Deep Application‑Driven Construction of Medical Knowledge Graphs: Methods, Models, and Case Studies

The session, hosted by Zhejiang Digital Medical Health Technology Research Institute, introduces the concept of knowledge graphs, distinguishing the broad definition (a suite of big‑data knowledge engineering technologies) from the narrow definition (large‑scale semantic networks of entities, concepts, and relationships).

It reviews international efforts such as UMLS and SNOMED CT, then details domestic initiatives including CUMLS, the Chinese medical health knowledge service system, traditional Chinese medicine knowledge graphs, and the OpenKG community.

Key domain features of medical knowledge are highlighted: terminology diversity, high precision requirements, and complex semantics, followed by typical application scenarios like clinical decision support, guideline recommendation, and data reporting.

The construction workflow of the institute’s medical knowledge graph is described in five major parts:

1. Model Establishment – a top‑down approach builds a schema (referencing UMLS, Schema.org, cnSchema) covering diseases, drugs, procedures, and examinations; the schema (72 semantic types, 493 relations) is publicly available.

2. "Qiqiaoban" Ontology Term Set – six steps: domain scope definition, source selection, term curation, relationship building, storage in relational tables, and platform support via the self‑developed CoWork system. The term set now contains 970,000 concepts, 1.23 million terms, and 2.92 million relationships.

3. "Huizhi" Knowledge Graph – five steps: source selection (guidelines, literature, encyclopedic data), entity and relation extraction (rule‑based NER plus expert review, semi‑supervised learning), knowledge fusion with the ontology, enriched triple storage (adding attribute groups and provenance), and platform support via CoWork. The graph currently covers seven domains with ~110 k entities and 820 k triples.

Practical applications are demonstrated:

Intelligent Alerts – using the graph to infer clinical risks such as hypokalemia or chest pain scenarios.

Guideline Recommendation – leveraging hierarchical ontology reasoning to suggest relevant guidelines, similar cases, and treatment pathways.

Data Direct Reporting – binding graph concepts to clinical information models for automated disease reporting and rule‑based alerts.

Images illustrating the ontology, graph structures, platform interfaces, and case studies are embedded throughout the original material and are retained here as tags.

Overall, the presentation showcases how a deep‑application‑driven methodology can build high‑quality medical knowledge graphs that support precision healthcare, research, and system integration.