Large language models for optimizing clinical trial recruitment in ICUs: application to ventilator-induced diaphragm dysfunction Authors: Kieffer Korvin, Zeno Loi, David Morquin, Boris Jung, Joris Pensier, Yassir Aarab, Nacim Benchabane, Eddine Bendiab, Sarah Colombani, Alain Lacampagne, Kevin Yauy, Stefan Matecki Published: 2026 ABSTRACT BackgroundVentilator-induced diaphragm dysfunction (VIDD) is a frequent and under-recognized consequence of prolonged mechanical ventilation in intensive-care unit (ICU) patients. Identifying eligible candidates for clinical trials targeting VIDD remains a major operational challenge. This study evaluates the use of large language models (LLMs) to automate patient prescreening from ICU discharge summaries and estimate recruitment capacity for a future phase 2 trial.MethodsWe developed an LLM-based prescreening pipeline to assess trial eligibility criteria from ICU discharge summaries, which was deployed to screen all 2024 ICU stays. Stays that were flagged as potentially eligible underwent expert adjudication. An enriched set of 50 ICU stays was independently annotated by six clinicians to define a reference standard, which was used to evaluate criterion-level model performances using F1-scores.ResultsThe best-performing model was GPT-OSS:120B with a criterion-level F1-score of 0.82. When applied to consecutive 1,342 ICU stays from Montpellier University Hospital in 2024, the selected model identified 532 patients with ≥ 3 days of mechanical ventilation. After applying exclusion criteria, 185 patients remained potentially eligible. Expert review confirmed 133 patients as eligible, resulting in a positive predictive value of 72% (95% CI 65–78). The LLM-assisted workflow resulted in an estimated 86% reduction in clinician review time. The LLM-based prescreening pipelines achieved criterion-level F1 scores ranging from 0.73 to 0.82, with GPT-OSS:120B demonstrating the highest performance.ConclusionsLLM-based prescreening offers a promising approach for identifying trial candidates in critical care, prioritizing candidates for clinician review. Future deployments should include targeted expert validation and ongoing monitoring to ensure safety and generalizability.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13054-026-06127-1. Research letter Recruitment failure remains a critical bottleneck in intensive care research, with nearly half of all acute care trials facing premature discontinuation, predominantly due to insufficient enrollment [1]. This frequent failure is often attributed to the overestimation of eligible patient pools and the complexity of real-time screening in high-workload environments [2]. To prevent such failures, it is crucial to accurately assess the inclusion capacity of investigating sites before definitively finalizing a trial’s inclusion and exclusion criteria. Indeed, overly restrictive criteria, while scientifically sound, often make it practically impossible to reach the required sample size. Before the advent of Artificial Intelligence, estimating this recruitment viability, to adapt the protocol accordingly, relied on the laborious and time consuming retrospective manual review of electronic health records (EHRs) over extended periods. Recently, Large Language Models (LLMs) have demonstrated strong capabilities in processing complex medical narratives. We report our experience in developing and validating an automated LLM-based prescreening pipeline designed to forecast recruitment viability for clinical trials. To illustrate this approach, we simulated a trial for ventilator-induced diaphragm dysfunction (VIDD), a condition affecting 60–80% of patients under mechanical ventilation. Its associated diaphragmatic failure delays liberation from ventilatory support and increases ICU length of stay, highlighting the necessity of upcoming clinical trials [3]. Our evaluation utilized retrospective data extracted from the Clinical Data Warehouse (eDOL) of Montpellier University Hospital, following approval by the Institutional Review Board (No. A034/2024-10-152/001), which waived the need for written informed consent due to the study’s retrospective nature. We retrieved free-text discharge summaries of patients admitted to the ICU between January 1 and December 31, 2024 who underwent invasive mechanical ventilation. The goal was to automatically determine patient eligibility based on a predefined set of realistic inclusion and exclusion criteria. The inclusion criterion was invasive mechanical ventilation (MV) for ≥ 3 consecutive days. To isolate the specific impact of ventilation and minimize confounding factors affecting diaphragmatic integrity independently of MV, eleven exclusion criteria were defined: heart failure, acute respiratory distress syndrome (ARDS), COPD, neuromuscular disease, congenital heart disease, active statin therapy, sepsis, pregnancy, current smoking, prone positioning, and vulnerable populations. For automated assessment, criteria were operationalized through decomposition into atomic subcategories, with final status assigned via predefined deterministic rules (Supplementary Table 1). Six clinicians independently annotated the evaluation set in