Some US Food and Drug Administration perspectives on data mining for pediatric safety assessment

Background: The US Food and Drug Administration's (FDA's) large spontaneous reporting database contains > 110,000 voluntary reports of adverse drug events (ADEs) observed during postmarketing pediatric practice and submitted to the FDA by manufacturers, health care providers, or consumers. These reports may provide evidence about known or unknown harm associated with single or combination drug treatments in pediatric patients. We recently implemented new Bayesian data mining tools to evaluate >2 million reports stored in this database to help systematically identify safety signals that appear with unexpectedly high frequency. Objective: The purpose of this paper was to describe the current status of the application of Bayesian data mining tools to screen pediatric reports of ADEs submitted spontaneously to the FDA. Methods: We applied DuMouchel's empirical Bayesian data mining algorithms to the FDA's spontaneous reporting database. These methods circumvent the lack of exposure denominators in passive surveillance data. The first method implemented, the Gamma-Poisson Shrinkage model (GPS) computer program, detects higher-than-expected associations between drugs and adverse events. The method currently being used, the updated GPS program implemented in a software program called MGPS, adjusts for the multiplicity of drugs and events per record in the data reported after October 1997 and generalizes to multiple combinations of drugs and events (eg, triples, quadruples) that occur together in reports with unexpectedly high frequency. MGPS also identifies how much these unusually frequent itemsets can be explained by pairwise associations. For this article, we used the MGPS program to analyze pairwise, higher-than-expected associations between drugs and adverse events in pediatric reports. Results: We illustrate the potential of the data mining techniques to improve the early detection and analysis of new adverse events involving unusually frequent drug-event pairs in pediatric reports and to verify the significance of known ADEs. Conclusions: New data mining techniques help skilled safety evaluators and epidemiologists at the FDA analyze reports of pediatric ADEs submitted to the FDA spontaneously. These tools are not intended to replace current pharma-covigilance techniques but to enhance them. In the near future, these techniques should facilitate our study of potential drug interactions, a serious problem for pediatric patients requiring complicated medication regimens.