Autoregressive Estimation of Short Segment Spectra for Computerized EEG Analysis

Abstract

The hypothesis that an electroencephalogram (EEG) can be analyzed by computer using a series of basic descriptive elements of short duration (1-5 s) has prompted the development of methods to extract the best possible features from very short (1 s) time intervals. In this paper, the merits of three alternative methods for estimating spectral features are compared to the fast Fourier transform (FFT). These procedures, based on autoregressive (AR) modeling are: 1) Kalman filtering, 2) the Burg algorithm, and 3) the Yule-Walker (YW) approach. The methods are reportedly able to provide high resolution spectal estimates from short EEG intervals, even in cases where intervals contain less than a ful period of a cyclic waveform. The first method is adaptive, the other two are not. Using Akaike's final prediction error (FPE) criterion, it was demonstrated that a fifth-order filter is sufficient to estimate EEG characteristics in 90 percent of the cases. However, visual inspection of the resulting spectra revealed that the order indicated by the FPE criterion is generally too low and better spectra can be obtained using a tenth-order AR model. The Yule-Walker method resulted in many unstable models and should not be used. Of two remaining methods, i.e., Burg and Kalman, the first provides spectra with peaks having a smaller bandwidth than the Kalman-flter method. Additional experiments with the Burg method revealed that, on the average, the same results were obtained using the FFT.