Cross-validation and neural network architecture selection for the classification of intracranial current sources

In the present paper, a new methodological approach, for the classification of first episode schizophrenic patients (FES) against normal controls, is proposed. The first step of the methodology applied is the feature extraction, which is based on the combination of the multivariate autoregressive model with the simulated annealing technique, in order to extract optimum features, in terms of classification rate. The classification, as the second step of the methodology, is implemented by means of an artificial neural network (ANN) trained with the backpropagation algorithm under "leave-one-out cross-validation". The ANN is a multilayer perceptron, the architecture of which is selected after a detailed search. The proposed methodology has been applied for the classification of FES patients and normal controls using as input signals the intracranial current sources obtained by the inversion of event-related potentials (ERP) using an algebraic reconstruction technique. Results implementing the proposed methodology provide classification rates of up to 93%.