Comparisons of Signal Subspace Methods for Estimating Visual Evoked Potentials

Extracting visual evoked potentials (VEPs) from
electroencephalogram (EEG) noise remains a challenging task
since the signal-to-noise ratio (SNR) involved is generally very low. In this work, filtering manipulations by means of subspace approaches that break the contaminated VEP signal space into the signal subspace and the noise only subspace are introduced. Out of the two mentioned subspace, only the former is selected for further processing. Specifically, two eigendecomposition based signal subspace methods containing unique basis and estimator matrices were developed and their efficiency and performance were compared between each other. These algorithms denoted as Signal Subspace Method 1 (SSM1) and Signal Subspace Method 2 (SSM2) are able to satisfactorily extract the P100, P200 and P300 peak latencies from artificially generated noisy VEPs subjected to SNRs from 0 to -10 dB. The simulation results show that the SSM1 estimator maintains an average success rate of 87.3 %, with average errors of 5.4 for
P100, 14.1 for P200 and 30.6 for P300. The SSM2 filter registers an average success rate of 93.3 %, with average errors of 9.5, 5.0 and 1.9 for P100, P200 and P300, repectively.