Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Language Models and RAG for Efficient Knowledge Discovery in Clinical Environments

Seokhwan Ko, Donghyeon Lee, Jaewoo Chun, Hyungsoo Han, Junghwan Cho

TL;DR

The paper addresses the challenge of privacy-preserving knowledge discovery in clinical environments by developing a locally deployable RAG system that relies on PubMedBERT embeddings for semantic retrieval and a lightweight LLaMA3.2 model for generation, all operating within a hospital network. It constructs an institutional knowledge base from PubMed records, uses a vector store and cosine similarity to retrieve relevant literature, and grounds the generative output with retrieved documents. The approach demonstrates feasibility and practical utility for identifying potential collaborators and summarizing topics, evidenced by high-quality retrieval and coherent synthesis in pilot examples. This work advances biomedical knowledge discovery under strict security constraints by showing that domain-specific encoders combined with lightweight, local LLMs can facilitate efficient interdisciplinary collaboration within healthcare institutions.

Abstract

Large language models (LLMs) are increasingly recognized as valuable tools across the medical environment, supporting clinical, research, and administrative workflows. However, strict privacy and network security regulations in hospital settings require that sensitive data be processed within fully local infrastructures. Within this context, we developed and evaluated a retrieval-augmented generation (RAG) system designed to recommend research collaborators based on PubMed publications authored by members of a medical institution. The system utilizes PubMedBERT for domain-specific embedding generation and a locally deployed LLaMA3 model for generative synthesis. This study demonstrates the feasibility and utility of integrating domain-specialized encoders with lightweight LLMs to support biomedical knowledge discovery under local deployment constraints.

Leveraging Language Models and RAG for Efficient Knowledge Discovery in Clinical Environments

TL;DR

The paper addresses the challenge of privacy-preserving knowledge discovery in clinical environments by developing a locally deployable RAG system that relies on PubMedBERT embeddings for semantic retrieval and a lightweight LLaMA3.2 model for generation, all operating within a hospital network. It constructs an institutional knowledge base from PubMed records, uses a vector store and cosine similarity to retrieve relevant literature, and grounds the generative output with retrieved documents. The approach demonstrates feasibility and practical utility for identifying potential collaborators and summarizing topics, evidenced by high-quality retrieval and coherent synthesis in pilot examples. This work advances biomedical knowledge discovery under strict security constraints by showing that domain-specific encoders combined with lightweight, local LLMs can facilitate efficient interdisciplinary collaboration within healthcare institutions.

Abstract

Large language models (LLMs) are increasingly recognized as valuable tools across the medical environment, supporting clinical, research, and administrative workflows. However, strict privacy and network security regulations in hospital settings require that sensitive data be processed within fully local infrastructures. Within this context, we developed and evaluated a retrieval-augmented generation (RAG) system designed to recommend research collaborators based on PubMed publications authored by members of a medical institution. The system utilizes PubMedBERT for domain-specific embedding generation and a locally deployed LLaMA3 model for generative synthesis. This study demonstrates the feasibility and utility of integrating domain-specialized encoders with lightweight LLMs to support biomedical knowledge discovery under local deployment constraints.
Paper Structure (10 sections, 6 equations, 3 figures)

This paper contains 10 sections, 6 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Embedding Representation
  4. Semantic Retrieval Using Cosine Similarity
  5. Prompt Construction for RAG
  6. Generative Synthesis With LLaMA3.2
  7. Results
  8. Conclusion
  9. Discussion
  10. Acknowlegment

Figures (3)

  • Figure 1: Overall workflow of the collaboration recommendation system, based on a Retrieval-Augmented Generation (RAG) architecture.
  • Figure 2: Example output for the query "deep learning prediction for medical images." Recommended researchers and topics are synthesized from retrieved PubMed publications.
  • Figure 3: Cosine similarity scores for the top retrieved documents given the query "Deep Learning Prediction of TERT Promoter Mutation Status in Thyroid Cancer Using Histologic Images." The exact matching publication is ranked first with the highest similarity score, confirming correct retrieval behavior.