2026 ECNP Congress

PS02-2039

Artificial intelligence in pharmacovigilance: advances in adverse event detection, signal management, and regulatory reporting

Aikaterini Stravoravdi — D. Spachos, P. Bamidis, G. Papazisis · School of Medicine- Aristotle University of Thessaloniki, Clinical Pharmacology & Clinical Research Unit, Thessaloniki, Greece

2026 ECNP Congress

Background

Pharmacovigilance (PV) databases such as VigiBase (40 million reports from 180+ WHO PIDM members, as of February 2025) and the FDA Adverse Event Reporting System (FAERS>30 million reports) have grown beyond the capacity of traditional manual methods for adverse event (AE) detection, signal management, and regulatory reporting. Artificial intelligence (AI)—including machine learning (ML), natural language processing (NLP), and large language models (LLMs)—offers computational approaches to address this challenge, yet the evidence supporting specific AI methods across PV tasks has not been systematically mapped.

Aims

This review aimed to (1) survey AI applications across the PV workflow, (2) evaluate the strength and comparability of the quantitative evidence base, and (3) identify priority gaps for future research.

Methods

We searched PubMed, Scopus, and Google Scholar for English-language articles published January 2020–December 2025 using pharmacovigilance terms combined with AI/ML/NLP/LLM/federated learning. Inclusion criteria were original research, systematic reviews, or scoping reviews reporting quantitative performance metrics on PV-relevant tasks, plus regulatory guidance documents from FDA, EMA, ICH, and CIOMS. The search yielded 42 articles for detailed review (25 cited). Performance metrics (F1-score, AUC, precision, recall) were extracted and compared descriptively across task types and model architectures. No formal meta-analysis or PRISMA screening was conducted.

Results

Across reviewed studies, performance varied substantially by task and evaluation methodology. For AE extraction from electronic health records, NLP systems achieved F1-scores of 0.66–0.97 (n=7 studies), with transformer-based models (BERT variants) consistently outperforming conditional random fields and rule-based approaches (F1: 0.82–0.97 vs. 0.66–0.85). For signal detection in FAERS, gradient boosting machines achieved AUC values of 0.93–0.97 (nivolumab and docetaxel), outperforming traditional disproportionality methods including reporting odds ratio, proportional reporting ratio, and Bayesian confidence propagation neural network. GPT-4 matched or exceeded prior state-of-the-art F1-scores for adverse reaction extraction from structured product labels without task-specific fine-tuning. A BERT-based triage model for adverse drug reaction report seriousness classification achieved F1-scores close to human-level performance at the Swedish Medical Products Agency. A federated learning proof-of-concept across regulatory agencies showed performance comparable to centrally-trained models (AUC gains of 0.04–0.09 over local baselines). However, temporal validation was performed in fewer than 20% of studies, external validation across databases was rare, and no shared evaluation benchmarks for PV-specific AI tasks were identified.

Conclusions

AI methods demonstrate strong performance on individual PV tasks—particularly signal detection (AUC >0.93) and ADE extraction (F1 up to 0.97)—but the evidence base is limited by heterogeneous evaluation, infrequent temporal and external validation, and the absence of standardized benchmarks. Regulatory frameworks from the FDA, EMA, and CIOMS converge on validation, transparency, and human oversight requirements, yet operational governance standards for generative AI remain underdeveloped. Priority research gaps include shared benchmarks, causal inference frameworks, multi-modal surveillance pipelines, and equity auditing of deployed systems.Realizing the potential of AI in pharmacovigilance requires not only technical advances in model performance but also progress in evaluation standardization, causal inference, federated learning deployment, and equity-centered design.The transition from reactive case management to proactive, real-time safety surveillance will depend as much on these infrastructural and methodological investments as on improvements in individual model accuracy.

References

  1. Ambaw AJ, Alemayehu B, Leontiadis GI, et al. Leveraging natural language processing and machine learning methods for adverse drug event detection in electronic health/medical records: a scoping review. Drug Safety. 2025;48(4):321–337
  2. Abernethy R, Bowers J, Derry S, et al. Evaluating generative large language models for extraction of adverse drug reactions from structured product labels. Drug Safety. 2025.
  3. Wang X, Altman RB. Leveraging FDA labeling documents and large language model to enhance annotation, profiling, and classification of drug adverse events with AskFDALabel. Drug Safety. 2025.
  4. Pacurariu AC, Touraine C, Lenaghan E, et al. BERT based natural language processing for triage of adverse drug reaction reports shows close to human-level performance. PLOS Digital Health. 2024;3(2):e0000409.
  5. Artificial intelligence in pharmacovigilance: a narrative review and practical experience with an expert-defined Bayesian network tool. International Journal of Clinical Pharmacy. 2025;47(5):932–944.