Noise ECG generation method based on generative adversarial network

Abstract

Over the years, significant hospitals have scanned paper electrocardiograms and saved them into electronic health records to form digital management of diagnosis and treatment records. Therefore, the electrocardiogram(ECG) scans have substantial sample diversity, correlate with the patient's past medical records, and have significant research value. However, the scanned ECG is noisy, and it is difficult to directly use it as the training data of the intelligent diagnosis algorithm, so it is necessary to preprocess it. In recent years, some researchers have proposed methods to extract ECG signals from noisy ECG using neural networks, but the results are not good enough due to the lack of paired noise ECG and noiseless ECG to train neural networks. This paper proposes an unsupervised noise ECG image generation method to overcome this difficulty. These generated images have the same ECG signals as the input ECG image, and the noise is similar to the actual ECG image. These two kinds of input images do not need to be paired. To evaluate the global and local quality of the generated images, our method uses a two-discriminator generative adversarial network and calls them global discriminator and local discriminator separately. The output vector of the global discriminator is fed into the generator when training the model, thereby helping the generator to improve the generated results in a targeted manner. Using the data generated by this method to train the neural network for ECG signal extraction can significantly improve its extraction performance. The Dice coefficient of the proposed network reaches 0.868, which is higher than 0.798 of the robust baseline model. Therefore, the proposed method can effectively solve the problem of lacking training data in the current ECG signal extraction network.