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

Preventable medication harm in oncology is often driven by drug-related adverse events (AEs) that trigger order changes such as holds, dose reductions, delays, rechallenges, and enhanced monitoring. Much of the evidence needed to make these decisions lives in unstructured clinical texts, where large language models (LLMs), a type of artificial intelligence (AI), now offer extraction and reasoning capabilities. In this narrative review, we synthesize empirical studies evaluating LLMs and related NLP systems applied to clinical text for oncology AEs, focusing on three decision-linked tasks: (i) AE detection from clinical documentation, (ii) Common Terminology Criteria for Adverse Events (CTCAE) grade assignment, and (iii) grade-aligned actions. We also consider how these findings can inform pharmacist-facing recommendations for order-level safety. We conducted a narrative review of English-language studies indexed in PubMed, Ovid MEDLINE, and Embase. Eligible studies used LLMs on clinical narratives and/or authoritative guidance as model inputs or reference standards; non-text modalities and non-empirical articles were excluded. Nineteen studies met inclusion criteria. LLMs showed the potential to detect oncology AEs from routine notes and often outperformed diagnosis codes for surveillance and cohort construction. CTCAE grading was feasible but less stable than detection; performance improved when outputs were constrained to CTCAE terms/grades, temporally anchored, and aggregated at the patient level. Direct evaluation of grade-aligned actions was uncommon; most studies reported proxies (e.g., steroid initiation or drug discontinuation) rather than formal grade-to-action correctness. While prospective, real-world impact reporting remained sparse, several studies quantified scale advantages and time savings, supporting an initial role as high-recall triage with pharmacist adjudication. Overall, the evidence supports near-term, pharmacist-in-the-loop use of AI for AE surveillance and review, with CTCAE-structured, citation-backed outputs delivered into the pharmacist's electronic health record order-verification workspace as reviewable artifacts. Future work must standardize reporting and CTCAE/version usage, and measure grade-to-action correctness prospectively, to advance toward order-level decision support.