Machine learning models to identify and predict delirium in hospitalized older adults - delirium, an acute disorder of attention and cognition, is a common and devastating syndrome in hospitalized…

The National Institute on Aging awarded The Feinstein Institutes For Medical Research $698,456 on September 1, 2026, under the Aging Research program (CFDA 93.866) to develop machine learning models that identify and predict delirium in hospitalized older adults.

The project addresses two clinical gaps: the high rate of missed delirium diagnoses in emergency departments (over 75% missed despite intensive screening) and the need to identify high-risk patients during hospital admission to enable timely prevention strategies. The recipient will develop and validate machine learning methods using two existing curated datasets with structured and unstructured electronic health record data to identify patients with emergency department delirium and predict imminent risk of delirium development during hospitalization. The project leverages prospective validation and a reference standard delirium assessment applied to a large, heterogeneous population.

Performance occurs in Manhasset, New York, with a period of performance from September 1, 2026, through May 31, 2031. The assistance type is a Project Grant.

Machine learning models to identify and predict delirium in hospitalized older adults - delirium, an acute disorder of attention and cognition, is a common and devastating syndrome in hospitalized older adults that has an underlying cause, is potentially preventable, and is associated with poor outcomes. The importance of identifying delirium starts in the emergency department (ED), where prevalence in older adults is 10-20%. Yet, consistent with previous studies, we found (R21AG075230) that, even with an intensive delirium screening program, ED delirium was missed in over 75% of patients. Missed ED delirium impedes the initiation of mitigation strategies (e.g., patient prioritization out of the ED). Thus, leveraging machine learning (ML) methods to identify patients with ED delirium allows for the timely initiation of mitigation strategies, leading to improved outcomes, while reducing clinician workload (aim 1). Further, after patients are admitted to the hospital, 20-40% will develop delirium. Although multicomponent nonpharmacologic strategies (e.g., addressing delirium risk factors) can prevent delirium in up to 40%, these strategies rarely can be performed effectively for all older adults, due to limited resources and competing priorities. Thus, the ability to leverage ML methods to predict which patients are at imminent risk of developing delirium during hospitalization will provide a critical time window for clinicians to initiate targeted prevention strategies to reduce delirium and improve outcomes, while optimizing resource utilization (aim 2). Despite the existence of delirium ML models, critical gaps remain. To address these weaknesses, we will: 1) use two existing (R21AG075230, R01NR020774) curated datasets that have structured and clinically rich, unstructured electronic health record data (e.g., clinician notes); 2) use a large, heterogenous population; 3) include prospective validation; and 4) use a reference standard delirium assessment (confusion assessment method). Further, we have previously developed ML models (R01NR020774) that employ cutting edge modelling techniques that incorporate continuous monitoring (CM) data via clinical wearable devices, to predict imminent clinical deterioration (i.e., delirium). Lastly, we will use stakeholder input to optimize delirium ML model integration into clinical practice (aim 3). Our interprofessional team with expertise in geriatrics, hospital medicine, delirium research, CM wearable data, ML methods, implementation science, and qualitative methods, is well-poised to complete these 3 aims: aim 1) develop an ML model to identify patients with delirium in the ED (1A), and prospectively validate (n=1800) the model (1B); aim 2) develop an ML model to predict which patients are at imminent risk of developing delirium during hospitalization (2A), and prospectively validate (n=1000) the model (2B); and aim 3) determine facilitators and barriers to integrating delirium ML models into clinical practice. Our research will result in delirium ML models that balance accuracy, robustness, and generalizability, and will lead directly to external validation studies. Our long-term goal is to implement the delirium ML models into clinical practice to improve patient outcomes and resource utilization while decreasing clinician workload.