Linear discriminant analysis is studied when the predictors are data of functional type and the response is a Bernoulli random variable. The aim of this work is to anticipate the prediction of the response earlier than the end of the observed stochastic process. Due to the infinite dimension of the predictor space, discriminant coefficient functions cannot be derived as in the classical way and partial least squares approach is proposed. Results of a simulation study as well as an application to kneading data are presented.