Siamese KG-LSTM: A deep learning model for enriching UMLS Metathesaurus synonymy

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• The Unified Medical Language System (UMLS) Metathesaurus is crucial for understanding biomedical language but faces challenges in construction due to rapid term growth.
• Current UMLS Metathesaurus development relies on error-prone and time-consuming lexical tools and human editors.

Purpose of the Study:

• To leverage deep learning for predicting synonyms and non-synonyms between biomedical terms within the UMLS Metathesaurus.
• To improve the efficiency and accuracy of UMLS Metathesaurus construction, focusing initially on the Disorders semantic group.

Main Methods:

• Utilized a Siamese Knowledge Graph - Long Short-Term Memory (KG-LSTM) deep learning architecture.
• Trained models on a subset of biomedical terms from the Disorders semantic group.
• Enriched model inputs with synonyms and hierarchical relationships from source vocabularies.

Main Results:

• The deep learning approach demonstrated excellent performance in detecting synonyms within the Disorders semantic group of the UMLS Metathesaurus.
• The study validates the potential of machine learning techniques for enhancing UMLS Metathesaurus construction.

Conclusions:

• The proposed Siamese KG-LSTM model offers a promising method for improving the accuracy and efficiency of UMLS Metathesaurus development.
• The approach has the potential for broader application across other semantic groups within the UMLS.