[FieldTrip] calculate DICS Beamformer on a group level?

[christian.merkel at med.ovgu.de]

Hello,


First of all, thanks for the in-depth input.

- Regarding the double-dipping of the pca-approach: The statistical test is a between-subject comparison. So I use the pca in order to identify the source for which i want to test whether its timecourse differs between two groups of subjects. So I average ALL subjects for one particular frequency, select the source and use the weights on each single subject to calculate the amplitude differences over time between groups. So the H0 is formulated in a way that double-dipping wouldn't be the issue.


- So but why do I want now those results in source space? Although I identified the source with my pca already I want to show that there are no artifacts explaining the effect. The frequency-range in question is a high-gamma range. Its source is bilateral occipital for a specific motion stimulus. I had lots of discussion about this effect and as you can imagine the most critical point are microsaccades potentially interfering. I applied initial ICA-algorithms of the raw datasets + p2p-artefact rejection. As one last step to show that this are no microsaccades I just wanted a describtive source distribution showing the bilateral occipital sources without any frontal contamination. Beamformer filters apparently inherently remove those kind of high-frequency artefacts (Hipp/Siegel, 2013).


Thank you again,


Christian



________________________________
Von: fieldtrip <fieldtrip-bounces at science.ru.nl> im Auftrag von Stephen Whitmarsh <stephen.whitmarsh at gmail.com>
Gesendet: Montag, 23. November 2020 14:40:09
An: FieldTrip discussion list
Betreff: Re: [FieldTrip] calculate DICS Beamformer on a group level?

Dear Christian,

Here are my 2 cents;

If I understand you correctly (*) you are asking whether you can beamform your PCA components? That does seem a redundant mix of source-localization techniques to me (of which PCA is one). Technically I would be at a loss explaining why I wouldn't expect this, but conceptually I would say that PCA already took care of the source-extraction, in this case a time-course + sensor-topography. That time-course is one 'side' of your source (the temporal part), and belongs to that topography - the other (spatial) side of that coin. Which is to say, you already did a spatial source analysis by manually selecting the spatial topography.

You could, for reasons of validation perhaps, beamform the power in a particular frequency / time period (the 'traditional way'), and see if it is also able to separate your source spatio-temporally in a similar way, but that would be a qualitative comparison between source-localization techniques.

Anyway, that's probably what you were thinking about anyway :)

Concerning the beamformer specifically; you do not beamform an average. The whole point is to use subject-specific headmodels (based on individual MRIs) and individual subject's CSDs to be able to estimate power on a gridpoint-by-gridpoint basis. This then also allows you to better 'average' or 'normalize' those individual brains and their activity-topographies, namely based on their individual headshape and headmodel, than a sensor-level averaging would do, given the undetermined relationship between brain and MEG helmet in the latter. However, your PCA method already got rid of this problem by removing the spatial part (the manual selection of the component).

So, you could do a 'traditional' DICS beamformer, but because you already done your statistics on your sources, that might be only for some qualitative source-location validation that might not be worth the effort?

Anyway, that's all the change I have in my pocket :)

Cheers,
Stephen

*) If you are doing your statistical test on the same effect as the one you used to select the PCA component, you are of course double-dipping.













Op ma 23 nov. 2020 om 13:48 schreef christian.merkel at med.ovgu.de<mailto:christian.merkel at med.ovgu.de> <christian.merkel at med.ovgu.de<mailto:christian.merkel at med.ovgu.de>>:
Hello,


I hope some of you can give me some advice regarding a more conceptual problem with localizing specific oscillations within MEG data:

I wanted to use the DICS-approach to localize a specific spectral effect in my MEG-data within the source space. Now I don't know whether this can be done on an averaged dataset across multiple subjects, since in this case the cross-spectral-density matrices of all the subjects would have to be averaged as well and I am not sure whether such an averaged CSD-matrix still makes sense.

In more detail, I have spectral timeseries (complex) of several seconds within one specific frequency-band (sens x time). I calculate the grand average across time to get the evolution of the topography of this effect over time. Additionally, I ran a PCA on that averaged timecourse to extract the specific source (component) in sensorspace that I am interested in. Then I used the resulting unmixing-matrix to multiply that components weights with each subjects timecourses (sens x time). Thus I have one overall topography of that effect in sensorspace and its timecourse + variance (which I can run statistics on). Now I thought I can also look for the same effect in sourcespace but I actually just have the topography without the CSD-Matrix. And even if one could average the single-subjects CSDs, how would I retain the CSD for a specific component after a pca?


Thanks for any help,


Christian M.

_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202