Integration of artificial intelligence into pharmacovigilance: A systematic review.
Iman El Sayed, Ramez Naguib Bedwany, Mai Salama et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review mapped artificial intelligence and machine-learning applications in post-marketing pharmacovigilance, drawing on 67 studies identified through searches of eight databases and supplementary sources. Safety-signal detection was the most frequent application (50.5%), followed by adverse-effect prediction (21%) and case-processing efficiency (15%). XGBoost and ensemble methods were common with structured data, while BERT-family models predominated for unstructured text. Evidence was largely retrospective; external validation was the weakest quality domain, with a mean APPRAISE-AI score of 52%, and prospective implementation studies were rare.
Key findings
- Safety-signal detection accounted for 50.5% of applications, followed by adverse-effect prediction (21%) and case-processing efficiency (15%).
- Drug-drug interaction detection represented 7.5% of applications, and high-risk patient identification 6%.
- External validation was the weakest APPRAISE-AI domain, with a mean score of 52%; prospective implementation studies were rare.
Why it matters
AI methods are being applied across several pharmacovigilance tasks, but promising model performance does not establish reliable benefit in routine safety monitoring. The review highlights gaps in independent validation and prospective implementation, as well as variation in methods that limits conclusions about generalizability and clinical impact. These gaps matter when assessing whether automated approaches can support patient-safety decisions.
What to keep in mind
The review reports heterogeneous methodologies, limited external validation, and few prospective implementation studies, restricting conclusions about generalizability and clinical impact.