[Eeglablist] Applying ICA weigthing matrix from one set to another

[Scott Makeig]

Sarah -

Applying ICA to 10-sec data epochs is problematic unless you are using
relatively few channels. For example, with 32 channels, the weight matrix
learned by ICA has 32x32 =1064 entries. This would be difficult to
learn accurately from only 256X10 = 2,560 data point examples. You might say
this is because in 10 s of data, the independence of each of the 32 ICA
sources from all the others is likely not well expressed...

If you are using ICA learned from a previous portion of the data for truly
online artifact removal, then the accuracy of the ICA filtering depends not
only on the spatial stability of the EOG artifact sources, but on the
spatialy stability of all other EEG and artifact sources in the data....
This makes online ICA filtering (with selective model updating, say) a
signal processing challenge. Audio solutions are being pursued agresssively
by some companies, I hear, but for 32+ source solutions as in EEG, not much
work may have yet been done.

For post-hoc (not online) analysis, it is better to decompose as much data
as possible, providing that the same electrodes are used and the subject
state is the same. Even then, however, different tasks can lead to different
spatial distribution of the dipolar ICA sources accounting for non-artifact
activity, something Julie Onton found in a preliminary exploration but
deserves more detailed study.

Scott



On 11/22/06, Sarah Hosni <sarah_m_hosni at yahoo.com> wrote:
>
> Hello,
>
> I read previous messages on : "*Applying ICA weigthing matrix from one set
> to another*"
> and I saw two replies: one saying that "Yes, this is correct since the
> data was recorded on the same day. *The best approach though is to
> concatenate all datasets and run ICA* (as ICA is more stable with more
> data)."
> the other reply was:"your procedure is methodolologically uncorrect. The
> results of the ICA, the components and the weight matrix, are always
> differents if you change datasets, because all the dataset are different,
> and the ICA works whith this data."
>
> *I want to ask*: if I'll concatenate all datasets and run ICA .. this will
> make ICA an offline method, meaning I have to run it on all data first. and
> if it's wrong to use same weight matrix for different sets.. then I can't
> see how ica can be really useful in removing EOG artifacts.
>
> In my experiments, I trained ICA on one 10 second trial for *each subject*and then used the resulting weight matrix to remove EOG from other trials
> for the same subject recorded in different sessions. and it worked.. SO my
> question is: was what I did wrong?
> I was thinking that the *eye independent component* for each subjects will
> remain the same.
>
> *another question*: if I obtained my weight matrix from training on a 10
> second EEG segment, can I use it to remove EOG from a smaller segment say
> an 0.5 second window of the same 10 second EEG instead of applying it on
> the whole data ( as ICA is more stable with more data but actually when we
> need to remove EOG is from small contaminated windows)
>
> thank you
>
>
> ------------------------------
> Sponsored Link
>
> Rates near 39yr lows. $420,000 Loan for $1399/mo - Calculate new house
> payment<http://www.lowermybills.com/lre/index.jsp?sourceid=lmb-9136-16418&moid=4116>
>
>
> _______________________________________________
> eeglablist mailing list eeglablist 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
>
>


-- 
Scott Makeig, Director and Research Scientist, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0961, http://sccn.ucsd.edu/~scott
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20061204/2b97ffeb/attachment.html>