Target-oriented augmentation privacy-protection domain adaptation for imbalanced ECG beat classification

Abstract

Computer aided diagnosis (CAD) systems based on ECG signals have become indispensable tools in the automatic detection of Arrhythmia, significantly reducing human effort. The rapid advancement in deep learning has ushered in a new era of such systems, showcasing promising results in ECG beat classifications. However, these systems grapple with domain shifts across different patients. Although Unsupervised Domain Adaptation (UDA) methods have shown potential in mitigating these shifts, they necessitate access to the source domain data, which poses a problem as ECG signals often contain sensitive patient information. This makes the need to enhance the performance of ECG-based arrhythmia detection CAD systems, while simultaneously respecting patient privacy, a pressing concern in clinical settings. Recently, source free domain adaptation (SFDA) methods, which exclusively use pre-trained models, have emerged as a solution to this privacy issue. Nevertheless, previous SFDA methods tend to overlook the problem of class imbalance in this setting. In response, a Target-oriented Augmentation Privacy-protection Domain Adaptation (TAPDA) framework has been developed. This method introduces a class-balance pseudo-label strategy, which selects an equal proportion of confident samples from each category. Data augmentation techniques are then applied to counteract class imbalance issues. The augmented data is provided with pseudo-labels. The selected and augmented data is used to fine-tune the pre-trained model. Then a two-step self-training process is employed to extract target-specific knowledge from the pseudo-label dataset. Numerical experiments confirm the effectiveness of our proposed method, surpassing other state-of-the-art SFDA methods.