The choice of features used to represent a domain has a profound effect on the quality of the model produced; yet, few researchers have investigated the relationship between the features used to represent text and the quality of the final model. We explored this relationship for medical texts by comparing association rules based on features with three different semantic levels: (1) words (2) manually assigned keywords and (3) automatically selected medical concepts. Our preliminary findings indicate that bi-directional association rules based on concepts or keywords are more plausible and more useful than those based on word features. The concept and keyword representations also required 90% fewer features than the word representation. This drastic dimensionality reduction suggests that this approach is well suited to large textual corpus of medical text, such as parts of the Web.