Novel Approach for Time-Varying Bispectral Analysis of Non-Stationary EEG Signals

Abstract

A novel parametric method, based on the non-Gaussian AR model, is proposed for the partition of non-stationary EEG data into a finite set of third-order stationary segments. With the assumption of piecewise third-order stationarity of the signal, a series of parametric bispectral estimations of the non-stationary EEG data can be performed so as to describe the time-varying non-Gaussian nonlinear characteristics of the observed EEG signals. A practical method based on the fitness of third-order statistics of the signal by using the non-Gaussian AR model, together with an algorithm with CMI is presented. The experimental results with several simulations and clinical EEG signals have also been investigated and discussed. The results show successful performance of the proposed method in estimating the time-varying bispectral structures of the EEG signals.