Improving Emergency Department Efficiency With Large Language Model-Guided Orthopaedic Triage for Proximal Humerus...

Area of Science:

• Orthopedic Surgery
• Artificial Intelligence
• Health Informatics

Background:

• Emergency departments (EDs) often receive consults for proximal humerus fractures that do not meet strict institutional criteria.
• This leads to potential delays in care for more severe cases and inefficient resource allocation.

Purpose of the Study:

• To assess the efficacy of large language models (LLMs) in identifying proximal humerus fractures that do not warrant orthopedic consultation.
• To determine if LLMs can reduce unnecessary ED consults for these specific fractures.

Main Methods:

• A retrospective review was conducted at a Level 1 trauma center.
• Generative Pre-trained Transformer-4o (GPT-4o) and o4-mini were used to analyze patient data (history, physical exam, X-ray reports) against institutional consult criteria.
• LLM performance was compared to a gold standard established by expert review and to real-world ED provider performance.

Main Results:

• LLMs demonstrated high accuracy: GPT-4o at 92.4% and o4-mini at 94.9% alignment with consult criteria.
• In contrast, ED providers showed only 32.7% alignment.
• Implementation of LLMs could have saved an estimated 179-183 consults, 295-302 wait hours, and significant work relative value units (wRVUs) over two years.

Conclusions:

• LLMs accurately identify uncomplicated proximal humerus fractures, effectively filtering cases that do not require specialist orthopedic consultation.
• The use of LLMs in the ED has the potential to significantly conserve resources and streamline patient management for proximal humerus fractures.