Obtaining clinical term embeddings from SNOMED CT ontology

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Traditional clinical term embeddings lack ontological knowledge.
• Graph-based methods capture concept embeddings but not term embeddings.
• Existing methods struggle with analogous or opposite terms.

Purpose of the Study:

• To develop a novel method for generating clinical term and word embeddings from the SNOMED CT ontology.
• To improve clinical term similarity and normalization tasks.
• To address limitations of existing corpus-based and ontology-based embedding methods.

Main Methods:

• Utilized a graph-based approach to obtain embeddings for clinical concepts from SNOMED CT.
• Trained a deep learning model to map clinical terms to concept embeddings.
• Introduced a pattern-learning method to differentiate same/different concepts for improved normalization.

Main Results:

• The novel embeddings outperformed corpus-based embeddings in predicting clinical term similarity across five datasets.
• The method achieved competitive accuracy on clinical term normalization.
• Supplementing normalization with learned patterns demonstrated performance improvements.

Conclusions:

• The developed method effectively integrates ontological knowledge into clinical term embeddings.
• The embeddings enhance performance on similarity and normalization tasks.
• Future work includes combining ontology-based and corpus-based embeddings for sentence-based tasks.