A Vector Similarity-based Knowledge Graph Embedding Method for Predicting Missing Hierarchical Relationships in SNOMED CT

SNOMED CT, as the largest clinical terminology set globally, serves as a vital knowledge source for numerous downstream applications. The accuracy and comprehensiveness of its domain knowledge significantly impact related applications. Among the crucial relationships within SNOMED CT, hierarchical relationships play a fundamental role in shaping the framework and structure of the terminology set. However, the inherent incompleteness of these relationships has been a limiting factor in the development of applications in the field. Therefore, automating the discovery of missing hierarchical relationships in this terminology set holds great significance and has garnered extensive attention from researchers. Most existing methods for this task primarily focus on the lexical features of concepts while overlooking the critical role of knowledge graph embedding methods. In this paper, we propose a knowledge graph embedding approach that can predict missing hierarchical relationships in the terminology set without relying on lexical features. To the best of our knowledge, our method is the first to introduce knowledge graph embedding to this task and demonstrate its effectiveness. Our model can learn low-dimensional vector representations that capture the specificity of nodes and relationships, as well as the joint embedding representations of nodes and relationships in the graph structure. We evaluate the proposed approach using a vector similarity prediction function on the US edition of SNOMED CT and achieve an F1 score of 97.80%. The experimental results demonstrate that our model can accurately predict missing hierarchical relationships between concepts, offering valuable guidance for completing the missing hierarchical relationships.