Generating Longitudinal Synthetic EHR Data with Recurrent Autoencoders and Generative Adversarial Networks

Siao Sun, Fusheng Wang, Sina Rashidian, Tahsin M. Kurç, Kayley Abell-Hart, Janos G. Hajagos, Wei Zhu, Mary M. Saltz, Joel H. Saltz

Abstract: Synthetic electronic health records (EHR) can facilitate effective use of clinical data in software development, medical education, and medical research without the concerns of data privacy. We propose a novel Generative Adversarial Network (GAN) approach, called Longitudinal GAN (LongGAN), that can generate synthetic longitudinal EHR data. LongGAN employs a recurrent autoencoder and the Wasserstein GAN Gradient Penalty (WGAN-GP) architecture with conditional inputs. We evaluate LongGAN with the task of generating training data for machine/deep learning methods. Our experiments show that predictive models trained with synthetic data from LongGAN achieve comparable performance to those trained with real data. Moreover, these models have up to 0.27 higher AUROC and up to 0.21 higher AUPRC values than models trained with synthetic data from RCGAN and TimeGAN, the two most relevant methods for longitudinal data generation. We also demonstrate that LongGAN is able to preserve patient privacy in a given attribute disclosure attack setting.