Large language models for optimizing clinical trial recruitment in ICUs: application to ventilator-induced diaphragm...

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Clinical Trial Methodology

Background:

• Ventilator-induced diaphragm dysfunction (VIDD) is a common complication in mechanically ventilated ICU patients.
• Identifying eligible patients for VIDD clinical trials is operationally challenging.
• This study explores using large language models (LLMs) for automated patient prescreening.

Purpose of the Study:

• To evaluate the efficacy of LLM-based prescreening for identifying clinical trial candidates.
• To estimate recruitment capacity for a phase 2 VIDD trial.
• To assess the impact of LLMs on clinician review time.

Main Methods:

• Developed an LLM pipeline to screen ICU discharge summaries for trial eligibility.
• Deployed the pipeline to screen 2024 ICU stays, followed by expert adjudication.
• Used expert-annotated data to evaluate model performance with F1-scores.

Main Results:

• The GPT-OSS:120B model achieved a criterion-level F1-score of 0.82.
• Prescreening identified 133 eligible patients from 1,342 ICU stays, with a 72% positive predictive value.
• The LLM-assisted workflow reduced clinician review time by an estimated 86%.

Conclusions:

• LLM-based prescreening is a promising method for identifying critical care trial candidates.
• This technology can prioritize patients for expert clinician review.
• Future implementation requires expert validation and ongoing monitoring for safety and generalizability.