Progress in pharmacovigilance research based on artificial intelligence and real-world data: a structured narrative review

Traditional pharmacovigilance (PV) relies heavily on spontaneous reporting systems (SRS), which remain vulnerable to under-reporting and limited mechanistic insight. Artificial intelligence (AI) and real-world data (RWD) offer opportunities to strengthen signal detection, risk prediction, and translational safety assessment, but most applications remain at retrospective development or early implementation stages.

We conducted a structured narrative scoping review with critical appraisal, reported using PRISMA-ScR. Literature from 2010 to 2026 was identified through expert review, reference chaining, and regulatory sources. English-language publications on AI or machine learning in clinical pharmacovigilance with real-world data were synthesized thematically across data sources, analytical methods, validation strategies, and implementation challenges.

RWD ecosystems-including electronic health records (EHRs), administrative claims, multi-omics, and global SRS-can be integrated with supervised ML, natural language processing (NLP), graph neural networks (GNNs), and federated learning (FL) for diverse PV tasks. Reported area under the receiver operating characteristic curve (AUROC) values often exceed 0.90 in internal validation, yet external validation, calibration, decision-curve analysis, and prospective implementation evidence remain inconsistently reported. Common data models such as the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) and FL architectures improve interoperability and privacy-preserving collaboration. Causal and explainable AI methods are emerging but require explicit assumptions and mechanistic confirmation before regulatory-grade causal claims.

AI-enabled PV is transitioning from association-focused surveillance toward prediction, mechanistic interpretation, and implementation science, yet widespread clinical or regulatory adoption requires transparent reporting (e.g., TRIPOD + AI, PROBAST), rigorous external and prospective validation, clear separation of association from causation, clinically interpretable outputs, sustained human oversight, and governance within harmonizing regulatory frameworks.