Enhancing cardiac disease recognition with generative model-based data augmentation

In this study, we propose a generative model-based data augmentation method to address the scarcity and imbalance of medical data, particularly Electrocardiography (ECG) signals. Our approach leverages a diffusion model with a text encoder as a conditional input, integrating time-adaptive weighting, modified Efficient Layer Aggregation Networks (ELANs), and multi-scale feature extraction networks to enhance generation quality. The diffusion model operates over 200 steps, progressively adding and removing Gaussian noise during training to produce realistic ECG signals. To evaluate our method, we augment hypertrophy data from the PTB-XL dataset and assess performance using the InceptionTime recognition model. Our augmented dataset improves the model’s overall Area Under the Curve (AUC) by 0.75% compared to training without augmentation, outperforming previous studies. These results demonstrate the effectiveness of our approach in enhancing recognition model performance with high-quality synthetic ECG data.