[FieldTrip] Problem with TFR calculation using multitaper

[Matt Craddock]

Dear Andreas and Frederic,

Another possibility would be a 60 Hz signal (a steady-state evoked 
potential to monitor refresh rate, for example) running throughout the 
trial, including in the baseline period. Given that you're using 11 Hz 
frequency smoothing, this would effectively be smeared across the 49-71 
Hz range.

You'd have a much higher baseline signal in the 49-71 Hz range (because 
of the 60 Hz signal appearing in all those frequencies), but the 
strength of the actual signal in those ranges wouldn't differ. So, the 
signal-to-baseline ratio would be much lower (because your 
signal-to-noise ratio is worse), and that'd created the appearance of 
boundaries at the edges of the frequency smoothing (because at 71 Hz you 
are still getting the 60Hz noise, but at 72Hz, you are not, so your 
signal-to-noise ratio dramatically changes).

I've attached a pic showing something similar caused by 50 Hz line noise 
(in EEG, with ~12Hz freq smoothing).

So I'd suggest checking the FFT of your data and possibly putting in a 
notch at 60 Hz if there's a big peak there.

Cheers,
Matt


On 16/11/2012 14:51, Frederic Roux wrote:
>
> Dear Andreas,
>
> if you used a 50 Hz notch filter in your
> preprocessing, that is probably the reason
> why you are seeing the 'artificial' cut
> between 30 and 70 Hz.
>
> My guess is that you ran your notch filter
> on segmented epochs of a few seconds (resulting
> in a low spectral resolution for your notch filter).
> This will wipe out a ~10Hz box (in your case I'm just guessing)
> around the line noise frequency.
>
> Running the multitaper on top of that may explain
> the pattern you are observing.
>
>
> Best,
>
> Fred
> ----- Original Message -----
> From: "Andreas Sauer" <sauer.mpih at googlemail.com>
> To: "FieldTrip discussion list" <fieldtrip at science.ru.nl>
> Sent: Friday, November 16, 2012 2:36:01 PM
> Subject: [FieldTrip] Problem with TFR calculation using multitaper
>
>
> Dear all,
>
> I have encountered a problem when trying to calculate the power spectrum of my data and I was wondering if anyone could help me to determine the origin of this problem.
>
> For the analysis I used multitapers (dpss) with a 11 Hz freq-smoothing and a time window with a length of 500 ms ending up with 10 tapers for the TFR calculation (the freq range is 20 - 200 Hz):
>
>
> cfg = [];
> cfg.output = 'pow';
> cfg.channel = 'MEG';
> cfg.method = 'mtmconvol';
> cfg.taper = 'dpss';
> cfg.keeptrials = 'no';
> cfg.foi = [20:1:200];
> cfg.toi = [-0.5:0.05:1.0]; % the preprocessed data length is 2 sec, from -0.8 to 1.2 s
> cfg.tapsmofrq = ones(length(cfg.foi),1).*11; % spectral smoothing = +/- 11 Hz
> cfg.t_ftimwin = ones(length(cfg.foi),1).*0.5; % length of time window = 500 ms
> cfg.pad = 'maxperlen';
>
> TFR_Mult_DPSS_10tapers_fmin20Hz_fmax200Hz = ft_freqanalysis(cfg,DataOut);
>
> After calculation I did a baseline normalization:
>
>
> cfg = [];
> cfg.baseline = [-0.4 -0.1];
> cfg.baselinetype = 'relchange';
>
> TFR_HighGamma{subject,condition} = ft_freqbaseline(cfg,TFR_file.TFR_Mult_DPSS_10tapers_fmin20Hz_fmax200Hz);
>
> averaged:
>
>
> cfg = [];
> cfg.channel = 'ParOcc'; % parietal and occipital channels
> GA_TFR_HighGamma{condition} = ft_freqgrandaverage(cfg,TFR_HighGamma{:,condition});
>
> and finally plotted the spectrum:
>
>
> time(1) = 3; % -400 ms
> time(2) = 27; % 800 ms
> indFreq = [20 200];
> freq = find(ismember(GA_TFR_HighGamma{i}.freq, indFreq) == 1);
>
> power = squeeze(mean(GA_TFR_HighGamma{i}.powspctrm,1));
> power = mean(GA_TFR_HighGamma{i}.powspctrm(:,:,time(1):time(2)),1);
> power = mean(power,1);
>
> fig = figure;
> set(fig,'PaperUnits','centimeters');
> set(fig,'Units','centimeters');
> set(fig,'PaperSize',[15 15]);
> set(fig,'PaperPosition',[0 0 14 14]);
> set(gcf,'Color',[1 1 1]);
>
> ha = gca;
> pcolor(GA_TFR_HighGamma{i}.time(time(1):time(2)),GA_TFR_HighGamma{i}.freq,squeeze(power));
> set(ha,'FontName','Arial','FontSize',10,'LineWidth',1.5);
> shading interp;
> xlim([-0.4 0.8]);
> ylim([40 120]);
> caxis([-.2 .2]);
> hold on;
> ha = gca;
> plot([0 0],[40 120],'k--','LineWidth',1);
> set(ha,'FontName','Arial','FontSize',10,'LineWidth',1.5);
> set(ha,'YTick',[40 60 80 100 120]);
> colorbar('SouthOutside');
> xlabel('Time [sec]');
> ylabel('Frequency [Hz]');
>
> The resulting plot looks strange to me since there is this artifical "cut" at 70 Hz and at 50 Hz. I checked the FAQs but my plot looks different from the "strange" ones described there.
> I also attached a screenshot of the MultiplotTFR and there you can see the strange "effect" clearly in channels MLO22 or MRO22 for example.
>
> Is something wrong with my TFR calculation? Or does the problem arise from an earlier step in the preprocessing?
>
> Any help or advice would be greatly appreciated!
>
> Best,
>
> Andreas
>


-- 
Dr. Matt Craddock

Post-doctoral researcher,
Institute of Psychology,
University of Leipzig,
Seeburgstr. 14-20,
04103 Leipzig, Germany
Phone: +49 341 973 95 44
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 50HznoiseFilt.png
Type: image/png
Size: 51346 bytes
Desc: not available
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20121116/85eb859c/attachment-0002.png>