Computing Scout Statistics after Time-Frequency Decomposition

Discussions
1

Hi Brainstorm Community,

I have a question regarding Time-Frequency decompositions and scouts. After calculating T-F decompositions (as in the resting omega tutorials), I get the power for different frequency bands. I would like to test for differences between groups (patient vs controls). I would like to use the scout function to test my hypothesis, which is something like this: in the orbitofrontal cortex (this ROI is an example. I will be testing multiple ROIs), patients have higher power in the alpha frequency band, when compared to controls.)

With that in mind, I've tried several things in the past to no avail. Here is a list of failed methods:

Method 1: I followed the resting omega tutorial and got my TF decompositions on the default anatomy (14 controls and 15 patients). I then computed the average for both groups, and computed a group difference (Test -> Difference (A-B) ). I got a group difference, but then wanted to test for significant regions of interest. I dragged all TF files per group to Process2 tab and ran a Patient-control parametric student's t-test with the following parameters:

t-test parameters.png426×700 23.1 KB

I got the following message (ignore the initial files. I got the same error when I used my data as well):

testing power values.png889×690 23.5 KB

I followed option two and extracted the measure (Magnitude) from complex values, and tried running the tests again, but this time Matlab gives the following message:

BST> Average corrected p-threshold: 0.0833333 (FDR, Ntests=6)
BST> All values are null. Please check your input file.

I should note that there isn't any indication of activity in the file produced. I went through the tutorial for Statistics, but they are mostly for ERPs (and not resting state) since they deal with time.

Method 2: I followed the Omega tutorial, but rather than computing brain-wide analysis, I decided to focus on ROIs. I selected 2 scouts on the desikan killiany atlas with the following options:

I immediately got the following error message:

project sources error.png890×692 20.3 KB

There were files computed in each subjects folder before the process was terminated, but the files were not projected to the default anatomy.

subject TF decomp.png702×594 30.7 KB

My question is would I still be able to compute Stat tests on the subject without using the standardized anatomy, or are Brainstorm's pre-defined scouts comparable between subjects (thereby negating the need to project to the default anatomy?)

Just to check what might happen if I ran a t-test on the files, I got the same error:

You are testing power values, while a more standard analysis is to test the magnitude (ie. sqrt(power)). Option #1: Recompute the time-frequency maps using the option "Measure: Magnitude". Option #2: Run the process "Extract > Measure from complex values", with option "Magntiude".

I have tried other methods, but for the sake of brevity and time, I'll forgo typing all of them here.

Thank you very much for your help

Best,

Aquila.

2

Hi Aquila,

I dragged all TF files per group to Process2 tab and ran a Patient-control parametric student's t-test with the following parameters:

The parametric t-test here is not adapted for testing power values. But you can use a non-parametric test instead, this is always valid.

BST> Average corrected p-threshold: 0.0833333 (FDR, Ntests=6)
BST> All values are null. Please check your input file.

Nothing wrong here, you simply have nothing significant. If you have only 6 subjects, there is no point in trying to get any significance measure, you should simply work with the average. You need a lot more sample to be able to estimate a robust average and standard deviation in your group.

Method 2: I followed the Omega tutorial, but rather than computing brain-wide analysis, I decided to focus on ROIs. I selected 2 scouts on the desikan killiany atlas with the following options: - I immediately got the following error message:

You can consider that the scout measures you obtain are already similar measures across subjects. You don't need to project them on a template (and you can't, as you noticed). Just run your test on the subject measures.
An alternative (and heavier) approach is to project the full sources maps on the template and then extract the scouts measures with the process Extract > Scouts time series.

Cheers,
Francois

3

Hi Francois,

Thank you for your quick response.

Using the non-parametric Fieldtrip sourcestatistics test on the TF files calculated using the method from the Omega tutorial, here is the result from the Matlab window:

cluster based non parametric test.png2959×994 529 KB

Matlab Window.png1087×893 44 KB

The same applies when I try with the permutation tests. I really cannot tell what I'm doing wrong because logically (in my head at least), this is what I'm trying to accomplish: compute the TF decomposition by frequency bands, and test for statistically significant differences between groups for each frequency. I can compute a difference, no problem, but I cannot compute any stats using the files I have. I would very much appreciate it if you could shed some light, because I might have tunnel vision since I've been working for so long on this dataset.

I get the same results even though I use all 29 subjects, as opposed to those 6 sample subjects I used to test out the protocol. I don't think it's a problem of sample size (although I do have a small sample size of 29 subjects, 14 controls and 15 patients), but rather a problem of the wrong input files for the Statistical commands.

