Introduction: Retrieval-Augmented Generation (RAG), an emerging AI architecture for large language models (LLMs), has been shown to improve model performance and reliability in medical applications by grounding responses in external, verified sources such as clinical practice guidelines. Our Level I trauma intensive care unit, which trains over 50 rotating residents from multiple specialties annually, exemplifies a clinical environment where rapid access to standardized information is critical. While institutional protocols and onboarding materials are accessible, they are housed as extensive documents within departmental intranet sites, rendering them cumbersome and impractical for point-of-care decision-making. Our goal was to evaluate the feasibility of deploying a centralized RAG platform to provide residents with immediate, verifiable answers to clinical queries sourced directly from our institutional protocols. Methods: Following a Plan-Do-Study-Act methodology, 26 institutional protocols were integrated into a free-to-use RAG platform to create a proof-of-concept LLM tool grounded exclusively in our uploaded source protocols. This prototype was then tested for retrieval speed and accuracy on 10 clinical queries representing a wide variety of clinical scenarios. Of note, AI was used to aid with grammar, spelling, and clarity in the preparation of this abstract. Results: The prototype demonstrated scalability by successfully indexing all 26 institutional protocols with significant capacity for additional source material. Validation queries (n=10) were answered in a mean of 6.1 seconds with 100% citation accuracy. Each answer was fully verifiable, providing a direct hyperlink to the specific source text within the original protocol. Conclusions: This study demonstrates the feasibility of developing a RAG-based AI platform to improve protocol accessibility for residents in a critical care setting. By providing rapid, verifiable answers, this tool addresses knowledge retrieval challenges in time-critical situations. Future iterations will use query analytics to identify knowledge gaps and inform targeted educational interventions relevant to the most queried topics. The platform is scalable as a reproducible framework for AI-augmented clinical decision support in other clinical environments.
Castanon et al. (Sun,) studied this question.