Artificial intelligence (AI) can help emergency department (ED) teams better anticipate which patients will need hospital admission, hours earlier than is currently possible, according to a multi-hospital study by the Mount Sinai Health System. This shows the strength of predictive models in this field.
By giving clinicians advance notice, this approach may enhance patient care and the patient experience, reduce overcrowding and “boarding” (when a patient is admitted but remains in the ED because no bed is available), and enable hospitals to direct resources where they’re needed most.
This is the largest prospective evaluation of AI in the emergency setting to date, the study appears in Mayo Clinic Proceedings: Digital Health, titled “Comparing Machine Learning and Nurse Predictions for Hospital Admissions in a Multisite Emergency Care System.”
For the study, researchers collaborated with more than 500 ED nurses across the seven-hospital Health System. Together, they evaluated a machine learning model trained on data from more than 1 million past patient visits. Over two months, they compared AI-generated predictions with nurses’ triage assessments to see whether AI could help identify likely hospital admissions sooner after the patient arrives.
The study, involving nearly 50,000 patient visits across Mount Sinai’s urban and suburban hospitals, showed that the AI model performed reliably across these diverse hospital settings. Surprisingly, the researchers found that combining human and machine predictions did not significantly boost accuracy, indicating that the AI system alone was a strong predictor.
By training the algorithm on more than a million patient visits, the researchers aimed to capture meaningful patterns that could help anticipate admissions earlier than traditional methods.
The output was that machine learning-based predictions outperformed triage nurse estimates for hospital admissions, displaying a high sensitivity and specificity in admission prediction. These findings suggest that an admission prediction system anchored by machine learning can perform reliably using data available at triage.
While the study was limited to one health system over a two-month period, the researchers hope the findings will serve as a springboard for future live clinical testing. The next phase involves implementing the AI model into real-time workflows and measuring outcomes such as reduced boarding times, improved patient flow, and operational efficiency.
This may trigger a new array of systems that allow AI to make complex predictions. The research is also of immediate, practical benefit. By predicting admissions earlier, the outcome can give care teams the time they need to plan, coordinate, and ultimately provide better, more compassionate care. It’s inspiring to see AI emerge not as a futuristic idea, but as a practical, real-world solution shaped by the people delivering care every day.
