[Eeglablist] ICA on lowpass / highpass filtered data

Krebber, Martin martin.krebber at charite.de
Mon May 20 14:39:31 PDT 2013


Hi,

you are right about there being only few brain sources apparent in the high pass data. In the datasets I checked there were usually less than 10 ICs (out of 128) that clearly showed brain activity. The rest was muscle artefacts, bad channels and a lot of ICs with "patchy" topographies that just looked like bad decompositions. I think there is just not enough meaningful activity in these data for ICA to find 128 truly independent components. And PCA doesn't appear to be a good solution to the problem, since most people advise against using it. However, I will try your suggestion with the pairwise mutual information (if I can figure it out). 

I might have to think a little more this data splitting approach. Seems like a lot of trouble right now.

Thanks again for your very helpful comments.

Martin

________________________________________
Von: Jason Palmer [japalmer29 at gmail.com]
Gesendet: Montag, 20. Mai 2013 22:33
An: Krebber, Martin; 'Simon-Shlomo Poil'; mmiyakoshi at ucsd.edu; eeglablist at sccn.ucsd.edu
Betreff: RE: [Eeglablist] ICA on lowpass / highpass filtered data

Hi Martin,

My take: I think a large part of the reason for the disappointing high-frequency decomposition is due to the extreme attenuation of the skull of high-frequency low-amplitude sources. The independent components usually extracted in ICA of EEG are primarily low frequency, plus muscle, eye, and line noise. One might think that a low frequency, say midline theta, component might contain gamma, possibly amplitude correlated to a certain phase of theta. Personally I haven’t had a lot of success with this in EEG. So there is a question of how much brain signal is left in the high-pass part. There is also a question of how intermittent the high-frequency signal, or how non-stationary relative to other intermittent high-frequency signals.

As for the computation time increase, I have noticed this kind of thing as well, and don’t know the reason, but have thought it had something to do with caching data in the CPU—the next number multiplied in the low-pass data might require fewer register or bit modifications.

I think your comments on PCA vs. ICA are right. ICA should extract independent components regardless of variance. I don’t think there is a way to estimate the number of ICs without actually performing ICA. You can then check pairwise mutual information of the sources and count the number of ICs having less than some threshold of mutual information with all other components, or the number of independent subspaces (with only pairwise mutual info within themselves.)

Best,
Jason

From: eeglablist-bounces at sccn.ucsd.edu [mailto:eeglablist-bounces at sccn.ucsd.edu] On Behalf Of Krebber, Martin
Sent: Friday, May 17, 2013 9:28 AM
To: Simon-Shlomo Poil; mmiyakoshi at ucsd.edu; eeglablist at sccn.ucsd.edu
Subject: Re: [Eeglablist] ICA on lowpass / highpass filtered data

Hi,

thanks a lot for your input. I tried Simons PCA approach on a couple of datasets and it seems like I need less than half of my components to explain 99% of variance in my high high frequency data.

I am wondering, though, if this approach is appropriate since PCA and ICA work differently. As far as I understood, PCA tries to maximize the variance explained of each component, whereas ICA tries to maximize the independence of the components. So if, for instance, 64 PCs explain 99% of my variance, 64 ICs might explain much less.

Another thing that seems worth mentioning is that when I run the same procedure over the low pass data or the original unsplit data I need even fewer PCs to explain the same percentage of variance. So this procedure does not really explain why my ICA on the high pass filtered data takes so much longer than the ICAs in low frequency or unsplit data.

Could it be that it's a question of independence rather than of variance explained? Is there a way to estimate how many independent sources there are in the data?

Thanks!

Martin



On 17.05.2013 00:25, Simon-Shlomo Poil wrote:

Dear Martin,

As Makoto says, you make the channels less independent of each other.
It might be resonable* to reduce using PCA. One way to determine the
number of relevant dimesions could be
[COEFF, SCORE, LATENT] = princomp(EEG.data');
tmp  = cumsum(LATENT);
nr=find(tmp/tmp(end)>0.975,1);

, which gives you the number of principle components explaining 97.5 %
of the variance.

*you can find previous mails discussing pro-/con-  of PCA reduction on
the this list (I remember there was even a paper in prep? I didn't see
it come out)

Best wishes
Simon

--
Simon-Shlomo Poil, Dr.

2013/5/16 Makoto Miyakoshi <mmiyakoshi at ucsd.edu><mailto:mmiyakoshi at ucsd.edu>:
> Dear Martin,
>
> If you apply a band-pass filter, your channel data become less independent
> of each other i.e. rank-reduced.
>
> Imagine you apply an extreme band-pass filter, say 10-11Hz. All of your
> channel data look very much like each other.
>
> Makoto
>
>
> 2013/5/16 Krebber, Martin <martin.krebber at charite.de><mailto:martin.krebber at charite.de>
>>
>> Hi all,
>>
>>
>> I am currently working on an analysis were I split the data into low and
>> high frequency portions using a lowpass (cutoff 35 Hz) and a highpass
>> (20 Hz) filter, respectively. The idea behind this approach is to do the
>> ICA artefact rejection seperately on low and high frequency data in
>> order to be better able to reject high frequency muscle artefacts and
>> obtain a clearer brain signal in the gamma range.
>>
>> My problem is that, especially with the highpass filtered data, ICA
>> takes a very long time (roughly 5-10 times the usual) and even then the
>> decomposition does not look very clean. I tried to reduce the
>> dimensionality of the data (from 128 to 96) by applying the PCA
>> parameter in pop_runica and it is way faster. Is it justified, or maybe
>> even recommended to reduce the data dimensionality after filtering out a
>> considerable portion of the signal? And if so, is there a rule of thumb
>> about how much to reduce the data dimensionality?
>>
>> Thanks for any suggestions!
>>
>> Regards,
>> Martin
>>
>> _______________________________________________
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>> To unsubscribe, send an empty email to
>> eeglablist-unsubscribe at sccn.ucsd.edu<mailto: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<mailto:eeglablist-request at sccn.ucsd.edu>
>
>
>
>
> --
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego



--





More information about the eeglablist mailing list