[Eeglablist] On ICA based artifact rejection

Joseph Dien jdien07 at mac.com
Sat Sep 15 14:56:06 PDT 2012


Hi Arno,
   it's an interesting paper.  Thanks for alerting me to it.  I'll have to disagree about the conclusion you drew from it.  David's finding is that one should indeed perform a mean correction on epoched data prior to ICA.  He simply concluded that it's better to use the mean of the entire epoch rather than of just the prestimulus baseline.  That said, his findings were not conclusive in this regard.  As he put it, "On the other hand, it could be that removing each epoch’s mean does not aid ICA so much as removing the prestimulus baseline impairs it."  So more work is needed.  All I can say is that in my own informal testing of my artifact correction routines, I found that I got better results with a baseline correction than with none at all (using EGI data, if it makes a difference).  I'll have to look into how results compare with using the entire epoch instead of just the baseline.  Have you had a different experience with your own datasets?

Also, to expand on Steve's comment about high-pass filtering, it's not just a matter of whether the data has been high-pass filtered (I think everyone does so) but also the degree to which it has been high-pass filtered.  My own experience is that if one high-pass filtered sufficiently to remove the slow drifts, it would also start attenuating some of the ERPs of interest, especially CNVs and the P600 and the like.  As David puts it, mean correcting the epoch operates like a leaky high-pass filter, and I think that its advantage is that it is one that does not have this unwanted side-effect.  Anyway, there is a lot more methodological work that needs to be done on this topic.  It's good that progress is being made on this important basic topic and that we're all thinking about it.  Let's go back to our computers and see what more we can come up with.

Cheers!

Joe


On Sep 15, 2012, at 2:32 AM, Arnaud Delorme <arno at ucsd.edu> wrote:

> Yes, Steve is right. Baseline correction might offset the potential of some electrodes compared to other electrodes, and if there is a source projecting linearly to several data channels (as it should), the projection might not be linear after baseline correction. Consistent with this, David Groppe who has done formal testing on ICA data quality decompositions and showed that ICA quality degrades when using short baseline prior to doing ICA. 
> 
> http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3062525/
> 
> Arno
> 
> On Sep 13, 2012, at 11:23 AM, Steve Luck wrote:
> 
>> Eren and Joe:  I'm not an expert on using ICA -- this is just the advice I received many years ago from Arno and Scott.  The idea is that baseline correction might change the scalp distribution of various components on each trial, which would prevent ICA from working properly.  However, I don't know whether baseline correction ends up being worthwhile on balance despite this potential problem.  It may, as Joe suggests, depend on factors such as whether the data have been high-pass filtered and the nature of the slow drifts.
>> 
>> In general, we need more sophisticated simulation studies to answer questions like this.  (This is a hint for you technically inclined graduate students and postdocs!)
>> 
>> Steve
>> 
>> 
>>> From: Joseph Dien <jdien07 at mac.com>
>>> Subject: Re: [Eeglablist] On ICA based artifact rejection
>>> Date: September 12, 2012 6:56:29 PM PDT
>>> To: "Gunseli, E." <e.gunseli at vu.nl>
>>> Cc: "eeglablist at sccn.ucsd.edu" <eeglablist at sccn.ucsd.edu>
>>> 
>>> 
>>> All due respect to Steve (who runs a great course), I'd need some persuading to follow this advice.  The entire reason we baseline correct is that without it we can't actually be sure what the true zero is due to slow drifts in the data.  Much of the relative values that are being removed we want to remove because they reflect these slow but, over time, substantial drifts.  In my own data, I've found that I get better results (in terms of clean removal of blink artifacts) with ICA artifact correction of epoched data if it is baseline corrected first because it helps minimize the effects of these drifts.  I haven't subjected this question to systematic evaluation however.  I imagine it would also depend on the characteristics of the recording equipment and filter settings and so forth.
>>> 
>>> Joe
>>> 
>>> On Sep 12, 2012, at 7:02 AM, "Gunseli, E." <e.gunseli at vu.nl> wrote:
>>> 
>>>> Dear all,
>>>>  
>>>> I have a question about the steps that should be taken before running ICA.
>>>>  
>>>> If we are going to epoch the data (step 5), why are we manually rejecting the extra noisy parts earlier (step 2).
>>>> Since these extra noisy portions are mostly at beginning and end of trial blocks, they will be gone away during epoching anyway.
>>>> So, epoching the critical time window can save a fair amount of time that else we would have spent on manual inspection.
>>>>  
>>>> At this point I have another question; I have read that, it is better to run ICA on continuous, non-epoched data.
>>>> One of the problems of running ICA on epoched data is that, “the baseline correction changes relative values across channels” (S. Luck, ERP Boot Camp Lecture Slides). But probably that is not the only reason to run ICA on continuous data because this problem can easily be overcome via removing the baseline after running ICAs.
>>>> So I guess there should be other problems related to running ICAs on epoched data.
>>>> Can anyone provide information about these potential problems?
>>>>  
>>>> Kind regards,
>>>> Eren
>>>>  
>> 
>> --------------------------------------------------------------------
>> Steven J. Luck, Ph.D.
>> Director, Center for Mind & Brain
>> Professor, Department of Psychology
>> University of California, Davis
>> Room 109
>> 267 Cousteau Place
>> Davis, CA 95618
>> (530) 297-4424
>> E-Mail: sjluck at ucdavis.edu
>> Web: http://mindbrain.ucdavis.edu/people/sjluck
>> Calendar: http://www.google.com/calendar/embed?src=stevenjluck%40gmail.com&ctz=America/Los_Angeles
>> --------------------------------------------------------------------
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>> _______________________________________________
>> 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


--------------------------------------------------------------------------------

Joseph Dien,
Senior Research Scientist
University of Maryland 

E-mail: jdien07 at mac.com
Phone: 301-226-8848
Fax: 301-226-8811
http://joedien.com//










-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120915/77bc322f/attachment-0001.html 


More information about the eeglablist mailing list