2021 FDA Science Forum
Spatio-Temporal Clustering of Adverse Events for Post-Market Approved Drugs
- Authors:
- Center:
-
Contributing OfficeOffice of the Commissioner
Abstract
Adverse side effects of a drug may vary over space and time due to different population, environment, and drug quality. Discovering all side effects during the development process is impossible. Once approved and available to the public, regulators rely on a combination of surveillance, reporting (by doctors and patients), and data mining to discover any post-market issues in approved pharmaceutical drugs. Our goal of this study is to find statistically significant spatio-temporal clusters among co-occurring adverse effects of U.S Food and Drug Administration (FDA) approved drugs. We use data obtained from FDA's Adverse Event Reporting System (FAERS) to explore the spatio-temporal distribution of combinations of adverse effects. This is done by computing, for each spatial region and for each year, the Top-k most frequent sets of adverse events using a frequent itemset mining approach. To assess the similarity of sets of adverse events between spatial regions, we employ Gestalt Pattern Matching between the textual representation of reported adverse effects. To find clusters of regions that exhibit similar adverse events we apply an agglomerative hierarchical clustering approach. Finally, we explore the resulting clusters of similar adverse events to discover patterns of spatial autocorrelation using Moran's I measure of spatial autocorrelation. In our experimental evaluation, we use adverse event records in Europe for three pharmaceutical drugs between 2014 and 2017. Our result show that the vast majority of mined clusters of regions having similar adverse events did not exhibit significant spatial auto-correlation, indicating that the adverse events within a clusters are not the result of spatial patterns or local effects. For a small number of clusters, we found significant spatial autocorrelation but after applying Bonferroni correction to account for the large number of tested hypotheses, we found no significant and interesting cases of spatial autocorrelation for three drugs studied. Yet, we note that our approach can be applied on other drugs to explore if other drugs exhibit spatially localized side-effects that may justify further investigation. We conclude our work by discussing future directions to discover spatial trends among adverse events that this study was not able to find.
