From Documents to Dynamic Graphs: Ontology-Driven Clinical Graph Construction with Neo4j and UMLS

Jarrod Marks will present an approach to dynamically constructing medical knowledge graphs from unstructured clinical documents, using an extended pipeline built atop the neo4j-graphrag package. This session highlights how the integration of ontology-driven term normalisation (using UMLS) enables graph continuity across disparate documents, creating a unified and query-able model.

In clinical settings, data often resides in loosely connected narrative forms, progress notes, discharge summaries, imaging, and pathology reports. The challenge lies in building a coherent, semantically-aligned knowledge graph that respects the variability of language while maintaining clinical context. Jarrod’s work enhances the neo4j-graphrag framework by introducing: (1) UMLS-based entity normalization for mapping clinical terms across notes, and (2) pipeline validation mechanisms to ensure quality and consistency of graph construction.

Attendees will learn how to:
• Extend the neo4j-graphrag architecture for healthcare-specific use
• Normalize and disambiguate clinical terms using the UMLS
• Validate and trace the provenance of graph nodes and relationships
• Create graphs that evolve in real time with new clinical data

This talk is ideal for developers, clinical informaticians, and researchers working with NLP, entity linking, or LLMs in clinical domains. It bridges structured ontologies and unstructured narratives using graph-native strategies.

Speaker: Jarrod Marks

Resources:
Get Started with Aura - https://bit.ly/3LOLrjh
Deployment Center - https://bit.ly/4jOelM3
Ground AI Systems and Agents with Neo4j - https://bit.ly/4oVsnyb


#nodes2025 #neo4j #graphdatabase #graphrag #knowledgegraph