[Eeglablist] FW: ICA Misinformation

[otte georges]

Dear Jason

 

As most of us clinicians, in my practice  I use a 19 ch EEG. Artefacts are not limited to blinks as also tremor, psychogalvanic, line noise, electrode pops, movement, emg, ECG etc can occur. Mostly we can contain this and we can select fairly large parts of EEG without artefact.

If I use ICA components of blinks, tremor, muscle EMG, pletysmo, etc .. to clean up the EEG recording then I will reconstruct parts of EEG that are completely artefact free. ICA is reconstructing parts that need not to be reconstructed. Maybe there is a solution to that (Phil Zeeman?). But as it is now this seems rather illogical to me. If I need to rebuild a wall in my house why should I tear down and rebuild all the other good parts? In analogy, why  change (by multiplying reconstructing) eeg segments when there are no blinks, tremor, muscle or pletysmo in those segments. I am old fashionedly devoted to Hanssens Theorema: “there is no substitute for clean data”.

 

If clinicians have access to “automatic methods” to clean their EEG’s they will no doubt use them without much thinking (they generally trust manufactures) and will get “nice” looking traces but if 5 or 6 ICA components are rejected from 19 ch can we honestly accept that a reconstruction will be as good as the original or even improve our EEG ? I doubt that. 

 

In some experiments that we did (trying to replicate Montefusco’s findings)unpublished but presented at a conference ,  we noticed that changes after reconstruction were mostly (but not only) at the origin of artefact and often in the same frequency band. The problem with 19 ch is that we can stumble over the overcomplete situation (more sources then sensors) and that artefact and genuine brain activity is lumped together. Removing IC components that also index for brain slow wave activity from frontal regions can be problematic in case of frontal tumor (glioma). Missing them  is one of my many (clinical fears and  concerns). 

 

Sincerely

 

PS I do not think that you have to explain ICA nor other signal analysis algorithms to Dr Thacher. 😊 but feel free to do so if that is your ambition.

 

 

 

From: Jason Palmer [mailto:japalmer29 at gmail.com] 
Sent: Sunday, June 18, 2017 9:53 PM
To: 'Robert Thatcher' < <mailto:rwthatcher2 at yahoo.com> rwthatcher2 at yahoo.com>
Cc: 'Iman Mohammad-Rezazadeh' < <mailto:irezazadeh at ucdavis.edu> irezazadeh at ucdavis.edu>; 'Ramesh Srinivasan' < <mailto:r.srinivasan at uci.edu> r.srinivasan at uci.edu>; 'Stefan Debener' < <mailto:stefan.debener at uni-oldenburg.de> stefan.debener at uni-oldenburg.de>; 'Arnaud Delorme' < <mailto:arno at ucsd.edu> arno at ucsd.edu>; 'eeglablist' < <mailto:eeglablist at sccn.ucsd.edu> eeglablist at sccn.ucsd.edu>;  <mailto:pnunez at tulane.edu> pnunez at tulane.edu; 'Georges Otte' < <mailto:georges.otte at telenet.be> georges.otte at telenet.be>
Subject: RE: [Eeglablist] ICA Misinformation

 

Dear Robert,

 

Thank you for your reply and data. I think it would be very helpful if you could explain how the data you are presenting is used by clinicians and why it causes them great concern. I do not find it surprising at all that measures are significantly changed when eye-blinks are removed. When you say “artfact free data” you are not referring to selected time periods without eye-blinks are you? If you are comparing periods containing eye-blinks to data reconstructed with eye-blinks removed, then clearly the measures will be different. The simple fact that measures are different is not surprising. As far as I am able to interpret the figures, the significant difference is largely over the frontal electrodes, where eye-blinks are influential. Again, that is entirely what one would expect. If removing eye-blinks from data did NOT cause any change in measures, then THAT would be surprising and seem to argue that removing eye-blinks is unnecessary at all.

 

Again, I would very much appreciate reference to some clinically relevant application beyond stating a simple difference in many measures. For example, if there were a study of beta activity in orbito-frontal cortex, and clinicians were finding a significant difference in populations at the channel level, perhaps using a more traditional eye-blink removal technique, but not using the ICA based eye-blink removal, then I think that would be very compelling.

 

Also, are you claiming that eye-blinks should not be removed by any method? That would seem to invite the objection that the results in a study of frontal regions, say, are potentially conflated by eye-blinks. If you are claiming that it is only the ICA method of eye-blink removal that causes the unacceptable data distortion, do you have a similar set of plots using a more traditional eye-blink removal method that do not show significant differences in the measures you presented?

 

Are the “Delorme” plots using EEGLAB instead of WinEEG? (Apologies if you have already discussed this previously.)

