Abstract 2755: Large scale extraction of adverse events by large language models uncovers latent structure of treatment toxicity.

Abstract Managing treatment toxicity is a major challenge in oncology. Large-scale data on treatment related adverse events (AEs) could help oncologists better understand and predict toxicity. However, most AE datasets are limited to small numbers or restricted types of events due to the labor-intensive manual reporting traditionally required to document AEs. Automated extraction of AEs from the electronic health record (EHR) could enable large-scale analysis but is a challenging task as information about most AEs is in diffuse, unstructured data such as clinical notes.We therefore developed and validated an agentic large language model (LLM) pipeline, AE-Extract, to extract and grade AEs from oncologists’ progress notes. The pipeline is highly sensitive with a recall of >95% for events in the same organ system and a precision of 67% compared to manual annotation. Lower precision relative to recall reflects uncertainty in the attribution of events to treatment effects as well as design choices to prioritize higher recall over precision. To comprehensively characterize AEs during cancer therapy, we subsequently applied AE-Extract to a set of 84,684 progress notes from an academic cancer center. We identified 279,457 total AEs, a mean of 3.3 per note. To identify recurrent patterns of treatment toxicity, we applied non-negative matrix factorization (NMF) to the AE data and identified 85 latent factors that capture patterns of toxicities in the dataset. These latent patterns capture co-occurrence of mechanistically related events such as neuropathy and falls as well as associations in different organ systems without known mechanistic connection such as pruritus and colitis in patients receiving immunotherapy or neuropathy and nail changes in patients receiving cytotoxic chemotherapy. As the latent factors quantify core patterns of AEs, we next studied how they could be used as robust feature sets for analysis of treatment toxicity. First, we found that the latent patterns identified in the first part of a patient’s treatment course can predict development of new AEs as well as worsening severity of existing AEs. We also found these patterns of treatment toxicity correlated with patient demographics and comorbidities: we recapitulated known associations such as the link between gender and cancer induced nausea/vomiting as well as novel relationships such as a link between ischemic heart disease and vestibulocochlear symptoms including tinnitus and hearing impairment.Overall, our work demonstrates a new method for accurate, high-throughput extraction of AEs from clinical notes. We show how comprehensive evaluation of treatment toxicity allows for better characterization of patterns of AEs, identifies robust associations between treatment toxicity and patient characteristics, and may serve as a starting point for building personalized models to predict treatment toxicity. Citation Format: John Lazar, Divneet Mandair, Catherine C. Smith, Travis Zack. Large scale extraction of adverse events by large language models uncovers latent structure of treatment toxicity [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2755.