Pharmacovigilance in the Era of AI and Big Data

The volume of safety data PV systems must manage has outgrown purely manual approaches: more than 25 million individual case safety reports now sit in VigiBase, and FAERS receives over 3 million reports per year. Natural language processing, machine learning, and generative AI now play an active role at every stage of the PV workflow — but the regulatory consensus is unambiguous: AI augments human judgment, it does not replace it.

NLP extracts drug names, adverse event descriptions, and causality language from unstructured text with 85–95% entity-extraction accuracy on well-defined fields. Generative AI drafts narratives and executive summaries in platforms including ArisGlobal NavaX, Veeva Vault Safety, and Oracle Argus Safety — but human medical review and sign-off remains mandatory for all regulatory submissions. The CIOMS Working Group XIV final report (December 2025), the FDA-EMA Joint AI Guiding Principles (January 2026), and EMA GVP Module IX Rev 2 collectively establish the same position: AI is a tool for enhancing PV efficiency and accuracy, not a substitute for qualified human oversight.

A practical way to think about where the line sits: AI is well-suited to tasks with a large volume of similar, well-defined decisions — flagging which literature articles are worth a human’s attention, suggesting a MedDRA code with a confidence score, drafting the first pass of a narrative. It is deliberately kept out of the final call on anything that requires clinical judgment about an individual patient — causality assessment, seriousness classification in an ambiguous case, or the decision to escalate a safety concern. That boundary isn’t a temporary limitation waiting to be engineered away; every current regulatory framework treats human judgment on safety-critical decisions as a permanent requirement, not a stopgap.