[Mne_analysis] noise covariance matrix

Mon Apr 10 06:50:35 EDT 2017

Hi Denis,

Thank you for your help.

I did the following as you suggested, but the whitening visualisation does not seem to change – the GFP is still <1 for the MEG data.

Noise_cov.plot(evoked.info)

Meg_picks = mne.pick_types(raw.info, meg =True, eeg=False)
Sss_rank = raw.estimate_rank(picks = meg_picks)
Rank = sss_rank
Mne.viz.evoked._plot_evoked_white(evoked, noise_cov[0], rank)

[cid:image004.jpg at 01D2B1F0.AA197C30]

Although the ‘shrunk’ method seems to be still problematic, is there any reason to suppose there are problems with the ‘empirical’ method for SSS-ed data? In other words, could I go ahead with ‘empirical’ method (I think it’s the standard?) or would I be better not using automated regularisation?

I would happily share a data file, but the MRC has a strict policy on this, so I just need to check.

Thank you.

Best wishes,

Lucy

From: mne_analysis-bounces at nmr.mgh.harvard.edu [mailto:mne_analysis-bounces at nmr.mgh.harvard.edu] On Behalf Of Denis-Alexander Engemann
Sent: 08 April 2017 15:09
To: Discussion and support forum for the users of MNE Software
Subject: Re: [Mne_analysis] noise covariance matrix

Hi Lucy,

your plots show that you're using SSS. We recently saw some rank estimation issues in some case at the level of the covariance matrix but not yet fully understand the situation.
Can I ask you to plot the covariance as follows:

```
cov.plot(evoked.info<http://evoked.info>)
```

then you could try to plot the whitening again, this time using our non public maintenance function.

```
meg_picks = mne.pick_types(meg=True, eeg=False)
sss_rank = raw.estimate_rank(picks=meg_picks)
mne.viz.evoked._plot_white(evoked, noise_cov, rank=sss_rank)
```

As a sanity check, the ```sss_rank``` should correspond to the prominent kink in your eigenvalue spectrum of the covariance.

If the display improves your file also belongs to the mysterious group of SSS'ed files for which our defaults aren't optimized.

I'm currently working on an improvement of our covariance computation with regard to the data rank. If you are happy to share an epochs or raw file with me privately I could use that as another test file for development.

Best,
Denis

On Fri, Apr 7, 2017 at 4:33 AM Lucy MacGregor <Lucy.MacGregor at mrc-cbu.cam.ac.uk<mailto:Lucy.MacGregor at mrc-cbu.cam.ac.uk>> wrote:
Hi Alex,

Many thanks for the suggestion.

I have tried again, now with a HPF of 1Hz (previously was 0.1Hz). The EEG look better (although I realise I have some noisy channels in there). For the MEG the GFP looks much better for empirical, with values around 1. However, for ‘shrunk’ the values are below 1. But ‘shrunk’ is chosen as the best method.

Please do you have an explanation for why the GPF<1 for the ‘shrunk’ method?
Do you have a suggestion as to what I could do (e.g. just choose ‘empirical’ or make some other changes)?

[image004.jpg]

Many thanks for your advice.

Best wishes,

Lucy

From: mne_analysis-bounces at nmr.mgh.harvard.edu<mailto:mne_analysis-bounces at nmr.mgh.harvard.edu> [mailto:mne_analysis-bounces at nmr.mgh.harvard.edu<mailto:mne_analysis-bounces at nmr.mgh.harvard.edu>] On Behalf Of Alexandre Gramfort
Sent: 06 April 2017 14:35
To: Discussion and support forum for the users of MNE Software
Subject: Re: [Mne_analysis] noise covariance matrix

Hi Lucy,

you seem to have some channels with very big drifts. That's why you see
so huge GFP values. If it's acceptable for your type of question you could
high pass a bit to fix this.

