Generative AI, specifically GPT-4, can significantly improve the prediction of emergency room admissions, even with minimal training on limited data, according to a recent study by the Icahn School of Medicine at Mount Sinai. The study, published in the Journal of the American Medical Informatics Association, highlights how AI can support clinical decision-making in high-volume settings.

The retrospective study analyzed data from over 864,000 ER visits across seven hospitals in the Mount Sinai Health System. Researchers utilized both structured data, such as vital signs, and unstructured data, like nurse triage notes. Out of these visits, 159,857 (18.5%) resulted in hospital admissions.

Researchers compared the performance of GPT-4 with traditional machine-learning models, including Bio-Clinical-BERT for text and XGBoost for structured data. They found that GPT-4, even with minimal training, adapted well to the ER environment and provided explanations for its decisions, a capability that sets it apart from conventional models.

“We aimed to test whether generative AI, like GPT-4, could enhance our ability to predict hospital admissions in busy ER settings,” said Dr. Eyal Klang, co-senior author and director of the Generative AI Research Program at Icahn Mount Sinai. “GPT-4's ability to explain its rationale offers a new dimension to AI in medical decision-making.”

This is hardly the first time GPT-4 has made waves in the medical community. Last year it passed a radiology board-style exam, highlighting the potential of large language models but also revealing limitations that hinder reliability, according to studies published in Radiology.

Traditional machine-learning models typically require millions of records for effective training. In contrast, large language models (LLMs) like GPT-4 can learn from fewer examples and incorporate predictions from traditional models, improving overall performance.

“Our findings suggest that AI could soon assist ER doctors in making rapid, informed admission decisions,” stated Dr. Girish N. Nadkarni, co-senior author and director of the Charles Bronfman Institute of Personalized Medicine at Icahn Mount Sinai. “While promising, this technology currently serves a supportive role, providing additional insights to enhance human decision-making.”

The study emphasizes the potential of integrating LLMs into healthcare systems to address complex clinical challenges in real time. “This research demonstrates how LLMs can be quickly trained to provide valuable insights in healthcare,” noted Dr. Brendan Carr, CEO of Mount Sinai Health System and study co-author. “Our work sets the stage for further AI integration in various healthcare domains, from diagnostics to administrative tasks.”

Funding for the study was provided by the National Heart, Lung, and Blood Institute under NIH grant 5R01HL141841-05.

Business Affairs Homepage