A method for ECG denoising based on generative adversarial networks.

Objectives: To tackle the challenge of noise in grayscale electrocardiograms (ECGs) transcribed from paper records, this study proposes a Generative Adversarial Networks (GANs)-based framework that includes both a denoising dataset construction method and an end-to-end denoising model for accurate waveform recovery.

Methods: We propose a GANs-based ECG waveform denoising approach, referred to Generative Adversarial Network Electrocardiogram Denoising. This method simultaneously trains a generator and a discriminator, enabling the generator to extract clean ECG waveforms in an end-to-end manner. Additionally, we propose an ECG Background Noise Generation technique, leveraging a GANs to generate diverse background noise patterns that closely resemble real-world transcribed ECGs, thereby providing a high-quality denoising dataset.

Results: We evaluated the similarity between denoised and clean ECG waveforms, along with the impact of various denoising methods on heart disease diagnostic accuracy. The proposed method achieved the best overall performance, with a Dice Similarity Coefficient of 92.72% and an Intersection over Union (IoU) of 86.58%. In terms of diagnostic utility, the dataset denoised using our method achieved an accuracy of 89% and 82.19%, significantly outperforming existing baseline approaches.

Conclusions: This study proposes a GANs-based approach for constructing an ECG denoising dataset and a GANs-based ECG denoising method. The joint application of these two approaches can effectively address the ECG denoising task.