irAE.AI: AI-powered exploration of real-world immune-related adverse events

Gabriela Fort; David Stone; Ching-Nung Lin; Arabella Young; Aik Choon Tan

doi:10.1093/jamiaopen/ooag094

irAE.AI: AI-powered exploration of real-world immune-related adverse events

JAMIA Open. 2026 Jun 12;9(3):ooag094. doi: 10.1093/jamiaopen/ooag094. eCollection 2026 Jun.

Authors

Gabriela Fort^{1

2}, David Stone^{1

2}, Ching-Nung Lin^{1

2}, Arabella Young^{1

3}, Aik Choon Tan^{1

2

4}

Affiliations

¹ Huntsman Cancer Institute, University of Utah, Salt Lake City, UT, 84112, United States.
² Department of Oncological Sciences, University of Utah, Salt Lake City, UT, 84112, United States.
³ Department of Pathology, University of Utah, Salt Lake City, UT, 84112, United States.
⁴ Department of Biomedical Informatics, University of Utah, Salt Lake City, UT, 84112, United States.

Abstract

Objectives: Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment, but their clinical benefit is limited by the onset of immune-related adverse events (irAEs). Real-world pharmacovigilance data, such as the FDA's Adverse Event Reporting System (FAERS) database, offer the scale needed to better characterize and understand these toxicities. However, extracting cohorts and normalizing complex fields requires expertise in data preprocessing, technical terminology harmonization, and programming.

Materials and methods: We curated an oncology-specific FAERS dataset of ICI-treated cases (2012Q4-2025Q3), standardized tumor, drug, and adverse event fields, deduplicated reports, and generated a flat-file resource (N = 71 175 irAE cases). To broaden accessibility, we developed a large language model (LLM)-powered analytics assistant that translates natural language queries into executable Python code for cohort filtering, visualization, and statistical testing. Our system also leverages retrieval-augmented generation (RAG) to return clinical guideline-grounded responses to questions about irAEs and their management for research purposes.

Results: We benchmarked the system on curated tasks (question intent classification, TableQA, statistics, plotting) using multiple LLMs and report model- and task-specific performance metrics. As a case study, the platform recovered established irAE trends including differences in irAE patterns between anti-CTLA-4 treatment and other immunotherapy regimens and tumor-specific differences in irAE profiles following anti-PD-1 treatment.

Discussion: Our results show that LLM-driven analytics can reliably translate natural-language queries into reproducible analyses, though model choice affects accuracy across tasks.

Conclusion: This platform provides a transparent, user-friendly approach for exploring real-world immunotherapy safety data to support hypothesis generation, candidate biomarker identification, and irAE risk assessment in immuno-oncology.

Keywords: artificial intelligence; drug-related side effects and adverse reactions; immune checkpoint inhibitors; large language models; real-world evidence.