Enterprise-level AI screening for oncology trials: Improving patient identification with OncoLLM

Quick Summary

Compared to traditional methods, OncoLLM resulted in a 40% increase in identified trial matches.

Link

JCO

Background: Comprehensive clinical trial screening for all oncology patients is challenging due to multiple factors. Traditional approaches rely on manual review, making enterprise-level trial screening infeasible. Centralized Clinical Trials Offices (CTO) prioritize interventional therapeutic trials, further limiting screening for interventional, non-therapeutic trials. Large language model (LLM)-based AI can enhance efficiency by systematically screening all patients against an institution’s full clinical trial portfolio before each visit. This study evaluated whether an LLM-based system could improve trial-to-patient matching using a representative clinical trial. Methods: This retrospective study assessed patients at the Medical College of Wisconsin Cancer Center from December 2024 to January 2025. OncoLLM, a fine-tuned LLM trained on institutional clinical data, cancer guidelines, and oncology textbooks, processed unstructured EHR data for patients scheduled in participating clinics to identify potential trial matches. OncoLLM is deployed using Triomics’ PRISM platform, employing a two-tiered approach: Primary screening: Identify relevant trials using disease site, histology, stage, and biomarker status. Deep screening: Evaluate full inclusion/exclusion (I/E) criteria, transforming each criterion into a structured query with a determination of eligibility. The primary outcome was the number of trial-to-patient matches to ALLIANCE-A222004-ANOREXIA (NCT04939090). Secondary outcomes included measuring LLM accuracy, defined as a recommendation accepted by the study team, and reasons for inaccuracy.

Results: A total of 2,277 patients with 38 tumor types were screened, representing 100% of patients seen during the study period. 49 patients were recommended as high-probability trial candidates for further assessment, and all recommended patients met key inclusion criteria, warranting deep screening by PRISM. A total of 22 of the 49 recommendations were accepted by the study team, resulting in patients being watch-listed for the study. Reasons for rejection included not meeting study-specific criteria, medical/health related, and logistical or health communication. Compared to traditional methods, OncoLLM resulted in a 40% increase in identified trial matches.

Conclusions: OncoLLM enabled enterprise-scale, AI-driven screening, identifying a greater number of potentially eligible patients than manual methods. OncoLLM demonstrated high criterion-level accuracy and feasibility for broad-scale clinical trial matching. As precision oncology continues to drive complexity in trial eligibility, AI-powered tools like OncoLLM are essential for democratizing trial access, optimizing enrollment, and reducing inefficiencies in oncology research.