Classification of multichannel EEG data using length/energy transforms

We propose the use of length and energy transforms in the classification of multichannel EEG data to identify different cognitive activity using a reduced set of recording electrodes. The length transform (ET) represents a temporarily smoothed time course of the data, while the energy transform (ET) can be interpreted as a short-term energy estimate. The transformation of the data in the length/energy domain allows to effectively preserving important data features when autoregressive (AR) models are used to reduce the dimension of the classification problem. We evaluate the performance of the ET and ET on the classification of real cognitive EEG data for the case when the optimal AR model is selected under the Schwarz's Bayesian criterion (SBC) and a Mahalanobis distance-based classifier is used. Our results show that accurate classification is achieved when the data is transformed through the ET or ET even for low-order AR models, having the ET slightly better performance.