Semantic classification of biomedical concepts using distributional similarity

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Ontological concepts require robust classification for effective use in natural language processing (NLP).
• Existing methods for classifying biomedical concepts can be time-consuming and lack high-throughput capabilities.
• The Unified Medical Language System (UMLS) provides a rich resource of biomedical concepts that can benefit from automated reclassification.

Purpose of the Study:

• To develop an automated, high-throughput, and reproducible method for reclassifying and validating ontological concepts.
• To enhance the utility of biomedical ontologies in natural language processing (NLP) applications.
• To create a scalable solution for managing and updating semantic classifications of medical terms.

Main Methods:

• A distributional similarity approach was employed to classify Unified Medical Language System (UMLS) concepts.
• Classification models were trained using contextual features derived from syntactic properties across two large corpora.
• Alpha-skew divergence was utilized as the similarity measure for concept classification.

Main Results:

• The developed method achieved low error rates, with the lowest estimated rates at 0.198 (top prediction) and 0.116 (top 2 predictions).
• Testing sets were automatically generated based on National Library of Medicine updates to UMLS semantic classifications.
• Misclassification analysis confirmed the method's performance and identified areas for potential improvement.

Conclusions:

• The distributional similarity approach effectively recommends high-level semantic classifications for biomedical concepts.
• The method is suitable for integration into natural language processing (NLP) pipelines requiring accurate concept categorization.
• This automated approach offers a reproducible and efficient solution for validating and reclassifying ontological concepts.