Intensive care units (ICUs) in the United States admit over five million patients annually, with mortality rates ranging from 10% to 29%. Given that a patient’s condition can change rapidly, early detection of deterioration is crucial for enabling timely interventions and improving outcomes. While existing systems provide some guidance, they are often manual and offer only a single snapshot of a patient’s health. Recent AI models have been developed to provide automated, real-time assessments, yet many models are limited and only provide an incomplete picture of a patient’s clinical status.
Miguel Contreras, a Ph.D. student in the Rashidi Lab, is the first author of a newly published study in Nature Communications that introduces a real-time predictive model designed to support patient care in the ICU. The paper, titled “Real-time prediction of intensive care unit patient acuity and therapy requirements using state-space modelling,” was co-authored by Parisa Rashidi, Ph.D., UF Foundation Professor, Co-Director of the Intelligent Clinical Care Center (IC3), and Director of the Intelligent Health Lab, and Azra Bihorac, M.D., M.S., FASN, FCCM, Director of the IC3, Director of the Precision and Intelligent Systems in Medicine Partnership (PRISMAP), and Senior Associate Dean for Research at the University of Florida College of Medicine.
The study presents a state-space modeling framework called APRICOT-M (Acuity Prediction in Intensive Care Unit-Mamba) that can anticipate patient acuity and therapy needs in real time, giving clinicians a tool to better predict and respond to critical changes in patient health. Trained on more than 140,000 ICU admissions from 55 hospitals, including UF Health Shands Hospital, APRICOT-M integrates vitals, labs, medications and patient data to predict deterioration and therapy needs up to four hours in advance. In its development and validation, APRICOT-M consistently outperformed or matched existing models in key metrics. During a clinical adjudication at UF Health Shands Hospital, physicians found the model’s alerts timely and actionable, underscoring its potential to drive earlier and more effective interventions. By capturing complex physiological signals and translating them into actionable predictions, APRICOT-M could improve clinical decision-making and patient outcomes in high-acuity settings.
Contreras led the development and testing of the model, working closely with collaborators to evaluate how it could be applied in clinical environments. The team’s research showcases key innovations in translating artificial intelligence research into impactful real-world solutions.
“This work reflects the potential of artificial intelligence to enhance clinical care,” Dr. Rashidi said, “We are excited about its promise for helping ICU teams make faster, more informed decisions.”
The UF Department of Biomedical Engineering (BME) and the UF College of Medicine’s (COM) continued partnership in advancing AI for health care underscores the importance of interdisciplinary collaboration in developing solutions that address urgent clinical challenges. The researchers have made the model’s code publicly available on GitHub to promote further research and collaboration in the field of clinical AI.