Source localizing theta oscillations: normalization?

Discussions
1

Hi Francois,

I hope you help me with a normalization issue.

The goal of my study is to source localize theta activity (resulting from a feedback stimulus) in EEG data. The time-frequency analyses were already performed previously in Vision Analyzer, but I want to use Brainstorm to compute the sources of the theta oscillations. I have four conditions and the epochs from these conditions have been successfully imported and baseline corrected in Brainstorm. I followed tutorial 12 with this respect and created the following pipeline:

  • Compute head Model (BEM)
  • Compute Noise Variance
  • Compute sources on single trial epochs (wMEM)
  • Average sources
  • Extract time (300-500 ms post stim) {!!! here I also compared maps from non-normalized vs z-score normalized sources}
  • Run time-frequency analyses (normally I run tf analyses on much longer epochs, is this correct to run time-freq on such short time frames?)

This pipeline (without any z-normalization) works fine, and seems to correspond with the steps in tutorial 12.

However, I believe somewhere in this pipeline I should perform a normalization method, and the problem is that I can’t perform time-frequency analyses on the z-score normalized source maps.

When comparing the source maps (see step between brackets: non-normalized vs normalized) I noticed a huge difference in source activation patterns, with the normalized maps showing activation that is very much in line with what I would expect. Thus, my guess is that I should use the data from the normalized sources, but how can I obtain the time-frequency representation of these normalized sources?

I hope you can help!

Cheers,
Melle

2

Hello,

Compute sources on single trial epochs (wMEM)

I think the wMEM is already localized in frequency: http://neuroimage.usc.edu/brainstorm/Tutorials/TutBEst
This inverse method is not linear and does not preserve the frequency contents of the signals, after running it you should not apply any additional frequency analysis. For frequency analysis, I would recommend you use non-normalized minimum norm maps (wMNE).

Extract time (300-500 ms post stim) {!!! here I also compared maps from non-normalized vs z-score normalized sources}

You don't need to normalize your maps for TF analysis.

Run time-frequency analyses (normally I run tf analyses on much longer epochs, is this correct to run time-freq on such short time frames?)

This is definitely too short. Take longer epochs, run the TF, then extract time blocks if needed.
If you click on the option box "Hide edge effects", you will see what are the values that could not be estimated correctly.

This pipeline (without any z-normalization) works fine, and seems to correspond with the steps in tutorial 12.
However, I believe somewhere in this pipeline I should perform a normalization method, and the problem is that I can't perform time-frequency analyses on the z-score normalized source maps.

Tutorial 12 is a bit outdated. We are in the process of writing a new set of tutorials but haven't reached the time-frequency chapter yet.
http://neuroimage.usc.edu/brainstorm/TutorialsNew
You should normalize your time-frequency maps indeed (not the source maps), in order to correct for the ~1/f decrease of the signal power. For this, you can run two processes on your time-frequency maps: Z-score or Event-related perturbation.

When comparing the source maps (see step between brackets: non-normalized vs normalized) I noticed a huge difference in source activation patterns, with the normalized maps showing activation that is very much in line with what I would expect. Thus, my guess is that I should use the data from the normalized sources, but how can I obtain the time-frequency representation of these normalized sources?

The normalization of the source maps just change the relative values of the signal at one vertex vs. the signal at its neighbors, but it doesn't change the signal itself. What you want to do is the following:

  • For displaying source maps: display Z-score(wMNE)
  • For TF decomposition of the source maps: Z-Score(TF(wMNE)) or ERDS(TF(wMNE))

Cheers,
Francois

3

Hi Francois,

Thanks for the quick reply.

I was indeed planning on calculating the sources with wMNE. Thanks for pointing towards the new tutorials. I’ll check them out. It seems to be working already :slight_smile:

A final question: I know the BEST toolbox is an external one, but do you know whether it is possible to run this method on single trials (since I am only able to get this working on averages of (a subset of) trials.

I would love to compare the results of BEST and wMNE.

Thanks!
Melle

4

I forwarded your message to the MEM developers.

5

Hi Melle,

The wMEM can be used to localize single-trial data, in fact it implements wavelet-based denoising, which improves the accuracy of source estimates.
As far as the code is concerned, there is no difference in localizing single-trials or averages, so the toolbox should run fine on single trials if you can already localize averages.

Best!

Younes

6

Hi younes,

Indeed, it works fine on single trials.

Cheers,
Melle