FHIR in the W.H.O.L.E.: Graphing Clinical Data for Smarter AI Systems
summary
Graph-based pipeline called FHIR in the W.H.O.L.E. converts FHIR JSON into semantically linked nodes and runs on Neo4j community edition to support Clinical AI and trial matching, claiming a reproducible graph RAG that maps hierarchy and cross-code linkage (SNOMED, LOINC, ICD). Demo shows four of twelve agents cooperating (diagnostics, risk, planner, documentation), end-to-end latency ~28,000 ms with intent ~3,000 ms, deployed on three 16GB-GPU Linux machines using fine-tuned models (deepseat, quen, Llama‑based RAG) and intent model Jella, code planned as open source.
AI-written summary · about
chapters (8)
- 0:00Introduction and setup
- 2:13What is FHIR and its limitations
- 5:49Introducing the WHOLE framework
- 9:22Hierarchy and ontology in healthcare
- 12:52Graph construction and agentic workflows
- 16:45Live demo of the system
- 24:17Clinical trial matching with FHIR
- 27:08Q&A and open sourcing plans
related talks
33mBridging large-scale FHIR to AI in production
48mAnalytics and AI on FHIR
39mAI, FHIR & the future of healthcare data
05mHow generative AI can support medical staff
30mFrom Code to Clinic — Building an Internal Developer Platform for Clinical AI
21mUsing FHIR Workflows to deliver personalized health experiences
21mMake sense of health data with Amazon HealthLake
15mGraph-Driven Knowledge Retrieval: Neo4j for Healthcare Building Codes