Reproduce “Reliable Generation of Privacy-preserving Synthetic Electronic Health Record Time Series via Diffusion Models” on OpenResearch

About this paper

Muhang Tian, Bernie Chen, Allan Guo, Shiyi Jiang, Anru R. Zhang

Researchers at Duke University developed TimeDiff, a diffusion model employing mixed sequence diffusion for continuous and discrete variables, to generate realistic and privacy-preserving synthetic Electronic Health Record (EHR) time series. It achieves high data fidelity, robust privacy protection, and improved computational efficiency compared to existing methods for medical data generation.

Key insights

• Directly modeling discrete variables with multinomial diffusion, rather than relying on Gaussian approximations and post-processing, significantly enhances the realism of generated synthetic EHR time series.
• Denoising Diffusion Probabilistic Models (DDPMs) can effectively capture the complex statistical characteristics and temporal dynamics of heterogeneous EHR data, including both continuous measurements and discrete events.
• Synthetic EHR time series generated by diffusion models can maintain sufficient utility to train downstream clinical prediction models, achieving performance comparable to models trained on real data while preserving privacy.

Summary by alphaXiv.