Related Experiment Video
Updated: May 12, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Personalized Case- and Evidence-Based TBI Prognosis with Small Language Models
Pranav Manjunath1, Syed M Adil2, Benjamin D Wissel2
1Dept. of Biomedical Engineering, Duke University, Durham, USA.
... IEEE-EMBS International Conference on Biomedical and Health Informatics. IEEE-EMBS International Conference on Biomedical and Health Informatics
|May 11, 2026
Summary
This study introduces a dual retrieval AI framework for traumatic brain injury (TBI) patient disposition, combining small language models with clinical guidelines and similar patient cases for improved accuracy and personalized decision-making.
Area of Science:
- Artificial Intelligence in Medicine
- Clinical Decision Support Systems
- Traumatic Brain Injury Management
Background:
- Emergency department disposition for traumatic brain injury (TBI) relies on complex data synthesis, often using heuristics leading to variable outcomes.
- Large language models (LLMs) offer potential for evidence-based practice but face limitations in size, cost, and privacy for clinical use.
Purpose of the Study:
- To develop a dual retrieval-augmented framework using efficient, on-premise small language models (SLMs) for personalized TBI patient disposition.
- To integrate evidence-based practice (guideline retrieval) with case-based reasoning (similar patient exemplars) for enhanced prediction.
Main Methods:
- Implemented a dual retrieval framework with two open-source SLMs (Phi-4-mini, Qwen-2.5) under 4B parameters.
- Modeled evidence-based practice by retrieving patient-specific guideline passages.
- Utilized case-based reasoning to retrieve similar patients as few-shot exemplars for personalized context.
Main Results:
- Similar patient exemplars consistently improved classification performance (sensitivity without sacrificing specificity) across both SLMs.
- Clinical guidelines had a smaller impact individually but shifted predictions towards guideline-consistent behavior when combined with exemplars.
- Clinician evaluations indicated that exemplars enhance accuracy, while guidelines improve clinical relevance and justification of AI outputs.
Conclusions:
- Targeted retrieval personalizes AI predictions and their rationale, improving performance, interpretability, and trustworthiness in clinical decision-making.
- The dual retrieval framework offers a viable approach for deploying AI in TBI disposition, balancing accuracy with clinical relevance.
