Improving medical term embeddings using UMLS Metathesaurus.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic Health Records (EHRs) contain valuable free-text clinical notes.
• Machine learning on EHR notes is crucial for medical informatics applications.
• Learning term embeddings from limited, specialized EHR data is challenging due to unique terminology.

Purpose of the Study:

• To develop a novel algorithm for learning medical term embeddings from limited EHR data.
• To leverage external medical term definitions to improve embedding quality.
• To address the challenge of rare or unseen terms in medical notes.

Main Methods:

• Propose definition2vec, an extension of the skip-gram algorithm.
• Incorporate textual definitions from the Unified Medical Language System (UMLS) Metathesaurus.
• Utilize the Medical Information Mart for Intensive Care (MIMIC-III) EHR dataset for evaluation.

Main Results:

• definition2vec successfully groups semantically similar medical terms in the vector space, even rare ones.
• Learned embeddings maintain semantic relationships for terms not frequently observed in the corpus.
• The method demonstrates utility in downstream medical informatics tasks.

Conclusions:

• Medical term definitions significantly enhance the learning of embeddings for rare or unseen terms.
• This approach is effective for small, specialized corpora like clinical notes.
• Improved embeddings facilitate advanced machine learning applications in healthcare.