[Eeglablist] Clustering ICs across sessions but within subjects

Tarik S Bel-Bahar tarikbelbahar at gmail.com
Fri Aug 11 10:40:20 PDT 2017

​Great Ghislain, Happy it helped to gain a little sanity. I often try to
play with/fool eeglab when I get weird errors, at least to figure out what
is going on. STUDY is a bit complex, and as I mentioned is getting updated
to be smarter.

Yes, I recommend using clustering and visual analysis/comparison of IC
properties, and looking for those that share substansial similarities. This
can also include statistical/quantitative comparison of IC properties. This
step should include closely looking at IC properties, and/or results with
some ICA classification tools (though you have some limitations such as
sparse channel montage).

Yes, if I saw that ICA solutions from within-sessions but different
conditions were similar, producing a near-common subset of neural ICs, then
I would merge the conditions for that subject (but not sessions). Another
reason for merging within a session is because ICA is hungry for "more time
samples" to do better source separation. Note that I would merge from
within one session to begin with, unless I had specific hypotheses about
substantially different nerual activity across the (within-session)
conditions (for example, jumping-rope vs. lying down, or pain stimulation
versus a working memory task). Overall, I have seen good results with ICA
for many conditions within-one-session, with ICA pulling out expected
neural sources that are related to the "different" and "multiple

******EXTRA NOTES FOR Ghislain:

I also recommend trying it both ways (one ICA per condition and one ICA for
all conditions) and comparing closely for several subjects, and then moving

Note that some researchers run different ICA solutions for resting versus
ERP data.

Note that usually researchers also cutoff some of the ICs by limiting the
IC clustering to only ICs with a certain level of residual variance. This
is done in the Edit Study, and usually requires dipoles I think. However in
your case, you could manually mark what are clearly artifactual or
low-value ICs before clustering/and/or/visual-review. These would also be
the ones you want to remove eventually.

Note that that not all subjects have great represenative neural ICs, and
that some decisions have to be made about which to keep (and consider real
across subjects).

Note that though it is not recommended, some researchers merge sessions
(different recordings) of the same subject. This sometimes lead to double
or triple ICs that are quite similar, but related only to each recording
sessions. These can then be merged if the researcher chooses to.

Note that the errors you got can also be investigated by examining the
STUDY data structure variables, and taking a close look at the complex but
important STUDY function details and code.​
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170811/daa0fb9c/attachment.html>

More information about the eeglablist mailing list