[Eeglablist] Q&A re independent component clustering

[Scott Makeig]

Carmen Canavier (Louisiana State University Health Sciences Center) and I
offer the following for the possible interest of other EEGLAB users. See my
> comments to her questions below.

Scott Makeig <smakeig at ucsd.edu>
Carmen Canavier <ccanav at lsuhsc.edu>

On Fri, Mar 20, 2020 at 11:01 AM Canavier, Carmen C. <ccanav at lsuhsc.edu>
wrote:

> Dear Scott,
>
>
>
> Tony in my lab has been using EEGlab to find dipolar independent
> components in human EEG data.
>
> I have three questions that you may be able to answer or refer me to
> someone who can help.
>
> Tony is clustering based on dipole location across individuals and several
> conceptual issues have arisen.
>
>

> 3) We are interested in alpha generators. We often get multiple IC from
> the same individual in a cluster across individuals. Tony selects the IC
> whose scalp map deviates the least from the scalp map for the average
> dipole for the cluster. It does not make sense to me that if we are
> identifying a functional circuit characterized by a dipole that is common
> across individuals that there would be more than one in a given individual.
> Has this issue come up before? Would it be reasonable to add alpha power to
> the space in which we are doing the clustering to try to get a unique
> contributor?
>

 > I'll respond to this question first. If you are looking for clusters of
alpha-producing source areas, then yes, it makes sense to cluster on
location and alpha power (though subject-normalized to avoid overweighting
Ss with large EEG [= thin skull??]). This is also the kind of approach
supported by Nima Bigdeley-Shamlo's Measure Projection Toolbox (MPT)
plug-in, though using different machinery for finding clusters (in MPT aka
"domains").
> Also, I too feel that a representative cluster across subjects makes most
sense when it includes just one source for each represented subject
(*non*-represented
subjects are a different question).

> Technically, however, there is an exception. What ICA does in fact is to
separate the scalp channel data into a sum of maximally independent
component *subspaces*. When a separated subspace has dimension 1, we call
it (itself) an independent component process. However, some (e.g.,
spatially non-stationary) sources may have more than one dimension. To see
this we may use pairwise mutual information (via an EEGLAB function
contributed by Jason Palmer, included in the post-AMICA plug-in) to find
the so-called IC 'dependent subspaces' (small sets of ICs with some
remaining mutual information). If an IC cluster contains one IC in such a
subspace, then excluding other subspace ICs that have naturally been
included in the same cluster does not quite make sense (though strictly
speaking, here the subject-wise statistics are a bit more complex - and It
does *not* make sense to me, by the way, to attempt to solve this problem
by just summing their projection to a single scalp channel, as different
ICs and IC subspaces project with varying strengths to different scalp
channels...).

1) It does not seem reasonable to cluster dipoles in different hemispheres
> in the same cluster.
>
> Is there a way to add a large distance penalty for crossing hemispheric
> boundaries? Alternatively,
>
> is it ad hoc or acceptable to separate clusters arbitrarily into left and
> right components.
>

> Since co-registering brains (even after skull co-registration) is not
exact, and the brain and skull midlines do not match exactly in ~anyone,
and several types of error contribute to equivalent dipole localization of
brain IC effective sources, we should expect some uncertainty as to
hemisphere in near-*medial* source clusters. As well, it is difficult or
impossible to fit a dual symmetric dipole model to a medial IC scalp map,
both for numerical reasons, I suspect, and because the strict geometric
symmetry we assume in the dual dipole fitting is not quite anatomically
accurate in anyone, and this problem may be exacerbated when the two
dipoles are very near each other.  So I would *not* exclude as
physiologically unrealistic clusters with an assigned dipole in the
(near-medial) opposite hemisphere.

>
>
> 2) It does not seem reasonable to cluster single dipoles and pairs of
> dipoles. Can a large penalty be added if the number of dipoles in an IC is
> different (ie 1vs 2)? Alternatively, is it ad hoc or acceptable to separate
> clusters arbitrarily into single and paired components.
>


>  > We (Piazza et al. 2020
> <https://www.frontiersin.org/articles/10.3389/fnhum.2020.00082/full>)
> have recently published a comparison between the prevalence and location of
> dual-symmetric ICs infants and adults (participating in passive auditory
> oddball experiment). Results showed that (i) decompositions of infant data
> found many more dual-dipole brain ICs, (ii) these could be clustered, and
> (iii) that the centroid locations of the clusters in infants and adults
> were near equivalent -- but that many of those clustered ICs in the adult
> data could be well modeled as having a much weaker dual dipole. (iv)
> Further, the hemisphere that was (slightly) dominant in an infant
> dual-dipole component cluster was strongly predominant in the adult data
> cluster. Although there is a lot more to explore and learn, I take these
> results to suggest that EEG/LSP interhemispheric coupling through corpus
> callosum (and/or the quite few other anatomic possibilities) is quite
> normal. (v) Zeynep Akalin Acar has just shown (from looking through only a
> few datasets so far) that in adults, the weaker hemisphere of the
> dual-dipole IC source process may also be less stable over time - she is
> using Fiorenzo Artoni's RELICA to see this.
>

> So, in sum, removing dual-symmetric dipolar ICs from consideration does
not seem a wise idea. Currently, I believe, the standard EEGLAB clustering
software considers only the position of the (right?) hemisphere dipole --
though this now should be the position of the *stronger* of the two
symmetric dipoles (Arno, please check!).

>
>
> Thank you in advance,
>
> Carmen Canavier
>

> Best wishes,  Scott Makeig


-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0559, http://sccn.ucsd.edu/~scott