Human-AI teaming to improve accuracy and efficiency of eligibility criteria prescreening for oncology trials: a randomized evaluation trial using retrospective electronic health records

This randomized noninferiority trial evaluated whether augmenting trained clinical research staff with a neurosymbolic AI language model could improve the accuracy and efficiency of eligibility criteria prescreening for oncology clinical trials, using retrospectively collected charts from 355 patients with non-small cell lung or colorectal cancer. The Human+AI approach demonstrated noninferior and statistically superior chart-level accuracy compared to human-alone prescreening (76.5% vs. 71.1%), with the greatest improvements seen in biomarker, staging, and response criteria. However, efficiency gains were not observed, with average chart review time remaining virtually unchanged between the two conditions, and automation bias emerged as a limiting factor in some domains.

These findings address a pressing bottleneck in clinical research: the manual, labor-intensive prescreening process that prevents many eligible patients from ever being identified for trial enrollment, contributing to the persistently low rates of adult cancer trial participation. While the accuracy improvements were modest, this large randomized evaluation demonstrates that AI language models can meaningfully augment human-driven prescreening workflows, laying the groundwork for future approaches that may better balance accuracy gains with efficiency improvements to ultimately broaden access to clinical trials.