False positive associations between drugs and health outcome of interest (HOIs) are a common problem in pharmacovigilance. One promising approach to reducing the occurrence of false positives is the calibration of p-values using drug-outcome pairs for which it is likely that no causal relationship exists. Previous studies using the approach have focused on only a few HOIs. In this paper, we describe an evidence base that will enable the method to be scaled to all possible drug-HOI pairs. Evidence about the potential relationship between drugs and HOIs from multiple disparate sources were integrated into a common schema and set of terminologies. Methods were developed to query for drug-HOI evidence and metadata across all sources.
The new evidence base integrates drug-HOI evidence from spontaneous reports (counts and statistical signals), scientific literature (PubMed and SemMedDB), American and European product labeling, and clinical trials (ClinicalTrials.gov). Machine learning methods that can predict negative drug-HOI associations are being built using features in the evidence base, and then tested for their performance characteristics.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.