ECG Synthesis and Utility Analysis - A Diffusion Model Based Approach

With the growing demand for privacy-preserving healthcare solutions, the generation of synthetic electrocardiograms (ECGs) offers a valuable alternative to using real patient data.

In this study, we present the adaptation of the SSSD-ECG diffusion model to generate high-quality synthetic 12-lead ECGs for Sinus Rhythm/Normal and Atrial Fibrillation (AF) conditions using 10-second recordings from the 12-lead MIMIC-IV ECG dataset.

We validate the utility of the generated ECGs through downstream classification tasks, with models trained on synthetic ECG features achieving an F1-score of 0.80 when tested on real data, and 0.91 when trained on real data and tested on synthetic data. Additionally, blind tests conducted by physicians at two university hospital sites demonstrated that the synthetic signals effectively mimic real ECGs in both morphology and key features.

This work establishes diffusion-based models as an effective tool for generating realistic synthetic ECGs, providing valuable resources for model development, supporting testing of clinical decision- making solutions, and enabling research in contexts where real data is scarce or not shareable.