[Eeglablist] removing eye channels before ICA?

[Chang Gu]

Hi Tarik & colleagues ,

Thanks a lot for the suggestion! As I mentioned, the data from eye channels
may raise 2 interesting questions:

(1) the range or scale of eye potentials are different from central channels
around vertex Cz before averaged re-reference. Eye data could be 100 micro V
with huge variance, whereas channels near Cz may less than 1 micro V, with
almost no variance because of the data range. Mathematically, ICA performs
linear transform to the EEG data and the data scale shouldn't be an issue.
But I'm curious about, computationally, if the ICA algorithm considered
about this multi-scale problem, that is, if the ICA training is independent
with the different scales of data.

(2) Based on our experience, eye channels are good at the beginning, but
sometimes become bad (dry out, not good contact.. ) later during the
recording. it's not practical to interpolate them. If, for example,  my data
are good for the first half of the recording, but not good for the 2nd half.
And I run the ICA on the whole data set, is it possible that the bad part
have negative  influence on ICA training across the whole data, not only
influences the 2nd part? From my understanding,  the inferior performance of
ICA  on bad parts could spread to the whole data set. In this case, removing
eye-channels might be even better than keeping them..

Looking forward to further clarification.

best,

On Wed, Jun 8, 2011 at 5:18 PM, Tarik S Bel-Bahar
<tarikbelbahar at gmail.com>wrote:

> 1. In egi data I usually dont see such a big difference. If you are
> collecting the extra emg via egi's custom setup for extra channels, you
> should check with them.
>
> 1.5. It's probable not enough work has been done as of yet to compare ic
> decompositions with extra noneeg.channels, and ic decompositions without
> extra noneeg channels.
>
> 2. I often get ic's that seem to pick out one bad channel, but if the.data
> is cleaned enough, and of appropriate size, for ica, and youve.removed
> enough bad channels and noisy data, you should get some components that seem
> brainbased and interpretable.
>
> 3. We usually keep in all or most.chamnels around the eyes, unless they are
> very noisy. Usually ica picks up the eyeblink and eyemovement artifact
> pretty clearly as an independent ic or two.
>
> 4. It will be good to know your final solution for your extra channels, and
> whether extreme data can bias the ica results.Please let us know.
> On Jun 8, 2011 1:05 PM, "Chang Gu" <chang.gu at vanderbilt.edu> wrote:
>



-- 
Chang Gu
Psychology & Human Development
Vanderbilt University
Nashville, TN
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20110609/7f097f73/attachment.html>