[Mne_analysis] MNE pca and SNR

Daniel Goldenholz daniel at nmr.mgh.harvard.edu
Tue Nov 18 12:02:00 EST 2008
Search archives:

I agree with you Yury. The safe bet is to discard any data that seems to
have a big artifact.
I look forward to the day when we can include those data with confidence
because we have a proper method of accounting for the artifact precisely.

Daniel Goldenholz MD, PhD

On Tue, Nov 18, 2008 at 10:55 AM, Yury Petrov <y.petrov at neu.edu> wrote:

> EEGLab, which is an extensive free Matlab package has an ICA extension.
> What you should think about is whether trying to remove eye-blinks, etc by
> using ICA is worthwhile. There is always a chance that it will corrupt your
> data. It is safer and, in my view, a better practice to simply collect more
> data and totally dump all the time intervals with artifacts using a simple
> threshold artifact detector.
> On Nov 18, 2008, at Nov 18, 2008 | 7:19 AM, Daniel Goldenholz wrote:
>  Marina
>> 1. The snr issue is actually a very deep one. You are telling MNE how to
>> set up the regularization parameter, and yes, it will change the resulting
>> signals. The reason that happens is that there are many more brain states
>> than there are MEG sensor states. (this is that "ill-posed problem" thing).
>> As a result, we do a pragmatic thing, which is simplify the problem and make
>> it solvable - we add the constraint of minimizing the current at every point
>> in the brain (loosely described). But when we do that, we need to make a
>> somewhat arbitrary choice about how much weight to give that constraint.
>> Because there is no simply physical interpretation of what we are doing,
>> physiology cannot easily guide us here. In practice, I have no suggestion. I
>> defer to Matti.
>> 2. I did some reading on the subject of artifact removal a few years ago.
>> I haven't kept up with the literature there. But I recall reading a
>> relatively compelling argument against using PCA for artifact removal that
>> Dan hinted at. Essentially PCA does not care if it also removes real brain
>> data along with artifact. One possible approach suggested out there was a
>> related technique, called ICA. I believe someone has written a matlab
>> toolbox for that. But the results there are still sometimes suspicious. If
>> you trust SSS, that may be an extremely powerful method to removal any
>> artifact not coming from the head. As for eye blinks and movements - you
>> could consider adaptive filtering. Another possibility is to fit several
>> dipoles at the anatomical location of the eyes, and fit data to them, and
>> remove them. In my playing around with these issues very briefly, I didn't
>> find one perfect solution. But many papers are out there saying they did.
>> Hope my vagaries are mildly useful
>> Daniel Goldenholz MD, PhD
>> --------------------------------------------------------
>> http://www.nmr.mgh.harvard.edu/~daniel<http://www.nmr.mgh.harvard.edu/%7Edaniel>
>> On Tue, Nov 18, 2008 at 6:13 AM, Daniel G. Wakeman <dgwakeman at gmail.com>
>> wrote:
>> Hi Marina,
>> 1.  I think the lack of reply generally reflects people's tendency to
>> leave it alone.  Please people chime in, if I'm wrong.
>> 2.  There is no function in MNE that does a PCA.  The closest thing within
>> MNE would be the ability of mne_browse_raw to create SSP projections.
>>  However, I would not recommend using this for that purpose; again please
>> chime in, if you disagree.  The next best possibility to work on this would
>> be to use the excellent MATLAB toolbox to read the fif files into matlab
>> remove the eyeblinks/cardiovascular artifacts using whatever technique you
>> like best and save the data back and continue processing.  Just be very
>> careful with your noise covariance matrices and your artifact extraction as
>> many of these tools have dangerous consequences for the true brain data.
>> Dan
>> On 7 Nov 2008, at 13:39, Marina Papoutsi wrote:
>>> Dear MNE users,
>>> I have two questions that I hope someone can answer.
>>> 1) The first is about SNR. Following an older post that I made on SNR
>>> (left unaswered :-( ) I will attempt to rephrase the question, hoping to
>>> hear back from someone.
>>> I am still wondering about how to best use the option --snr in
>>> mne_make_movie. I have played around with the data a bit and changing the
>>> SNR value seems to make a difference in the stc files.
>>> So what is the best way to chose an SNR?
>>> a) use the default value (3) independent of whether this is reflected in
>>> your data
>>> b) estimate the mean SNR value (over the epoch time), which will be
>>> different for each subject and condition?
>>> At the moment, I have calculated the SNR for each subject and condition.
>>> The SNR value is between 1 and 2 and it is different for subjects and
>>> conditions. Will that be a problem when I need to do group-level statistics
>>> comparing different conditions?
>>> 2) The second question is about PCA. Is there any function in MNE that
>>> would allow one to do a PCA on the *.fif files to remove eye-blinks or
>>> cardiovascular related oscillations?
>>> Thank you in advance for the help,
>>> Marina
>>> _______________________________________________
>>> Mne_analysis mailing list
>>> Mne_analysis at nmr.mgh.harvard.edu
>>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis
>> _______________________________________________
>> Mne_analysis mailing list
>> Mne_analysis at nmr.mgh.harvard.edu
>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis
>> _______________________________________________
>> Mne_analysis mailing list
>> Mne_analysis at nmr.mgh.harvard.edu
>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20081118/0e4eca93/attachment.html 

More information about the Mne_analysis mailing list