CODER: Knowledge-infused cross-lingual medical term embedding for term normalization.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Biomedical Data Science

Background:

• Medical term normalization is crucial for accurate data analysis.
• Existing embedding methods often lack domain-specific knowledge.
• Integrating medical knowledge can improve representation learning.

Purpose of the Study:

• To propose knowledge-aware embedding for medical term normalization.
• To develop a novel cross-lingual knowledge-infused medical term embedding model named CODER.
• To enhance the performance of medical term representations.

Main Methods:

• Developed CODER using contrastive learning on the Unified Medical Language System knowledge graph.
• Utilized both terms and relation triplets from the knowledge graph for similarity calculations.
• Incorporated medical knowledge into embeddings through relation training.

Main Results:

• CODER demonstrated superior performance in zero-shot term normalization.
• Evaluated effectiveness on semantic similarity and relation classification benchmarks.
• Outperformed state-of-the-art biomedical word, concept, and contextual embeddings.

Conclusions:

• CODER embeddings effectively capture medical concept semantic similarity and relatedness.
• CODER can be used for embedding-based medical term normalization and as machine learning features.
• The model can be fine-tuned for various downstream tasks, offering flexibility.