Advisor(s)
Sandeep Bodduluri
Committee Member(s)
Carlos Cardenas
Ryan Godwin
Document Type
Thesis
Date of Award
6-18-2026
Degree Name
Master of Science (MS)
School
School of Health Professions
Department
Health Professions
Abstract
In acute trauma care, clinicians require rapid access to precise thresholds, activation criteria, and decision pathways embedded within institutional protocols. However, navigating static, unindexed PDF documents is time-consuming, fragmented, and error-prone in time-critical environments. This work presents the design, architecture, and prototype implementation of a grounded Retrieval-Augmented Generation (RAG) system developed as a protocol-access acceleration layer for University of Alabama at Birmingham (UAB) trauma protocols. The system employs a local full-stack architecture consisting of a Next.js frontend, FastAPI backend, Qdrant vector database, and on-device inference via Ollama (gpt-oss:20b). Institutional protocols are ingested using a two-pass parsing pipeline, deduplicated through content-derived hashing, and segmented via regex-based chunking to preserve structural context. Dense vector semantic search retrieves relevant protocol segments, which are synthesized by the large language model into structured, scan-oriented outputs organized as “Bottom Line,” “Key Actions,” and “Cautions.” To support patient safety and foster clinical trust, the system prioritizes strict grounding over autonomous reasoning. Anti-hallucination guardrails constrain outputs exclusively to retrieved content, and all generated responses are explicitly linked to source material through inline citations. A dedicated source panel enables rapid verification against original PDF excerpts, page numbers, and extracted figures. Although the current system remains a prototype requiring formal quantitative evaluation prior to clinical deployment, it demonstrates strong potential to reduce time-to-guideline, standardize access to institutional protocols, and mitigate nonadherence in the trauma bay.
Keywords
Artificial Intelligence in Medicine;Clinical Decision Support;Grounding Transparency;Natural Language Processing;Retrieval-Augmented Generation;Trauma Protocols
Recommended Citation
Fairburn, Stevan Craig, "Grounded Rag Model For Guideline Implementation In The Clinical Workflow" (2026). ETDs from 2020-2029. 198.
https://digitalcommons.library.uab.edu/etd-2020s/198
