ProEGAN-MS: A Progressive Growing Generative Adversarial Networks for Electrocardiogram Generation

Electrocardiogram (ECG) is a physiological signal widely used in monitoring heart health, which is of great significance to the detection and diagnosis of heart diseases. Because abnormal heart rhythms are very rare, most ECG datasets have data imbalance problems. At present, many algorithms for ECG anomaly automatic recognition are affected by data imbalance. Conventional data augmentation methods are not suitable for the augmentation of the ECG signal, because the ECG signal is one-dimensional and their morphology has physiological significances. In this paper, we propose a ProGAN based ECG sample generation model, called ProEGAN-MS, to solve the problem of data imbalance. The model can stably generate realistic ECG samples. We evaluate the fidelity and diversity of the data generated by the model and compare the data distribution of the original and generated data. In addition, in order to show the diversity of the generated ECG data more intuitively, we manually checked the diversity and calculate the statistics of the data. The results show that compared with other ECG augmentation methods based on GANs, the ECG data generated by our model has higher fidelity and diversity, and the distribution of generated samples is closer to the distribution of original data. Finally, we established neural network models for arrhythmia classification, and used them to evaluate the improvement of the classification model performance by ProEGAN-MS. The results show that augmented data by ProEGAN-MS can effectively improve the insufficient sensitivity and precision of the classification model.

Automated summary

This study proposed a ProGAN-based ECG sample generation model, ProEGAN-MS, to address data imbalance issues, demonstrating higher fidelity and diversity of the generated data compared to other GAN-based ECG augmentation methods.