Generating Multi-label Discrete Patient Records using Generative Adversarial Networks

TL;DR

This work tackles privacy barriers in sharing electronic health records by generating realistic synthetic EHR data. It introduces medGAN, a hybrid autoencoder–GAN that learns distributions over high-dimensional discrete clinical codes, using minibatch averaging and architectural enhancements to produce diverse, plausible records for binary and count variables. The approach achieves close-to-real performance on distribution statistics and predictive tasks while exhibiting limited privacy risk, suggesting practical utility for research while mitigating re-identification concerns. Overall, medGAN provides a scalable, generalizable framework for privacy-preserving synthetic health data with potential to accelerate biomedical research and collaboration.

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.

Figures (11)

• Figure 1: Scatterplots of dimension-wise probability results. Each dot represents one of 615 codes. The x-axis represents the Bernoulli success probability for the real dataset A, and y-axis the probability for the synthetic counterpart generated by each model. The diagonal line indicates the ideal performance where the real and synthetic data show identical quality.
• Figure 2: Scatterplots of dimension-wise prediction results. Each dot represents one of 615 codes. The x-axis represents the F1-score of the logistic regression classifier trained on the real dataset A. The y-axis represents the F1-score of the classifier trained on the synthetic counterpart generated by each model. The diagonal line indicates the ideal performance where the real and synthetic data show identical quality.
• Figure 2: Qualifying ICD-9 codes for heart failure
• Figure 3: Boxplot of the impression scores from a medical expert.
• Figure 4: a,b: Sensitivity and precision when varying the number of known attributes. The total number of attributes (i.e. codes) of dataset B is 1,071. c,d: Sensitivity and precision when varying the size of the synthetic dataset. The maximum size of the synthetic dataset $S \in \{0,1\}^{N \times |\mathcal{C}|}$ is matched to the size of the training set $R \in \{0,1\}^{N \times |\mathcal{C}|}$.