Detecting Underspecification in SNOMED CT Concept Definitions Through Natural Language Processing

Quality assurance and audit issues play a major role in maintening large biomedical terminology, such as SNOMED CT. Several automatized techniques have been proposed to facilitate the identification of weak spots and suggest adequate improvements.In this study, we address a well-known issue within SNOMED CT: Albeit the wording of many free-text concept descriptions suggests a connection to other concepts, they are often not referred to in the logical concept definition.To detect such inconsistencies, we use a semantic indexing approach which maps free text onto a sequence of semantic identifiers. Applied to SNOMED CT concepts without attributes, our technique spots refinable concepts and suggests appropriate attributes, i.e., connections to other concepts. Based on a manual analysis of random samples, we estimate that approximately 18,000 refinable concepts can be found.