The screening procedure consists in building a specification region CX in the d-dimensional space such that a future individual with a characteristic vector in CX has higher probability of being a success (that is a response variable Y of interest belongs to a well defined set CY). In the Bayesian predictive approach, CX is obtained considering an optimality criteria based on the maximization of P(Y in CY|X in CX; D) constrained to the class of regions CX of size α, that is, with predictive probability of screening α fixed. Parametric modeling is a usual way to obtain the predictive distributions required for the formulation of the screening problem. Such modeling often implies the specification of a certain number of assumptions which are difficulty to verify in practice. We relax the parametric assumption by proposing a Bayesian nonparametric screening methodology.   The model is illustrated with simulated and real data.