Continuous Patient-Centric Sequence Generation via Sequentially Coupled Adversarial Learning.

Analyzing massive patient-centric Electronic Health Records (EHRs) becomes a key to success for improving health care and treatment. However, the amount of these data is limited and the access to EHRs is difficult due to the issue of patient privacy. Thus high quality synthetic EHRs data is necessary to alleviate these issues. In this paper, we propose a Sequentially Coupled Generative Adversarial Network (SC-GAN) to generate continuous patient-centric data, including patient state and medication dosage data. SC-GAN consists of two generators which coordinate the generation of patient state and medication dosage in a unified model, revealing the clinical fact that the generation of patient state and medication dosage data have noticeable mutual influence on each other. To verify the quality of the synthetic data, we conduct comprehensive experiments to employ these data on real medical tasks, showing that data generated from SC-GAN leads to better performance than the data from other generative models.