Specialized ECG data augmentation method: leveraging precordial lead positional variability

Abstract

Deep learning has demonstrated remarkable performance across various domains. One of the techniques contributing to this success is data augmentation. The essence of data augmentation lies in synthesizing data while preserving accurate labels. In this research, we introduce a data augmentation technique optimized for electrocardiogram (ECG) data by focusing on the unique angles between precordial leads in 12-lead ECG, considering situations that may occur in a clinical environment. Subsequently, we utilize the proposed data augmentation technique to train a deep learning model for diagnosing atrial fibrillation or atrial flutter, generalized supraventricular tachycardia, first-degree atrioventricular block, left bundle branch block and myocardial infarction from ECG signals, and evaluate its performance to validate the effectiveness of the proposed method. Compared to other data augmentation methods, our approach demonstrated improved performance across various datasets and most tasks, thereby showcasing its potential to enhance diagnostic accuracy. Additionally, our method is simple to implement, offering a gain in total training time compared to other augmentation methods. This study holds the potential to positively advance further development in the fields of bio-signal processing and deep learning technology, addressing the issue of the lack of optimized data augmentation techniques applicable to ECG data in the future.