Progressive Generation of Electrocardiogram Signals Based on Generative Adversarial Networks

Electrocardiogram (ECG) signals, as a tool for recording cardiac activity, are widely used for the detection and diagnosis of heart diseases. However, challenges in data imbalance limit the performance of automatic detection models. Generative Adversarial Networks (GANs), as a data augmentation technique, can effectively address these issues. In this paper, we propose a progressive training strategy aimed at preserving the details of ECG signals while ensuring the stability of model training. This strategy helps progressively generate signals by gradually increasing the resolution of template signals, from coarse to fine. Additionally, we employ GANs with the proposed strategy to generate atrial fibrillation (AF) signals and evaluate the similarity between the generated signals and real signals. We further build an ECG signal classification network and evaluate the impact of the generated signals on the performance of the classification network.