[Eeglablist] A second ICA after removing components?

[Scott Makeig]

Yes, the comment of Ronald (below) is correct:  Running ICA a second time on
the back-projection of an IC subset (or on the IC subset waveforms
themselves) will 1) need to be a lower-dimension decomposition (in the first
case using PCA data dimension reduction on the channel data; in the second,
on the reduced number of IC waveforms directly), and 2) should return the
same results (IC maps and time courses) as the original decomposition
(within numeric/statistical error bounds).

Running ICA on data from which time points at which non-stereotyped 'noise'
epochs or stretches of data have been removed, on the other hand, can indeed
result in a better ICA decomposition -- in two senses:  a) the resulting IC
time courses will be more independent, and b) more of the ICs produced will
have a ('dipolar') scalp map taht can be associated with activity in a
single (localizable) cortical patch...

We will soon include in EEGLAB a plug-in function by Jason Palmer to measure
numerically the reduction in mutual information effected by an ICA
decomposition (i.e., info on measure a) above); the number of 'dipolar' ICs
returned by a given decomposition can be determined now by properly fitting
equivalent dipole models to the IC scalp maps, and measuring the number that
are fit with low residual variance...

Scott Makeig



On Tue, Aug 16, 2011 at 12:15 AM, Ronald Phlypo <Ronald.Phlypo at ugent.be>wrote:

>  Dear Max,
>
> the problem lies in the maximally allowed number of components in both
> decompositions. The first decomposition may allow for as many sources as
> sensors (15 in your case). However, once 3 artefacts have been removed, your
> data dimension reduces to 12, which allows to estimate a maximum of 12
> sources only. To circumvent this problem, I suppose what the wiki means is
> to do local decompositions first (trial by trial or epoch by epoch) and then
> concatenating the trials/epochs again after their correction and before the
> second "joint" decomposition. Since in this case the artefact removal is
> nonlinear, it does not reduce the dimension of your concatenated data.
>
> Hope this helps,
>
> Ronald
>
> PS: you might also want to have a look at
> http://www.hindawi.com/journals/cin/2007/075079/cta/ where short time and
> long term windows are used jointly for artefact removal. The text refers to
> literature on the mean duration of electrophysiological processes to
> motivate this decision.
>
> Le 15/08/2011 18:09, Maximilien Chaumon a écrit :
>
> Hello all,
>
> I'm currently cleaning data before working with components.
>
>    - I cut my dataset into epochs
>    - reject epochs where signal is bad
>    - run an ICA
>    - find blink and muscle components, reject them
>    - run an ICA again, and look at my components.
>
> ... as I understood was suggested at the bottom of this page
> http://sccn.ucsd.edu/wiki/Chapter_01:_Rejecting_Artifacts
>
> Then the ICs look very nice but come in pairs of extremely similar
> topographies with different time courses, as shown on this picture<http://oszilla.hgs.hu-berlin.de/public/2ICAs.png>
> .
> I am wondering what happened here. I can imagine that rejecting components
> before running the second ICA is what went wrong... But why did I read that
> on the wiki?
>
> Thanks a lot for any advice.
> Max
>
>
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request 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
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>



-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation & Adj. Prof. of
Neurosciences, University of California San Diego, La Jolla CA 92093-0559,
http://sccn.ucsd.edu/~scott
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20110817/478d3ea2/attachment.html>