MVPA analyses on EEG activity / source

Discussions
,
1

Hi everyone,
I tested MVPA analyses in brainstorm, I have several questions:

-I used the Matlab SVM classifier to distinguish two conditions. As a result, I get a decoding curve across time. However, I’m not able to see how the results differ across electrodes. Is it possible to have the classifier run on every electrode ? Does it runs only on an average of all electrodes ?

-I tried to do the same with sources but the decoding option stays in gray. Is it possible to run the classifier in the source space ?

Thanks for your help.

Thomas

2

Hi Thomas,

MVPA stands for MultiVariate (or Multi-Voxel) pattern analysis, it needs the values across all the electrodes (the “pattern”).
So no, you cannot have a MVPA classifier running on a single electrode.

This decoding process was not implemented in source space. In theory it could work on the 15000 source time series instead of the few dozens of electrodes, but in practice it is not very useful.
What we try to do here is to determine if the input data contains enough information to discriminate two experimental conditions. And all the data is already contained in the recordings. The projection in source space helps understanding and explaining the recordings but doesn’t add any information, it smooths a bit the recordings so it would even tend to destroy information. If the classifier cannot discriminate between two conditions in sensor space, it won’t be able to do it in source space.
The main interest of these MVPA approaches is to run blindly on the recordings: it’s fast and efficient.

Cheers,
Francois

3

Hi François,
Thanks for these info, this is clearer now.

Cheers,
Thomas

4

Hi Francois,

I’m employing Brainstorm’s Decoding for two conditions (almost 700 trials each). Using Matlab’s SVM, 10 permutations and 10 bins.

My question is how can I test for statistical significance (i.e. get the p values and correct for multiple comparisons), and what is the best approach? Is FDR ok?

Thanks

5

Hello,
I transferred your question to experts in this field. They will get back to you, but with maybe a delay due to holidays. Don’t hesitate to post again your question if you don’t get any answer by mid-august.
Francois

6

Thanks Francois.

Cheers,
Sebastian

7

Hi Francois,

I’ve been able to solve the MCP issue.

I would like to ask you for some references regarding decoding in sensor space vs. decoding in source space.
For instance, a reference regarding the fact that projections in source space may destroy information relevant for decoding.

Thanks
Sebastian

8

Hi Sebastian,

I think decoding should be used on the most unprocessed data you have access to, so that you keep all the features of the signals that may help the classifiers. So decoding in sensor space seems a better idea than decoding in source space.

You can refer to the publications at the end of the decoding tutorial:
http://neuroimage.usc.edu/brainstorm/Tutorials/Decoding#References

Francois