Dual-branch fusion network with mutual learning for 12-lead electrocardiogram signal classification

Abstract

At present, the vast majority of methods use 12-lead electrocardiograms as a two-dimensional array as network input, and use deep neural networks to extract inter-lead correlation features. However, extracting intra-lead differential features is particularly important, as not every lead’s feature carries equal significance for classification. In this paper, we propose a dual-branch fusion network with 12-lead separation and combination, integrating the idea of mutual learning. The dual-branch network extracts differentiated and correlated features respectively and fuse them for classification. Each branch network is not only supervised by the ground truth but also referenced the learning experience of another branch network to further improve its classification ability. To address data imbalance, we introduced a category weighted binary focal loss to increase the attention of the network to the samples with few classes. We validated the proposed method on two publicly available multi-label datasets. Compared to the baseline model, our model has significantly improved in performance, demonstrating strong competitiveness and validating the effectiveness of our method. The experimental results show that our proposed method surpasses existing methods and achieves state-of-the-art performance. The method enables lightweight deployment on wearable devices, such as 12-lead ECG garments and smartwatches, facilitating real-time arrhythmia monitoring.