[Eeglablist] is there pca in eeglab

Arnaud Delorme arno at ucsd.edu
Tue Apr 30 17:54:29 PDT 2019


Hi Joe,

Yes, my original comment was about spatial ICA. I was not arguing of the usefulness of PCA in signal processing, especially in terms of dimension reduction. I still remain to be convinced of its superiority for extraction of physiological features as the orthogonality constraint prevents PCA axis from matching the data native distribution - a constraint that ICA is not subject to. At least in the spatial domain, we have showed that PCA performed very poorly in recovering components that could be interpreted as brain sources 

https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0030135

My intuition is that this will be the same in the time domain (this is what is done when applying ICA to fMRI data - usually ICA (not PCA) is applied to the time domain BOLD signal). We have actually tried it once in a collaboration with Sylvain Baillet where we used MEG data inverse solution at each time point, then used complex ICA in the time domain. We obtained meaningful ICA components but did not purse to publish it. Below is a link to the student report in case you are interested.

http://sccn.ucsd.edu/mediawiki/images/7/79/Memoire_M2R_RIM_denis_julien.pdf

Best wishes,

Arno

> On Apr 22, 2019, at 7:51 PM, Joseph Dien <jdien07 at mac.com> wrote:
> 
> Hi Arno,
>    I agree that this was a very interesting paper but that it doesn’t rebut my point.  My counter-argument goes as thus:
> 
> 1) Time exists.  Events occur in the external world.  Light reflected from these events enter the retina where it is transduced into electrical signals.  These signals are then conducted to the brain, where it cascades through cortical networks.  If what one is interested in is the temporal ordering of this information flow, then it is largely irrelevant whether the ERP signals originate as singular spikes or as phase resets of ongoing oscillations.  Either way the scalp recorded signals we use in EEG research are only useful summaries of the true complexities of the neural activity.  My results indicate that temporal PCA-Promax does a better job at describing the nuances of the temporal aspects of the data than Spatial ICA and it is not surprising that it does, based on first principles.
> 
> 2) Makeig et al (2002) specifically examined the N1.  Your conclusion that it "showed that ERPs often arise from phase synchronization in different trials - and true ERPs are rarely observed.” is unsupported by this empirical data.  Indeed, these authors take pains to state that "Although we do not suggest all features of averaged ERPs are necessarily generated by partial phase resetting of EEG processes without concurrent energy increases” (p. 693).  
> 
> Even the N1 findings could use more examination as it relied in part on the observation that the high alpha trials correlated with a larger N1 amplitude and as we all know causation does not imply causation and relied otherwise largely on the appearance of oscillations in the ICs that could also have been produced by misallocation of variance from the superimposed alpha waves.  Thus one could also argue that the absence of oscillations in the PCA-Promax solutions is a reflection of improved separation from the alpha activity.  So be careful not to make a circular argument here.  Unfortunately this study only had 31 channels, which restricted the extent to which fine-grained distinctions could be made in the data, including separating alpha waves from superimposed ERP components.
> 
> I’m not saying that phase-resets do not play a part, just that your statement is not supported by the data.  For example, I do think that that evidence for frontal theta phase resets is quite intriguing.  But I’m unconvinced that it explains all or even most other ERP features.
> 
> As for your proposed test, I’m not averse to further studies.  Going to real data does have the problem of ground truth though.  A method could result in improved p-values for the wrong reasons (conflating separate ERP components) as the late positive complex has a number of overlapping members.  My own preferred criterion is convergence between source analyses of ERP data and fMRI data, as in:
> 
> Dien, J., Spencer, K. M., & Donchin, E. (2003). Localization of the event-related potential novelty response as defined by principal components analysis. Cognitive Brain Research, 17, 637-650. 
> 
> There it was a comparison of Varimax and Promax, not Infomax though.
> 
> Anyway, more study is definitely called for but please do heed my call to be more measured about the wording of your recommendations.  Your opinions carry a great deal of weight in the EEG community.
> 
> Respectfully,
> 
> Joe
> 
>> On Apr 20, 2019, at 00:34, Arnaud Delorme <arno at ucsd.edu> wrote:
>> 
>> Hi Joseph,
>> 
>> I have had a look at your paper, and, correct me if I am wrong, it seems that all of your simulations assume that the true ERP is hidden in background EEG noise. Scott’s 2002 paper (Makeig S, Westerfield M, Jung T-P, Enghoff S, Townsend J, Courchesne E, Sejnowski TJ. Dynamic brain sources of visual evoked responses. Science, 295:690-694) showed that ERPs often arise from phase synchronization in different trials - and true ERPs are rarely observed. For some of the ERP, the inter-trial reliability (as measure using inter-trial coherence) can be as low as 0.1 (N400) and even P300 rarely goes above 0.5 (1 would be a true ERP as in your simulations). I would therefore argue that your simulations might not reflect true EEG/ERP activity.
>> 
