[Eeglablist] Copy ICA weights from 1Hz highpass to 0.1Hz highpass

Fri Apr 1 09:07:05 PDT 2016

Dear Isaiah and Makoto,

the problem is presumably related to spectral leakage due to (a) the Hamming window used by default in pwelch combined with (b) larger offsets and/or drifts in the data and possibly (c) zero padding.
(a) The Hamming window type does not end at zero amplitude at the edges, thus, it allows a certain amount of spectral leakage. Purpose is to improve spectral resolution by suppressing the first side lobes.

I will add some demo code below demonstrating the problem. Depending on your EEGLAB version either the upper left or the upper right panel will resemble the corresponding EEGLAB spectopo plot. In the recent EEGLAB versions (at least the git developer version; not sure whether there is a public version yet) zero-padding is not applied unless necessary/requested. Hann or more elaborate window types as used in the two lower panels will efficiently suppress leakage at the cost of spectral resolution (obvious for Blackman-Harris).

Makoto, this issue is reappearing on the quite regularly (monthly?). I would like to suggest some changes:
* The issue should be mentioned in the help text of (pop_)spectopo, optimally with a reference to a FAQ or something similar (e.g. http://sccn.ucsd.edu/wiki/TIPS_and_FAQ) explaining the problem in more detail.
* A window type parameter could be added to spectopo (at least allowing to check whether the observed effects disappear). I quickly scanned through the code and if I didn’t miss something major the necessary adjustments should be limited.
* In theory it should be possible to automatically check whether offset and/or drift exceed a certain threshold and either issue a warning or even automatically adjust the applied window type. However, I would consider this as a major endeavor. Not really sure whether it’s worth (or even works).

Best,
Andreas

%%
EEG = eeg_emptyset;
EEG.srate = 500; EEG.pnts = 5000; EEG.trials = 1; EEG.nbchan = 3;
t = ( 0:EEG.pnts - 1 ) / EEG.srate;
EEG.data( 1, : ) = 2 * sin( 2 * pi * 50 * t ); % 50 Hz sine
EEG.data( 2, : ) = 2 * sin( 2 * pi * 50 * t ) + t; % with drift
EEG.data( 3, : ) = 2 * sin( 2 * pi * 50 * t ) + sum(t) / EEG.pnts; % or offset

pop_spectopo(EEG, 1, [ 0 0 ], 'EEG', 'freqrange', [ 0 100 ], 'electrodes', 'off' )

figure
h( 1 ) = subplot( 2, 2, 1 );
pwelch( EEG.data', hamming( 512 ), 0, 1024, EEG.srate );
title( 'Hamming win w/ zero-padding' )
legend( 'Sine', 'w/ drift', 'w/ offset' )

h( 2 ) = subplot( 2, 2, 2 );
pwelch( EEG.data', hamming( EEG.srate ), 0, EEG.srate, EEG.srate );
title( 'Hamming win w/o zero-padding' )

h( 3 ) = subplot( 2, 2, 3 );
pwelch( EEG.data', hann( EEG.srate ), 0, EEG.srate, EEG.srate );
title( 'Hann win w/o zero-padding' )

h( 4 ) = subplot( 2, 2, 4 );
pwelch( EEG.data', blackmanharris( EEG.srate ), 0, EEG.srate, EEG.srate );
title( 'Blackman-Harris win w/o zero-padding' )

set( h, 'XLim', [ 0 100 ], 'YLim', [ -150 20 ] )
%%

> Am 25.03.2016 um 14:25 schrieb Isaiah Innis <isainnis at indiana.edu>:
> 
> That is good to know, thank you. I'm relieved it's not a problem with copying the ICA weights. 
> 
> 2016-03-22 5:33 GMT-04:00 Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
> Dear Isaiah,
> 
> There is a known problem that if you have a big low frequency drift, EEGLAB's spectrum analysis fails. Once I understood how this happens when I discussed a related issue with Andreas, but now I don't remember it. Anyway, if you want to make the spectrum plot work, excessive low-frequency drift should be removed.
> 
> Makoto
> 
> On Fri, Feb 19, 2016 at 6:02 AM, Isaiah Innis <isainnis at umail.iu.edu> wrote:
> Hello eeglablist:
> 
> 
> Some of our datasets have displayed a large amount of slow frequency drifts. See here for an example: https://onedrive.live.com/redir?resid=A1398B0A4C8F85DE!2199&authkey=!AD6I0359QqJH5eA&v=3&ithint=photo%2cpng
> 
> In order to address this, we have tried the suggestion mentioned on this list of:
> 1 Filtering each dataset with a 1Hz highpass 
> 2 Run ICA on this dataset
> 3 Copy the ICA weights from the 1Hz dataset and apply them to our normal 0.1Hz filtered dataset
> 
> However when viewing the components in the 0.1Hz dataset (the one we copied weights onto), I have noticed jagged spectra in the properties of several components.
> Here are a few examples (the 1Hz component is on the right in the image, and the 0.1Hz is on the left):
> 
> https://onedrive.live.com/redir?resid=A1398B0A4C8F85DE!2191&authkey=!ALiODmU2Hqc5HZE&ithint=folder%2cpng
> 
> Is there something wrong here? These steps were suggested in vol 130 issue 5 of the eeglab digest for reference. 
> 
> Thank you in advance, 
> 
> -- 
> Isaiah Innis
> Indiana University '13
> EEG Technician, IUB IRF
> 
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
> 
> 
> 
> -- 
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
> 
> 
> 
> -- 
> Isaiah Innis
> Indiana University '13
> EEG Technician, IUB IRF
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu