[FieldTrip] PLV, wPLI: Discripancy using powandcsd and fourier
Schoffelen, J.M. (Jan Mathijs)
jan.schoffelen at donders.ru.nl
Tue Aug 4 08:08:34 CEST 2020
Hi Akshay,
Please check the consistency of the order of the channels in your data. It is notoriously difficult to keep this consistent, and in general it is not recommended to rely on the fact that this will not change (sometimes by accident, for instance due to accidental alphabetical ordering during bookkeeping operations (e.g. matlab’s intersect/unique etc. may return alphabetically ordered lists). When converting to ‘cmbrepresentation’ ‘full’ with ft_checkdata, there was no way for the code to know whether there is a preferred ordering of the channels, since the input data structure in this case only contains ‘labelcmb’, and no ‘label’. For this reason, I suspect that that the order of the channels in the output is in alphabetical order.
You will see that the different data structures that you consider to be different numerically, have their channels ordered differently. If you’d compare the corresponding row/column pairs for the different data structures, you’d see that there’s no issue.
Best wishes,
Jan-Mathijs
On 4 Aug 2020, at 00:30, Akshay Ravindran <akshay.s.ravindran at gmail.com<mailto:akshay.s.ravindran at gmail.com>> wrote:
Thank you, yes so I do get the same figures when I plot the two using ft_connectivityplot while feeding the EEG data. However, my intent is to get the actual numerical PLV values rather than the plot so I believe I should be calling plv3 = ft_checkdata(plv3, 'cmbrepresentation', 'full'); to get the channel x channel x frequency representation. The discrepancy happens when I evaluate the numerical values using this method.
I did some exploration as I understand it is very difficult to explain the anomaly considering that as you said it should not matter irrespective of whether the data being fed in is EEG or random numbers as the function just sees them as numbers. Upon following the lines of code for ft_connectivityplot I came to know that inside the function there are instances wherein it first adjusts the data using ft_checkdata and also does some operations on intersecting the channels to make them align. And I suspected there are some internal corrections being done to align the values and suspected it had to do with the naming of the labels.
So, when I run your script with the following additional lines I do get the same answers in plv3 and plv1 printed out ( changed nchan = 5).
plv3 = ft_checkdata(plv3, 'cmbrepresentation', 'full');
disp(plv1.plvspctrm(1,:,1));
>>1.0000 0.1000 0.1000 0.5000 0.3000
disp( plv_csd.plvspctrm(1,:,1));
>>NaN 0.1000 0.1000 0.5000 0.3000
Then I replaced the following lines accordingly and reran the same set of code
for k = 1:nchan
if k==1
data.label{k} = sprintf('chan%02d',k);
elseif k==2
data.label{k} = sprintf('Nchan%02d',k);
else
data.label{k} = sprintf('N1chan%02d',k);
end
end
...
...
...
Here, the output is different for both powandcsd as well as fourier when the label size is different.
disp(plv1.plvspctrm(1,:,1));
>> 1.0000 0.1000 0.1000 0.2000 0.3000
disp( plv_csd.plvspctrm(1,:,1));
>> NaN 0.1000 0 0 0.1000
So by default, the EEG labels are not having the same size, some are two-letter names(Cz) whereas others are 3 letter names (FCz). So there is an influence on the channel labels on how the results come about and I suspect that is causing this anomaly.
Now this is with my understanding that I am supposed to call plv3 = ft_checkdata(plv3, 'cmbrepresentation', 'full'); after powandcsd to get the data in correct dimension (ch x ch x freq) and I might be wrong if this is not the way to extract the values in the format. If there is another way to get the data, if you could kindly inform me, that will be greatly appreciated.
On Mon, Aug 3, 2020 at 11:18 AM Schoffelen, J.M. (Jan Mathijs) <jan.schoffelen at donders.ru.nl<mailto:jan.schoffelen at donders.ru.nl>> wrote:
Hi Akshay,
I cannot reproduce your problem. A data structure is a data structure, so I don’t see why a FT-structure that some numbers in a trial cell-array should behave different when it is obtained from EEGLAB or from ft_connectivitysimulation. For that matter, it shouldn’t make a difference whether you just put in random numbers. Using the below snippet of code I can reproduce your observation that it shouldn’t matter whether plv is computed from singletrial powandcsd, or from singletrial fourier.
Best wishes,
JM
% create some data
nsmp = 1000;
ntrl = 20;
nchan = 3;
for k = 1:ntrl
trial{1,k} = randn(nchan,nsmp);
time{1,k} = (0:(nsmp-1))./1000;
end
data = [];
data.trial = trial;
data.time = time;
for k = 1:nchan
data.label{k} = sprintf('chan%02d',k);
end
% spectral transformation
cfg = [];
cfg.method = 'mtmfft';
cfg.output = 'fourier';
cfg.pad = 1;
cfg.tapsmofrq = 1;
freq1 = ft_freqanalysis(cfg, data);
cfg.output = 'powandcsd';
freq2 = ft_freqanalysis(cfg, data);
cfg.keeptrials = 'yes';
freq3 = ft_freqanalysis(cfg, data);
% connectivity estimation
cfg = [];
cfg.method = 'plv';
plv1 = ft_connectivityanalysis(cfg, freq1);
plv2 = ft_connectivityanalysis(cfg, freq2); % this one is not correct, since wrong normalisation
plv3 = ft_connectivityanalysis(cfg, freq3);
% these should be the same
cfg = [];
cfg.parameter = 'plvspctrm';
ft_connectivityplot(cfg, plv1, plv3);
On 3 Aug 2020, at 15:38, Akshay Ravindran <akshay.s.ravindran at gmail.com<mailto:akshay.s.ravindran at gmail.com>> wrote:
Hi Jan,
Thanks for the explanation. However yes so I do do understand the normalization done in PLV and for the same reason I had done keep_trials='yes' for powandcsd. I forgot to write it in the line of code below. I was getting that difference when using powandcsd though.
