Improvements in screening precision with AI-enhanced clinical trial matching using EMR source data for cancer clinical trials.

603 Background: Trial enrollment in the US remains low despite calls for more US representation in cancer trials. Efforts to proactively screen patients for eligibility are needed, but manual screening is inefficient. Tech-enabled clinical trial matching (CTM) using algorithmic matching over structured patient data has been deployed but efficiency gains over manual screening have been incremental. Emerging technologies, including large language models (LLMs), may address CTM challenges, but little is known about the benefit of AI used for this purpose. Methods: Two CTM approaches were evaluated by research staff at New York Cancer & Blood Specialists (NYCBS), a large community oncology practice, for 4 therapeutic, interventional clinical trials: 1) an algorithmic, rule-based platform using structured clinical data and pre-structured data extracted from clinical documents; and 2) an AI-native platform using a structured data prefilter, LLM-based reasoning of patient unstructured documents from the EHR, and natural language summaries of trial eligibility criteria. The AI-native platform supports broader, more nuanced eligibility criteria to identify a more targeted set of patients at the optimal point in their care journey. Patients identified by each method for the 4 trials were reviewed by NYCBS research staff. Results: AI-native CTM improved precision following initial human screening over a rule-based CTM approach. In 3 trials, the AI-native approach reduced identified patients for screening by 31-87%, while in a fourth trial the patients identified increased by 32% but resulted in a doubling of eligible patients after human screening. Upon review, the difference in eligible patients with the rule-based approach appeared to be mostly driven by patients who, while potentially eligible in the future, were far from a treatment decision-point relevant to the trial. Conclusions: Deployment of AI increased precision over a previous structured data-based CTM solution in 4 cancer clinical trials across different diseases at an oncology practice. Limitations include lack of data on final eligibility determination and ongoing measurement of false negative rate (screening out potentially eligible patients). Trial CTM Approach Patients Identified by CTM for Screening Patients Screened by Research Staff (Ongoing) Eligible Patients* Patient Discovery Precision HER2+ Metastatic Breast Cancer Rule-based 752 572 227 40% AI-native 96(87% reduction) 59 43 73% Metastatic Non-Small Cell Lung Cancer Rule-based 466 231 46 20% AI-native 203(56% reduction) 74 23 31% Metastatic Breast Cancer Rule-based 1,336 875 422 48% AI-native 916(31% reduction 399 395 99% Metastatic Colorectal Cancer Rule-based 252 189 20 11% AI-native 332(32% increase) 66 40 61% *Eligible patients defined as those considered still eligible or likely to become eligible after initial research staff review.