Generating Electronic Health Records with Multiple Data Types and Constraints.

Area of Science:

• Health Informatics
• Artificial Intelligence
• Data Privacy

Background:

• Sharing electronic health records (EHRs) at scale poses significant privacy risks.
• Generative adversarial networks (GANs) offer a promising approach to mitigate these risks by simulating EHR data.
• Existing GAN-based methods are limited in handling diverse EHR data types and feature constraints.

Purpose of the Study:

• To introduce an improved GAN-based method for simulating multi-type EHR data.
• To address limitations of previous methods by incorporating multiple data types and feature constraints.
• To evaluate the utility and privacy-preserving capabilities of the proposed simulation technique.

Main Methods:

• Refinement of the generative adversarial network (GAN) architecture.
• Development of techniques to account for inter-feature constraints within EHR data.
• Integration of key utility measures to assess the quality of generated synthetic EHRs.
• Validation using a large-scale EHR dataset from Vanderbilt University Medical Center.

Main Results:

• The proposed model successfully simulates EHRs comprising multiple data types (e.g., demographics, diagnoses, procedures, vital signs).
• The method effectively retains statistical properties, cross-feature correlations, and structural patterns from real EHR data.
• The generated synthetic data accurately represents feature constraints and associated patterns.
• Privacy is maintained without significant loss of data utility.

Conclusions:

• The novel GAN-based approach enhances the simulation of complex, multi-type EHR data.
• This method offers a robust solution for generating privacy-preserving synthetic EHRs.
• The findings support the use of advanced GAN frameworks for secure health data sharing and research.