>> I think a test that would convince me of the superiority of PCA over ICA in the time domain would be to extract P300 components (in the temporal domain), and assess how many trials are needed to reach significance when using each method as in Kappenman and Luck article https://www.ncbi.nlm.nih.gov/pubmed/20374541.
>> 
>> Cheers,
>> 
>> Arno
>> 
>>> On Apr 19, 2019, at 3:37 PM, Joseph Dien <jdien07 at mac.com> wrote:
>>> 
>>> Hi Arno,
>>>   with the greatest respect for all the wonderful work you’ve done for the EEG community with EEGlab, I have to disagree with this statement.  As I’ve shown empirically in my papers, ICA is indeed very good in the spatial domain (with electrodes as the variables) and is my preferred method for eyeblink correction, but it is not as good as PCA-Promax at ERPs in the temporal domain.  This follows naturally from the nature of the rotational criteria and the characteristics of the data.  Which is needed for an analysis depends on the analysis goals.  A couple such citations follow.
>>> 
>>> Dien, J., Khoe, W., & Mangun, G. R. (2007). Evaluation of PCA and ICA of simulated ERPs: Promax versus Infomax rotations. Human Brain Mapping, 28(8), 742-763. 
>>> 
>>> Dien, J. (2010). Evaluating Two-Step PCA Of ERP Data With Geomin, Infomax, Oblimin, Promax, And Varimax Rotations. Psychophysiology, 47(1), 170-183. 
>>> 
>>> Also, thanks Tarik for bringing Olav’s website to my attention.  This is indeed my Toolkit that he is posting to github.  I have no idea why he is doing this.  I’ll have to have a word with him.  I would appreciate folks downloading the EP Toolkit directly from my own sourceforge site as I use the download figures to seek grant funding, just as the eeglab team does.
>>> 
>>> Joe
>>> 
>>>> On Nov 23, 2018, at 19:25, Arnaud Delorme <arno at ucsd.edu> wrote:
>>>> 
>>>> There is also a PCA plugin in one of the workshop lectures (page 15).
>>>> 
>>>> https://sccn.ucsd.edu/mediawiki/images/9/95/EEGLAB2018_scripting5.pdf
>>>> 
>>>> We do not recommend using PCA which does not capture the structure of the data as ICA does. The paper Tarik mentioned is a good one
>>>> 
>>>> Artoni, F., Delorme, A., & Makeig, S. (2018). Applying dimension reduction to EEG data by Principal Component Analysis reduces the quality of its subsequent Independent Component decomposition. NeuroImage, 175, 176-187.
>>>> 
>>>> See also
>>>> 
>>>> Delorme A, Palmer J, Onton J, Oostenveld R, Makeig S. (2012) Independent EEG sources are dipolar.PLoS One, 7(2).
>>>> https://www.ncbi.nlm.nih.gov/pubmed/22355308
>>>> 
>>>> Best,
>>>> 
>>>> Arno
>>>> 
>>>> 
>>>> 
>>>>> On Nov 23, 2018, at 12:17 PM, Tarik S Bel-Bahar <tarikbelbahar at gmail.com> wrote:
>>>>> 
>>>>> Hello A S,
>>>>> Some brief notes below. When you reach a solution that satisfies your needs, please share it with the list so that other users can benefit from it.
>>>>> 
>>>>> *There are functions named runpca and runpca2 in the eeglab distribution. Review their documentation and test them out, as they may not be regularly used.
>>>>> *There is also Joe Dien's PCA toolkit: https://sourceforge.net/projects/erppcatoolkit/
>>>>> *There is also the following, which I believe uses Dien's tools or is a replica of them https://github.com/krigolson/MATLAB-EEG-PCA-Toolbox 
>>>>> *There are also some PCA functions in the Fieldtrip LIte folder that is part of the eeglab distribution (search for m files with pca in their title)
>>>>> 
>>>>> Also, ICA in eeglab is not just for removing artifacts. Many researchers analyze the ICs themselves as indexes of unique neural sources.
>>>>> From my understanding, eeglab developers strongly recommend ICA and NOT PCA, you can google "eeglablist + ICA + PCA" for past posts about that.
>>>>> There is a PCA flag in the runica function in eeglab, but it will essentially run ICA on PCA-reduced data. 
>>>>> The following recent article is interest,findable on google scholar: Artoni, F., Delorme, A., & Makeig, S. (2018). Applying dimension reduction to EEG data by Principal Component Analysis reduces the quality of its subsequent Independent Component decomposition. NeuroImage, 175, 176-187.
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> On Fri, Nov 23, 2018 at 12:00 PM A S <eng.emetsasa at gmail.com> wrote:
>>>>> Hi all,
>>>>> I know there's ICA in EEGLAB to remove artifacts. However I want to
>>>>> use PCA (Principal Components Analysis) to reduce the electrodes to
>>>>> spatio-temporal information according to the regions of interest. I
>>>>> can't find the PCA. Is there PCA in EEGLAB?
>>>>> _______________________________________________
>>>>> 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
>>>>> _______________________________________________
>>>>> 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
>>>> 
>>>> _______________________________________________
>>>> 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
>>> 
>>> --------------------------------------------------------------------------------
>>> 
>>> Joseph Dien, PhD
>>> Senior Research Scientist
>>> Department of Human Development and Quantitative Methodology
>>> University of Maryland, College Park
>>> E-mail: jdien07 at mac.com
>>> Cell Phone: 202-297-8117
>>> http://joedien.com
>>> 
>> 
>> _______________________________________________
>> 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
> 
> --------------------------------------------------------------------------------
> 
> Joseph Dien, PhD
> Senior Research Scientist
> Department of Human Development and Quantitative Methodology
> University of Maryland, College Park
> E-mail: jdien07 at mac.com
> Cell Phone: 202-297-8117
> http://joedien.com
> 



More information about the eeglablist mailing list