SIR 2024
General IR
Brendan Cline, MD
Assistant Professor of Radiology
Duke University Medical Center
Financial relationships: Full list of relationships is listed on the CME information page.
Brian P. Triana, MD, MBA
Resident Physician
Duke University Medical Center
Disclosure information not submitted.
Alexander Miller, MD
Fellow - Interventional Radiology
Duke University
Disclosure information not submitted.
Nicholas T. Befera, MD
Assistant Professor
Duke University Medical Center
Financial relationships: Full list of relationships is listed on the CME information page.
Within large hospital systems, difficulty with routing procedure requests to the appropriate team and covering provider can delay patient care and cause frustration for both radiologists and ordering clinicians. Furthermore, the heterogeneity of interventional radiology practices further increases complexity for procedure requests between non-vascular interventional teams or procedure teams from other specialties. Artificial intelligence (AI) large language models (LLMs) enable a wide range of capabilities across industries. This work demonstrates a proof-of-concept, LLM-based tool to route procedure requests to the appropriate teams.
Materials and Methods:
At a large academic hospital, existing teams, pager/phone numbers, and schedules were used to create text-based rules for procedure requests. Using the OpenAI application programming interface (API) with Python, an LLM-based assistant was created to route procedure requests at specific days and times to the appropriate teams. Using GPT-3.5 Turbo and GPT-4 models, 270 procedure requests were tested using randomly generated days and times. The estimated cost of each API request was recorded.
Results:
The assistant correctly routed 82.2% of procedure requests using GPT-3.5 Turbo and 96.3% of procedure requests using GPT-4. The routing was performed at an average cost of $0.00068 per request for GPT-3.5 Turbo and $0.013 per request for GPT-4. The most common errors for both models were in early morning requests, times at which multiple subspecialty division procedure services are covered by overnight resident phones. The GPT-3.5 Turbo model demonstrated lower accuracy with routing post-pyloric feeding tube placements, frequently routing them incorrectly to the IR service, a common error among clinicians in our clinical experience.
Conclusion:
This work demonstrates the feasibility of an accurate, low-cost AI-powered assistant to appropriately route procedure requests in a large, academic hospital system. Given the free-text input, the rules and teams can easily be adapted to different coverages or hospital systems. A similar approach may be used to help clinicians navigate a radiology phone tree, or as a tool to help reading room coordinators route requests effectively with decreased training.