AI-driven real-time patient identification for randomized controlled trials.

Abstract

e13565 Background: In India, clinical trial recruitment is a challenging process, often relying on the time- and labor-intensive process of investigators reviewing handwritten charts to identify potential candidates. We developed an AI-driven, real-time solution to address this issue and demonstrated its feasibility by identifying patients that would have been candidates for a 2017-2018 study of metastatic breast cancer (mBC) patients who failed prior chemotherapy. The original trial recruited 46 patients across 16 sites over 7 months in India. Methods: The study protocol was reviewed and applied to six years (2017-2022) of retrospective electronic medical records (EMR) data from a cancer hospital in India’s National Capital Region. Trial inclusion and exclusion criteria were translated to data elements which were extracted from both structured (e.g., lab) and unstructured (e.g., clinical notes) EMR data. For the unstructured data specifically, natural language processing algorithms were developed and implemented over the span of three weeks to extract the specific data elements. The veracity of the extracted data was confirmed via manual review by an oncologist. To assess the sensitivity of the results to the study criteria and provide a range of the potential candidate population size, the criteria were applied in two ways – as a strict application (e.g., ECOG must be reported and must be < = 2) and as a relaxed application (e.g., ECOG may be unreported or, if reported, must be < = 2). Results: Among the initial cohort of 112,239 females between 18 and 65 years of age, 3,203 (3%) were identified as having mBC and failure of prior chemotherapy or relapse. 537 patients remained after applying all inclusion criteria (ECOG status < = 2, LVEF > = 50%, no prior radiotherapy, prior chemo completed 4+ weeks prior, adequate bone marrow, renal, and hepatic function). Exclusion criteria (prior taxane exposure, 2+ prior chemotherapies, cardiac conditions, CNS lesions, motor/sensory neurotoxicities, HBV/HCV/HIV infection) were then applied, resulting in 37 patients that would be potential candidates for the trial, per the strict application of criteria. The relaxed version of the analysis yielded a candidate pool that was five times larger than the strict application, which was largely driven by cases where information was unavailable in the EMR. Conclusions: We demonstrated that the use of an AI-driven platform to apply clinical trial criteria to Indian real-world data can provide an efficient solution to identifying candidate trial patients. Once the criteria are defined and algorithms are developed, patient identification can happen in real time and reduce patient recruitment timelines by months. Future applications will include other cancer types and expand the breadth of data elements extracted to enable the rapid assessment of additional inclusion and exclusion criteria.