The Nan and 1, I understood came from the way the implementation was done but my concern was the discrepancy in the off diagonal element which was the same for the artificial data whereas for actual EEG, they came out to be quite different. Is there something I can do to make it replicable? Maybe can I send the data?
On Mon, Aug 3, 2020, 4:03 AM Schoffelen, J.M. (Jan Mathijs) <jan.schoffelen at donders.ru.nl<mailto:jan.schoffelen at donders.ru.nl>> wrote:
Hi Akshay,
I understand your question, but honestly the pasted code is not sufficient for me to try and reproduce. Nonetheless, a few things can be said:
PLV is a measure (as is wPLI) that requires single trial spectrally resolved data as an input. This is because the algorithmic steps that are needed to compute the connectivity measure include a per-trial normalisation of the fourier coefficients by their magnitude (plv), or some other fancy non-linear manipulation of the single trial phase difference estimates (wPLI). If you ask ft_freqanalysis to output ‘powandcsd’, without explicitly specifying cfg.keeptrials = ‘yes’, then the single trial dimension is lost for posterity, and the consequent computation of the PLV/wPLI is nonsense (unless you used as an input a single epoch’s spectrally transformed data with massive multitapering, keeping the information of the individual tapers, but that’s a different story altogether). In other words, where you mention that you think that ‘fourier’ is the right way to implement these, you are right, but as an alternative you could have used ‘powandcsd’ + ‘keeptrials = ‘yes’’.
The NaNs versus 1 are easily explained by the fact that with ‘powandcsd’ the diagonal elements are explicitly set to NaN (i.e. they are not computed), whereas in the ‘fourier’ case they are, and they trivially result in an all(ones).
Best wishes,
Jan-Mathijs
On 2 Aug 2020, at 23:35, Akshay Ravindran <akshay.s.ravindran at gmail.com<mailto:akshay.s.ravindran at gmail.com>> wrote:
Good Evening,
I tried estimating the functional connectivity using different phase-based measures like PLV and wPLI_debiased. I tried testing the two measures using cfg.output set to either 'fourier' or 'powandcsd' as I have seen people implementing PLV with both. However, I am getting quite varying results using either of these on an EEG data saved in .set file loaded and run using the commands shown at the end of this question.
Even though I think 'fourier' is the right way to implement these, to test for the validity of it, I tested this on the artificial data created using the ft_connectivitysimulation function ( 3 channel, 200 Hz Fs, the same number of trials as EEG, default noise covariance and params as in the example). In this case however, I am getting the very same results using either output (off-diagonal elements are the same in either; diagonal elements are NAN in using powandcsd whereas 1 while using Fourier).
My question is mainly the following
1) Am I missing something while testing on EEG data? If not, why is there a discrepancy found while testing on EEG but not on the simulated data?
2) Is there any recommended instance of using 'powandcsd ' over 'fourier' in any of the connectivity measures?
% EEG sampled at 200 Hz, 60 channels with events. The epoching is fine as the ERP was validated after epoching.
data = ft_connectivitysimulation(cfg);
cfg = [];
cfg.channel = ['all'];
cfg.dataset = 'Dataset_Epochs.set';
cfg.trialfun = 'ft_trialfun_general'; % this is the default
cfg.trialdef.eventtype = 'trigger';
cfg.trialdef.eventvalue = 'Onset';
cfg.trialdef.prestim = 0.2; % in seconds
cfg.trialdef.poststim = 0.6; % in seconds
cfg = ft_definetrial(cfg);
data= ft_preprocessing(cfg);
cfg = [];
cfg.pad = 'nextpow2';
cfg.method = 'mtmfft';
cfg.taper = 'hanning';
cfg.output = 'fourier';; % Tried replacing this with powandcsd and both yielded different results
cfg.keeptrials = 'yes'
cfg.tapsmofrq = 1;
spectral_decomp = ft_freqanalysis(cfg, data);
% Compute the functional connectivity using PLV
cfg = [];
cfg.method = 'plv'
fc = ft_connectivityanalysis(cfg, spectral_decomp);
To further test if it is something with the particular dataset, I also tried on separate continuous EEG data which was segmented in fieldtrip. However this yielded similar quite varying results when using either output
% cfg.dataset = 'Dataset_RS.set';
% [data] = ft_preprocessing(cfg)
%cfg = [];
%cfg.length = 6;
%cfg.overlap = 0;
%data_segmented = ft_redefinetrial(cfg, data);
Thanks,
--
Regards,
ASR
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202
--
Regards,
Akshay Sujatha Ravindran
Graduate Research Assistant
Department of Electrical & Computer Engineering
Laboratory for Noninvasive Brain-Machine Interface Systems
University of Houston
Engineering Building 2, E 413
President | BRAIN Student Group (UoH)
Vice President | Graduate and Professional Students Association (UoH)
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20200804/28979635/attachment.htm>
More information about the fieldtrip
mailing list