Deep Application‑Driven Construction of Medical Knowledge Graphs

The session, hosted by Dr. Xue Lan from the Zhejiang Digital Medical Health Technology Research Institute, introduces the achievements of the institute in knowledge‑graph construction and shares experiences and insights.

1. Knowledge‑Graph Concepts – A broad definition views a knowledge graph as a suite of big‑data knowledge‑engineering technologies, while a narrow definition treats it as a large‑scale semantic network of entities, concepts, and relationships.

2. International Developments – Examples include UMLS (a comprehensive medical language system with over 3 million concepts) and SNOMED CT (covering 350 k concepts and 1.1 M relationships).

3. Domestic Initiatives – Projects such as CUMLS, the Medical and Health Knowledge Service System, Chinese Medicine Knowledge Graphs, and OpenKG illustrate the rapid growth of Chinese medical KGs.

4. Domain Features and Application Needs – Medical terminology diversity, high precision requirements, and complex semantics demand customized solutions and robust data integration.

5. Application Scenarios – Include intelligent early‑warning systems, guideline‑based recommendations, and direct disease reporting, each illustrated with case‑study images.

6. Institute’s KG Construction

6.1 Model Establishment – The institute follows a top‑down approach, designing a schema inspired by UMLS, Schema.org, and cnSchema, covering diseases, drugs, procedures, and examinations. The schema (released August 2019) contains 72 semantic types and 493 relations.

6.2 "Seven‑Piece" Ontology Terminology Set – Built in six steps: domain definition, source selection, term extraction, relationship establishment, storage & browsing (using relational databases), and platform support via the self‑developed CoWork system. The ontology now holds 970 k concepts, 1.23 M terms, and 2.92 M relations.

6.3 "Huizhi" Knowledge Graph – Constructed through five steps: source selection (clinical guidelines, literature, etc.), entity and relation extraction (rule‑based NER plus expert review, semi‑supervised learning), knowledge fusion with the ontology, enriched storage (adding attribute groups and provenance), and platform support via CoWork. The graph currently covers 7 domains, ~110 k entities, and 820 k triples, with quarterly releases.

7. Real‑World Applications

7.1 Intelligent Alerts – Knowledge graphs combined with rules enable comprehensive clinical reasoning, exemplified by potassium‑deficiency and chest‑pain alerts.

7.2 Guideline Recommendations – Hierarchical reasoning over the ontology enriches recommendation results, demonstrated with Coarctation of the Aorta cases.

7.3 Direct Data Reporting – Mapping KG concepts to clinical information models facilitates data binding and automated disease reporting.

Additional uses include intelligent coding and research analytics. The institute’s collaborative KG project has run for five years, involving hundreds of volunteers and numerous enterprises.