[Eeglablist] Coherence analysis on ICA components

[Marco Congedo]

Hello,

  neither averaging weights nor averaging source time series is a valid
approach,
and this is true in general, not just for coherence analysis.
Note that ICA weights and time series have arbitrary scaling,
thus averaging them makes no sense.
Either you perform coherence analysis for pairs of components,
or, if you are inetrested in coherence of clusters of components,
you have to use multivariate coherence measures.

See the following paper for multivariate coherence definitions:
http://arxiv.org/abs/0711.1455

A paper on BSS coherence analysis can be found here:
http://hal.archives-ouvertes.fr/index.php?action_todo=search&view_this_doc=hal-00423717&version=1&halsid=jbd4j01nu66tdhdoaoc37feqt6


A software for computing (group) BSS and pairwise (instantaneous and
lagged) coherence as described in the latter paper is available here:
https://sites.google.com/site/marcocongedo/software/nica

Hope this helps,

_______________________________________________________________

Marco CONGEDO,



Research Scientist,

Centre National de la Recherche Scientifique (cnrs) and Grenoble University.



Team ViBS (Vision and Brain Signal Processing)

GIPSA-lab (Grenoble Images Parole Signal Automatique)



11 rue des Mathématiques

Domaine universitaire - BP 46 -

38402, Grenoble, France.



tel: +33 (0)4 76 82 62 52

fax: +33 (0)4 76 57 47 90



http://sites.google.com/site/marcocongedo

_______________________________________________________________





On Wed, Mar 7, 2012 at 3:24 AM, Agatha Lenartowicz <alenarto at ucla.edu>wrote:

>
> Dear Susann.  The coherence bit seems like you have some feedback on.
>  However note that how you combine ICs is not trivial.  Imagine you have 4
> occipital components: left, right, dorsal and ventral topography. You want
> to combine them bc they all show similar alpha responses.  One approach is
> to average the time series. Another - that you propose - is to average the
> mixing weights and retrieve new activations.  In my example an average of
> the mixing weights would place the topography centrally -- where none of
> the original had weighted.  So my new activations would be weighted heavily
> by locations that weakly contributed to the uncombined ICs.  This is all to
> say - keep an eye on what your combining is doing to your data.  I've
> struggled with this - have no clear solution - but am favoring very simple
> averaging where a single best IC is not available. Agatha
>
>
>
> Sent from my phone.
>
> On Mar 4, 2012, at 14:29, Susann Sgorzaly <susann.sgorzaly at st.ovgu.de>
> wrote:
>
> > Dear All,
> >
> > I am working on my master thesis on EEG data and would like to do an
> > coherence analysis on ICA components. So my question is, has anyone
> > experience how to analyse it best?
> >
> > My approach was:
> > (1) Run ICA and identify task-related components.
> > (2) If there are more than one component: average weights of these
> > components and recalculate activation matrix
> > (3) Run coherence analysis on this new component with all other ICA
> > components
> > (4) Perform a cluster analysis on those ICA components which are most
> > coherent with the averaged component off step (2) to see         if they
> > have a similar topography.
> >
> > Is there a better way to perform coherence analysis on ICA components?
> >
> > Thanks
> >
> > Susann
> > _______________________________________________
> > 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
>



-- 
Marco Congedo
http://sites.google.com/site/marcocongedo
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120307/db1fe256/attachment.html>