[Eeglablist] Clustering ICs across sessions but within subjects

Tarik S Bel-Bahar tarikbelbahar at gmail.com
Thu Aug 10 10:22:22 PDT 2017

Hello Ghislain, some comments below. Best wishes


0. This error sometimes comes up and usually means that there is something
about the data that STUDY is not happy with. If it's replicable, and it
occurs when you use the eeglab tutorial data, I would say please put a note
about in the eeglab bugzilla. However it's most likely a bug on your end,
in your data and it's structure or setup.

*Note that ICA for sparse-EEG systems is not recommended. Trustworthy ICA
solutions usually occur with more than 32 channels (it's a spatial method),
though some researchers have used 32 channels or less.

*overall, forget about design for a minute,making it just one condition,
and just try to achieve clustering. Just tell study all 4 files are from 4
IDs, and from one condition.

******extra notes below

1. To begin with, try just the following: just load 1 file that has the ICs
for this subject from one condition, and then just load 1 file that has the
ICs for this subject from another condition. Name them as different
subjects. Then create study. Then attempt the clustering. What you want to
see, manually/visually, one step at a time, is some sanity/okayness
regarding your IC solutions (what is their quality, are there clear neural
ICs, how is this pattern different across subjects and conditions) .... and
then move to attempt statistical and study analyses.

2. Alternatively, load one file with ICA per subject.Note that you could
just get one ICA solution for all files from  single subject, and use these
as common ICs across all condition files for that subject. This would make
things easier for you than your current strategy.

3. Alternatively, load one file with ICA, for each unique ICA
decomposition. Then create study and cluster the ICs. You may have to trick
study into thinking each one file that is loaded is from a different
subject. Later, after you have clusters and know which ICs are usable from
each ICA decomposition, you can create a more formal study.

 4. The idea here is to FIRST figure out what the common ICs are for a
subject, for a condition, and across subjects and conditions. Then you
probably want to clean each data file based on that information (dropping
non-neural ICs) and then attempt analyses. Alternatively, you can find the
common neural ICs within subjects, conditions, and the whole study, and
then go from there.

5. The clustering is of better use for looking at ICs across subjects. So
for that kind of clustering, load 1 instance of each different subject's
ICA solution (just one ICA solution per subject for the moment). Then
re-attempt clustering.

6. Note that many researchers choose not to use study but rather compute
single-subject single-condition metrics and go from there, essentially
making their own study computations. However, there is much that can be
done with STUDY. However, STUDY can't do everything, and it often helps to
build out your own "averaging procedures". ....In terms of ICs, this would
mean at least having a list of "good ICs" and "common ICs" within-subjects,
across-subjects, and across-conditions.

7. Note that you will have to decide how/which ICs within one subject
across different conditions are the same or similar. This can/should be
done first by examining the IC properties from different conditions from a
single subject. I recommend trying this as a kind of sanity check.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170810/d03a62ff/attachment.html>

More information about the eeglablist mailing list