[Paper Review] Generating Multi-label Discrete Patient Records using Generative Adversarial Networks

Edward Choi, Siddharth Biswal+4 more|arXiv (Cornell University)|Mar 19, 2017

Machine Learning in Healthcare42 references341 citations

TL;DR

medGAN combines an autoencoder with a GAN to generate high-dimensional, multi-label discrete EHR records (binary and count variables). It uses minibatch averaging to reduce mode collapse and improves training with batch normalization and shortcut connections, achieving close-to-real data quality and assessing privacy risks.

ABSTRACT

Access to electronic health record (EHR) data has motivated computational advances in medical research. However, various concerns, particularly over privacy, can limit access to and collaborative use of EHR data. Sharing synthetic EHR data could mitigate risk. In this paper, we propose a new approach, medical Generative Adversarial Network (medGAN), to generate realistic synthetic patient records. Based on input real patient records, medGAN can generate high-dimensional discrete variables (e.g., binary and count features) via a combination of an autoencoder and generative adversarial networks. We also propose minibatch averaging to efficiently avoid mode collapse, and increase the learning efficiency with batch normalization and shortcut connections. To demonstrate feasibility, we showed that medGAN generates synthetic patient records that achieve comparable performance to real data on many experiments including distribution statistics, predictive modeling tasks and a medical expert review. We also empirically observe a limited privacy risk in both identity and attribute disclosure using medGAN.

Motivation & Objective

• Motivate synthetic EHR data generation to mitigate privacy risks while enabling research.
• Generate high-dimensional discrete variables (binary and count) from EHRs.
• Leverage an autoencoder to enable GAN to model discrete records and decode to synthetic outputs.
• Improve GAN training stability for discrete, multi-label data via minibatch averaging and architectural enhancements.
• Evaluate realism, utility for predictive tasks, and privacy risks of the synthetic data.

Proposed method

• Define EHR data as fixed-size vectors of discrete counts or binaries.
• Use an autoencoder to learn salient features of discrete variables and decode to discrete outputs.
• Train a GAN where the generator outputs go through the pre-trained decoder to produce synthetic discrete records.
• Discriminator distinguishes real records from synthetic outputs without explicit rounding.
• Introduce minibatch averaging to expose the discriminator to minibatch statistics and mitigate mode collapse.
• Enhance the generator with batch normalization and shortcut connections to balance D and G during training.

Experimental results

Research questions

• RQ1 Can medGAN reproduce the distributions of high-dimensional discrete EHR features (binary and count) compared to real data?
• RQ2 Does medGAN preserve inter-feature relationships and enable accurate predictive modeling using synthetic records?
• RQ3 What is the practical impact of minibatch averaging on mode collapse and data realism?
• RQ4 How realistic are the generated records to clinical experts?
• RQ5 What are the privacy risks (presence and attribute disclosure) associated with medGAN-generated data?

Key findings

• medGAN and its enhanced variants achieve close-to-real distribution and predictive-task performance for discrete EHR data.
• Minibatch averaging significantly improves dimension-wise probability and prediction realism, while minibatch discrimination offers no clear benefit in this setting.
• Batch normalization and shortcut connections in the generator improve learning efficiency and data realism.
• Medical expert review indicates synthetic count data are broadly indistinguishable from real data, with some outliers due to domain-specific issues.
• Privacy analysis shows limited presence and attribute disclosure risks under the evaluated scenarios, suggesting some privacy benefits of synthetic generation.