[FieldTrip] 'rule of thumb' for defining how many independent components to analyze for EEG?

Simon-Shlomo Poil poil.simonshlomo at gmail.com
Wed Nov 7 14:21:13 CET 2012


Dear all,

My experience with cleaning high-density EEG (i.e. >=129 channels) is that
PCA reduction is necessary before ICA. If you do not do the PCA, you (1)
spread out your artifactual source across multiple components,  (2) you get
a lot of components representing single channels, and (3) you need long
recordings to get enough data for so many component (i.e. ~30*129^2 as
lower limit for 129 components, a rule-of-thumb I have from the EEGLAB
website).

This also means that we, e.g., recommend our students using the
Neurophysiological Biomarker Toolbox to clean data to reduce their
129-channel data to rank 15 (
http://www.nbtwiki.net/doku.php?id=tutorial:compute_independent_component_analysis#.UJpfS4az70E)!
It is simply easier for them to understand 15 components than 129
components...

However, my opinion is only based on my non-systematical observations.. I
don't know if, e.g., reducing a 129 channel Signal to a PCA rank 30 Signal
potentially could remove low power neuronal signal (e.g. gamma
oscillations)?

I look forward to see the report on PCA dimension reduction effects on ICA!

Back to Eleanor's question: I would also recommend you to do a full rank
ICA for 64 channel data. Good luck with the cleaning.

Best regards,
Simon-Shlomo Poil






2012/11/7 Matt Craddock <matt.craddock at uni-leipzig.de>

> On 06/11/2012 19:44, Arnaud Delorme wrote:
>
>> Dear Ellie,
>>
>> you should decompose your full rank matrix. PCA dimension reduction may
>> seriously affects and bias ICA solutions (I have seen a yet-to-be published
>> report on that).
>>
>> Best,
>>
>> Arno
>>
>
> Dear all,
>
> Uh oh. If PCA is off the table, what should be done if the data is not
> full rank? e.g. if it's in average reference or multiple channels are
> bridged etc. Is the bias introduced by PCA worse than trying to decompose a
> matrix which is not full rank as if it were full rank?
>
> Wouldn't
>
> cfg.runica.pca  =rank(your_number_of_channels)
>
> always set it to 1, and thus tell runica to reduce to only a single
> component? Or does setting it to 1 tell runica to detect rank-deficiency
> and suggest an appropriate reduction? I don't normally run ICA via
> fieldtrip, so am not sure how to tell it to detect the rank and reduce
> accordingly; EEGLAB at least asks you about this when it detects rank
> deficiency (or at least, should do - i've found it a little temperamental
> on this point, but this is the Fieldtrip list so...).
>
> Regards,
> Matt
> --
> Dr. Matt Craddock
>
> Post-doctoral researcher,
> Institute of Psychology,
> University of Leipzig,
> Seeburgstr. 14-20,
> 04103 Leipzig, Germany
> Phone: +49 341 973 95 44
>
> ______________________________**_________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/**mailman/listinfo/fieldtrip<http://mailman.science.ru.nl/mailman/listinfo/fieldtrip>
>



-- 
-- 
Simon-Shlomo Poil

Center of MR-Research
University Children’s Hospital Zurich

Mobile number: +41 (0)76 399 5809
Office number: +41 (0)44 266 3129
Skype: poil.simonshlomo
Webpage: http://www.poil.dk/s/ and http://www.nbtwiki.net and
http://www.kispi.uzh.ch/Kinderspital/Medizin/mrzentrum_en.html
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20121107/9015cc13/attachment-0002.html>


More information about the fieldtrip mailing list