Artifact rejection for improving the performance of evoked potential neural network classifiers

This paper is aimed at improving, through artifact rejection, the performance of neural network evoked potential (EP) classifiers designed to detect match/mismatch conditions. A cluster analysis approach is formulated to identify artifacts that occur in the signals used for training the neural network classifiers. The clustering based artifact detection algorithm uses a distance measure resulting from a nonlinear alignment procedure designed to optimally align EP signals. Match and mismatch EPs collected for network training are clustered and the identified artifact signals are excluded from the training set. Artifacts that occur during testing are also identified and rejected by including an additional output in the neural net classifier for the artifact class. Preliminary experiments conducted show significant improvements in classification accuracy when the proposed artifact rejection methods are incorporated in the training and testing phases of a neural network EP classifier.