[Eeglablist] ICA decomposition with 128 versus 64 channels?

[Arnaud Delorme]

>I've heard from one EEGLAB user that 128 channels don't confer much advantage
>over 64, since inputs must be spatially downsampled in order to be processed
>practically on typical computing hardware
>
Standard computers with 1Gb of RAM can process a standard 100 to 200Mb 
dataset with 128 channels (and run ICA). It is the lower limit though.

>, and since the independent components
>of interest (those from neural sources) don't become much cleaner with 128
>inputs as compared to 64.
>
They do become cleaner. It is important to have as many channels as 
possible if you want to perform source localization. For instance, in 
EEGLAB, DIPFIT will return residual variance on the collection of all 
data channels. If you can model one component with a single dipole and 
the residual variance is 2%, it is much more convincing if you have 128 
channels than if you have 64. The reason is a single dipole (5 numbers) 
was able to explain a scalp map containing 128 numbers (in contrast to 
64). Since ICA does not take into account the spatial localization of 
channels (and thus there is no reason for scalp maps to end up being 
modeled with a single equivalent dipoles), it is much more convincing to 
conclude than an independent component represent the activity from a 
single part of cortex when you have 128 channels than when you have 64.

Also, testing (not published) I have done with 256 channels show that 
the (single) dipole localization for ICA component becomes less stable 
when using less than 70 channels (in term of spatial localization). I 
reduced gradually the number of channels (using standard montages) and 
then performed dipole localization. The exact localization did not 
change much between 256, 128, and 70, but broke down below 70, and was 
dramatically different with 32 channels (more than 1 cm off). This would 
have to be validated in a larger study though.

Best,

Arno