Application of generative AI in healthcare

Abstract

Graph databases are gaining popularity due to their ability to efficiently and intuitively represent relationships and structures. This project investigates the construction of an end-to-end pipeline that leverages a Neo4j-based biomedical Knowledge Graph, integrated with data from SNOMED CT Ontology and NHANES, to support natural language question answering in clinical and population health domains. The system comprises of five key components: concept extraction and normalisation using LLM guided synonym resolution; Cypher query generation via prompt-engineered LLM; structured querying over Neo4j; agentic evaluation using LangChain to assess response completeness; and synthesis of user-facing explanations. Meta’s LLaMA-3 8b model was used for the entirety of the project. Findings show that the system is able to effectively handle simple queries, but struggles at more complex ones. Results also underscore the importance of hybrid reasoning architectures to combine symbolic precision and semantic flexibility for effective natural language access to complex medical data. The developed system therefore is a great starting point to enhanced integration with medical graph databases, and provides a basis for the integration of further and multiple database technologies and LLMs.

The full implementation, including setup instructions and sample queries, is available at: https://github.com/jtan573/FYP_CCDS24-0108/tree/main.