Large Language Models for Drug-Related Adverse Events in Oncology Pharmacy: Detection, Grading, and Actioning

Area of Science:

• * Clinical Informatics
• * Artificial Intelligence in Medicine
• * Pharmacovigilance

Background:

• * Preventable medication harm in oncology frequently results from drug-related adverse events (AEs).
• * Clinical decision-making regarding AEs relies on evidence often found in unstructured clinical texts.
• * Large language models (LLMs) offer advanced capabilities for extracting and reasoning from clinical text data.

Purpose of the Study:

• * To synthesize empirical studies on LLMs and NLP systems for oncology AE detection, grading, and action assignment.
• * To evaluate the potential of LLMs in identifying AEs from clinical documentation.
• * To inform pharmacist-facing recommendations for improving order-level medication safety.

Main Methods:

• * Conducted a narrative review of English-language studies indexed in PubMed, Ovid MEDLINE, and Embase.
• * Included studies utilizing LLMs on clinical narratives and/or authoritative guidance.
• * Excluded non-text modalities and non-empirical articles; nineteen studies met inclusion criteria.

Main Results:

• * LLMs demonstrated potential in detecting oncology AEs from routine notes, often surpassing diagnosis codes for surveillance.
• * Common Terminology Criteria for Adverse Events (CTCAE) grading was feasible but less stable than detection, improving with constrained outputs and patient-level aggregation.
• * Direct evaluation of grade-aligned actions was uncommon; studies reported proxies, highlighting a need for prospective impact reporting.

Conclusions:

• * Evidence supports near-term, pharmacist-in-the-loop use of AI for AE surveillance and review.
• * CTCAE-structured, citation-backed AI outputs can be integrated into pharmacist workflows for review.
• * Future research should standardize reporting, CTCAE usage, and prospectively measure grade-to-action correctness for enhanced decision support.