EEG modelling using neural network and enchament averaging of brain evoked potentials

Abstract

In this thesis, a new parametric model for Evoked Potential (EP)estimation has been developed and implemented. It is assumed that prestimulusEEG data can be modeled by an implicit nonlinear autoregres~;ive(NAR) model. The N.AR model has been realized USing a multilayered neuralnetwork having a single hidden later and a single output neuron. Theconventional backpropagation learning law has been applied to estimate t.heparameters of the network.The mode I obtained using pre-stimulus data has been used toforecast post-stimulus signals. The forecast errors have been interpreted astrle EPs. The EP:3 tfiUS obtained have been compared favorat>1y /;vitrl UK,iSeobtained USing conventional averaging methods WhlCh requJre considerablymore trials.To test the validity of the model Hle autocorrelation of thepredlction error was computed. This error should be white if the rnodel isadequate.The software implementing the proposed method is developed onIBM PC/MS DOS environment USing` Turbo C 2.0` programming language.Keywords: Evoked Potential, Neural Networks, Backpropagation.