Pharmacovigilance Data Mining Techniques

Abstract

Pharmacovigilance data mining is central to post-marketing drug safety surveillance, enabling systematic detection of adverse drug reaction (ADR) signals from vast spontaneous reporting databases and complementary electronic health record (EHR) sources. This study conducted a comprehensive comparative evaluation of eight signal detection algorithms -- proportional reporting ratio (PRR), reporting odds ratio (ROR), information component (IC/BCPNN), multi-item gamma Poisson shrinker (MGPS), empirical Bayes geometric mean (EBGM), neural network classifier (NNet), random forest (RF), and BERT-based text mining -- across three data sources: the FDA Adverse Event Reporting System (FAERS, 2012-2022; n = 8.4 million reports), EudraVigilance (2012-2022; n = 6.1 million reports), and anonymised EHR data (n = 2.8 million patient-years). Signal detection performance was benchmarked against 428 drug-ADR pairs from the WHO-UMC reference set with known true-positive and true-negative status. EBGM and MGPS achieved the highest sensitivity-specificity balance (EBGM: AUC = 0.874 on FAERS; MGPS: AUC = 0.868), outperforming classical PRR (AUC = 0.784) and ROR (AUC = 0.796). The BERT-based text mining model applied to unstructured EHR clinical notes identified 84 novel drug-ADR associations not detected by any disproportionality method, with 61.9% subsequently confirmed by clinical review -- demonstrating the complementary signal discovery value of NLP-based pharmacovigilance. Multi-source integration (FAERS + EudraVigilance + EHR) improved overall AUC to 0.912 using an ensemble weighting approach, identifying 12 priority signals warranting regulatory review. These findings establish evidence-based criteria for algorithm and data source selection in regulatory pharmacovigilance programmes.