Real-time on-line brain (BCI) applications will require automated analysis and classification of single trial or continuous of Electroencephalographic (EEG) and/or Magnetoencephalographic (MEG) data. Single trial or continuous analysis of this data requires the extraction of minute signals of interest from noisy environments in which the signal to noise ratios (SNR) can actually have negative values (the noise is more powerful than the signals of interest). Classical analysis of EEG/MEG data has dealt with these noise issues by means of signal averaging, trial rejection and regression analysis. With respect to BCI applications, all these techniques are highly problematic.
Independent Component Analysis (ICA) is one of the more modern approaches to single trial EEG/MEG analysis. ICA separates the raw data into independent components by seeking a decomposition that maximizes the independence between the extracted components.
The "classical" ICA approach to the noise problem assumes that the number of sources matches the number of detection sensors. Unfortunately, this assumption is rarely valid and usually there is a mismatch between the number of sources and the number of sensors. It should be noted that, under these circumstances, if forced to extract as many "sources" as there are sensors, the ICA algorithm will generate a set of sources that are as independent as possible, but they are not truly independent: they should be considered "pseudo-independent". When such a pseudo ICA decomposition is used to removed noise for BCI purposes, inherent mixing can also lead to misclassification of the cognitive process/state.
So, determining the proper number of sources to extract becomes a crucial issue in using ICA for EEG/MEG analysis.
Last modified: Friday, 16-Jul-2004 16:31:42 NZST
This page is maintained by Caroline Wills.