Title: Detecting Misaligned and Missing Concepts in SNOMED CT using Structural and Lexical Patterns

Objective: Quality assurance of large ontological systems such as SNOMED CT is an indispensable part of the terminology management lifecycle. We introduce a hybrid structural-lexical method for scalable and systematic discovery of novel anomalies in SNOMED CT. The structural component is based on shared isa relations to other concepts. The lexical component leverages shared words in descriptions between concepts. Material and Methods: All non-lattice subgraphs (the structural part) in SNOMED CT are exhaustively extracted. Four types of lexical patterns (the lexical part) are identified among the concepts involved in non-lattice subgraphs. Non-lattice subgraphs exhibiting such lexical patterns are often indicative of misaligned and missing concepts. Results: Applying our hybrid structural-lexical method to the September 2015 version of SNOMED CT (U.S. edition), we extracted 171,011 non-lattice subgraphs, among which 6,801 matched the lexical patterns. A subset of 2,046 small non-lattice subgraphs with sizes 4 to 6 with lexical patterns was obtained. A random sample of 100 of these subgraphs was selected, visualized and manually reviewed by two domain experts. Of these, 59 (59%) revealed errors confirmed by the experts. The most frequent type of error was missing isa relations due to incomplete or inconsistent modeling of the concepts. Discussion: The combined non-lattice and lexical-based anomalies have not been uncovered by other existing ontology quality assurance approaches known to date. Non-lattice subgraphs of sizes 4, 5 and 6 can be easily visualized for manual inspection by experts. It also makes sense to investigate them first, because they are often included in larger subgraphs. Conclusions: Our hybrid structural-lexical method is innovative and effective in detecting SNOMEDCT anomalies that have escaped existing quality assurance processes. References1. Zhang GQ, Zhu W, Sun M, Tao S, Bodenreider O, Cui L. MaPLE: A MapReduce Pipeline for Lattice-basedEvaluation and Its Application to SNOMED CT. Proc IEEE Int Conf Big Data. 2014 Oct;2014:754-759. (PMID:25705725)