Ensuring electronic medical record simulation through better training, modeling, and evaluation.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Privacy

Background:

• Electronic medical records (EMRs) are valuable for research but pose privacy risks.
• Generative adversarial networks (GANs) are used for EMR simulation but often yield suboptimal results.
• Existing GAN approaches lack principled methods for EMR data simulation.

Purpose of the Study:

• To improve the simulation of electronic medical records (EMRs) using a novel generative adversarial network (GAN) pipeline.
• To enhance the utility and privacy of simulated EMR data for medical research.
• To address limitations in current GAN-based EMR simulation techniques.

Main Methods:

• Developed a novel GAN pipeline incorporating Wasserstein divergence and layer normalization.
• Introduced two utility measures for assessing structural similarity between real and simulated EMRs.
• Implemented a filtering strategy to improve GAN training for rare clinical concepts.
• Evaluated the model using over 1 million EMRs from Vanderbilt University Medical Center.

Main Results:

• The proposed GAN model significantly outperformed existing methods in retaining the characteristics of real EMRs.
• Improved prediction performance and structural properties of simulated EMRs were achieved without compromising privacy.
• The filtering strategy enhanced data utility, especially with smaller EMR training datasets.

Conclusions:

• EMR simulation using GANs can be substantially enhanced through refined training, modeling, and evaluation.
• The novel pipeline offers a more principled and effective approach to EMR data simulation.
• Improved EMR simulation facilitates broader and safer data sharing for medical research.