CWANet: Class-wise alignment for enhanced cross-patient generalizability in ECG classification

The electrocardiogram (ECG) is the gold standard for the diagnosis of arrhythmias and serves as a fundamental diagnostic tool in clinical practice. Although numerous deep learning (DL) methods have achieved impressive results in ECG classification, cross-patient variability remains a significant challenge that limits the generalizability of DL models. To address this issue, we propose a novel Class-Wise Alignment strategy integrated into a deep learning-based framework (CWANet) for multi-label ECG classification. CWANet is designed to enhance generalizability across different patients by aligning high-dimensional deep neural network (DNN) features. The solution includes a multi-branch DNN enhanced representation module and a joint optimization with multiple loss functions to improve the representational capability of the model. In addition, a class-wise fusion strategy is employed at the decision level to further enhance the performance. We conducted experiments on two large-scale public ECG datasets (Chapman–Shaoxing and CPSC2018 dataset). The proposed approach achieved F1-score of 0.6911 and 0.8398 on the respective test sets, outperforming previous work, as well as several benchmark models. Ablation studies further confirmed the substantial improvements offered by our method. These results underscore the effectiveness of our approach and highlight its potential to assist in real-world application.