A prospective study comparing AI-based clinical trial eligibility screening with traditional EMR-based screening.

Abstract

e13636 Background: For clinical trials of novel oncology therapies, screening patients for eligibility relies on research staff manually abstracting complex eligibility criteria, which is time-consuming and subject to inconsistent interpretations. AI algorithms based on structured data fields may facilitate automated eligibility criteria extraction, assure more complete screening for the benefit of all patients, improve workflow for research staff, and assure no bias in who is reviewed. AI algorithms can support abstraction and imputation approaches to assure the greatest number of clinical records get reviewed. Prospective validation of such algorithms and approaches for sponsored registrational clinical trials is lacking from literature. Methods: We conducted a prospective controlled observational study of an AI model to prioritize potentially eligible patients, as part of a phase 3, randomized, multi-site therapeutic trial for Relapsed or Refractory Multiple Myeloma. In this observational study, we studied eligibility criteria focused on measurable disease and prior anti-myeloma treatments. The AI eligibility model extracted structured eligibility criteria to predict likelihood of meeting study eligibility. A single CRC (Clinical Research Coordinator) allocated 40 hours to screening in both control and intervention phases of the study. In the control phase, the CRC screened myeloma patients’ electronic health records in alphabetical order by last name. In the intervention phase, the CRC screened a different set of patients ranked by the eligibility likelihood model. Descriptive statistics were used to compare primary outcomes of efficiency (percentage of patients watch-listed) and speed (average time to screen a patient) between control and intervention phase. Patients marked as watch-listed meet eligibility criteria but need additional testing or disease progression before becoming eligible for the study. Results: The AI Eligibility method captured the same share (19%1) of screened patients as the EMR method, with a 3.3x (12.5 versus 41 minutes) improvement per screened patient. Conclusions: The AI eligibility method allows research staff to screen three times more patients with similar outcomes as compared to routine CRC EMR screening. Future work can leverage AI to identify and predict patient progression, further enhancing screening efficiency. [Table: see text]