Structural indicators for effective quality assurance of snomed ct

The Standardized Nomenclature of Medicine -- Clinical Terms (SNOMED CT -- further abbreviated as SCT) has been endorsed as a premier clinical terminology by many national and international organizations. The US Government has chosen SCT to play a significant role in its initiative to promote Electronic Health Record (EH R) country-wide. However, there is evidence suggesting that, at the moment, SCT is not optimally modeled for its intended use by healthcare practitioners. There is a need to perform quality assurance (QA) of SCT to help expedite its use as a reference terminology for clinical purposes as planned for EH R use. The central theme of this dissertation is to define a group-based auditing methodology to effectively identify concepts of SCT that require QA. As such, similarity sets are introduced which are groups of concepts that are lexically identical except for one word. Concepts in a similarity set are expected to be modeled in a consistent way. If not, the set is considered to be inconsistent and submitted for review by an auditor. Initial studies found 38% of such sets to be inconsistent. The effectiveness of these sets is further improved through the use of three structural indicators. Using such indicators as the number of parents, relationships and role groups, up to 70% of the similarity sets and 32.6% of the concepts are found to exhibit inconsistencies. Furthermore, positional similarity sets, which are similarity sets with the same position of the differing word in the concept’s terms, are introduced to improve the likelihood of finding errors at the concept level. This strictness in the position of the differing word increases the lexical similarity between the concepts of a set thereby increasing the contrast between lexical similarities and modeling differences. This increase in contrast increases the likelihood of finding inconsistencies. The effectiveness of positional similarity sets in finding inconsistencies is further improved by using the same three structural indicators as discussed above in the generation of these sets. An analysis of 50 sample sets with differences in the number of relationships reveal 41.6% of the concepts to be inconsistent. Moreover, a study is performed to fully automate the process of suggesting attributes to enhance the modeling of SCT concepts using positional similarity sets. A technique is also used to automatically suggest the corresponding target values. An analysis of 50 sample concepts show that, of the 103 suggested attributes, 67 are manually confirmed to be correct. Finally, a study is conducted to examine the readiness of SCT problem list (PL) to support meaningful use of EHR. The results show that the concepts in PL suffer from the same issues as general SCT concepts, although to a slightly lesser extent, and do require further QA efforts. To support such efforts, structural indicators in the form of the number of parents and the number of words are shown to be effective in ferreting out potentially problematic concepts in which QA efforts should be focused. A structural indicator to find concepts with synonymy problems is also presented by finding pairs of SCT concepts that map to the same UMLS concept