LLM / NLP
Medical Disease Knowledge Graph
Extracted diseases, symptoms and relations from medical text; designed a graph schema to support retrieval and assisted diagnosis.
Background and problem
Medical knowledge is scattered across medical records, guidelines and literature, making retrieval costly and relationships hard to see. Building a disease knowledge graph structures relations such as “disease — symptom — examination — treatment”, supporting retrieval and assisted judgement.
Constraints
- Medical text is highly specialised, with inconsistent entity boundaries and wording.
- The data is sensitive and had to be processed in a controlled environment, with no raw text provided externally.
- The graph structure had to accommodate clinical usage habits as well as future extension.
Approach and architecture
The process has three steps: medical text preprocessing and entity recognition, relation extraction and verification, and design of the graph’s node and relation structure. Extraction results were manually spot-checked before being loaded into the graph.
Key decisions
- Ontology first: define entity types and relation types before extraction, avoiding changing things mid-extraction.
- Manual spot checks: the medical domain tolerates little error, so extraction results must be sampled and verified.
- Extensible structure: relation types leave room for extension, making it easy to add examination and treatment information later.
Results and measurement
Completed information extraction and organisation of medical text, contributed to designing the graph’s node and relation structure, and provided a structured foundation for knowledge retrieval and downstream applications.
Project details and data scope are subject to the partner’s public information.
Technical details
- Performed entity recognition and relation extraction on medical text, unifying disease/symptom naming.
- Designed the schema for nodes, relations and attributes, constraining the graph’s query paths.
- Used graph retrieval to connect symptoms, diseases and assisted-diagnosis knowledge.
Reflection
The hard part of a medical knowledge graph is not the algorithm but ontology design and quality verification. If I did it again, I would move manual verification earlier, into the annotation stage, and establish more detailed entity guidelines.
Scope and security
This is a static case study. The page does not run project code or collect uploads, logins or visitor data. Metrics come from project records and local experiments; reproduce details from the source or report. Read the site security model →
