[Eeglablist] What should be said about subjects that do not contribute to a given independent component cluster?

[Scott Makeig]

Scott & all -

> If a subject does not contribute to a given cluster, are we to interpret
that the subject simply does not possess certain sources (~they're missing
a brain part !?).

Yes, this is an important question. My thinking about it is as follows:

First, it is possible that finding the component is hindered by the ICA
equivalent of 'signal to noise ratio' (SNRO, namely that the number of
appreciable effective sources (including non-brain processes, aka artifact)
is larger than the number of channels, meaning that the projected
activities of some sources (at least) need to be spread across some or all
of the returned ICs. The way to mitigate this is to more strictly reject
periods of 'messy' artifact in the data. This can be augmented by running
an AMICA decomposition of the cleaned dataset using more than one ICA model
(as many as 20 have been used by us, with long, 'complicated' datasets).
Each AMICA model then has been trained on, and thus fits some *portion* of
the data points (the post-AMICA plug-in has good representations of these),
rather than fitting th whole data. If the time point 'domain' of the model
containing the IC of interest includes all or most of the time points of
interest, then the model IC can be added to the cluster (though effects on
standard EEGLAB statistical functions, etc. are unknown territory...).

But also, the failure to find a given cluster ICA for some subject may
reflect some difference in information cortical processing in that subject.
Julie Onton and I found a clear example of that when a lateral occipital
cluster accounting for the P1 (and part of the N1) peak(s) in the visual
item ERP in one study was missing 9 or the 25 subjects. Searching for the
IC anatomically closest in dipole source location in those folks, we found
that these 'missing subject' ICs showed a consistent difference in location
from the clustered subject ICs (e.g., they were further lateral), and also
showed distinct event-related time-frequency measures (ERSP, ITC) from the
clustered ICs - thus forming in effect a consistent 9-subject (sub)cluster.
The differences in IC source location and visual ERP were small enough to
have been overlooked in the previous visual ERP literature.  Were they
'real', i.e., did they reflect some substantive (reproducible,
generalizable) subgroup difference?  We had no way to fin out as we did not
have in depth information about the subjects to find correlations to the
physiological difference.

Do remember that even after gross (Talairach or MNI) co-registration,
brains have different physiognomies -- e.g., equivalent functional areas of
cortex may be separated by cm's in brain template space, and may also have
different orientations, meaning their scalp projections may differ
strongly. Arthur Tsai built a quite convincing simulation of this using
MR-based brain models for 5 subjects - see Figure 1 here: A Tsai, T-P Jung,
V Chien, AN Savostyanov, S Makeig. Cortical surface alignment in
multi-subject spatiotemporal independent EEG source imaging.
*NeuroImage* 87:i297-310,
2014 https://sccn.ucsd.edu/~scott/pdf/Tsai_etal_2014.pdf. I (still) hope to
work with Arthur to build EEGLAB software for the advances he proposes
there, as I believe it is the way forward for EEG source imaging.

How to treat source-resolved EEG measures in statistical analyses?  Here
Tim Mullen found this web page (h
<https://urldefense.com/v3/__https://www.uvm.edu/*statdhtx/StatPages/More_Stuff/Missing_Data/Missing-Part-Two.html__;fg!!Mih3wA!WJU9sCpgYj9cuxWMut9Ki4yvG4sbxAjD507Reox4fXIvRvv3aK-9BpUFzmh0zgzEXMz6cQ$>
https://urldefense.com/v3/__https://www.uvm.edu/*statdhtx/StatPages/Missing_Data/Missing-Part-Two.html__;fg!!Mih3wA!XMbCu6Ma_JJDjzBRaoLFoiY5cvcjoeqAo2hFWntxhp5Ffq36SnjiQNRK-eYExoS8JU2qMw$ 
<https://urldefense.com/v3/__https://www.uvm.edu/*statdhtx/StatPages/Missing_Data/Missing-Part-Two.html__;fg!!Mih3wA!WJU9sCpgYj9cuxWMut9Ki4yvG4sbxAjD507Reox4fXIvRvv3aK-9BpUFzmh0zgzUVRzuAg$>)
in
which a statistician explains the available approaches to statistical
analysis with missing data. The 0th way to to remove the subjects from the
analysis altogether. The 1st way is to replace the 'missing' measures with
the group-mean measure. The 2nd and more modern way is to estimate the
measure for each subject based on other measures that are correlated in
some way with it (this is, in effect, what advertisers do when they select
ads to show you on the web). The author gives more details ...

