[Eeglablist] ICA questions
Ronald Phlypo
ronald.phlypo at ugent.be
Fri Aug 24 23:24:42 PDT 2007
?????? <knyazevv at ngs.ru> a écrit :
> Dear eeglab Experts,
>
> I am just starting to acclimatize myself to eeglab and
> would be most grateful if you can answer a couple of
> questions that a crucial for selection of strategy of my
> future work.
>
> 1. Does it make sense to use ICA if only 32 channels (2
> oculomotor) are available? What is the minimal number of
> channels which still allows to do ICA?
>
ICA can be used from 2 channels on under the usual conditions that the
underlying sources are independent, the spatial topographies differ
-i.e. distinct spatial filters- and there is a maximum of one Gaussian
source. One can use the extra conditions that the mixing should be
linear and (over-)determined if you want to use straightforward linear
BSS algorithms.
After all, the sought after information should be available in your
data, and the criteria for separation that are used should not be
violated (see e.g. the above conditions).
The interpretation of the components however remains an issue of
discussion, since a mathematical sound model does not necessarily
return physiological estimates (although the assumption that "natural"
sources can be thought of as mutually independent in many cases). Good
post processing and careful interpretation of results is the major
care for a good ICA application!
> 2. Is it possible to compare groups of subjects or perform
> a correlational analysis (say, correlate eeg and
> personality measures) using ICA (given that in each
> subject IC are individual)? I realize that clastering may
> help to combine components of a sample of subjects into
> chimerical "subjects" and thus get round the problem, but
> it does not help when individual differences are of
> interest. Could you please suggest a possible solution or
> give relevant references? Receiving a pdf copy of relevant
> paper would be particularly appreciated.
>
Correlation analysis as explained above could be done by ICA if
dependency is sought, although one should take care whether or not the
correlation already gives sufficient information. Suggestions would be
to add the measures as an extra channel and perform ICA on the whole
set, the "augmented topographies" of the sources of interest would
then show a large projection onto both the channels of interest and
the added channel(s).
I do not know of any reference within this field, but I am almost sure
there should exist some.
Kind regards,
Ronald
> Thank you in advance,
>
> Gennady Knyazev
> _______________________________________________
> eeglablist mailing list eeglablist at sccn.ucsd.edu
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
>
More information about the eeglablist
mailing list