding results are 97.37% and 84.40%. Compared with known state-of-the-art methods, the proposed method exhibited the highest multilabel classification performance across various experiments.