Right Patient, Right Specialist, Right Time: Retrieval Augmented Generation for Specialty Referral Routing.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Physician clinical questions require accurate routing to specialists for timely and relevant responses.
• Manual routing of eConsults is administratively burdensome and can lead to care delays.
• Specialist-curated question templates ensure appropriate context for clinical queries.

Purpose of the Study:

• To develop and validate an embedding-based retrieval system for automatically directing physician clinical questions to the most relevant specialist-curated template.
• To reduce administrative burden on physicians and minimize delays in care delivery.

Main Methods:

• Utilized MPNet, a transformer-based model, to generate dense vector embeddings for clinical questions and 24 predefined clinical templates.
• Employed cosine similarity to compute the relevance between query and template embeddings.
• Validated the system against retrospective eConsult data across five medical specialties.

Main Results:

• The system accurately identified the most relevant template for clinical questions in 87% of cases (success@1).
• The correct template was ranked within the top three results 99% of the time (success@3).
• Demonstrated high performance across multiple specialties.

Conclusions:

• Automating the selection of specialty and referral of clinical questions significantly enhances efficiency.
• The developed system minimizes administrative workload for physicians and reduces patient care delays.
• Improved context through accurate template matching leads to better specialist responses.