Experts' reasoning selects the final diagnosis from many candidates by using hierarchical differential diagnosis. In other words, candidates give a sophisticated hiearchical taxonomy, usually described as a tree. In this paper, the characteristics of experts' rules are closely examined from the viewpoint of hierarchical decision steps and and a new approach to rule mining with extraction of diagnostic tax- onomy from medical datasets is introduced. The key ele- ments of this approach are calculation of the characteriza- tion set of each decision attribute (a given class) and one of the similarities between characterization sets. From the relations between similarities, tree-based taxonomy is ob- tained, which includes enough information for hierarchical diagnosis. The proposed method was evaluated on three medical datasets, the experimental results of which show that induced rules correctly represent experts' decision pro- cesses. Keywords Rough sets, data mining, taxonomy, granular computing.