Diffusion-Based ECG Augmentation

Generate synthetic ECG signals via style transfer between Atrial Fibrillation (AFib) and Normal Sinus Rhythm classes using a disentangled diffusion model.

How it works: The model separates ECG signals into content (beat morphology) and style (rhythm), then generates new ECGs by transferring the style of one class onto another using SDEdit denoising.

Key finding: Replacing 33% of real training data with these synthetic ECGs produces statistically equivalent classifier performance (TOST, p = 0.007).

GitHub | Model Hub | University of Peradeniya + SimulaMet

Target Class
0.1 0.9
10 100
1 5