Predictive K-means is a predictive clustering algorithm which allows to describe and pre- dict simultaneously. Unlike supervised classification and traditional clustering, the perfor- mance of this type of algorithm is closely related to its ability to achieve a good tradeoff between both the prediction and the description. Yet, to our knowledge, an analytical criterion to measure this compromise does not exist. In this paper, we propose SDB a modified version of Davies-Bouldin index to evaluate the performance quality of the predictive K-means. This modification is based on the integration of a new dissimilarity measure to build a relation- ship between the closeness of observations in terms of distance and their class membership. The experimental results has shown that our proposed criterion allows to measure the descrip- tion/prediction compromise from the results obtained by the predictive K-means approach.