[Eeglablist] ICA "adds" noise?

Fri Jan 18 11:39:05 PST 2013

Another thought occurs to me.  I have indeed noticed a tendency for increased noise to show up in my own ICA-based artifact correction routine in the EP Toolkit (Tim Curran first reported it to me).  I've never worked out why.  I ended up implementing a trial-by-trial workaround wherein the eyeblink factors are removed from a given trial only when it reduces the overall variance of the trial.  In other words, when the benefit outweighs the cost.  The increased noise that I see is small enough that it gets averaged out for ERPs so has not been an issue.  Could be an issue for frequency-based measures though.  I need to look into this further.  Anyway, what you're reporting seems more severe than anything I've observed so perhaps something different.

Joe

On Dec 21, 2012, at 6:27 AM, Kristina Borgström <kristina.borgstrom at psychology.lu.se> wrote:

> Dear Tarik,
> 
> I really appreciate your thorough reply and attempt to help! I have received several good suggestions here and will try some of the different approaches people have suggested.
> 
> I think the first thing I will check is to run the original vertex-referenced data throught the ICA, so that the data is still full rank. I will also try keeping the average mastoid ref but removing one of the channels during ICA. I will also try the average reference while taking into consideration the rank issue with that. The reason I have not gone with average reference from the beginning is that I want to compare with some other data that was mastoid-referenced. 
> 
> To answer some of your questions: I am actually running Matlab v 10, so I don't think the software version is an issue.
> 
> I have only tried ICA on a few subjects so far, but this issue only appeared in 2 subjects, so it doesn't seem to be a general problem, although it's too early to know how widespread the problem is.
> 
> Regarding your points on cleaning the data before ICA, I have done both some automatic artifact detection, but also a complete visual check to throw out any extreme artifacts other than the ocular artifacts. So I'm pretty confident that I'm feeding the ICA with as much clean data as possible.I want to use the ICA primarily to correct the eye artifacts so I don't have to throw out those trials completely. 
> 
> Thanks for all the tips on different ways to handle the number of channels to use. I will definitely try some different alternatives and compare the results. 
> 
> I have considered using ICA for the main analysis as well as artifact correction, i think it's a matter of trying to see what works out the best. I definitely want to see what the actual ERP data looks like though. 
> 
> I won't be able to work with this until after the Christmas holiday, but I will post an update to the eeglablist as soon as I've tried some new techniques to let you know they've worked!
> 
> Thanks to all who have gotten involved and contributed with valuable thoughts!
> 
> Best regards,
> Kristina
> 
> 
> On Dec 21, 2012, at 2:03 AM, Tarik S Bel-Bahar wrote:
> 
>> Greetings Kristina,
>> 
>> First off, thanks for your lucid and informative email. You provide a lot of
>> useful detail that makes it easier to make some suggestions.
>> Working with precious pediatric data of course has some unique issues,
>> and I think you can easily find out about these approaches
>> by looking at existing work with infant and pediatric eeg (including #
>> of trial issues).
>> 
>> ..I'm not sure if you specified you had a problem with rank in your
>> data, but as Matt suggested
>> it is useful to make sure you are operating and interpreting your ICA correctly.
>> 
>> ...I would urge to check that this is a pattern across all subjects or
>> only some.
>> If across all, then it could a hardware or child development issue.
>> If not, then it could be a subset of your 2 year olds where doing
>> something special.
>> 
>> I don't have any other useful direct thoughts, but further below
>> I suggest some things that came to mind that you might want to try out,
>> having worked with egi systems a bit.
>> 
>> Best wishes, and let us know what your final solution/interpretation is!
>> 
>> Tarik
>> 
>> 
>> ********more thoughts here
>> consider the following:
>> dropping those bad or noisy channels (you can safely drop down to 90 channels
>> if you have to, as long as the channels are not all contiguous or take
>> out Major scalp regions
>> (such as all the occipital channels). Further, the validity and
>> cleanness of the ICA decomposition
>> depends partly on how much data that is free of extreme artifacts
>> that you have given to ICA  (ideally close to an hour or more).
>> I would suggest making sure you have really cleaned as well as you
>> can, using some
>> combination of functions and visual checking.
>> However it seems that your focus is on using ICA as a cleaning technique,
>> and you are not using doing your analyses on ICs themselves.
>> 
>> As an additional quick stringent test,
>> you can try to remove u to 40 channels, and leaving at least 30 minutes of data.
>> See if you can get interpretable information under those extreme conditions.
>> 
>> As an additional strategy (in case total data timepoints are too few)
>> is to reduce your electrode density
>> to 64 and restart your processing from there. If you recorded at
>> higher than 250 samples per second, you
>> might also consider reducing the sampling rate. Both strategies would
>> make things easier for ICA to
>> show you what might be going on in the eeg.
>> 
>> Regarding your channels, I've seen that kind of noisiness, very
>> similar to your first pic,
>> in infant, adolescent, and adult egi data, but not always. There are
>> multiple possible sources
>> for this, including muscle tension and electrode contact, and possibly
>> impedance as well.
>> 
>> Regarding removing components, yes you might be removing  some eeg
>> dynamics related to cognition.
>> 
>> Last, have you considered using the ICs themselves as your main
>> analysis points rather than
>> the reconstructed EEG data after removal of artifact ICs ? On the
>> other hand, there are many
>> studies that have successfully used ICA for cleaning. Cheers!
> 
> 
> _______________________________________________
> 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/20130118/c6d65b5f/attachment.html>