Delirium is a frequent and serious complication in the intensive care unit, affecting up to one in three patients and contributing to longer hospital stays, higher costs, and worse outcomes. Early identification of patients most at risk remains challenging, particularly in busy ICU environments where clinicians must rapidly interpret large volumes of clinical data.

A multidisciplinary University of Florida team led by first author Miguel Contreras, a Ph.D. student in the Rashidi Lab, has developed a new artificial intelligence model to help address this challenge. Their study, published in Scientific Reports, was co-authored by collaborators from the College of Medicine and the Herbert Wertheim College of Engineering, including senior authors Parisa Rashidi, Ph.D., and Azra Bihorac, M.D.

Read the paper: https://www.nature.com/articles/s41598-025-22634-7

The team introduces DeLLiriuM, a large language model that uses structured electronic health record data from the first 24 hours of ICU admission. These data — including vital signs, laboratory results, medications, assessments, demographics and comorbidities — are converted into a text-based format designed for large language models. Using this information, DeLLiriuM estimates a patient’s risk of developing delirium during the remainder of their ICU stay.

Trained and validated on more than 100,000 ICU admissions across 195 hospitals from three major databases — eICU, MIMIC-IV and UF Health’s Integrated Data Repository — the model consistently outperformed traditional machine learning and deep learning approaches. It achieved strong performance across multiple metrics, including AUROC and AUPRC, during external validation. DeLLiriuM also provides case-level explanations that highlight which clinical features contribute most to predicted risk, supporting more transparent and clinically interpretable decision support.

Contreras and collaborators worked closely to design, implement and evaluate the model, demonstrating how large language models applied to structured ICU data can support earlier identification of delirium risk and inform prevention and management strategies. The study highlights the impact of interdisciplinary collaboration at UF in translating advanced AI methods into tools with potential to improve patient care in high-acuity settings.