[FieldTrip] ICA subspace reduction with PCA eigenvalues

Arnaud Delorme arno at cerco.ups-tlse.fr
Fri Nov 14 16:41:03 CET 2014

Hi Nehad,

This is Arno from UCSD (developer of EEGLAB, the software which popularized ICA).

All commercial software preprocess the data using PCA to reduce the dimensionality. This can not only bias ICA but also introduce artifacts in the data. ICA should only be done full rank if possible.

You might want to contact Stepehn Shall Jones (shall-jones at infoscience.otago.ac.nz) which student Carl Leichter has written a thesis on that subject "Eigenspecters: Information Misallocation in Lossy Principal Component Reductions of Magnetoencephalographic and Electoencephalographic Data". He shows which artifacts are introduced by doing PCA before ICA.



On Nov 14, 2014, at 4:18 AM, Nenad Polomac <polomacnenad at gmail.com> wrote:

> Dear all,
> I have one question regarding ICA calculation subspace. I have high pass filtered MEG data (> 30 Hz) on which I calculated ICA.
> I know there are different methods to limit data dimensionality by ICA calculation, usually performed by PCA prior to ICA. E.g. taking 99% of variance explained or selecting first N components...
> However I did it by taking specific eigenvalue (2.5^27) and assuming that all components which have eigenvalue smaller than this represented noise. Argument for this would be that eigenvalue of one component represent scale of the variance explained by this component (for example, muscle artifacts component would have high eigenvalues while noise component oscillate in narrower range, hence variance and eigenvalue would be smaller). I have found this value to result in a stable number of components of around 45 per participant which is large enough to contain most of the signal and small enough to obtain stable and meaningful ICA components. 
> Does anybody know any studies that used this constant eigenvalue approach to limit ICA subspace?
> Thank you very much in advance? 
> all the best!
> Nenad
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20141114/b5f95640/attachment-0002.html>

More information about the fieldtrip mailing list