Atrial Fibrillation Signal Generation Based on 2D Image Generation Model

Due to the scarcity of abnormal cases, electrocardiogram data often exhibit a high degree of imbalance. Computer-aided diagnosis systems have become an important part of the medical field, but the training of these systems usually requires a large amount of training data. The improved denoising diffusion model has been proven in previous work that the quality of the images it generates is superior to that of the current cutting-edge generative models. In order to explore its ability in synthesizing ECG, this paper proposes a new architecture, which uses BiLSTM-CNN GAN and improved DDPM to synthesize atrial fibrillation signals. Additionally, this paper employs DTW, Euclidean distance functions, and PRD for quantitative performance measurement. In addition, five different methods of generating beat evaluation are proposed and applied. Building upon these metrics, we establish threshold parameters, acceptable beat criteria, and productivity indices to construct an integrated analytical framework. The results show that DDPM can successfully generate acceptable heartbeats with highly similar morphological features to a certain extent, and the model has the potential to be used to enhance unbalanced data sets. For diagnostic validation, we developed an evaluation framework incorporating 1D convolutional neural networks and random forest classifiers for atrial fibrillation detection. The results indicate that using synthetic electrocardiogram signals to augment imbalanced datasets can significantly enhance classification performance.