Imbalanced domain adaptation Net for multi-class Electrocardiography signals classification

Arrhythmia detection using Electrocardiogram (ECG) signals plays a crucial role in cardiovascular disease diagnosis. Deep learning has shown great promise in developing diagnostic models for ECG classification. However, maintaining the performance of these models remains challenging in cross-domain scenarios, where ECG data might be collected from different devices, patient populations, or environments. Although existing unsupervised domain adaptation (UDA) methods have been utilized to improve the generalization of the model, many consider the discrepancies in feature distributions across domains but overlook variations in label distributions, also known as label shift. To tackle this problem, an imbalanced domain adaptation network (IDANet) is proposed for multi-class ECG diagnosis, accounting for both changes in feature distributions and variations in label distributions. Specifically, to address label shift, a class-sensitive re-weighting method with label-distribution-aware re-margining regularization is utilized which effectively enables the extraction of class-invariant features and facilitates the establishment of differentiated classification boundaries. Meanwhile, semantic alignment is applied to reduce the category-wise feature distribution discrepancies. Experiments on four ECG datasets show a considerable improvement (up to ∼ 16 % in accuracy and ∼ 17 % in F1 score) over other state-of-the-art UDA methods. These results highlight the promising potential of the proposed approach for future medical diagnostic applications.