A question about time-frequency (Morlet wavelets )

SBeumer · March 31, 2020, 12:03pm

Hi Neuro,

Do you want to know it at any time point or just over the whole signal in general. In the second case wouldn't it be better to do a frequency analysis and leave the time out?

In principle this would work though yes

Kind regards,
Steven

neuro151 · March 31, 2020, 12:16pm

Hi SBeumer,
Thank you for your answer. Actually, I want to know it at a particular time period in whole time process (such as, 0--500ms in -500--1000ms). If I want to leave time out and do a frequency analysis, is the step similar with time-frequency?

SBeumer · March 31, 2020, 12:36pm

Hi Neuro,

The steps would be more or less the same, you could window your frequency analysis to achieve what you would like. Though I do not have a lot of experience with this myself. So it might be that someone else chimes in with a better answer.

Kind regards,
Steven

neuro151 · March 31, 2020, 12:54pm

Thanks a lot.

neuro151 · April 1, 2020, 11:11am

Hi Steven,
I have a question when I followed the previous steps to calculate the power ratio of one band to all bands, in my understanding, all ratios in the TF matrix after spetrum normalization should add up to 1, but it’s not. So, I'm a little confused about how to calculate the ratio of one band to the whole band from the result of spectrum normalization. Should I calculate it directly from the results of the time-frequency? For example, add all the values of theta band in the matrix and divided by the sum of all the values in the TF matrix.

SBeumer · April 1, 2020, 11:31am

Hi Neuro,

Can you be a bit more specific in your calculations and to what it adds up etc.? I'm not an expert on doing this calculations so maybe someone else can chime in.

By the way you could always calculate the powers and then in Matlab sum up to the total power and divide the power of one band by that total power. There are multiple options

Kind regards,
Steven

neuro151 · April 1, 2020, 12:05pm

Hi Steven,
I did the same things that you mentioned below，sum and division based on the results after spectrum normalization. Maybe I should calculate it based on the results of the time-frequency？What's the difference between the ratio that calculated by the spectrum normalization and the ratio that calculated by the method you mentioned below(calculate the powers and then in Matlab sum up to the total power and divide the power of one band by that total power).

SBeumer · April 1, 2020, 12:20pm

Hi,

I'm not a brainstorm developer or associated to the team. I'm at the moment a PhD student with a bit more time due to the quarantaine trying to help . I guess you could check the source code but I assume internally it does more or less the same.

Kind regards,
Steven

neuro151 · April 1, 2020, 12:31pm

Thank you all the same.

neuro151 · April 1, 2020, 12:45pm

Hi Francois,
I want to how to calculate the power ratio of one band to all bands, I followed the step suggested by SBeumer, That is, compute spectrum normalization, and get the results. Then I export it to matlab, get the TF matrix, I want to know how to calculate the ratio of one band to the whole band from the result of spectrum normalization.
Maybe I should calculate it based on the results of the time-frequency？Just sum up to the total power and divide the power of one band by that total power?

Francois · April 1, 2020, 2:28pm

Actually, I want to know it at a particular time period in whole time process (such as, 0--500ms in -500--1000ms). If I want to leave time out and do a frequency analysis, is the step similar with time-frequency?

If you want an average value for all the time definition: use the PSD process instead of a time-frequency approach.
If you want the average of a shorter time window, compute the TF of the whole epoch (-500,1000ms), then compute the average over the period of interest (eg. 0-500ms) with the process "Average > Average time" or "Extract > Extract values" (with the averaging option selected).

Maybe I should calculate it based on the results of the time-frequency？Just sum up to the total power and divide the power of one band by that total power?

This is exactly what the process "Standardize > Spectrum normalization" with the option "Relative power" does:

github.com

brainstorm-tools/brainstorm3/blob/master/toolbox/process/functions/process_tf_norm.m#L132


            % Reshape to have the scaling values in the third dimension
            Factor = reshape(Factor, 1, 1, []);
        case 'relative'
            % If measure is not power/magnitude/log
            if ~ismember(Measure, {'power', 'magnitude'})
                errorMsg = ['Values with measure "' Measure '" cannot be normalized with this process.'];
                TF = [];
                return;
            end
            % Divide by the total power
            Factor = 1 ./ sum(TF,3);
        otherwise
            errorMsg = ['Invalid normalization method: ' Method];
            TF = [];
            return;
    end
    % Apply multiplication factor
    if ~isempty(Factor)
        TF = bst_bsxfun(@times, TF, Factor);
    end
end

If you want to apply another measure, you can do it either with the process "Pre-process > Run Matlab command" or by exporting the files to Matlab, processing them the way you want, and re-importing the modified values:
https://neuroimage.usc.edu/brainstorm/Tutorials/Scripting#Custom_processing

neuro151 · April 28, 2020, 2:55pm

Hello, Francois,
I want to calculate temporal-spectral evolution. Is there any processing step similar to the temporal-spectral evolution（TSE） of BESA in brainstorm?

Francois · April 29, 2020, 9:03am

The time-frequency methods available in Brainstorm are described in this tutorial:
https://neuroimage.usc.edu/brainstorm/Tutorials/TimeFrequency

If you install FieldTrip (or use the compiled version of Brainstorm), you would also be able to use the FieldTrip multitaper approach (function ft_mtmconvol):
http://www.fieldtriptoolbox.org/tutorial/timefrequencyanalysis

I'm not sure how this relates to the TSE method available in BESA, you might need to read the methods articles published about this method.
@Sylvain @hossein27en @leahy Any suggestion?

Sylvain · April 29, 2020, 1:07pm

TSE is a broad term for tracking frequency-specific changes in signal power or amplitude over time. This is topically equivalent to the time-frequency decompositions featured by Brainstorm.

neuro151 · April 29, 2020, 2:24pm

Thank you, Sylvain. If I want to do TSE in brainstorm, what exactly should I do? Are there any tutorials on time-frequency decompositions? If so, could you please send me the link?

kind regards

Francois · April 29, 2020, 2:50pm

Time-frequency tutorial:
https://neuroimage.usc.edu/brainstorm/Tutorials/TimeFrequency

Francois · April 29, 2020, 2:51pm

If you haven't done this yet, we recommend you also read all the other introduction tutorials in the section "Get started":
https://neuroimage.usc.edu/brainstorm/Tutorials

neuro151 · April 30, 2020, 1:43am

Thank you for your help.