[Eeglablist] Question regarding ICA on continuous vs epoched data

[Scott Makeig]

Danielle -

Actually, ICA has no 'idea' of time relationships of the different frames
(time points) of data it trains on.  It also has no 'idea' of a head, of
electrode channel positions, ...
It simply finds a model of the data as linear mixtures of independent
sources of temporal information contributing to the data - with no regard
for the order of the data points in time, nor of the channels in space.
Therefore, it does not matter in what temporal order you submit the data
for decomposition.

- If decomposing data epochs, do not remove the epoch baseline from each
epoch -- as this in effect adds random sources to the data, complicating
the goal of ICA.
- Better to decompose more rather than less data -- iff the data do not
contain 'non-stereotyped' non-brain 'noise' -  my classic example: the
participant scratching their head, something giving a series of spatial
(noise) patterns (from electrode movement and capacitance changes) that do
not reappear in the data... That is, 'longer is better' - except when the
cognitive state of the participant changes (e.g., they fall asleep, change
task focus, etc.), in which case multi-model AMICA ICA decomposition may be
applied.
- Better to highpass filter the data above ?1? Hz, so as to reduce the
influence of source process contributions to the data that a) may be quite
large, and b) may not represent static brain source activity (e.g.,
galvanic skin response, organized as slow waves of moving changes in skin
conductance associated with skin moisture).

Scott Makeig

On Mon, Dec 11, 2023 at 6:36 PM Danielle Farrar via eeglablist <
eeglablist at sccn.ucsd.edu> wrote:

> Hello,
>
> I am trying to process some VEP data using a checkerboard paradigm that
> flips every 0.5 - 1.5 seconds, using a 64 lead cap to record.  There's
> about 30 minutes total of data, with breaks of checkerboard flips every few
> minutes and lasting about 40 seconds.  Given the density of cap, it seems
> that I should be using all of the data, otherwise I will have roughly 240
> short epochs total.   I have been wondering about using ICA and ICLabel
> rejection to remove artifacts -- I see that a lot of papers report
> performing ICA on epoched data, but it seems that the epochs are longer.
> If I epoch the data, it seems that it will either yield very short epochs
> or overlapping epochs, both of which seem not ideal.  I'm wondering if I
> should choose the overlapping epochs, choose short epochs, or just use the
> continuous data to get an accurate analysis.  Any thoughts or suggestions
> would be much appreciated.
>
> Thank you,
> Danielle
> _______________________________________________
> 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