[Eeglablist] ICA "adds" noise?

[Matt Craddock]

Hi Kristina,

I've recently hit on similar behaviour (and in some cases even *worse*) 
with some of my datasets. For me, this seemed to be because of 
decomposing the data as if it were full rank (i.e. all the channels are 
approx linearly independent from each other) when it wasn't (there are 
also previous posts along these lines by Maximilien Chaumon - see 
http://sccn.ucsd.edu/pipermail/eeglablist/2011/004326.html). Looking in 
your specific case, I'd be quite cautious about removing those "noisy" 
components because they look to me like they also contain some genuine 
brain activity (though note that the two are basically identical, just 
opposite sign, which makes it look like the same problem as above).

There are much more expert people than me on this list who can tell you 
much better why these noisy components happen, but what I can tell you 
is interpolation and rereferencing reduce the rank of the data, and this 
can, occasionally, make ICA do things like you're seeing. I'd suggest 
running PCA first to reduce the dimensions in the data accordingly. I've 
seen Arnaud Delorme suggest on another list you should avoid using PCA 
first, but I'm not sure that advice applies when the data is not full-rank.

Now, if you run ICA through the menus in EEGLAB, it should detect that 
your data is not full rank, suggest an appropriate number of components 
to return, and then run PCA before ICA. However, if you run the ICA as 
part of a script, it'll probably be set up in such a way that it skips 
this step of suggesting an appropriate number of components and asking 
if it should reduce the data first. You may see a message along the 
lines of "fixing rank computation inconsistency probably because you're 
on linux 64 bit matlab"; this appears even if you're on Windows, and 
always selects the higher number of components returned by two different 
methods of calculating rank, which, in my experience, means it always 
decomposes the data as if it were full rank. So I'd suggest, if you're 
running ICA as part of your script, calculating the rank of your data 
and then passing that to the pop_runica function yourself.

Cheers,
Matt

On 19/12/2012 10:26, Kristina Borgström wrote:
>
>> Hi,
>> I have an issue regarding ICA for artifact correction that I really
>> would appreciate some help with.
>> Here is some background information: I have recorded child data (2
>> years old) with EGI, 128 channels. Before export to EEGLab, the data
>> has been band-pass filtered 1-30 Hz, epoched, clearly bad epochs (with
>> more artifacts than just eye artifacts) were removed, and bad channels
>> in the remaining epochs were interpolated. The data was rereferenced
>> to average mastoid reference. I then imported into EEGlab, which
>> treats the data as continuous, but has all the event information. I
>> have then performed ICA in order to correct for eye artifacts.
>>
>
> *The problem (please see the image files located at following links):*
>> **https://dl.dropbox.com/u/7016081/DataPlots.jpg
>
> https://dl.dropbox.com/u/7016081/ICAComponents.jpg
>>
>> In (at least) two data files, besides some clear eye artifact
>> components (components 1 & 11: blink; component 7: horizontal eye
>> movement), the ICA also found two components that look like pure high
>> frequency noise (components 2 & 3).
>> When I remove the eye artifact components (1,7 & 11), the eye
>> artifacts are in fact removed, BUT the data looks generally “noisier”,
>> i.e. the channels overall are fuzzier (see the image “DataPlots” for
>> comparisons). When I calculated individual averages with this data, it
>> indeed contained massive amounts of high frequency noise that was not
>> present in the averages where I did not use ICA at all but instead
>> removed all epochs containing eye artifacts.
>> I then continued and tested removing the “noise components” (2 & 3),
>> and the data then looked like it did originally, minus the eye
>> artifacts. It didn’t seem to have a major effect on the ERP components
>> either, but of course removed the high frequency noise.
>>
>> *My main questions are:*How can noise be “added” to the data, after
>> removal of certain components? How can I determine what type of noise
>> components 2 & 3 consist of? I’ve looked at the frequency plots, but I
>> don’t think it’s very clear. Can it be line noise, or EMG? The scalp
>> topographies are very widespread, and EMG is usually more laterally
>> located. Should it be ok to just remove these two components when they
>> appear, or is there a risk that they contain cognitive components?
>> Many thanks for any input you can give me!
>> Regards,
>> Kristina Borgström
>> PhD Student
>> Department of Psychology
>> Lund University
>> Sweden
>> +46-46-2223638

-- 
Dr. Matt Craddock

Post-doctoral researcher,
Institute of Psychology,
University of Leipzig,
Seeburgstr. 14-20,
04103 Leipzig, Germany
Phone: +49 341 973 95 44