In many application data are imperfect, imprecise or more generally uncertain. Many classification methods have been presented that can handle data in some parts of the learning or the inference process, yet seldom in the whole process. Also, most of the proposed approach still evaluate their results on precisely known data. However, there are no reason to assume the existence of such data in applications, hence the need for assessment method working for uncertain data. We propose such an approach here, and apply it to the pruning of E 2 M decision trees. This results in an approach that can handle data uncertainty wherever it is, be it in input or output variables, in training or in test samples.