The inversion of late Event-Related Potentials to intracranial current source activities provides a method to observe brain phenomena related to information processing mechanisms. The use of current source waveforms, as inputs in classification systems, may provide robust performance for Decision Support Systems in Psychiatry. In the present work a new method for the classification of intracranial current sources is proposed, combining the Multivariate Autoregressive model with the Simulated Annealing technique, in order to extract optimum features, in terms of the classification rate. The classification is implemented with a three-layer NN trained with the back-propagation algorithm. The system was applied in the classification of normal controls and schizophrenic patients, providing classification rates of up to 90%. Furthermore, the clustering of intracranial source locations providing best classification performance may indicate relationships between the brain areas corresponding to these locations and pathological mechanisms.