Understanding Unrealized Trial Enrollments Following Patient-to-Trial Matching with OncoLLM in Surgical Oncology

Quick Summary

At Medical College of Wisconsin, Triomics' OncoLLM screened over 500 GI surgical oncology patients over a six-month study period, resulting in eight trial enrollments.

Link

Surgery

Introduction: Clinical trials are essential to improving cancer care, however identifying patients in surgical clinics can be challenging due to the manual and time-intensive nature of enrollment tasks, which are difficult to integrate into the fast-paced workflow of surgical practice. Artificial intelligence (AI) tools, including those using Large Language Models (LLM), can potentially improve this by automating aspects of clinical trial matching. The objective of this study was to identify why patients did not enroll on a study after receiving a trial screening recommendation from an LLM-based platform for clinical trial matching.

Methods: This retrospective study included patients seen by surgeons participating in an LLM-based platform implementation within gastrointestinal (GI) surgical oncology clinics (July - December 2024). OncoLLM is a fine-tuned LLM trained using institutional clinical data, cancer guidelines, and oncology textbooks. Weekly, all patients scheduled in participating clinics are processed using OncoLLM, and potential trial matches are forwarded to provider teams for review. The primary outcome was "unrealized trial enrollment" defined as a patient-to-trial match that did not result in a trial enrollment. Secondary outcomes included LLM-evaluated patients and reasons for unrealized trial enrollment.

Results: 514 (100%) patients during the study period were successfully evaluated using OncoLLM, resulting in 33 patients (6.4%) with 35 patient-trial matches. Of these, 27

matches (77.1%) represented unrealized trial enrollments, while 8 of these matches resulted in enrollment (22.9%). Among the unrealized trial enrollments, 9 were ultimately

ineligible (33.3%), 5 declined participation (18.5%), 4 did not enroll based on provider discretion (14.8%), and the reason for non-enrollment was not documented for 9 of these

matches (33.3%). Of the ineligible matches, 6 remained potential candidates pending change in clinical status and 3 were model inaccuracies.

Conclusion: OncoLLM can successfully automate the process of clinical trial matching in surgical clinics, with over 500 patients screened for eligibility using AI. This study

highlights the potential for an AI-based platform to automate the labor-intensive process of manual clinical trial matching by systematically screening all patients. Identifying and addressing reasons for unrealized trial enrollments can optimize accrual, reduce disparities, and advance cancer care.