Building a Healthcare & Life Sciences Knowledge Graph with Synthea & ChatGPT — Franz, Inc | KGC 2023

Healthcare and life-sciences teams can move from “ask-and-forget” chat answers to usable, queryable intelligence by pairing a hospital-ready knowledge graph with large language models—then validating and correcting outputs with web search and structured prompts. The core pitch is a plug-and-play healthcare knowledge graph built for hospitals, plus a set of workflows that let ChatGPT-like systems extract causal chains, generate clinical-style notes from patient context, and populate a graph with evidence-backed links.

The talk frames a practical bottleneck: hospitals rarely build knowledge graphs because the work is too hard to start and too risky to deploy. To address that, the team created a plug-and-play knowledge graph preloaded with realistic healthcare data and a library of example questions and methods, delivered in a way that can be explored quickly via notebooks. Under the hood is an “entity-event” knowledge graph model: a patient is the core entity, and everything that happens over time—demographics, encounters, procedures, observations, claims-related events—becomes typed event objects with start/end times and key-value attributes. Those events are organized and terminated in taxonomies so the graph can support consistent querying and downstream analytics.

The knowledge graph’s value proposition is operational. Queries become far simpler to write and interpret, and retrieving “everything about a person” stops being a scavenger hunt across many databases. The talk also emphasizes security and governance: because personal health information can’t be freely shared, the team builds the graph using Synthea, a synthetic patient dataset, as a safe stand-in that hospitals can demonstrate internally to leadership and colleagues.

Beyond structure, the graph is designed to integrate unstructured medical text—an important point given that most medical data is not neatly formatted. The system links clinical notes and literature by mapping entities into UMLS, then uses entity extraction (including BioWordVec-based extraction) and an internal UMLS mapping approach to connect PubMed articles, clinical-trial text, and Synthea’s structured outputs into one system. That enables cross-domain queries such as linking patients to relevant clinical trials, or connecting medical records to conditions and evidence in the literature.

Large language models enter as both a productivity layer and a reasoning layer, but with heavy emphasis on validation. For causal-chain extraction, the workflow prompts an LLM to produce structured outputs (JSON or Python lists) and then checks whether the referenced links actually exist; the speaker notes that naive generation can fabricate or break half the links, and that web search plus correction can repair many of those failures. Another use case uses the graph plus prompting to generate narrative clinical reports from patient context (last conditions, procedures, medications, observations, and encounters), producing clinician-style notes and recommendations—while acknowledging that some outputs may be artificial and still require human oversight.

Finally, the talk extends the same pattern to a wine knowledge graph built automatically from an ontology and then enriched with LLM-generated descriptions, taste profiles, and (optionally) web-verified information. The closing message is a workflow principle: intelligent applications work best as a collaboration between humans, an LLM, and a knowledge graph—because the graph stores structured knowledge that can be queried repeatedly, while LLM answers alone are fleeting and harder to audit.