Classifier-Guided Diffusion Model for Generating Disease-Specific ECG Data: A Case Pilot Study on Inferior Myocardial...

As a non-invasive diagnostic tool, electrocardiography (ECG) is one of the most widely used cardiac screening methods. However, due to privacy concerns and limitations in medical resources, the scarcity of high-quality ECG data with clear signals and accurate clinical annotations presents a common challenge for researchers. Although synthetic ECG generation offers a solution, existing disease-specific methods rely on costly labeled datasets. To reduce this cost, we propose a method that employs a classifier-guided diffusion model to generate ECG data for a specific disease without requiring extensive annotations. We applied our proposed method to the public PTB-XL dataset and generated ECG data for inferior myocardial infarction (IMI). To evaluate the quality of the generated data, we conducted two analyses: classification performance evaluation using a classifier trained on the synthetic dataset and a statistical analysis of the generated IMI ECGs. In the first analysis, we trained classifiers to distinguish between IMI and normal ECGs. The only difference between these classifiers was the training dataset: one was trained on fully real data, while the others were trained on merged datasets containing both real and synthetic data. The classifiers trained on the merged datasets achieved performance comparable to that of the classifier trained solely on real data, indicating that synthetic data can partially replace real data without degrading model performance. In the second analysis, we conducted a statistical distribution analysis of a key IMI ECG feature: Q wave depth. This analysis revealed that the distribution of the generated IMI ECGs closely resembled that of the real IMI ECGs. These results validate the high quality and reliability of the generated data and confirm the effectiveness of our proposed method.