Benchmark and Best Practices for Biomedical Knowledge Graph Embeddings.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Knowledge Representation

Background:

• Biomedical data often exists as discrete text and medical codes.
• Existing knowledge bases and ontologies are underutilized in machine learning due to a lack of effective knowledge representation methods.
• Biomedical concept embedding methods have lagged behind general natural language processing advancements.

Purpose of the Study:

• To investigate the efficacy of knowledge graph embedding models for the biomedical domain.
• To develop and benchmark state-of-the-art knowledge graph embedding techniques using SNOMED-CT.
• To highlight the importance of multi-relational knowledge graph learning for biomedical concept representation.

Main Methods:

• Trained several state-of-the-art knowledge graph embedding models.
• Utilized the Systematized Nomenclature of Medicine - Clinical Terms (SNOMED-CT) knowledge graph.
• Performed benchmarking and comparison against existing methods.

Main Results:

• Demonstrated the potential of knowledge graph embeddings for learning biomedical concept representations.
• Provided a comparative analysis of different embedding models on SNOMED-CT.
• Established best practices for applying these models in the biomedical domain.

Conclusions:

• Knowledge graph embeddings offer a promising approach to unlock the value of biomedical knowledge bases for machine learning.
• Leveraging the multi-relational structure of knowledge graphs is crucial for effective biomedical knowledge representation.
• The developed embeddings, code, and materials will be publicly shared to foster further research.