Literature-Based Discovery for Explainable GraphRAG in the Medical Domain

Abstract

To improve retrieval quality while maintaining explainability, we developed a GraphRAG system that incorporates a strategy from a related field, Literature-Based Discovery (LBD), into the retrieval pipeline. Our approach uses the Linking Term Count (LTC) metric to identify higher-order relationships from a knowledge graph derived from the Unified Medical Language System (UMLS), restricted to the Medical Subject Headings (MeSH) vocabulary. The system is evaluated on the MedMCQA dataset using accuracy and an LLM-Judge for assessing reasoning quality. We aimed to answer two core questions: (1) how LBD-based triple extraction impacts performance compared to existing RAG methods and (2) how this approach affects downstream explainability. Results forthcoming.