A hybrid knowledge-based and data-driven approach to identifying semantically similar concepts.

An open research question when leveraging ontological knowledge is when to treat different concepts separately from each other and when to aggregate them. For instance, concepts for the terms "paroxysmal cough" and "nocturnal cough" might be aggregated in a kidney disease study, but should be left separate in a pneumonia study. Determining whether two concepts are similar enough to be aggregated can help build better datasets for data mining purposes and avoid signal dilution. Quantifying the similarity among concepts is a difficult task, however, in part because such similarity is context-dependent. We propose a comprehensive method, which computes a similarity score for a concept pair by combining data-driven and ontology-driven knowledge. We demonstrate our method on concepts from SNOMED-CT and on a corpus of clinical notes of patients with chronic kidney disease. By combining information from usage patterns in clinical notes and from ontological structure, the method can prune out concepts that are simply related from those which are semantically similar. When evaluated against a list of concept pairs annotated for similarity, our method reaches an AUC (area under the curve) of 92%.