[Eeglablist] ICA clustering

[Scott Makeig]

Adam -

Performing ICA decomposition on your entire subject task data is not so
misguided as you seem to think. I like to say that ICA identifies source
activities (we call them activations) by removing from the data all the
other (max independent) sources. It does this by learning the inverse
weights (runica) or the source maps themselves (AMICA) - they are just
inverses of each other. Most IC sources of interest - in particular
distinct non-brain sources (eye movements, scalp muscle activities, line
noise artifact) are in common to both task conditions (unless they are very
different from each other). Therefore, we routinely decompose the whole
task data and expect that ICA will use the larger training data to separate
the components (and learn the component scalp projection maps) more
exactly.

Now, one way to see whether source space non-stationarity is created by
change in task conditions is to run AMICA decomposition on the whole data
to give two models - which compete with each other during decomposition for
data points, thereby separating the data into two models. A post-AMICA tool
will plot the log likelihood of each measure through the dataset (using
some smoothing length you can set). Does this plot show that each condition
receives its own model?   Very likely not, or only for a few subjects, as
the models compete for each data frame (time point) separately.
Unfortunately, we have not yet found a good method for learning what
exactly produces the model separation - different brain or non-brain
sources, different activity patterns? ...

A second possibility is to a) decompose the whole data, b) remove
interpretable non-brain sources, then separately decompose the two
dimension-reduced portions corresponding to the two conditions?  Do they
produce different brain components of interest?  And how do you know they
are different? (Here you might use a repeated decomposition approach such
as RELICA uses).   Though this may require a lot of programming to be made
statistically sound, simply trying it as an exploratory step should not be
difficult...

Scott Makeig

On Wed, May 24, 2023 at 10:57 AM Adam Grinberg via eeglablist <
eeglablist at sccn.ucsd.edu> wrote:

> Hi,
> I am interested in performing source reconstruction on independent
> components.
>
> I have epoched data of ERP experiment, with two conditions (Both
> conditions are on the same dataset). I already ran ICA on each dataset
> (i.e., each participant) and now wish to create an EEGLAB study for all
> participants. I first tried to create a study and add "code" as an
> independent variable in the study design, to differentiate between the
> trials. EEGLAB just ignores this, it seems... The result is that the PCA
> clustering is performed on all trials regardless of condition. Not very
> useful.
>
> Another option is to divide the datasets into two, a separate dataset for
> each condition. This is actually a requirement when creating a "simple ERP
> STUDY".
> Alternatively, I can create one study for each condition.
>
> However, when I separate the conditions into two datasets, the ICA weights
> from the original dataset are retained. Therefore, when I cluster the
> components, the plots look the same for both conditions. Does anyone have
> any insights? How can I solve this problem?
>
> Thank you in advance.
>
> /Adam
>
> ______________________________________________________________________________________
> Adam Grinberg, PT, PhD
> Department of Community Medicine and Rehabilitation, Physiotherapy
> Umeå University
> Umeå, Sweden
> Tel: +46 (0) 90 786 96 30
>
> _______________________________________________
> 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