[Eeglablist] Computing a second ICA on a dataset

[Germán Gómez-Herrero]

Hi Jan,
Applying ICA in several consecutive runs is usually not a good idea unless
you take special precautions to avoid using rank-deficient data in your ICA
estimation. Also, running ICA several times will not let you (in general)
obtain sources that you did not obtain already in the first run. The
problems that you experience are a typical result of trying to apply ICA on
rank-deficient data.

If you had a 20-channels EEG data and in the first ICA run you removed 10
components, this means that in the second run you will not be able to
extract more than 10 independent components (even you still have 20 EEG
channels). This is because in the first ICA run you already removed 10
spatial components out of the 20 that you needed to describe perfectly your
original EEG dynamics. If you try to apply ICA and extract 20 independent
components out of this "rank-deficient" data the result will be inaccurate
in the best case and non-sense in the worst (e.g. an imaginary separating
matrix or just a crash). Probably the easiest solution to this problem is to
always run PCA before ICA so that you ensure that your data covariance
matrix is well-conditioned. Check the attached main.m file to see what I
mean.

I would suggest EEGLAB's developers to enforce this pre-conditioning step in
the function runica to avoid other users running into the same problems.

Hope that helps,
Germán




---------------------------------------------------------------------
Germán Gómez-Herrero
M. Sc., Researcher
Tampere University of Technology
P.O. Box 553, FI-33101, Tampere, Finland
Phone:   +358 3 3115 4519
Mobile:  +358 40 5011256
Fax:     +358 3 3115 4989
http://www.cs.tut.fi/~gomezher/index.htm

> -----Original Message-----
> From: eeglablist-bounces at sccn.ucsd.edu [mailto:eeglablist-
> bounces at sccn.ucsd.edu] On Behalf Of Jan R. Wessel
> Sent: 07 November 2008 18:01
> To: eeglablist at sccn.ucsd.edu
> Subject: [Eeglablist] Computing a second ICA on a dataset
> 
> Hey everybody,
> 
> I'm analyzing an EEG dataset on an eye-movement task. In order to cope
> with the massive eye-movement artifacts, I computed a normal ICA on the
> dataset and got rid of all the oviously eye-movement related
> components.
> Now, I'm trying to compute a second ICA in order to clean up the data
> even more, because there is still considerable saccadic activity in
> most
> of the datasets.
> 
> When I'm doing this, I'm running into a series of problems. Seemingly
> depending on the memory of the machine I am running it on, a) either
> the
> ICA takes days to compute for a single subject, b) or Matlab terminates
> without any prompt, leaving me with a "segmentation fault" in the unix
> terminal or c) if it runs until the end, it seems to give me (and here
> comes the really weird part) complex numbers (!) such as
> "0.009259085778863 + 0.013638045843933i" as ICA weights. When I display
> the component activations, I get some funny activation spots in the
> bottom left corner of the plot, while the rest of the plot is blank.
> This effect is regardless of which epochs I display.
> 
> So obviously, the second ICA doesn't really seem to work out. Does
> anybody have any experience with consecutive ICA analyses or has even
> come across the same problem and can maybe help me out here?
> 
> Thanks in advance,
> best,
> Jan
> 
> --
> Jan R. Wessel, Dipl.-Psych.
> Cognitive Neurology Research Group
> Max Planck Institute for Neurological Research
> Gleueler Straße 50
> D-50931 Cologne
> Phone: 0221 - 4726 343
> Mail:  Wessel at nf.mpg.de
> 
> _______________________________________________
> 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
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: main.m
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20081113/834b4810/attachment.ksh>