Synthesizing Multimodal Electronic Health Records via Predictive Diffusion Models

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Biomedical Data Science

Background:

• Synthesizing electronic health records (EHR) is crucial for addressing data scarcity, enhancing data quality, and ensuring fairness in healthcare applications.
• Current generative models like GANs and VAEs often fail to adequately capture temporal dependencies and time information in EHR data, limiting generation quality.
• Existing methods struggle with effective visit representation learning due to simplistic mapping functions, compromising the fidelity of synthetic EHR data.

Purpose of the Study:

• To propose EHRPD, a novel diffusion-based model for generating synthetic electronic health records (EHR) that accurately models temporal dynamics and time intervals.
• To enhance the quality and diversity of synthetic EHR data by introducing a time-aware visit embedding module and a predictive denoising diffusion probabilistic model (P-DDPM).
• To optimize the P-DDPM model using a predictive U-Net (PU-Net) for improved EHR data synthesis.

Main Methods:

• Developed EHRPD, a diffusion-based model predicting the next patient visit and estimating time intervals.
• Introduced a time-aware visit embedding module to improve representation learning.
• Implemented a predictive denoising diffusion probabilistic model (P-DDPM) optimized by a predictive U-Net (PU-Net).

Main Results:

• EHRPD demonstrated superior performance in generating synthetic EHR data across fidelity, privacy, and utility metrics on two public datasets.
• The proposed model effectively addresses limitations of existing methods by accurately modeling temporal dependencies and time information.
• Experimental results confirm the efficacy of the time-aware embedding and PU-Net in enhancing generation quality and diversity.

Conclusions:

• EHRPD offers a significant advancement in synthetic EHR data generation, overcoming key limitations of current generative techniques.
• The model's ability to capture temporal dynamics and time intervals enhances the utility and reliability of synthetic EHR data.
• EHRPD provides a promising solution for data scarcity and fairness challenges in healthcare through high-quality synthetic data generation.