The P600 component is part of the late components of Event-related Potentials (ERPs), which have been related to working memory (WM) mechanisms. The relation of psychiatric illnesses to deficits in WM may manifest itself as a differentiation at the level of the ERP scalp measurements. In the present work, in order to test the robustness of a classification system under various levels of Gaussian noise, ERP activity at 15 leads was simulated creating two sets of templates representing the normal control and patient classes. From these templates a number of representatives of the two classes were produced. Independent Component Analysis (ICA) was applied as a pre-processing step. From the ICA-reconstructed ERPs, in the time window corresponding to the P600 component, five morphological features were extracted and they were used as input features to a Probabilistic Neural Network (PNN) classifier. Results indicate acceptable tolerance of noise, corresponding to overall classification performance levels higher than 80%, up to levels of 20% noise. In most of the best feature combinations and noise level tested, the standard deviation of the amplitude of the P600 component was present, indicating the possible significance of this feature for discrimination in the case of noise-corrupted data.