Agentic AI-enabled exploration of real-world immune-related adverse events

Immune checkpoint inhibitors (ICIs) have transformed cancer therapy, but their clinical benefit is often limited by the onset of immune-related adverse events (irAEs), which can be severe and lead to treatment interruption or discontinuation. A deeper understanding of irAEs is urgently needed to identify patients at highest risk of developing adverse events and to guide strategies for irAE prevention and risk management. The FDA's Adverse Event Reporting System (FAERS) database contains over 32 million adverse event reports submitted to the FDA to support drug safety surveillance. However, although a public dashboard exists to perform basic exploration of FAERS data, extracting immuno-oncology-related safety events or executing complex, multiparameter queries of this data still requires substantial technical expertise, programming skills, and a familiarity with the underlying database structure. To address this gap and to enhance the accessibility of this public resource, we downloaded all FAERS reports from 2012-2025 and systematically filtered for cancer cases treated with ICIs. We curated a high-quality, oncology-specific irAE dataset and generated a standardized flat-file resource for downstream data exploration and analysis. To further facilitate access and enable non-programmers to easily explore these data, we developed an agentic AI-driven interface and workflow that allows natural language querying of the dataset. Our system uses open-source large language models with specialized prompting to classify user intent and generate executable python code for complex analytical tasks including filtering, visualization, and statistical analyses, returning results in real time. This framework enables interactive and flexible exploration of irAE patterns across tumor types, drug classes, and other clinical features. Preliminary analyses recapitulate known irAE associations in cancer (e.g. elevated endocrine and cutaneous toxicities compared to other irAEs in anti-PD-1-treated patients) and reveal potential tumor and treatment-specific irAE profiles that warrant further investigation. In summary, this platform provides a transparent, scalable, and user-friendly approach for mining real-world immunotherapy safety data that may be leveraged to inform biomarker discovery, fuel hypothesis generation, and/or guide irAE risk mitigation strategies in immuno-oncology.