AI-Assisted Clinical Trial Recruitment for Devices

Abstract

Clinical trial recruitment is the most frequent cause of device trial delay and the most preventable. More than 80% of device trials fail to meet original enrolment targets on schedule, extending development timelines by a median of 8-14 months and generating costs that fall disproportionately on smaller device developers without the buffer capital to absorb slippage. AI-assisted recruitment addresses the root causes of this problem: slow patient identification, high screen failure rates from manual eligibility review, and inadequate representation of the real-world device use population in trial cohorts. This study presents the AI-Assisted Device Trial Recruitment Framework (AADTRF), evaluating five AI recruitment approaches -- NLP-based electronic health record eligibility screening (NLP-EHR), machine learning patient matching from device registries (ML-Reg), digital outreach and social media recruitment (DO-SM), federated multi-site cohort identification (Fed-CI), and predictive retention and dropout risk modelling (PR-DRM) -- across four device trial contexts: cardiac implant trials, orthopaedic device trials, neurostimulation device trials, and diagnostic imaging trials. Performance was scored using the Recruitment Effectiveness Score (RES), a weighted composite of enrolment rate improvement (0.30), screen failure reduction (0.20), time to enrolment completion (0.20), population representativeness (0.15), and cost efficiency (0.15). NLP-based EHR screening achieved the highest RES (0.890), improving enrolment rates by a mean 47% and reducing screen failure from 38.4% to 16.2% across four trial contexts. ML registry matching ranked second (0.888) with the best screen failure reduction (0.920) in device-specific contexts. Federated cohort identification (0.877) demonstrated unique value for rare device trial indications, generating eligible patient pools 3.4 times larger than single-site approaches. Digital outreach led on cost efficiency (0.940) and population representativeness (0.920) but produced the highest screen failure rates (0.800) due to self-selection bias in volunteer populations.