[Eeglablist] ICA and Connectivity On Combined Data From Multiple Subjects

Makoto Miyakoshi mmiyakoshi at ucsd.edu
Wed Apr 3 10:06:47 PDT 2019


Dear Nathan,

> It is good to know that a group ICA can be interpreted as a grand-mean
spatial filter.

It's not exactly identical to the (Riemanian?) grand-mean spatial filter
though. It can be 'interpreted' as such.

> I was worried that the discontinuities caused by individual differences
and cap placements would somehow confuse the ICA algorithm.

Yes it does confuse ICA. The individual difference is identical to
non-stationarity problem when applied to a single-person EEG, but
distinctively worse--because this is as if a brain (and a head and sensors
altogether) swaps before and after the boundary! This should be the
ultimate non-stationarity ha ha.

> I think the group ICA approach makes sense for my group (Human Factors
and Ergonomics) in which sample sizes are relatively low.

Using the low number of subjects does not justify the use of group ICA.
Instead, group ICA's ICA would find something like the greatest common
divider across everyone.

There is an easy way to find out what ICA finds here. Perform AMICA instead
of runica()'s infomax. AMICA results come with log likelihood plot (see
postAmicaUtility() plugin). If the single-model AMICA's log likelihood
shows good value only for one/subsets of the subjects, it means the ICA
result is dominated by a single/small subset of the subjects. If it shows
relatively good value across subjects (but you should expect to see some
level of fluctuation across subjects), then you can safely says that the
group ICA worked successfully.

This way, you can escape from the never-ending qualitative discussion of
the validity of group ICA. When we can answer a question by quantifying,
quantifying gives us the best answer!

%%%%%%%%%%%%%%%%%%%%%%%

> I have also been looking forward to the release of groupSIFT. I hope
development is proceeding well. Let me know if you could use another person
to help test it.

Thank you for your interest in groupSIFT. I know people are greatly
interested in it. I got stuck with some really stupid problems, mainly due
to Matlab GUI design instability across OSs! Basically groupSIFT works
perfectly, IF you use Matlab 2013a! If you want to try it out, let me know.

The first paper in which I USED groupSIFT is coming out soon from
NeuroImage Clinical, if the manuscript survives the second round of the
revision. But it is not a dedicated technical paper. Nor does the paper
have a decent Appendix that explains the technicality of the solution. The
algorithm is explained rather bluntly in the Supplementary Materials,
unfortunately. I'll find a chance to finish the project. Thank you for your
encouraging comment.

Makoto



On Wed, Apr 3, 2019 at 9:28 AM Nathan Sanders <nesander at ncsu.edu> wrote:

> Hey Makoto,
>
> Thanks for your kind response! It is good to know that a group ICA can be
> interpreted as a grand-mean spatial filter. I was worried that the
> discontinuities caused by individual differences and cap placements would
> somehow confuse the ICA algorithm. I think the group ICA approach makes
> sense for my group (Human Factors and Ergonomics) in which sample sizes are
> relatively low. I had been looking at a clustering approach like the one
> outlined in this post
> <https://sccn.ucsd.edu/pipermail/eeglablist/2016/011386.html>, and was
> sadly aware of the drawback you described. I have also been looking forward
> to the release of groupSIFT. I hope development is proceeding well. Let me
> know if you could use another person to help test it. By the way, you've
> done a great job with the SCCN Wiki and your preprocessing pipeline has
> been an invaluable resource.
>
> Thanks again,
>
> Nathan
>
> On Mon, Apr 1, 2019 at 8:59 PM Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> wrote:
>
>> Dear Nathan,
>>
>> Good question!
>> First of all, it is NOT wrong to concatenate all subject's data first
>> then run a single ICA: this approach is called group ICA. The assumption is
>> that ICA gives you a group's grand-mean spatial filter. Meanwhile, SCCN's
>> approach is to run ICA for each individual: The assumption is that the
>> resultant ICA weight matrix represents individual subject's spatial filter
>> i.e., functionally independent stationary cortical region to scalp sensors.
>> In a sense, SCCN's approach allows individual differences to be maximally
>> manifested. For comparison, consider Fz, Cz, and Pz monitoring some brain
>> electric activity. A spatial relation between an active brain region and a
>> scalp electrode varies from subject to subject, because of anatomical
>> variance across population. But ICA can scoop the brain region that emits a
>> unique signal than others. This is as if ICA would provide a tailor-made
>> stethoscope for each individual to monitor the active cortical region.
>>
>> However, even if you use an ICA for individual subjects, in the end
>> you'll need to summarize all the individual data into much lower
>> dimensional data that represents the group's representative values
>> (typically average). This dimension reduction is the whole purpose of the
>> group level analysis. SCCN's approach keeps individual differences all the
>> way down to the final clustering stage. Other approaches, like group ICA,
>> collapse individual differences in earlier stages. Again, these approaches
>> are equally valid, and not that one is good and the other is bad/wrong.
>>
>> This is related to a meta-level question of what we want to do with a
>> group-level analysis. One of my colleagues criticized the current SCCN's
>> approach for separating individual subject analysis and the group-level
>> analysis. His point is that if we know we will collapse all the subject's
>> data in the end, why don't we optimize individual data in terms of
>> group-level summary. I think such a criticism is valid. This idea will be
>> demonstrated soon.
>>
>> An even larger problem here is what is individual difference to us. The
>> current inferential statistics treats it as an error. However, geneticists
>> take it granted that there are always subtypes/subgroups in a group, so
>> individual differences are readily justified. If you follow SCCN
>> IC-clustering approach, there is a chance that you encounter an emerging
>> subtypes/subgroups in your result (if you have a large database), which
>> makes an interesting advantage of this approach. But, if you have rather
>> small (n=15 or less) data sets, then you may suffer from the individual
>> differences, since you may find only 60% of unique subjects in a given
>> cluster, for example...
>>
>> Makoto
>>
>>
>>
>>
>> On Mon, Apr 1, 2019 at 5:21 PM Nathan Sanders <nesander at ncsu.edu> wrote:
>>
>>> Hey Group,
>>>
>>> My understanding based on reading the SCCN wiki is that ICA should be
>>> performed on a per subject (per session) basis. For example, Makoto's
>>> preprocessing pipeline says that "To apply a single ICA, your subject
>>> cannot take off the electrode cap!
>>> <https://sccn.ucsd.edu/wiki/Makoto%27s_preprocessing_pipeline#General_tips_for_performing_ICA_.2806.2F26.2F2018_updated.29>".
>>> However, one approach to group connectivity analysis that I have run across
>>> involves first combining data sets from all subjects, then performing ICA
>>> on the aggregate data. This doesn't seem right to me, but I am having a
>>> hard time articulating why, mostly because I am new to the field and still
>>> only have a basic understanding of how ICA works. Can anyone speak to the
>>> validity of applying ICA to combined data sets? Can you recommend any
>>> papers to help me understand?
>>>
>>> Thanks,
>>>
>>> Nathan
>>> _______________________________________________
>>> 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
>>
>>
>>
>> --
>> Makoto Miyakoshi
>> Assistant Project Scientist, Swartz Center for Computational Neuroscience
>> Institute for Neural Computation, University of California San Diego
>>
>

-- 
Makoto Miyakoshi
Assistant Project Scientist, Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20190403/d7b32d54/attachment-0001.html>


More information about the eeglablist mailing list