Logical definition-based identification of potential missing concepts in SNOMED CT

Area of Science:

• Biomedical Informatics
• Knowledge Representation
• Ontology Engineering

Background:

• Biomedical ontologies are crucial for cross-disciplinary research, providing standardized terminology.
• Quality issues, such as missing concepts, can impede the effective use of ontologies like SNOMED CT.
• Identifying and addressing missing concepts is essential for enhancing the completeness and utility of biomedical knowledge bases.

Purpose of the Study:

• To develop and evaluate a logical definition-based approach for identifying potential missing concepts in SNOMED CT.
• To generate both logical definitions and fully specified names for these potential missing concepts.
• To validate the identified missing concepts using external biomedical resources.

Main Methods:

• Intersecting logical definitions of unrelated concepts in non-lattice subgraphs to derive definitions for missing concepts.
• Fine-tuning the PEGASUS text summarization model to predict fully specified names from logical definitions.
• Validating identified concepts against the Unified Medical Language System (UMLS), PubMed literature, and newer SNOMED CT versions.

Main Results:

• Identified 30,313 unique logical definitions for potential missing concepts from SNOMED CT (March 2021 US Edition).
• The fine-tuned PEGASUS model achieved ROUGE scores of 72.83 (ROUGE-1), 51.06 (ROUGE-2), and 71.76 (ROUGE-L).
• Identified 23,031 potential missing concepts, with 2,312 (10.04%) automatically validated against external resources.

Conclusions:

• The proposed logical definition-based approach demonstrates promise for identifying missing concepts in SNOMED CT.
• The method successfully generates logical definitions and predicts names for potential missing concepts.
• Further improvements in concept naming based on logical definitions are warranted.