SMOOTH-GAN: Towards Sharp and Smooth Synthetic EHR Data Generation

Abstract

Generative adversarial networks (GANs) have been highly successful for generating realistic synthetic data. In healthcare, synthetic data generation can be helpful for producing annotated data and improving data-driven research without worries on data privacy. However, electronic health records (EHRs) are noisy, incomplete and complex, and existing work on EHR data is mainly devoted to generating discrete elements such as diagnosis codes and medications or frequent laboratory values. In this work, we propose SMOOTH-GAN, a novel approach for generating reliable EHR data such as laboratory values and medications given diagnosis codes. SMOOTH-GAN takes advantage of a conditional GAN architecture with WGAN-GP loss, and is able to learn transitions between disease stages with high flexibility over data customization. Our experiments demonstrate the model’s effectiveness in terms of both statistical similarity and accuracy on machine learning based prediction. To further demonstrate the usage of our model, we apply counterfactual reasoning and generate data with occurrence of multiple diseases, which can provide unique datasets for artificial intelligence driven healthcare research.

A Appendix

1.1 A.1 Binary Data Distribution

As GANs were known to struggle with generating binary values, we added Fig. 4 to illustrate dimension-wise probability for medications comparing real versus synthetic data.

Fig. 4.

Dimension-wise probability performance for binary values.

1.2 A.2 Is Training Data Memorized by the GAN?

For ensuring privacy and discovering whether the GAN is generating new cases or memorizing the training set, we followed the footsteps of [8] by measuring maximum mean discrepancy (MMD) and applying the three-sample test [4, 11]. MMD can answer whether two sets of samples were generated from the same distribution. If the synthetic data is memorized then MMD(synthetic, training) would be significantly lower than MMD(synthetic, test). For this reason, we state the null hypothesis as GAN has not memorized the training set, and consequently MMD(synthetic, test) $\le$ MMD(synthetic, training). We sampled from these three datasets 35 times and calculated MMDs and p-values for the hypothesis. The mean p-value with its standard deviation is $0.26\pm 0.15$ which means we cannot reject the null hypothesis and we can establish that GAN did not memorize from the training set.