[Eeglablist] Query regarding ICA and data length

[Scott Makeig]

Rebecca -

Although we have no principled or carefully researched  understanding of
what k = time_points / #channels^2 should be, with 256 channels
(dimensions), ICA decomposition does to us seem to require more data. A
quantitative investigation should now use mutual information reduction as a
metric of ICA effectiveness, as in Delorme et al., 2012 (ref below).

Fiorenzo Artoni investigated the use of PCA dimension reduction preceding
ICA decomposition (Artoni et al., 2016), concluding that PCA should not be
used (except to reduce the number of channels to its rank). If not, then
what?

My first thought would be to acquire longer data, if you really want to
take advantage of 256-dimension data.  You might ask subjects to
alternate rest periods with e.g. music listening, etc. The brain and
non-brain ('artifact') sources mixed in the EEG data will not be much
different, and the participant can be comfortable (ask them what music they
would like to restfully listen to?).

If this is somehow not possible, another approach could be to pick channel
subsets to decompose. Nima Bigdely-Shamlo wrote a command line function to
pick spatially evenly-disperse channel subsets, which should in the EEGLAB
distribution.  For example, you might decompose 90-channel subsets (each
subset spatially dispersed). This would lead to a component clustering
problem ... with no guarantee of real success.

A simpler approach would be to just acquire a smaller amount of data using
only say 64 electrodes. This would be less work. Unless you would be able
to perform quite careful brain source localization, the value to your
analysis of more channels than this is to me questionable. See the
simulations by Zeynep Akalin Acar (2013) on this point.

Scott Makeig

Delorme, A., Palmer, J., Onton, J., Oostenveld, R. and Makeig, S., 2012.
Independent EEG sources are dipolar. *PloS one*, *7*(2), p.e30135.

Artoni, F., Delorme, A. and Makeig, S., 2018. Applying dimension reduction
to EEG data by Principal Component Analysis reduces the quality of its
subsequent Independent Component decomposition. *NeuroImage*, *175*,
pp.176-187.

Akalin Acar, Z. and Makeig, S., 2013. Effects of forward model errors on
EEG source localization. *Brain topography*, *26*(3), pp.378-396.

Scott

On Tue, May 3, 2022 at 3:05 PM Reh, Rebecca <rebareh at psych.ubc.ca> wrote:

> Hello,
>
> I have several questions regarding best practices for ICA artifact
> detection on high density EEG data. I have 256 channels of resting state
> data, collected for 10 minutes per subject, downsampled to 250 Hz.
> According to the best practices outlined in Makoto’s preprocessing
> pipeline, the following calculation is recommended to determine the amount
> of data necessary for a good decomposition of the data using ICA: number of
> electode^2 * k (with k at 20 to 30 but potentially higher for high density
> recording). Using this formula, I’ve calculated I would need over an hour
> of data from each subject in order to run the full ICA (256 ICs). Reading
> through the EEGlab documentation, it looks like I can either run a PCA to
> reduce the dimensionality of the data before running the ICA, or I can
> exclude some of the channels from the analysis. I’m curious what best
> practices are in this case, and also how the field generally handles having
> less data than would be ideal for ICA (given that I don’t think most people
> are collecting over an hour of resting state data).
>
> Thank you in advance for any guidance/advice!
>
> Rebecca K Reh, PhD
> Department of Psychology
> 2136 West Mall
> The University of British Columbia
> Vancouver, BC Canada V6T 1Z4
>
> UBC Point Grey campus sits on the traditional, ancestral, and unceded
> territory of the Musqueam First Nation
>
>
>
>
> _______________________________________________
> 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