In supervised classification, data generally comes from a basic sampling scheme : simple random, or a stratified sampling plan according to the categories of the response variable. In this paper we deal with the case of a binary response and quantitative predictors where data are collected by means of an orthogonal or near orthogonal design, built on the predictors (eg. fractional factorial, response surface design, ?), often with repeated measurements. In this framework, the estimation and use of the usual classifiers, such as Fisher?s LDA, logistic regression, quadratic or SVM, have some specific features we will develop.