Synthetic EHR data

Synthetic electronic health record (EHR) data are generated rather than collected. A generative model, such as a diffusion model, learns the joint distribution of a real dataset and then samples new records that resemble it statistically without copying any individual patient.

Synthetic data can make it easier to share datasets, prototype analyses and augment small cohorts. It is not automatically private: a model can memorize rare records, so synthetic datasets should be tested with membership inference attacks and checks for near-copies.

Quality is judged on fidelity (do distributions and relationships match?), utility (does a model trained on synthetic data perform well on real data?) and privacy.