AI-Driven Curation and Analysis of Immune-Related Adverse Events in Cancer Immunotherapy Using Secondary Immune-Oncology Databases

About This Grant

ABSTRACT Background and Significance Immune checkpoint inhibitors (ICIs) have revolutionized cancer therapy by improving outcomes for patients with various malignancies. However, ICIs can induce immune-related adverse events (irAEs) that result in treatment interruptions or discontinuation, thereby diminishing therapeutic effectiveness and compromising patient outcomes. Meanwhile, there remains an unmet need for a comprehensive data registry to support research on irAEs. The curation of such data often requires extensive manual labeling across multiple sources including clinical notes. Therefore, developing an AI-assisted curation tool to automate irAE information extraction and a predictive model to identify patients at risk early in treatment is critical to advancing precision immunotherapy. Innovation This project leverages the Georgetown Lombardi Comprehensive Cancer Center’s Immune-Oncology (IO) registry, encompassing over 2,500 immunotherapy-treated patients from ten MedStar hospitals, and harmonized data from Hackensack Meridian (HMH) Health’s John Theurer Cancer Center with 1,300 patients. Guided by oncologists’ clinical expertise, this proposal presents a framework that integrates natural language processing (NLP) and hybrid machine learning to automate the extraction, prediction, and interpretation of irAEs using a comprehensive spectrum of clinical data. The resulting tools and models will be shared to advance data curation and treatment management within the cancer research community, supporting the broader goal of precision oncology. Approach and Expected Outcomes We will first develop a large language model (LLM)–based NLP pipeline to extract irAE evidence, type, and grade from de-identified clinical notes, creating a reusable, annotated corpus (Aim 1). Next, we will integrate structured features including demographics, treatment regimens, tumor characteristics, and laboratory trends, into hybrid predictive models that combine static and longitudinal information (Aim 2). The model will be interpreted using explainable AI methods, enabling clinical validation of key predictive features in collaboration with oncologists. External validation using HMH data will test model generalizability across institutions (Aim 3). This project will generate an accurate and interpretable predictive model for irAE prediction and an NLP tool for automated curation. Both resources will be made publicly available to advance cancer information retrieval and enhance clinical decision-making in cancer treatment management.