A Scalable Predictive Modelling Approach to Identifying Duplicate Adverse Event Reports for Drugs and Vaccines

The practice of pharmacovigilance relies on large databases of individual case safety reports to detect and evaluate potential new causal associations between medicines or vaccines and adverse events. Duplicate reports are separate and unlinked reports referring to the same case of an adverse event involving a specific patient at a certain time. They impede statistical analysis and mislead clinical assessment. The large size of such databases precludes a manual identification of duplicates, and so a computational method must be employed. This paper builds upon a hitherto state of the art model, vigiMatch, modifying existing features and introducing new ones to target known shortcomings of the original model. Two support vector machine classifiers, one for medicines and one for vaccines, classify report pairs as duplicates and non-duplicates. Recall was measured using a diverse collection of 5 independent labelled test sets. Precision was measured by having each model classify a randomly selected stream of pairs of reports until each model classified 100 pairs as duplicates. These pairs were assessed by a medical doctor without indicating which method(s) had flagged each pair. Performance on individual countries was measured by having a medical doctor assess a subset of pairs classified as duplicates for three different countries. The new model achieved higher precision and higher recall for all labelled datasets compared to the previous state of the art model, with comparable performance for medicines and vaccines. The model was shown to produce substantially fewer false positives than the comparator model on pairs from individual countries. The method presented here advances state of the art for duplicate detection in adverse event reports for medicines and vaccines.