Yale School of Medicine researchers have developed an artificial intelligence prediction model that can identify emergency department patients at risk of becoming agitated before symptoms develop, representing a significant advance in proactive emergency care safety.
The study, published in JAMA Network Open and led by Associate Professor Ambrose Wong, analyzed over 3 million emergency department visits across nine hospitals in the northeastern United States between 2015 and 2022. The machine learning model successfully predicts patient agitation risk using 50 key features routinely collected during ED visits.
Addressing a Critical Safety Challenge
Emergency departments have experienced a surge in visits for mental health conditions over the past decade, frequently involving episodes of patient agitation or aggressive behavior. Once agitation occurs, clinicians often resort to physical restraints or intramuscular medications to manage patients safely. However, these interventions carry significant risks including blunt chest trauma, respiratory depression, and even sudden death.
“In general, patients that experience psychiatric emergencies or behavioral emergencies first land in our doors,” said Wong, who is also director of simulation research at the Yale Center for Healthcare Simulation.
Machine Learning Approach
The research team evaluated nearly 700 potential risk factors before refining their model to focus on the most predictive elements. These include patient age, insurance status, past medical history, current medications, reason for the ED visit, and whether the patient has a primary care provider.
Notably, the researchers intentionally excluded race and ethnicity from the prediction model despite including this data in their fairness analysis, specifically to prevent algorithmic bias in clinical decision-making.
The strongest warning signs for future agitation included frequent past ED visits, abnormal vital signs, relevant medical history, and any previous use of restraints or sedation.
Prevention Over Reaction
“What we don’t want is for the model to be self-fulfilling but instead predict agitation early in a visit before symptoms even develop,” Wong explained. “Our ultimate goal is to better allocate critical and limited resources to those that are in most need or that have the greatest benefit.”
Senior author Andrew Taylor, associate professor adjunct of biomedical informatics and data science at Yale School of Medicine, emphasized the dual benefits for patients and healthcare providers.
“On the patient side of things, physical restraint or chemical restraint can be a pretty dehumanizing experience,” Taylor noted. “On the staffing side, when people become violent or aggressive, there can be both emotional trauma and physical trauma to staff, too.”
Implementation and Next Steps
The prediction model integrates with existing electronic health record systems, making deployment feasible without requiring additional data collection. The research represents a proof-of-concept that demonstrates the viability of using machine learning for mental health prediction in emergency settings.
“We know the model works, but we really have to get it implemented,” Taylor said. “We have to build up that clinical support around it and look at the pragmatics of when to deploy it and how to deploy it so that it best fits with our staff’s workflow.”
The research team is now focusing on scaling the model, securing clinical buy-in, and developing implementation strategies for hospital systems. This work represents part of a growing trend toward prediction models in modern medicine, which are increasingly being deployed across cardiology, diabetes care, and chronic kidney disease management.
The study received support from grants from the National Institute of Mental Health, the National Institute of Nursing Research, the National Institute on Minority Health and Health Disparities, and the Patient-Centered Outcomes Research Institute.
Clinical Significance
This research addresses a significant gap in mental health prediction modeling within emergency medicine. By shifting focus from reactive restraint use to proactive risk identification, the AI tool has potential to improve both patient experience and clinical outcomes while reducing the physical and emotional trauma associated with aggressive episodes in emergency settings.
The model’s development using diverse, multi-institutional data enhances its generalizability across different hospital systems and patient populations, suggesting broad applicability for emergency departments seeking to enhance patient and provider safety.