Optimum Order Selection Criterion for Autoregressive Models of Bandlimited EEG Signals

Autoregressive (AR) model is commonly used in many areas of signal processing and is particularly significant in Electroencephalogram (EEG) related studies owing to its enhanced resolution, smoother spectra and its potential to be used on short segments of a signal. When fitting an AR model to an EEG multivariate system, selection of model order ( p) is of critical importance. While the lower model orders provide inadequate representation of the signal, higher orders drastically increase noise. Therefore, identification of optimum AR model order is an open challenge. Conventional methods for estimating model orders include Akaike Information Criterion (AIC), Final Prediction Error (FPE), Bayesian Information Criterion (BIC), and Hannan Quinn (HQ). In this paper, we show how these criteria fail to determine optimal order of multivariate EEG signals that undergo mandatory filtration to separate bands and noise. Consequently, we present a novel, yet simple and effective technique to find the AR optimum order to model such systems. Extensive application of proposed method on different EEG dataset indicates that the new method gives better signal estimation with minimum possible order; thereby, improving the accuracy of reconstructed signal at reduced computational cost.