Clustering WHO-ART terms using semantic distance and machine learning algorithms

WHO-ART was developed by the WHO collaborating centre for international drug monitoring in order to code adverse drug reactions. We assume that computation of semantic distance between WHO-ART terms may be an efficient way to group related medical conditions in the WHO database in order to improve signal detection. Our objective was to develop a method for clustering WHO-ART terms according to some proximity of their meanings. Our material comprises 758 WHO-ART terms. A formal definition was acquired for each term as a list of elementary concepts belonging to SNOMED international axes and characterized by modifier terms in some cases. Clustering was implemented as a terminology service on a J2EE server. Two different unsupervised machine learning algorithms (KMeans, Pvclust) clustered WHO-ART terms according to a semantic distance operator previously described. Pvclust grouped 51% of WHO-ART terms. K-Means grouped 100% of WHO-ART terms but 25% clusters were heterogeneous with k = 180 clusters and 6% clusters were heterogeneous with k = 32 clusters. Clustering algorithms associated to semantic distance could suggest potential groupings of WHO-ART terms that need validation according to the user's requirements.