Diagnosis system for predicting bladder cancer recurrence using association rules and decision trees

Abstract

In this work we present two methods based on Association Rules (ARs) for the prediction of bladder cancer recurrence. Our objective is to provide a system which is on one hand comprehensible and on the other hand with a high sensitivity. Since data are not equitably distributed among the classes and since errors costs are asymmetric, we propose to handle separately the cases of recurrence and those of no-recurrence. ARs are generated from each training set using an associative classification approach. The rules’ uncertainty is represented by a confidence degree. Several symptoms of low intensity can be complementary and mutually reinforcing. This phenomenon is taken into account thanks to aggregate functions which strengthen the confidence degrees of the fired rules. The first proposed classification method uses these ARs to predict the bladder cancer recurrence. The second one combines ARs and decision tree: the original base of ARs is enriched by the rules generated from a decision tree. Experimental results are very satisfactory, at least with the AR’s method. The sensibility rates are improved in comparison with some other approaches. In addition, interesting extracted knowledge was provided to oncologists.