Simulating drug effects on blood glucose laboratory test time series with a conditional WGAN

Abstract

The unexpected effects of medications has led to more than 14 million drug adverse events reported to the Food and Drug Administration (FDA) over the past 10 years in the United States alone, with a little over 1.3 million of them linked to death, and represents a medical and financial burden on our healthcare. Laboratory tests have the potential to capture inter-individual variability in drug responses, but a significant portion of the patient population has unique treatment pathways that impedes forecasting and optimal decision making. Generative Adversarial Networks (GANs) are flexible implicit generative models that have demonstrated their ability to capture complex correlations in field like computer vision and natural language. Their latent representation capacity is an opportunity for drug effect simulation on laboratory test trajectories. In this paper, we developed and evaluated conditional GANs on glucose laboratory tests in patients exposed to drug combinations and showed a proof of concept for these models in the simulation of unseen drug combinations. By using conditional Wasserstein GANs (WGANs) to simulate drug effects in laboratory tests, we hope to pave the way for novel clinical decision support (CDM) systems and enable the development of better predictive models for rare cohorts of patients.