[Eeglablist] ICA on multi-epoch data

[arno]

Dear Peng,

>   Thank you for your reply. Do you mean that it would get the same
> (similar) result if I link all the 120 epochs together (make it look
> like one epoch), then perform ICA, then divide the results into 120
> epochs at the right time point? I mean within Infomax or FastICA
> algorithm.

Yes, that's exactly right and that's what EEGLAB is doing.

>   Another issue is about the ica algorithm. I'm now using 128-channel
> EEG system, which would produce 128 components by default settings of
> Infomax algorithm. That would make the computing very slow... And,
> biologically, there may be less brain sources (eg. we anticipated to
> get less clusters). So I wish to set less components (say, 10-20) for
> "runica". However, I am not sure whether it would reduce the accuracy
> of ica results.

Biologically you can expect to have as many sources as neurons in the 
brain! so the more sensor you have the better. When you use more 
channels though, some ICA components could split into several components 
(because the additional channels allow them to express their 
independence). This means that you may have to regroup their activity 
(manually or using the EEGLAB clustering interface), and this makes the 
analysis tedious.

However, we like to use all ICA components even if it is longer. The 
main reason is that from a theoretical point of view, reducing the 
dimensionality of the data using PCA may introduce artifacts (because 
you might affect several independent components by removing one PCA 
component). All the test we have done here show that reducing the data 
using PCA does not alter dramatically the decomposition, so this 
theoretical limitation might not apply.

I guess that an sensible approach could be to decompose your data with 
only a few components (20) just to have an idea of what you can get. 
Then you have a clear idea and have found an hypothesis to test, you 
could perform and use full decompositions.

>   And more on ica: If there were 10 "sources", and we pre-set 11
> components for our ica program. Then what would the one more component
> stand for? Did it mean one source would be interpreted as two
> components; or there're possiblities that all 10 "sources" could be
> mixed together?

This is an ideal case and I do not know exactly what would happen with 
artificial data. You should send an email to the ICA mailing list. I 
guess it depends on the type of algorithm. In EEG data, we do not expect 
this would happen because you can expect that there is a (very) large 
number of sources which contribute differently to the data (some sources 
may contribute a lot, and some others not at all, but it is not like 
that there is a discrete finite number of sources).

Hope this helps,

Arno