Increased Confidence in Deduplication of Drug Safety Reports with Natural Language Processing of Narratives at the US Food and Drug Administration

Abstract

The US Food and Drug Administration (FDA) receives millions of postmarket adverse event reports for drug and therapeutic biologic products every year. One of the most salient issues with these submissions is report duplication, where an adverse event experienced by one patient is reported multiple times to the FDA. Duplication has important negative implications for data analysis. We improved and optimized an existing deduplication algorithm that used both structured and free-text data, developed a web-based application to support data processing, and conducted a 6-month dedicated evaluation to assess the potential operationalization of the deduplication process in the FDA. Comparing algorithm predictions with reviewer determinations of duplicates for twenty-seven files for case series reviews (with a median size of 281 reports), the average pairwise recall and precision were equal to 0.71 (SD ± 0.32) and 0.67 (SD ± 0.34). Overall, reviewers felt confident about the algorithm and expressed their interest in using it. These findings support the operationalization of the deduplication process for case series review as a supplement to human review.