37AI-assisted Clinical Trial Matching for Cholangiocarcinoma Patients
Yin Fang, Amanda Nottke, Melinda Bacchini, Lourdes Rocha-Nussbaum, Kari Ramage, Juan W Valle, Shubo Tian, Qiao Jin, Zhiyong LuAbstract
Background
Identifying appropriate clinical trials for oncology patients is challenging due to complex eligibility criteria, heterogeneous clinical documentation, and evolving trial landscapes, particularly in rare cancers such as cholangiocarcinoma. Manual trial screening is time consuming and may overlook relevant opportunities. This study presents an AI-assisted trial matching system customized to support efficient and clinically meaningful trial identification for cholangiocarcinoma patients.
Methods
We developed an AI-assisted trial matching system that supports real-world cholangiocarcinoma patient information in different formats, including both structured EHR-based data and unstructured narrative clinical summaries, and prioritizes the most relevant candidate trials for clinician review. For a given patient, the system analyzes clinical context and matches it against hundreds of cholangiocarcinoma-related trials, including biomarker-selected trials that allow cholangiocarcinoma. Subsequently, the system produces a ranked list of candidate trials with structured explanations (including eligible reasons, ineligible reasons, and missing information). System performance was manually assessed by clinicians across 19 patient cases.
Results
We evaluated the system on 9 structured and 10 unstructured cholangiocarcinoma patient cases. Eligibility agreement with clinician review was 73.4% (structured) and 77.8% (unstructured). Among clinician-confirmed eligible trials, 56.3% (structured) and 73.8% (unstructured) were judged clinically appropriate to recommend. The system consistently prioritized clinically meaningful trials and was effective at identifying biomarker-specific studies embedded within broader solid tumor protocols, which clinicians noted are challenging to identify through manual screening alone. Clinicians also found the explicit explanation of missing or uncertain eligibility information valuable for guiding follow-up eligibility review. From a clinician perspective, the Al-generated explanations were efficient to review, typically requiring 1-2 minutes per trial.
Conclusion
This study demonstrates the real-world usability and effectiveness of using AI-assisted trial matching for cholangiocarcinoma trial screening. Future directions include further refinement to improve performance, and additional testing in a real-life advocacy setting.