Predicting Agitation Events in the Emergency Department Through Artificial Intelligence

Area of Science:

• Emergency medicine
• Artificial intelligence in healthcare
• Clinical prediction modeling

Background:

• Agitation events are increasing in emergency departments (EDs), posing significant safety risks to both patients and healthcare providers.
• The complex nature of agitation, with diverse clinical causes and behavioral patterns, makes accurate prediction challenging in the emergency setting.

Purpose of the Study:

• To develop, train, and validate an artificial intelligence (AI) model specifically designed to predict agitation events.
• The model leverages a large and diverse dataset of past emergency department (ED) visits to identify predictive factors.

Main Methods:

• A retrospective cohort study utilized electronic health record (EHR) data from 9 ED sites within a large urban health system.
• Data from over 3 million ED visits (patients aged 18+) between 2015 and 2022 were analyzed.
• The primary outcome, agitation, was defined by orders for intramuscular chemical sedation or violent physical restraint. Model performance was assessed using AUROC and PR-AUC metrics.

Main Results:

• The final AI model incorporated 50 predictors and demonstrated strong predictive performance with an AUROC of 0.94 and PR-AUC of 0.41 in cross-validation.
• Key predictors included prior ED visit frequency, initial vital signs, medical history, chief complaint, and previous sedation/restraint events.
• Model calibration was robust across predicted probability ranges, indicating reliability in predicting agitation risk.

Conclusions:

• The developed prediction model accurately identifies patients at risk for agitation in the ED, showing high accuracy and broad applicability across diverse populations.
• Clinical implementation of this model can facilitate proactive de-escalation strategies and potentially prevent agitation events.
• This approach supports enhanced patient-centered care by enabling timely and targeted interventions.