Synthesizing electronic health records using improved generative adversarial networks.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Electronic Health Records (EHRs) are crucial for medical research but access is limited.
• Existing methods like medical Generative Adversarial Network (medGAN) have limitations in generating realistic synthetic EHR data.

Purpose of the Study:

• To develop and evaluate novel Generative Adversarial Network (GAN) models for generating realistic synthetic EHR data.
• To improve upon the capabilities of existing medGAN models for EHR data synthesis.

Main Methods:

• Modified medGAN to create two new models: medical Wasserstein GAN with gradient penalty (medWGAN) and medical boundary-seeking GAN (medBGAN).
• Trained and generated synthetic EHRs using medGAN, medWGAN, and medBGAN on MIMIC-III and NHIRD datasets.
• Compared model performance using statistical tests (Kolmogorov-Smirnov test) and machine learning tasks (association rule mining, prediction).

Main Results:

• The proposed models (medWGAN and medBGAN) demonstrated superior performance in generating synthetic EHR data compared to medGAN.
• medBGAN consistently outperformed both medGAN and medWGAN across all evaluated metrics.
• The models effectively learned the data distribution of real EHRs, producing highly realistic synthetic data.

Conclusions:

• The developed medWGAN and medBGAN models are effective for generating realistic synthetic EHR data.
• These advanced models can overcome limitations of EHR data accessibility, accelerating medical informatics research.
• The findings highlight the potential of medBGAN as a leading method for synthetic EHR generation.