ROC curve for prostate cancer classification (prostate cancer vs. BPH). ROC curves for prostate cancer classifcations (prostate cancer vs. BPH cases) were constructed for the proposed GDFS method (blue) and rpart (orange). The results for CFS are not included here, as the method selected no features in all but one of the cross-validation runs. In each case, the posterior probability of belonging to the prostate cancer group was calculated for each observation j using features selected, and model parameters estimated, with observation j omitted (leave-one-out cross-validation). ROC curves were constructed from these posterior probabilities using the ROCR package in R . ‘X’ marks the 0.5 classification threshold on each ROC curve.