A comparison of hierarchical structures generated using different combinations of similarity metrics (cosine, correlation, Yule and Yule2) and clustering algorithms (average-link (AL), complete-link (CL) and ward-link (WL)) was performed in this study. The purpose is to identify the most appropriate hierarchical structure for the visualization and exploration of medical interventions. The greater the number of hierarchy levels, the longer it takes for a user to browse and search for a topic in a hierarchy. 10 poorly-formulated therapy questions were processed and 120 hierarchies were constructed. A lower number of hierarchy levels were discovered from the WL clusterings. The best clusters appear on average at level 4-5 of the WL clusterings and at level 5-10 of the AL and CL clusterings. Up to 98% of the best clusters can be obtained from the top 10 levels of the WL clusterings. By fixing the hierarchy level to 5, approximately 50-60% and 25-27% of relevant document, respectively, were identified from the best clusters of WL and AL clusterings. The Yule2-WL clusterings outperform other types of clusterings, in terms of mean average precision, 11-point interpolated precision, R-precision and precision at k. Similar analyses were applied to 10 well-formulated therapy questions and 30 Yule2-based hierarchies were constructed. No significant difference was found between the performance of AL, CL and WL algorithms. The overall results indicate that the combination of Yule2 similarity metric and WL clustering algorithm provides the most efficient structure for the visualization and exploration of medical interventions generated from both poorly- and well-formulated therapy questions.

average-link, complete-link, hierarchical clustering, medical interventions, ward-link.