Key takeaway
Europe received 1.77 million adverse reaction reports in 2025, but only 60 confirmed signals reached PRAC. How AI speeds up signal detection, and why medical judgment still decides.
Short answer: Yes, at the detection and prioritisation stages. AI can screen literature, remove duplicate reports, cluster patterns across millions of cases and flag rare events earlier than classical statistics. It cannot decide what a signal means. Causality, clinical interpretation and benefit–risk decisions stay with qualified medical experts, and regulators now expect that split to be documented.
Case volumes keep growing, products are more complex, and safety data now arrives from many more places: electronic health records, claims databases, registries, patient-reported outcomes, wearables and social media. Most of it is unstructured narrative text. Classical signal detection was designed for lower volumes and linear workflows. It is methodologically robust, but largely manual, retrospective and expensive to scale.
The numbers show the funnel. Europe's safety database took in 1.77 million reports in 2025. EMA reviewed 1,201 potential signals, and only 60 confirmed signals reached PRAC for assessment. Every step in between depends on expert time, and false positives and duplicate cases consume much of it.
At the 5th Pharmacovigilance & Drug Safety Conference in Madrid in April 2026, Dimitris Zampatis, Director Product Patient Safety at Sandoz International, presented "Signal Management in the AI Era", sharing his personal views. His conclusion was that AI changes where people spend their time, not whether they are needed:
"AI augments signal management by shifting human effort from detection to interpretation."
Dimitris Zampatis, 5th Pharmacovigilance & Drug Safety Conference, April 2026
He also noted that signal management remains a key inspection focus, and that regulators assess AI as part of the signal management system, not as a standalone tool.
Zampatis walked through five published examples of AI applied across the signal lifecycle:
Day to day, the most common uses are less dramatic: automated literature screening, machine-learning duplicate detection and prioritisation dashboards that tell reviewers where to look first. EMA is moving the same way. Its 2025 report describes new signal analytics tools as "a significant step towards an automated and integrated system for signal detection".
Zampatis was equally clear about the limits. Models are biased towards frequent events, so rare but critical signals can be under-detected. They cannot fully adjust for confounding in real-world data, they learn poorly from small early-lifecycle datasets, and they can rank signals on data artefacts rather than clinical relevance.
That is why five responsibilities stay with qualified medical professionals: causality assessment, biological plausibility, clinical interpretation, benefit–risk integration and regulatory decision-making. In his hybrid model, AI handles detection, pattern recognition and prioritisation; people handle validation, medical assessment and regulatory action.
The expectations are converging. Zampatis listed four: proactive and timely surveillance, clearly defined human accountability for AI-supported decisions, documented validation with ongoing performance monitoring, and traceable audit trails, including overrides. The EMA–FDA principles ask for proportionate validation and oversight, and for life cycle management that monitors models and re-evaluates them to address data drift. The CIOMS report distinguishes human-in-the-loop, where a person makes the final decision, from human-on-the-loop, where people monitor automation continuously.
The common inspection risks follow from that: weak AI governance, bias in training data, poor change control when models are retrained, and over-automation without documented medical oversight.
Header photo: 1981 Digital on Unsplash.
Signal detection is the search for new or changed safety information about a medicine, such as a possible new adverse reaction, in reports, literature, clinical trials and real-world data. A signal is then validated, prioritised and assessed before any regulatory action.
No. AI can screen and prioritise data at scale, but causality assessment, clinical interpretation and benefit–risk decisions require qualified medical professionals, and regulators expect clear human accountability for AI-supported decisions.
In 2025, EMA reviewed 1,201 potential signals and PRAC assessed 60 confirmed signals, 26 of which led to updates of product information.
Documented validation, ongoing performance monitoring, change control for model updates, traceable audit trails including overrides, and defined human oversight, in line with the EMA–FDA good AI practice principles and the CIOMS XIV report.
The Pharmacovigilance & Drug Safety Conference, organised by UBIQ Events, brings together pharmacovigilance and drug safety leaders to discuss AI in case processing, signal management and inspection readiness.
