[Eeglablist] How does transferring ICA matrixes between same-subject data sets affect further processing?

[Marius Klug]

Hey Duncan and all!

Yeah, you really should ask Prof. Gramann's group ;-) (Duncan was doing an
internship with us for a month, I am one of Prof. Gramann's PhD
students...) We were discussing this at some point, I saw this mail and
thought "Ah that sounds similar to what Duncan asked me about, let's check
what the EEGLAB Pros answer!" and only later realized that it was actually
you asking this question! So, great to have this discussion here!

I'd also say just try out everything and check, this is going to be
interesting. You know that we're still exploring this ourselves, so well...
just try out and report to us and the list ;-) We're experiencing heavy
artifacts of all kinds in our datasets, which is also due to the Phasespace
Mocap system and the HTC Vive VR headset... Basically, we're in the stage
of finding suitable solutions for data cleaning and source separation and
do it similarly, trying out what works best at this point. Evelyn
Jungnickel just told me that she repeatedly did ICA, then checked where the
algorithm had troubles separating the sources (e.g. by checking if there
are time periods in which all the components have strong activity), then
cleaned those parts out of the original data set and then repeated the
process until the components appeared comparably clean. ICA and denoising
is a delicate process for sure... Data cleaning and source separation is
the toughest frontier in MoBI research at the moment I'd say, and hopefully
ever more sophisticated solutions will emerge on the research horizon! On
another note there's also a more general thing "Joint Decorrelation" (
https://www.ncbi.nlm.nih.gov/pubmed/24990357), which might turn out to be
useful in some cases. But as I said, we're still on this ourselves... and
there's also the valid point you raised once that the more you clean the
data the more signal you are also prone to lose in the process. No perfect
solutions unfotunately!

As a question from my part: I thought that the ICA components combined
would always contain the whole original dataset, just split up differently,
and if you subtract all components you get zero at all channels. I take
that is true for the data set from which the ICA was calculated? So, if I
understand this correctly, you say that's not the case any more if you take
the weights and apply them to longer/other data sets? Really interesting.
But I'm afraid I don't fully understand this... Makoto you wrote that the
parts unaccounted for by the ICA would be spread through all components, so
if all components are subtracted, there should again be nothing left, or
where am I on the wrong track here? Also, the backprojection from
components to channels is just a matrix multiplication with the remaining
components' weights inversed, but if there's no component left, this would
be an empty matrix, or how would this work mathematically? I might have
gotten something wrong here, though...

Cheers,
Marius



2016-10-21 15:02 GMT+02:00 Tarik S Bel-Bahar <tarikbelbahar at gmail.com>:

> Hi Duncan, some quick notes about your questions below, hope they are of
> use. Best wishes, Tarik
>
> Mobile brain imaging (EEG in not just "passive and sitting-quietly"
> conditions) is still a developing field. You may benefit from contacting a
> few researchers who work in this area, including Klaus Gramman and Daniel
> Ferris, so as to garner their opinions. There are several other groups that
> should be findable via google. As you're at the frontier of this area,
> it's better to test a range of options and make the best informed
> decision, as there are few consolidated guidelines for this kind of
> research. There's a range of issues and options, just a few thoughts below
> for now.
>
> If you don't give some subset of data to ICA, it's of course not
> considering it. If you have several ICA decompositions for one person, then
> there should definitely be similarities across the decompositions. Thus,
> it's not clear that you would end up with, after subtractions, just, as you
> said, "  remaining signal consisting of all the activity that lacked in
> the passive conditions?". This expectation makes sense, but it's just not
> published about enough to be sure.
>
> And yes, generally, if you are trying to subtract ICs representing neck
> muscle activity from a dataset that has very little (or a lot of) activity
> like that, the basic assumptions about subtracting the ICs may not hold.
>
> In your case, consider yourself first in an exploratory stage. You are
> considering several important points, so that's good. The field needs to
> know more about mobile brain imaging problems and how to deal with them.
> Overall this is an excellent opportunity to make a contribution to real
> problems in the field.
>
> You should be able to run ICA on the non-passive conditions and get
> interpretable ICs that reflect real brain sources. However, they will
> likely be more mixed with artifact activity than the ICs you get from the
> passive condition.
>
> I would recommend running an ICA across all conditions per person, as well
> as a (separate) ICA for each condition, and review the results for yourself
> (from single-condition and all-condition ICAs). You may get more
> "prototypical artifact ICs that are applicable across the conditions. If
> you have "weak artifact" IC from one condition and apply it to a "receiving
> set"
> it should remove some of that artifact activity, but not necessarily most
> of it, especially if the IC information from the "donor set" does not
> well-characterize the artifact dynamics in the "receiving set".
>
> Another option for future consideration, is to ask people to make a  range
> of stereotyped movements with their bodies, faces, and arms, and use that
> as a kind of glossary of artifacts for a specific person.
>
> Another option is to compare several conditions, with each condition
> having "more artifacts". For example, really slow walking, normal walking,
> and fast walking.
>
> Some of your issues like cable sway and so on from movement may be
> resolved with wireless caps available from several manufacturers, though I
> would not recommend dry systems nor "non-research" consumer systems.
>
> You may also want to think about using ASR in eeglab to clean up the data
> beforehand. See an example of the effects of ASR on EEG with active
> movement here:
> https://www.youtube.com/watch?v=qYC_3SUxE-M
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
> _______________________________________________
> 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...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20161028/d82b3b76/attachment.html>