Drug-drug similarity comparisons play an important role in clinical research. Ontology based semantic similarity measures enable automated systems to reason about, query, or form clusters from, drug treatment histories in clinical data. Ontology pruning removes unneeded concepts from an application domain, thereby improving the performance of application. Pruning, however, may also affect semantic similarity measures among drugs, many of which are graph-based. We present a pruning strategy for drug ontologies and evaluate a set of semantic similarity measures against an expert derived ontology for three separate clinical domains – congestive heart failure, hypertension and HIV. We show that our pruning approach results in drug-drug similarity measures that are closer to the expert derived measures than from the full ontology hierarchy. We believe that this finding may result in standard, re-usable drug-drug similarities matrices useful for a number of clinical research applications.
