Pharmacovigilance in Large-Scale Clinical Settings

Abstract

Adverse drug reactions (ADRs) are estimated to cause 197,000 deaths annually in the European Union alone and rank as the fourth leading cause of death in hospitalised patients, yet post-marketing surveillance systems detect fewer than 10% of ADRs that actually occur. The traditional pharmacovigilance pipeline -- spontaneous reporting to national agencies, periodic safety update reports, and signal detection through disproportionality analysis -- was designed for an era when drug exposure data was limited to clinical trial populations of a few thousand patients. Today, with electronic health records covering hundreds of millions of patients and real-time prescription data flowing through pharmacy networks, the opportunity exists to transform pharmacovigilance from a reactive, report-dependent system into a proactive, data-driven safety surveillance platform. We evaluated five pharmacovigilance approaches across 4.2 million patient-drug exposure episodes at six European hospitals (Zurich, Rome, Madrid, Paris, Stockholm, and Tallinn) from 2015 to 2021: spontaneous reporting analysis, disproportionality-based signal detection, EHR-based active surveillance, NLP extraction from clinical narratives, and ML-powered predictive pharmacovigilance. Our Pharmacovigilance Effectiveness Score (PVES) integrates signal detection sensitivity, detection timeliness, false signal rate, population coverage, and actionability. ML-powered predictive pharmacovigilance achieved the highest PVES (0.924) by identifying safety signals a median of 8.4 months before regulatory action, while NLP extraction (PVES 0.906) uncovered 3.2 times more ADR events than coded diagnoses alone. These findings support transitioning European pharmacovigilance toward active, EHR-integrated surveillance augmented by machine learning