[Eeglablist] Subjects comparison when ICA components rejection was not the same
Tarik S Bel-Bahar
tarikbelbahar at gmail.com
Tue Aug 25 14:16:08 PDT 2015
Hello Emmanuelle, Hoping all is well. some thoughts below that should help.
For your purposes, I would suggest sticking with one set of methods for all
your participants, as it is not usually acceptable to process some
participants one way, and some another. I would recommend doing normal
(non-ICA) cleaning of all participants, and showing the results that way.
After that, run the analyses within ICA, for only those participants who
did have good ICs. It all depends on how many participants total you have,
and how many you can "drop" for one reason or another. Finally , you may
want to confirm for yourself that statistics actually change for one
participant with and without the ICA cleaning.
-if people don't blink or move their eyes during eyes closed rest, then we
should not expect ICA to "find" such things. In the future, consider having
people give you a series of sample blinks, movements, etc.. This will make
it easier for ICA rejection, or similar methods that need some "information
-some researchers take an IC-artifact template from other studies and
"apply it" to new data. This works to some degree because at least eye
artifacts are quite stereotypical.
You're right, ICA methods and their deployment are still in development,
and usually only the beginning of the process is well-delineated. The most
advanced or current methods can usually be found in articles, or within
in-house methods in various labs.
You might also want to look at automated processing pipelines discussed on
this list, such as the TAPEEG toolbox, and another recent one from Kothe,
Bigdely-Shamlo et al. See also the wide range of ICA-rejection toolboxes
including ICMARC and SASICA.
-Some data just does not give great ICA decompositions, for various
-The small amount of time you have for ICA is likely impacting the quality
of your ICA decompositions. It's better to have closer to a half hour or
more, though many people try to use ICA on short periods.
-It also depends on how many channels you have. 64 chans or better is
recommended for ICA, though many people have published ICA and microstate
results with 32 channels.
On Mon, Aug 24, 2015 at 4:10 PM, Emmanuelle Renauld <
emmanuelle.renauld.1 at ulaval.ca> wrote:
> There are many discussions on ICA, but I can't find many on what happens
> I am now working on data where the subjects had their eyes open and looked
> at a computer screen.
> I am trying to use ICA to remove eye blinks and eye movements, but it is
> very difficult: I have no EOC, only 7 components, and only
> approximately 3-5 minutes of data.
> For most subjects, it works well, but sometimes data is not well divided,
> of course (ex: no component has the right topoplot, or their spectrum
> contain alpha waves). For some subjects, I have tried many different
> options, but I can never really remove eyeblinks and eye movements.
> So now if we want to compute statistics (ex, Fourier spectrum), how do I
> compare subjects where it worked and subjects where these artefacts were
> not removed!? Sometimes I feel it would be better to forget artefact
> rejection, so that, at least, all subjects are more or less the same...
> Any idea?
> Thak you very much!
> 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
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the eeglablist