[Eeglablist] Clustering components not all participants included

[Marius Klug]

Dear Martin,

this is a significant and known problem with IC clustering. Since the
decomposition is never the same for different subjects (or even the same
subject on a different recording), clusters will not contain a 1:1 mapping
of ICs. Another big problem is that the algorithm does not find the same
solution if clustering repeatedly, which is very bad for replication. I
played around with this problem for a good while and created an algorithm
for repeated clustering which provides the best solution for a particular
region of interest, based on several quality measures. Please refer to this
poster:

https://www.researchgate.net/publication/326690351_The_BeMoBIL_Pipeline_-_Facilitating_Mobile_BrainBody_Imaging_MoBI_Data_Analysis_in_MATLAB

You can find the code online on my github, but be aware that it is work in
progress and the usability is not yet very good.

https://github.com/MariusKlug/bemobil-pipeline

The functions of interest for you are: bemobil_precluster() and
bemobil_repeated_clustering_and_evaluation(),
the help function should guide you through the process. You can define a
ROI in the somatosensory cortex in talairach coordinates and try it out. I
recommend no less than 100 repetitions. You will get the best 5 solutions
plotted in the end, if they look very similar the algorithm converged and
you can trust that somebody will likely be able to replicate your results
using the same weight values with the same data.

The algorithm was applied in this paper:
https://www.biorxiv.org/content/early/2018/09/14/417972
For creating the study I used the first 70 ICs of each subject (which is
about the better half of all ICs) and relied on the negative weight for
residual variance in my clustering. You find all relevant values in method
section of the paper.

I have also tried the measure projection toolbox, but was not able to
replicate their own results with their example dataset and then dropped the
project again in favor of the repeated clustering. If you can use measure
projection effectively please drop me a line!

I'm happy to help with questions and hope this helps you in your research!
Marius

Am Sa., 20. Okt. 2018 um 23:11 Uhr schrieb Martin Simoneau <
Martin.Simoneau at kin.ulaval.ca>:

> Dear EEGlab experts,
>
>
>
> I have clustered the components of all the participants. Let says that I
> am targeting components located in the somatosensory cortex. Lucky enough,
> one cluster mean is in the somatosensory cortex. However, not all
> participants have a component in this cluster. What do I do with the
> participants that do not have a component in this cluster? If there is any
> paper (documentation…) related to this issue, please let me know as I will
> be happy to read it.
>
>
>
> When performing K-mean clustering, is there a minimum or a maximum number
> of clusters that one should select? Again, if there is any paper
> (documentation…) related to this topic, please let me know as could not
> find any.
>
>
>
> Finally, does the measure projection toolbox solved this problem of having
> a cluster with component from some participant not included?
>
>
>
> Thanks for your help,
>
>
>
> Martin
>
>
>
>
> _______________________________________________
> 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



-- 
Marius Klug
Research Associate / PhD Student at TU Berlin, Germany
Department of Biological Psychology and Neuroergonomics
+49 (0)30 314-79 514
bemobil.bpn.tu-berlin.de
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20181025/2af3724e/attachment.html>