Clustering of Adverse Events of Post-Market Approved Drugs

Abstract

Adverse side effects of a drug may vary over space and time due to different populations, environments, and drug quality. Discovering all side effects during the development process is impossible. Once a drug is approved, observed adverse effects are reported by doctors and patients and made available in the Adverse Event Reporting System provided by the U.S. Food and Drug Administration . Mining such records of reported adverse effects, this study proposes a spatial clustering approach to identify regions that exhibit similar adverse effects. We apply a topic modeling approach on textual representations of reported adverse effects using Latent Dirichlet Allocation. By describing a spatial region as a mixture of the resulting latent topics, we find clusters of regions that exhibit similar (topics of) adverse events for the same drug using Hierarchical Agglomerative Clustering. We investigate the resulting clusters for spatial autocorrelation to test the hypothesis that certain (topics of) adverse effects may occur only in certain spatial regions using Moran’s I measure of spatial autocorrelation. Our experimental evaluation exemplary applies our proposed framework to a number of blood-thinning drugs, showing that some drugs exhibit more coherent textual topics among their reported adverse effects than other drugs, but showing no significant spatial autocorrelation of these topics. Our approach can be applied to other drugs or vaccines to study if spatially localized adverse effects may justify further investigation.