DS4U-Net: A Diffusion Model-Based Approach to Multi-Lead ECG Signal Generation

The electrocardiogram (ECG) signals are a safe, convenient, and effective method for diagnosing cardiac diseases. However, due to the high privacy of ECG signals, some categories in ECG datasets have low sample sizes. Computerized automated medical-assisted diagnosis typically requires a large amount of clinical data to train models. Therefore, generating high-quality ECG signals is crucial for assisting in the diagnosis of cardiac disease. This paper proposes a method for generating multi-lead ECG signals based on a diffusion model. The method uses a fusion of a structured state-space sequence model and the U-Net network architecture as an additive and denoising model for the diffusion process. The structured state space sequence model solves the problem of capturing long-time sequence information instability and, at the same time, combines with the U-Net to enhance the model's ability to learn complex features. The experiments demonstrate that the signals generated by the model exhibit spectral features similar to those of actual signals. The method generates signals that are used to train a classifier, which is then tested on real data. The results show good classification performance and confirm the effectiveness of the method. Comparison with several other generative models also indicates that this method outperforms them.