Natural Language Processing for Extracting Adverse Drug Reactions Information from Electronic Health Records: Systematic Review

Abstract

Adverse drug events (ADEs) contribute to 5-10% of hospitalizations and cost approximately USD 30 billion annually, yet spontaneous reporting systems capture only 5-10% of actual ADEs due to severe underreporting. This systematic review analyzed 60 peer-reviewed studies (2019-2025) on natural language processing (NLP) methods for extracting ADE information from electronic health record (EHR) clinical notes, following PRISMA 2020 guidelines across five databases (PubMed, IEEE Xplore, ACL Anthology, Scopus, Web of Science). Results demonstrate that transformer-based models, particularly BioBERT and ClinicalBERT, represent the state-of-the-art with F1-scores of 0.85-0.92 on benchmark datasets (n2c2 2018, MIMIC-III, MADE1.0), significantly outperforming rule-based systems (+15-20%) and traditional machine learning methods (+8-12%). Domain-specific pre-training on clinical text proved crucial, improving performance by 3-5% over general BERT models. However, critical challenges persist: negation and speculation detection (30-40% of medical mentions require contextual disambiguation), temporal reasoning for determining ADE onset relative to drug exposure, ambiguous medical abbreviation resolution, and causality assessment. A significant lab-to-clinic gap of 10-15% performance degradation was identified, with only 8% of studies reporting actual clinical deployment experiences. Reproducibility remains problematic, with merely 23% of studies sharing code and 15% providing trained models. Future priorities include developing few-shot learning approaches to address limited labeled data (~5,000 annotated clinical notes publicly available), enhancing model interpretability through explainable AI methods, conducting multi-center external validation studies, and establishing standardized evaluation protocols. This review provides evidence-based guidance for researchers developing NLP methods, practitioners implementing ADE detection systems, and policymakers formulating standards for NLP-based pharmacovigilance.