Part 4 — Computing Graph Embeddings

Once the SNOMED CT concept graph is ready, we can train embeddings that capture semantic proximity between medical concepts. The most common techniques include Node2Vec and DeepWalk.

Recommended Steps

Apply Node2Vec to the concept graph for random walk–based feature extraction.
Use weighted edges to emphasize strong semantic relationships.
Store vectors as a dense matrix with conceptId as the key.
Export results to CSV or directly to Oracle 23ai Vector Store for semantic search.

These embeddings serve as the foundation for intelligent retrieval, clustering, and synonym discovery in clinical applications.

Part 5 — Term and Synonym Embeddings

Beyond concept-level vectors, SNOMED CT descriptions provide linguistic richness. Each conceptId has multiple description.term entries — synonyms, variants, or preferred labels. Embedding these terms captures semantic nuances for NLP applications.

Steps

Generate sentence embeddings for description.term entries using models like sentence-transformers/all-MiniLM-L6-v2.
Aggregate embeddings for terms sharing the same conceptId (e.g. mean or weighted average).
Store both concept-level and term-level vectors for hybrid search scenarios.
Optionally normalize and upload vectors to Oracle 23ai for semantic search.

Combining graph and linguistic embeddings provides a complete semantic layer — connecting conceptual relationships with natural language expressions used by clinicians.

Support BlindPhysio

Help us develop BlindPhysio — an AI-driven diagnostic tool for blind and visually impaired massage therapists and physiotherapists. Your contribution directly supports accessibility-focused medical innovation.

Donate