Adaptive Filtering and Pattern Recognition of Evoked Potentials

Abstract

<p>The problem of estimating evoked potentials and its pattern recognition and classification is addressed in this thesis. After providing the relevant physiological background and reviewing the various methods of processing the evoked potential, we propose the method of adaptive noise cancellation for estimating the evoked potential without stimulus repetition. A new weighted exact least squares lattice algorithm is derived for this purpose. The variable weighting factor can be used to make the algorithm robust. Its performance is compared to that of unnormalized and normalized exact least squares lattice algorithms and is shown to be superior. One example of using adaptive noise cancellation to estimate evoked potential without stimulus repetition is presented. Pattern recognition of evoked potentials is achieved by syntactic methods. We derive a finite-state grammar to represent the normal evoked potential. Suitable preprocessing using a zero-phase bandpass filter, parsing and attribute checking are the steps in this classification procedure. A database of normal evoked potentials and optimized acceptance criterion are used for checking the attributes. Detailed training and test runs are performed to demonstrate the performance of this classifier.</p>