Vector autoregressive model selection in multichannel EEG

The objective of this paper is to present a methodology for the selection of vector autoregressive (VAR) models for multichannel electroencephalogram (EEG) data. This technique is based on the minimization of the Kullback-Leibler discrepancy index, which gives a measure of the dissimilarity between the unknown true model and a sample-based model. An experiment was performed by modeling the EEG corresponding to various sleep stages using 4 channels of sleep EEG segments. Two estimators, AIC and HQ, of the discrepancy were used. HQ produced smaller model orders than AIC. No characteristic order was associated with the models of each sleep stage represented in the EEG segments.