Event-related EEG and MEG data averages, called event-related potentials (ERPs) or fields (ERFs), are a natural ICA application domain since summation at scalp electrodes (or SQUID sensors) is near linear and instantaneous, and since Infomax and other ICA algorithms require relatively few data points. Event-related averaging normally reduces the number of spatially distinguishable sources of ERP/ERF data to on the order of ten (plus EEG and artifact sources). These most probably index activations of (local or distributed) brain systems involved in cognitive information processing or motor control. Artifact sources separate activity arising from muscle, eye, heart, and line noise, and ICA provides an excellent method for removing them. Residual EEG signals are produced by spontaneously active brain systems not time- and phase-locked to the events of interest. Decomposition of spontaneous EEG using ICA involves issues of spatial nonstationarity not usually encountered in analysis of evoked responses.

EEG and ERP applications of ICA were first explored in Makeig et al. (1996), and the first detailed analysis of ERP data using ICA was presented in Makeig et al. (1997). Current projects include decomposition of P300 and memory-related responses.

A difficulty for psychophysiological research using ICA is the number of components to be examined and the complexity of their relations to phenomena of interest in the data. To assist researchers, a Matlab toolbox of 50 ICA decomposition, evaluation and visualization routines has been made available via this web site, and has been downloaded (as of 3/3/98) by nearly 500 researchers from around the world. Plans for this toolbox include new three-dimensional and single-trial visualization tools and a wider variety of ICA algorithms, A discussion of issues involved in applying ICA to psychophysiological data is contained in a frequently-asked questions page.

Scott Makeig