[Eeglablist] Best Approach for Running ICA on EEG Data for Classification(ME & MI) (Pragati Dode)

[japalmer29 at gmail.com]

Hi Pragati and others,

Keeping the data together is a sound approach, but I think this is really a
research issue, dependent on the application. For example, we can easily
imagine a context where the brain sources differ with condition,
particularly in the topographic map / dipole location. In this case, the
unmixing matrices will differ with condition. If a classifier is trained on
specific source activity, you may get better results using the condition
specific source versus the unified sources (and a multiclass classifier). On
the other hand, if the condition differences involve differences in source
activity from a spatially fixed brain source, then you would be better off
doing ICA on the combined data as this will give a better estimate of the
spatially fixed source. As we generally don't know what the differences are
a priori, it's probably best to try both, with cross-validation, and asses
the performance.

Additionally, I have found at least with ECoG data that improved
classification can be achieved by doing band-pass filtering prior to ICA,
using band specific source power as features.

Best,
Jason

-----Original Message-----
From: eeglablist <eeglablist-bounces at sccn.ucsd.edu> On Behalf Of Yamil Vidal
via eeglablist
Sent: Monday, March 17, 2025 6:49 AM
To: EEGLAB List <eeglablist at sccn.ucsd.edu>
Subject: Re: [Eeglablist] Best Approach for Running ICA on EEG Data for
Classification(ME & MI) (Pragati Dode)

Dear Pragati,
Undoubtedly, you need to keep all your data together for running ICA, and
actually for all of your preprocessing. You should only divide the data in
conditions as a last step for the purpose of running the classification.
This is because if you first divide your data in conditions and then do
further preprocessing, be this ICA cleaning or anything else, you are likely
to introduce differences in the data across conditions which your classifier
might later pick.
In this way, you might find differences between conditions, even if there
are no real differences.
One way to think of it is that you are likely going to use ICA to remove
blinks, cardiac and muscle artifacts, which should not differ across
conditions and therefore you don't need to divide the data beforehand.
Another reason not to divide the data is that ICA benefits from having
plenty of data.
I hope this is helpful.
Best,
Yamil

Dr. Yamil Vidal
Predictive Brain Lab
Donders Centre for Cognitive Neuroimaging Nijmegen the Netherlands
_______________________________________________
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
or visit https://sccn.ucsd.edu/mailman/listinfo/eeglablist  .