Purpose: To examine the robustness of a pattern recognition system under the influence of noise in Event Related Potentials (ERP) signals in discriminating controls from patients. Material and Methods: ERP recordings were simulated by generating two series of signals, based on real signal templates from normal controls and patients, with various levels of added-on Gaussian noise. From the resulting signals, a number of waveform characteristic quantities were calculated and they were used as input to an ensemble classification structure, which consisted of three different classifiers, namely the Bayesian classifier, the k-Nearest Neighbor (kNN) and the Probabilistic Neural Network (PNN), and following the majority vote rule. Results: The classification accuracies of individual classifiers were over 80% for the PNN and over 75% for the kNN and the Bayesian. The ensemble structure improved classification precision resulting in an overall accuracy of over 87% for all noise levels tested. Conclusion: Results provide an estimation of the robustness of the developed ensemble classification scheme, which may be of value to the clinician, given that ERP signals are usually corrupted by noise in clinical practice.