CODER

CODER produces cross-lingual, knowledge-infused medical concept embeddings by applying contrastive learning to Unified Medical Language System (UMLS) term similarities and knowledge graph relation triplets.

Key Features:

• Cross-lingual embeddings: Generates embeddings that represent medical concepts across languages.
• UMLS knowledge integration: Integrates structured knowledge from the Unified Medical Language System (UMLS) knowledge graph into the embedding space.
• Contrastive learning objective: Uses contrastive learning to align term-level similarity signals with relational information.
• Relation triplet training: Trains on relation triplets from the UMLS knowledge graph to encode inter-concept relations.
• Term- and relation-level fusion: Combines term-based similarities and KG relational triplets when learning embeddings.
• Pretrained-model compatibility and fine-tuning: Supports fine-tuning analogous to pretrained language models for task adaptation.
• Benchmark evaluation: Evaluated on zero-shot term normalization, semantic similarity, and relation classification benchmarks with superior performance versus biomedical word, concept, and contextual embeddings.
• Feature provision for ML: Produces embeddings intended for use as features in downstream machine-learning models.

Scientific Applications:

• Zero-shot term normalization: Enables mapping of diverse surface forms to standardized UMLS concepts without task-specific supervision.
• Semantic similarity assessment: Supports quantitative evaluation of semantic similarity between medical concepts.
• Relation classification: Encodes relational signals useful for classifying relations among medical concepts.
• Embedding-based terminology mapping: Facilitates embedding-driven mapping of terminologies to UMLS concepts.
• Feature extraction for downstream models: Supplies knowledge-aware embeddings as inputs to machine-learning pipelines.

Methodology:

Applies contrastive learning on UMLS knowledge graph data, training with relation triplets and term-based similarity signals to produce knowledge-infused embeddings and supports fine-tuning similar to pretrained language models.