I understand that the scouts are similar across subjects. However, I'm not certain how I'd be able to test my hypothesis with extracting the time series. As a reminder, I would like to use the scout function to test my hypothesis, which is something like this: in the orbitofrontal cortex (this ROI is an example. I will be testing multiple ROIs), patients have higher power in the alpha frequency band, when compared to controls. Based on what you suggested, my guess would be to compute the sources, project to default anatomy, extract the scout time series, and get an image like this:

Extract Scout time series.png1595×943 110 KB

Then you compute the PSD on the extracted scout series, and compute the stats on the PSD files?

I'm not sure if there's a bug in the program, but extracting the Magnitude from the complex values does not make a difference when I test the extracted files.

On a related note, how would I extract the power values for the scout so I can correlate with behavioural measures? For example, I'd like to see if there's a correlation between an increase or decrease in behavioral scores and an increase or decrease in scout power for a certain frequency.

As usual, Francois, thank you very much for taking the time to respond to my inquiries.

I appreciate it.

Best,

Aquila.

4

Using the non-parametric Fieldtrip sourcestatistics test on the TF files calculated using the method from the Omega tutorial

Start working with only the process Test > Permutation test: Independent, Files A=15 files (one for each patient), Files B=14 files (one for each control subject).

The same applies when I try with the permutation tests.

What do you mean by "the same"? The is no problem of significant clusters if you don't use clusters.
When you display the results of the test, you can edit the p-value threshold used to assess significance. Set it the threshold to something high (0.9 or 1), with no correction for multiple comparison, and you will see the value of the t-statistic for all the vertices. Play with this threshold to explore the significance of your results.
Most likely, you just don't have any significant effect between your two groups for thee measure you computed. Maybe the hypothesis or the processing pipeline are to blame instead of the statistical test.

I get the same results even though I use all 29 subjects, as opposed to those 6 sample subjects I used to test out the protocol. I don't think it's a problem of sample size

29 is a small sample size.

Based on what you suggested, my guess would be to compute the sources, project to default anatomy, extract the scout time series

Do not project to the template. Select directly the source files in the Process1 box, run the process "Power spectrum density" with the option "Use scouts" selected.

I'm not sure if there's a bug in the program, but extracting the Magnitude from the complex values does not make a difference when I test the extracted files.

It should not make much difference with non-parametric statistics. But it would with parametric statistics.
I recently removed this warning about using the magnitude vs the power when using non-parametric tests.

On a related note, how would I extract the power values for the scout so I can correlate with behavioural measures?

There are no tools for this in Brainstorm yet. You would need to read the values directly from the files in a script and compute the correlations yourself.

5

Hi Francois, Aquila,

I am attempting to do something similar with my resting state EEG data. Our research question is to see where in the brain are there differences in alpha power between patients and controls (specific ROI is in the occipital cortex). I have also followed through the OMEGA tutorial, but used OpenMEEG to compute my head models and used sLORETA to compute my sources. I then executed the proposed pipeline as in the OMEGA tutorial to produce power maps for each subject, which I then averaged for each group. I was successfully able to produce the average power maps as displayed in the tutorial.
I would now like to run a statistical analysis to compare the power in the alpha and theta frequency bands between my control group and my patient group.

Thus far I have tried the following:
i) compute group difference (Test --> Difference (A-B) ) using my intra-subject Avg: Raw| Band | notch files
– I can view the new file this produces (Avg: raw| band | notch (28) - Avg: | raw | band | notch (52) ) and observe the power distribution projected on to the cortex, but how I do I interpret these result statistically?

ii) I have run a non-parametric Fieldtrip sourcestatistics test with all control subjects in File A, and all patient subjects in File B. The output found 13 positive clusters and 11 negative clusters. What do I do from here to observe statistical differences in power according to the frequency bands I am interested in?

Any help would be much appreciated.

6

i) compute group difference (Test --> Difference (A-B) ) using my intra-subject Avg: Raw| Band | notch files
– I can view the new file this produces (Avg: raw| band | notch (28) - Avg: | raw | band | notch (52) ) and observe the power distribution projected on to the cortex, but how I do I interpret these result statistically?

This is a difference of average, you won't get a p-value from this. This might give you indications on what is going on in your data, but won't tell you what is significant.

ii) I have run a non-parametric Fieldtrip sourcestatistics test with all control subjects in File A, and all patient subjects in File B. The output found 13 positive clusters and 11 negative clusters. What do I do from here to observe statistical differences in power according to the frequency bands I am interested in?

I'm not sure I understand your question. It seems to me that you already have your answer.
I would start with a permutation test for getting actual thresholded maps that might be easier to interpret.
https://neuroimage.usc.edu/brainstorm/Tutorials/Statistics#Nonparametric_permutation_tests