A vector-based semantic relatedness measure using multiple relations within SNOMED CT and UMLS.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Ontology Engineering

Background:

• Accurate semantic relatedness measures are crucial for biomedical data analysis.
• Existing methods often rely on external corpora or limited semantic relations.
• There is a need for relatedness metrics that leverage intrinsic ontology structure.

Purpose of the Study:

• To propose a novel vector-based semantic relatedness metric.
• To derive word vectors from the intrinsic structure of biomedical ontologies.
• To evaluate the metric's performance against human judgments and existing methods.

Main Methods:

• Utilized SNOMED CT from UMLS as the testbed ontology.
• Generated concept vectors based on attribute-value, descendant, and is_a relations.
• Computed semantic relatedness using cosine similarity of averaged concept vectors.

Main Results:

• Achieved Spearman's rank coefficients of 0.655 (physicians), 0.744 (coders), and 0.742 (experts).
• The proposed method demonstrated comparable performance to word-embedding techniques.
• Outperformed path-based, information content-based, and other multiple relation-based metrics.

Conclusions:

• Integrating attribute relations with is_a hierarchies in SNOMED CT improves relatedness accuracy.
• The approach is robust to ontological design inconsistencies.
• This method effectively utilizes intrinsic ontology structure, reducing reliance on external textual resources.