a blinded manner using Label Studio, with final labels established by majority vote. The core of our automated Clinical Research Assistant was built upon five existing open-weight models : openai/gpt-oss-120b, mistralai/Mistral-Large-Instruct-2407, mistralai/Magistral-Small-2506, google/medgemma-4b-it and deepseek-ai/DeepSeek-R1-Distill-Llama-70B. No additional training or fine-tuning was performed. To ensure high reliability and mitigate the risk of hallucinations, we implemented a Self-Consistency Assessment (SCA) strategy [4]. For each clinical note, each criterion was assessed over 11 independent iterations and the final criterion status was determined by a majority vote across these iterations. The pipeline was prompted via a zero-shot classification approach, where the LLM was asked to act as a hospital text classifier, returning its predictions for each criterion and justifications in a standardized JSON format. Across configurations, the prompt template, inference parameters, and self-consistency procedure were kept constant; only the underlying foundation model varied. All inferences were conducted on secure, on-premise servers isolated from external networks to ensure strict compliance with patient data privacy regulations (Supplementary Appendix S1-S6). Our prescreening app is publicly available at https://github.com/ERIOS-project/EriosClassifier. A central challenge in evaluating AI systems for clinical research lies in defining an appropriate reference standard [5]. The pipeline’s performance was validated against a gold standard established by a panel of six physicians (ICU and infectious disease specialists) who independently and blindly annotated a reference set of 50 ICU stays, purposively selected to be enriched with rare exclusion criteria, thereby ensuring a robust assessment of criterion-level extraction rather than a simple estimation of population-level prevalence. Among the evaluated LLM configurations, the OpenAI GPT-OSS:120B model combined with the SCA strategy achieved the best criterion-level performance, with a macro-F1 score of 0.82 (95% CI 0.67–0.87), a sensitivity of 0.89 (0.71–0.92), and a specificity of 0.81 (0.68–0.89). (Supplementary Table 2). Performance varied across eligibility criteria, with an overall high specificity and more heterogeneous sensitivity. The inclusion criterion of invasive mechanical ventilation for at least 3 days was classified with excellent accuracy (F1, 0.966). Among exclusion criteria, performance was highest for pregnancy (F1, 1.000), COPD (F1, 0.941), ARDS (F1, 0.909), and prone positioning (F1, 0.909). Lower performance was observed for neuromuscular disease (F1, 0.500), vulnerable populations (F1, 0.667), sepsis (F1, 0.698), and heart failure (F1, 0.759) (Supplementary Table 3). The best-performing configuration (GPT-OSS:120B) was subsequently applied to all 1,342 ICU stays from 2024.The selected model identified 532 stays with ≥ 3 days of invasive mechanical ventilation, and after applying exclusion criteria, 185 stays were flagged as potentially eligible. Manual expert review of the 185 candidates confirmed eligibility to the planned phase 2 trial for 133 patients, corresponding to 25% of stays with MV ≥ 3 days and 9.9% of all ICU stays screened. This corresponded to a patient-level PPV of 72% (95% CI, 65–78). The automated screening of the entire dataset resulted in an estimated 86% reduction in clinician review time. The criterion-level results suggest that the selected model could handle several forms of negation and narrative complexity. However, lower performance in identifying sepsis (F1: 0.698) and heart failure (F1: 0.759) likely reflects the syndromic nature of these conditions, which often rely on implicit documentation and nuanced clinical interpretation rather than explicit labeling. Unlike objective parameters such as ventilation duration, these criteria are subject to higher inter-observer variability, making them more susceptible to misclassification when the model’s extraction logic differs from the physician’s interpretive judgment. The observed difference between criterion and patient-level PPV is due to error accumulation across multiple classification tasks. This workflow aims to prioritize clinical review and forecast trial feasibility rather than provide autonomous adjudication. Acknowledging that only flagged cases were expert-reviewed, false negatives at the cohort level cannot be definitively ruled out. This experience highlights the potential of advanced LLMs in critical care research. By leveraging routine free-text clinical narratives, automated pipelines can provide a rapid, highly accurate “stress-test” of trial protocols before their initiation. This allows trial designers to dynamically adjust inclusion and exclusion criteria to ensure realistic recruitment forecasting, thereby mitigating the risk of future trial discontinuation. While further validation across multiple centers and prospective clinical workflows is warranted to confirm these findings, the integration of LLM-based assistants represents a major step toward more efficient and viable clinical research in the ICU. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.