To extract the independent components of the evoked potentials in the EEG using ICA

Abstract

The aim was to develop a reliable method of extracting the independent components of single trial evoked potential (EP) signals to derive features for the subject’s bioprofile, for diagnostic, prognostic, and monitoring purposes. Single trials are of interest, because conventional averaging conceals trial-to-trial variability and hence information. Independent Components Analysis (ICA) is a technique for Blind Source Separation (BSS) to recover N temporally independent source signals s = {s1(t), ... sN(t)} from N linear mixtures (the observations), x = {x1(t), ... xN(t)} obtained by multiplying the matrix of unknown sources s by an unknown mixing matrix A, (x = A.s). ICA seeks a square unmixing matrix W such that s = W.x. Difficulties arise for short duration, relatively low amplitude EPs, which have sparse ICs. The effectiveness of different algorithms was compared. Problems associated with more sources than measurement electrodes and with the generation by the algorithms of artefactual components were investigated. Ways of extracting the true EP components were considered. Component grouping was applied to obtain reliable groups, which could be explored for any clinical interpretations. Here we describe the recommended approach as developed by our virtual research group.