HTH
Alex

On Thu, Apr 6, 2017 at 6:29 AM, Lucy MacGregor <Lucy.MacGregor at mrc-cbu.cam.ac.uk<mailto:Lucy.MacGregor at mrc-cbu.cam.ac.uk>> wrote:

Dear MNE users,

I would very much appreciate your advice on the results I am getting from calculation of the noise covariance matrix. I’m using the “method” option for mne.compute_covariance to do automated regularisation.

Data were collected with Neuromag 306 Vectorview system. My responses are time locked to the onset of the average of ~300 auditorily-presented sentences. I have used the silent (baseline) period -500-0ms before sentence-onset as the time period from which to estimate the noise.

##################
event_id = None
tmin, tmax = -0.5, 5.5
reject_tmin, reject_tmax = -0.5, 1.5
bmin, bmax = -0.5, 0

epochs = mne.Epochs(raw, events, event_id, tmin, tmax, reject_tmin = reject_tmin, reject_tmax = reject_tmax, picks=picks, baseline=baseline, reject=reject, preload=True, add_eeg_ref=True)
noise_cov = mne.compute_covariance(epochs, method =(‘shrunk’, ‘empirical’), tmin=bmin, tmax=bmax, return_estimators = True)
###################

The plot below is for a single subject (but all my subjects show similar-looking output) for a period -500 to 5000ms covering the duration of my sentences.

I have compared my output with that for the examples:
http://martinos.org/mne/stable/auto_examples/visualization/plot_evoked_whitening.html#sphx-glr-auto-examples-visualization-plot-evoked-whitening-py
http://martinos.org/mne/stable/auto_tutorials/plot_compute_covariance.html

The result tells me that “shrunk” is the best method, but from looking at the output from whitening I’m unsure how this is the case, and in fact whether either method is working as it should.

[image003.jpg]

Evoked signals for all channels:
For the MEG, during the baseline the values are generally within the +/- 1.96 indicated by the red dotted line, so I think this is OK.
Data look quite noisy

GFP plots for MEG:
For ‘empirical’, the baseline values > 1 whereas for ‘shrunk, the baseline values <1. As I understand it values should be around 1 and therefore both methods look problematic.

My question is therefore:
when the baseline GFP is > or < 1 then is this due to problems with regularisation and where should I go from here?

With thanks for your thoughts and advice.

Kind regards,

Lucy MacGregor

_______________________________________________
Mne_analysis mailing list
Mne_analysis at nmr.mgh.harvard.edu<mailto:Mne_analysis at nmr.mgh.harvard.edu>
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis

The information in this e-mail is intended only for the person to whom it is
addressed. If you believe this e-mail was sent to you in error and the e-mail
contains patient information, please contact the Partners Compliance HelpLine at
http://www.partners.org/complianceline . If the e-mail was sent to you in error
but does not contain patient information, please contact the sender and properly
dispose of the e-mail.

_______________________________________________
Mne_analysis mailing list
Mne_analysis at nmr.mgh.harvard.edu<mailto:Mne_analysis at nmr.mgh.harvard.edu>
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis

The information in this e-mail is intended only for the person to whom it is
addressed. If you believe this e-mail was sent to you in error and the e-mail
contains patient information, please contact the Partners Compliance HelpLine at
http://www.partners.org/complianceline . If the e-mail was sent to you in error
but does not contain patient information, please contact the sender and properly
dispose of the e-mail.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170410/f51f8082/attachment-0001.html 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.jpg
Type: image/jpeg
Size: 37510 bytes
Desc: image001.jpg
Url : http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170410/f51f8082/attachment-0003.jpg 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image002.jpg
Type: image/jpeg
Size: 37936 bytes
Desc: image002.jpg
Url : http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170410/f51f8082/attachment-0004.jpg 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image004.jpg
Type: image/jpeg
Size: 80893 bytes
Desc: image004.jpg
Url : http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170410/f51f8082/attachment-0005.jpg 