Scott

On Thu, Aug 20, 2020 at 4:10 AM Scott Burwell <burwell at umn.edu> wrote:

> Greetings,
>
> I am working with a large resting-state EEG dataset and using Measure
> Projection Analysis to combine and cluster independent components' spectra
> across multiple subjects. While MPA takes some of the guesswork out of more
> traditional methods of clustering independent components (e.g., k-means),
> it still leaves the possibility that some subjects do not contribute to a
> given cluster (or in MPA-speak, "domain"). For example, if a "medial
> frontal theta" cluster is found in MPA but a given subject does not have an
> independent component dipole within close proximity of that cluster, the
> subjects' contribution to the cluster is undefined or NaN.
>
> For these subjects that do not contribute to certain clusters / domains, I
> am curious how others might interpret the results? I know Scott Makeig and
> others have written about EEG independent components as reflecting "EEG
> effective sources," suggesting that only certain sources are sufficiently
> "effective" (e.g., generating large signals, acceptable orientation /
> distance from sensors) to contribute to the scalp EEG and ICA
> decomposition.
>
> If a subject does not contribute to a given cluster, are we to interpret
> that the subject simply does not possess certain sources (~they're missing
> a brain part!?). Or, perhaps with more channels and better coverage, we
> might be able to better resolve the source of interest (~we don't presently
> have enough measurements to faithfully capture the source of interest)?
> Other explanations? Any suggested readings that discuss these issues in the
> context of EEG/ICA would be greatly appreciated.
>
> Thank you in advance.
>
> Best,
> Scott
>
> --
> Scott J. Burwell, PhD
> He/Him/His
> NIDA T32 Postdoctoral Research Fellow
> Department of Psychiatry & Behavioral Sciences
> University of Minnesota, Minneapolis, MN


On Thu, Aug 20, 2020 at 4:10 AM Scott Burwell <burwell at umn.edu> wrote:

> Greetings,
>
> I am working with a large resting-state EEG dataset and using Measure
> Projection Analysis to combine and cluster independent components' spectra
> across multiple subjects. While MPA takes some of the guesswork out of more
> traditional methods of clustering independent components (e.g., k-means),
> it still leaves the possibility that some subjects do not contribute to a
> given cluster (or in MPA-speak, "domain"). For example, if a "medial
> frontal theta" cluster is found in MPA but a given subject does not have an
> independent component dipole within close proximity of that cluster, the
> subjects' contribution to the cluster is undefined or NaN.
>
> For these subjects that do not contribute to certain clusters / domains, I
> am curious how others might interpret the results? I know Scott Makeig and
> others have written about EEG independent components as reflecting "EEG
> effective sources," suggesting that only certain sources are sufficiently
> "effective" (e.g., generating large signals, acceptable orientation /
> distance from sensors) to contribute to the scalp EEG and ICA
> decomposition.
>
> If a subject does not contribute to a given cluster, are we to interpret
> that the subject simply does not possess certain sources (~they're missing
> a brain part!?). Or, perhaps with more channels and better coverage, we
> might be able to better resolve the source of interest (~we don't presently
> have enough measurements to faithfully capture the source of interest)?
> Other explanations? Any suggested readings that discuss these issues in the
> context of EEG/ICA would be greatly appreciated.
>
> Thank you in advance.
>
> Best,
> Scott
>
> --
> Scott J. Burwell, PhD
> He/Him/His
> NIDA T32 Postdoctoral Research Fellow
> Department of Psychiatry & Behavioral Sciences
> University of Minnesota, Minneapolis, MN
> View my profile on: LinkedIn
> <
> https://urldefense.com/v3/__https://www.linkedin.com/in/scottjburwell/__;!!Mih3wA!R4dgkVnIcGO6T84r_m0RvI7mxl1b9LFXDtYlLb-clkuXCeGQaQ-zVO7LXlA7A1XTCJR4Bg$
> > | Google
> Scholar <
> https://urldefense.com/v3/__https://scholar.google.com/citations?user=S-DUpywAAAAJ&hl=en__;!!Mih3wA!R4dgkVnIcGO6T84r_m0RvI7mxl1b9LFXDtYlLb-clkuXCeGQaQ-zVO7LXlA7A1XZjHLVLA$
> > |
> GitHub <
> https://urldefense.com/v3/__https://github.com/sjburwell__;!!Mih3wA!R4dgkVnIcGO6T84r_m0RvI7mxl1b9LFXDtYlLb-clkuXCeGQaQ-zVO7LXlA7A1X_KgsTJA$
> > | ResearchGate
> <
> https://urldefense.com/v3/__https://www.researchgate.net/profile/Scott_Burwell__;!!Mih3wA!R4dgkVnIcGO6T84r_m0RvI7mxl1b9LFXDtYlLb-clkuXCeGQaQ-zVO7LXlA7A1WF_OCm8Q$
> >
> _______________________________________________
> 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
>


-- 
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