Periodic Spatial Filter for Single Trial Classification of Event Related Brain Activity

Abstract

Because of the small amplitudes of event related potentials (ERPs), they are usually hidden in electroencephalogram (EEG) recordings. This is particularly a problem when analyzing single-trial data. A spatial filtering method for P300 detection in oddball paradigm is proposed in this paper which is based on the assumption that brain responses to the same stimulus look the same (or at least do not change significantly over trials). Therefore, the sequence generated by concatenating all the responses to the same type of stimulus has a hidden periodicity. Enhancing the periodic structure of this sequence, a transformation is found to project the data into a lower dimensional subspace. Experiments show that even with a small subspace of the projected data, the classification performance in single-trial P300 detection is still high.

Joint spatial and spectral filter estimation for single trial detection of event related potentials

“…First, in the πSF method, a fixed FIR filter is assumed and the objective is to find the optimum spatial filters. The details of the πSF method have been presented previously in [7,8], however, since πSSF is based on πSF, this method is briefly introduced here. Next, in the πSSF method, we alternate between estimating the two filters by fixing one and optimizing the other.…”

“…We construct the matrices y T and y S by horizontally concatenating all instances ofŷ Ti andŷ Sj , respectively. Assuming that ERPs are deterministic [7,24], in an ideal case each row of y T and y S is a periodic signal with the period equal to the ERP window length, τ = m e .…”

“…Inspired by the periodicity measure defined in [25], the periodic spatial filter (πSF), proposed in [7], is based on enhancing the periodic structure of the data by minimizing…”

“…Other spatial filters are specifically designed for eventrelated potentials (ERPs), e.g., the methods in [4][5][6][7][8]. The proposed method in [4] maximizes the sum of the squared distances of the components of the mean feature vectors.…”

“…In [6], tensor decomposition is used to improve the features and reduce dimensions of the data. Recently, periodic spatial filter (πSF) method is proposed in [7,8]. The method is simple and efficient and is based on enhancing the periodic structure of the data created by concatenating ERP windows of the same type.…”