Structural analysis and auditing of SNOMED hierarchies using abstraction networks

SNOMED is one of the leading healthcare terminologies being used worldwide. Due to its sheer volume and continuing expansion, it is inevitable that errors will make their way into SNOMED. Thus, quality assurance is an important part of its maintenance cycle. A structural approach is presented in this dissertation, aiming at developing automated techniques that can aid auditors in the discovery of terminology errors more effectively and efficiently. Large SNOMED hierarchies are partitioned, based primarily on their relationships patterns, into concept groups of more manageable sizes. Three related abstraction networks with respect to a SNOMED hierarchy, namely the area taxonomy, partial-area taxonomy, and disjoint partial-area taxonomy, are derived programmatically from the partitions. Altogether they afford high-level abstraction views of the underlying hierarchy, each with different granularity. The area taxonomy gives a global structural view of a SNOMED hierarchy, while the partial-area taxonomy focuses more on the semantic uniformity and hierarchical proximity of concepts. The disjoint partial-area taxonomy is devised as an enhancement of the partial-area taxonomy and is based on the partition of the entire collection of so-called overlapping concepts into singly-rooted groups. The taxonomies are exploited as the basis for a number of systematic auditing regimens, with a theme that complex concepts are more error-prone and require special attention in auditing activities. In general, group-based auditing is promoted to achieve a more efficient review within semantically uniform groups. Certain concept groups in the different taxonomies are deemed “complex” according to various criteria and thus deserve focused auditing. Examples of these include strict inheritance regions in the partial-area taxonomy and overlapping partial-areas in the disjoint partial-area taxonomy. Multiple hypotheses are formulated to characterize the error distributions and ratios with respect to different concept groups presented by the taxonomies, and thus further establish their efficacy as vehicles for auditing. The methodologies are demonstrated using SNOMED’s Specimen hierarchy as the test bed. Auditing results are reported and analyzed to assess the hypotheses. With the use of the double bootstrap and Fisher’s exact test (two-tailed), the aforementioned hypotheses are confirmed. Auditing on various complex concept groups based on the taxonomies is shown to yield a statistically significant higher proportion of errors