Knowledge Graph Boosts AI Extraction of Drug Side Effects from Patient Narratives

Extracting adverse drug reactions from the way patients actually talk about their medications has long been a stubborn problem for biomedical text mining. Patient narratives are informal, highly variable in wording, and often far removed from the standardized medical terminology used in drug safety databases. A new open-access study published in BioData Mining on 17 August 2026 presents GraphRAG-ADR, a retrieval-augmented generation framework that pairs a large language model with a directed attributed knowledge graph to close this gap between colloquial expressions and formal medical concepts.

The knowledge graph at the heart of the system combines three sources: the hierarchical terminology of MedDRA, clinically validated drug–side-effect relationships drawn from SIDER, and patient-reported adverse drug reaction expressions from the CADEC dataset. By integrating these, the graph represents both formal medical concepts and the real-world language patients use to describe side effects, giving the language model structured context rather than relying on textual representations alone.

A hybrid retrieval mechanism selects and prioritizes relevant medical concepts using both semantic similarity and graph structure. These retrieved concepts are supplied as context to the large language model, improving the alignment between a patient’s own phrasing and the standardized adverse drug reaction concepts that regulators and clinicians rely on.

GraphRAG-ADR achieved a strict F1-score of 73.02% and a relaxed F1-score of 83.11%, outperforming existing approaches under the same evaluation settings.

According to the study, these results demonstrate that integrating structured biomedical knowledge with large language models improves both accuracy and robustness in adverse drug reaction extraction. Because the article is being shared early as peer-reviewed, accepted research with a permanent DOI, the published version may still undergo further edits before the final Version of Record replaces it. As with any single-dataset evaluation, broader validation across other patient populations and narrative sources would be needed to confirm how well the approach generalizes beyond the PsyTAR benchmark.

Subject of Research: A knowledge graph–enhanced retrieval-augmented framework for extracting adverse drug reactions from patient narratives.

Article Title: GraphRAG-ADR: a knowledge graph–enhanced approach for extracting Adverse Drug Reactions (ADRs) from patient narratives

Article References: Elbiach, O., Nfaoui, E. H., & Grissette, H. (2026). GraphRAG-ADR: a knowledge graph–enhanced approach for extracting Adverse Drug Reactions (ADRs) from patient narratives. BioData Mining. https://doi.org/10.1186/s13040-026-00597-3

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00597-3

Keywords: adverse drug reactions, pharmacovigilance, knowledge graph, large language models, retrieval-augmented generation, MedDRA, SIDER, CADEC, PsyTAR, natural language processing, BioData Mining, GraphRAG-ADR