Automatic correction of artifact from single-trial event-related potentials by blind source separation using second order statistics only

Abstract 

Event-related potentials (ERP) are in general masked by various kinds of artifacts. To attenuate the effects of artifacts, various schemes have been introduced, such as epoch rejection, electro-oculogram (EOG) regression and independent component analysis (ICA). However, none of the existing techniques can automatically remove various kinds of artifacts from a single ERP epoch. EOG regression cannot handle artifacts other than ocular ones. ICA incorporating higher order statistics (HOS) normally requires data with large number of time samples in order that the solution is robust. In this paper we blindly separate the multi-channel ERP into source components by estimating the correlation matrices of the data. Since only second order statistics (SOS) is involved, the process performs well at the single epoch level. Automatic artifact identification is performed in the source domain by introducing objective criteria for various artifacts. Criteria are based on time domain signal amplitude for blink and spurious peak artifact, scalp distribution of signal power for eye movement artifact and power distribution of frequency components for muscle artifact. The correction procedure can be completed by removing the identified artifactual sources from the raw multi-channel ERP.

Keywords: Electroencephalogram (EEG), Event-related potential (ERP), Automatic artifact correction, Blind source separation (BSS), Second order statistics (SOS), Independent component analysis (ICA), Higher order statistics (HOS)