 

Thanks, best,

Jason

 

 

 

From: Robert Thatcher [mailto:rwthatcher2 at yahoo.com] 
Sent: Monday, June 19, 2017 4:17 AM
To: Jason Palmer
Cc: Iman Mohammad-Rezazadeh; Ramesh Srinivasan; Stefan Debener; Arnaud Delorme; eeglablist; pnunez at tulane.edu <mailto:pnunez at tulane.edu> ; Georges Otte
Subject: Re: [Eeglablist] ICA Misinformation

 

Jason,

    Thank you and I understand your strong feelings since you stated:  " I have not seen any convincing evidence that there is a clinically significant alteration caused by judicious use of ICA for artifact removal and data reconstruction"

   

 

 I am attaching a .zip file that shows about 75% to 90% of EEG metrics are statistically significantly different from the original EEG with eye movement deleted vs post ICA reconstruction.  Also, there are edf files of the original, and post ICA by clinicians in Australia (they used WinEEG) and the Post ICA by Arno.   As you can see they are all different from each other.

 

I am not the person who first noticed the adulteration of artifact free segments of the original EEG by ICA reconstruction.   Our company has over 3,000 clinicians using our software and it the users of our software who complain about the adulteration.  You and Arno are going to have to convince these hunderds of hard working clinicians that ICA does not distort the original EEG.   Clinicians are seeing patients with serious clinical disorders and this is not an issue that can be dismissed because you have not seen any evidence of adulteration and also because you have not taken the time to compare the original EEG recording after only analysing parts of the record with no eye movement to the same artifact free part of the record after ICA reconstruction.  If there was no adulteration then the digital values and phase and coherence would be identical.  Instead they are statistically different at P < 0.0001 in well over half of the comparsions.

 

WinEEG was first incorporated aroun 1998 and has over 5,000 clinicians that use their software ICA.   I recommend that you do an internet search for WinEEG to learn more.  This is not a trivial number of individuals who are having problems achieving stable assessments of their patients and who also complain about ICA reconstruction.

 

Respectfully,

 

Robert

 

 

On Sunday, June 18, 2017, 3:04:17 PM EDT, Jason Palmer <japalmer29 at gmail.com <mailto:japalmer29 at gmail.com> > wrote:

 

 

Hi Robert, 

 

The statement that "the jury is in and the verdict is Do Not Use ICA Reconstruction" sounds incredibly presumptuous and unjustified. I believe Arnaud has admirably refuted your alpha example and I have not seen any convincing evidence that there is a clinically significant alteration caused by judicious use of ICA for artifact removal and data reconstruction. Your claim that ICA reconstruction is decoupled from the brain, etc., is incorrect. The back-projection of the source data, after components like eye-blinks are removed, is entirely "coupled" to the brain and the original data recording. The reconstruction / back-projection is NOT based on z-score ICA component activations, but on the simply linearly transform based activations. It is simply not true that the ICA reconstructed data is divorced from the brain. The same phase and amplitude relationships and generating source locations will exist in the reconstructed channel data without artifacts like eye-blinks, line-noise, scalp muscles, etc. as exist in the original channel data.

 

The notion that it is wrong to use ICA for eye-blink removal because "every data point is changed" again does not sound like a reasonable statement. It is equally correct to say that the EM potential added to the EEG by eye-blinks themselves "change the data". There are several papers that demonstrate the superiority of ICA based eye-blink removal over the "many other methods" that exist, and thus it is incorrect to say that there is no justification to use ICA over other methods. If you turned your polemic against other methods of eye-blink removal, such as regression, I have no doubt that you would find similar or greater grounds for objection based on you stated reasons.

 

I have not used or heard of WinEEG. Citing a specific implementation as a basis for rejection of a method is again unfair.

 

I hope that before issuing general method censure, you will take the time to use the freely available EEGLAB software to demonstrate any let alone a clinically important example of ICA 
"phase distortion". Your extreme warnings about the "phase distortion" caused by ICA do not seem to be founded in a correct understanding of the way ICA works. It is a linear transformation, creating a simple linear decomposition of the signal into constituent components, including eye-blinks. Theorems from differential geometry are irrelevant. Reconstructing the data with the eye-blink component left out will NOT cause any phase distortion in ICA. The statement that "every data point is changed" is not scientifically compelling. ICA is INCREDIBLY USEFUL FOR ARTIFACT REJECTION AND DATA RECONSTRUCTION as demonstrated by countless published papers. Cautioning against the use of automatic component rejection is one thing, but the issuance of general injunctions against using ICA for data reconstruction at all is clearly unwarranted. 

 

I apologize for the strident tone of this email, but there seems to be a somewhat cavalier and patronizing tone with a predetermined conclusion in the arguments I have seen made against the use of ICA thus far, and the polite responses of Arnaud, Iman, Stefan, et al. compared to your forceful, and I believe as yet unjustified by compelling evidence beyond what has been already addressed by Arnaud, or hearsay and reference to people from a particular community largely influenced by yourself, will likely be misleading to the general reader of this discussion.

 

The "jury" may be in, but the verdict is not as you claim. There are some who hold out a "guilty" verdict against ICA, but they are far from a majority.

 

Regards,

Jason Palmer

 

P.S. The yahoo.com <http://yahoo.com>  email you are using is consistently being put in "spam" by gmail, and not even showing up in my Microsoft Outlook mail (even in junk mail). Gmail claims that the address "fails yahoo.com <http://yahoo.com> 's required tests for authentication". It would be very helpful and appreciated for this discussion, assuming there is no ready solution to the yahoo.com <http://yahoo.com>  issue, if you might consider using an alternative email.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170620/cf6ce61f/attachment.html>