Extending a Generative Adversarial Network to Produce Medical Records with Demographic Characteristics and Health System Use

Generative adversarial networks use machine learning to generate synthetic data that is similar to real data. This has been widely applied to image data, and is now being applied to electronic medical records. Synthetically generated medical records are promising for many applications where privacy and security issues make using real medical records too risky. This includes software and systems development, training, and health research. Developing upon previous work, we have extended the MEDGAN system to generate records with eight additional variables, including demographic and health system use factors. The records generated are similar in distribution to the underlying dataset for all of these added variables. Finally, we discuss our future plans, with an emphasis on privacy-protecting approaches.