[Eeglablist] ICA Misinformation

Robert Thatcher rwthatcher2 at yahoo.com
Wed Jun 14 17:13:14 PDT 2017

Ramesh,    Myself and hundres of others use automatic deletion of artifact that is objective and efficient.  There are three levels: 1- the use of the Hilbert transform to compute instantaneous time pont-by-time point Z scores with respect to a fully artifact free reference that has been scrutinized by numerous individuals and national institutions like NIH and FDA and 2- The selection of 10 seconds of an artifact free template that is then matched one to one or by adjustment to the remainder of the record to select artifact free samples and, 3- as a last resort manual selection of artifact free EEG based on expert judgement.   No matter what an expert needs to review all artifact free EEG samples to confirm that there is no artifact. 
Because the artifact rejection is very fast one can repeat the automatic procedure and produce multiple artifact free selections and determine that they are essentially 0.95 the same and the clinical interpretation is unchanged with multiple independent selections.  One can do this in less than one minute with multiple selections of artifact free EEG and show the same clinical conclusions to a court.   Many users of Neuroguide do this in front of judges and attorneys in court trials and as a consequence QEEG is admitted in court in almost all states in the USA using this method.  In contrast ICA reconstruction is not accepted in court because the expert cannot repeat or replicate his reconstructed EEG and the opposing attorney prevails.
Here is a url to a You Tube that describes how multiple artifact free selections can be made showing very little change in phase differences and other metrics are the same independent of the selections: http://youtu.be/FQeHQ6u_nXg
I also do not agree that all EEG contains artifact - artifact is produced by non-brain sources that obey the basic physics of external sources that have voltage gradients and magnitudes and shapes that are very different than EEG produced by sources inside of the skull.

On Wednesday, June 14, 2017, 7:31:18 PM EDT, Ramesh Srinivasan <r.srinivasan at uci.edu> wrote:

Hi Bob 
I certainly think your suggestion to "delete the one second of data" is preferable, if thats all there were to it.  But there are two additional things to consider - 
(1) Many EEG studies are in populations, or use experimental paradigms, where inevitably there will be an eye-blink/eye-movement on a large fraction of trials.  
(2) I believe manual editing as you describe is also highly subjective, as is the selection of ICA components to remove.  We usually only remove the eye-movement component.  
In my opinion, there really is no such thing as artifact-free data, except for narrow-band signals like SSVEPs.   
 On 06/14/2017 03:15 PM, Robert Thatcher wrote:
  Ramesh,      Thank you for your post and I agree that artifact is broad-band and superimposed on many if not all of the EEG channels.  Reconstruction therefore will necessarily change relative phase which can be seen in the waves themselves and is accumulative in the average phase differences between channels.  As for your concern "It's not obvious to me to prefer the original relative phase with the artifact components." I believe that you should have no concern because the original phase differences that are artifact free are real and produced by the underlying physiology and represent the summation of LFP due to synaptic rise times and synaptic integration times and conduction velocities between groups of neurons in networks of the brain.  The original phase differences must be preserved and not altered in any manner if one wants to study brain networks and dynamics. 
  The simple solution is to not use ICA for artifact rejection and instead use algorhythms to delete the parts of the record that have artifact and retain the parts of the original record with no artifact.  Because of the stochastic and nonstationarity of the EEG as one increases the sample size then one converges toward the stable and reproducabe average of the instantaneous phase differences between channels that is not corrupted by artifact.  ICA reconstruction alters an entire 5 minute EEG recording even if there is ony a single 1 second of eye movement artifact.   Why not simply delete the one second artifact and then work with the remaining 4 minutes and 59 seconds? 
  ICA is excellent for feature detection and can serve as "seeds" to guide further cross-spectral analyses only if the phase differences in the original recording are preserved. 
    On Wednesday, June 14, 2017, 4:46:34 PM EDT, Ramesh Srinivasan <r.srinivasan at uci.edu> wrote: 
Hi All - 
I think Bob is right that the relative phase will be changed by deleting 1 or 2 artifact components.   Any artifact is broad-band and hence has components in each frequency bin.  When reconstructing the (in this example, 19) channels, the relative phases will change because some of the signal in each frequency  bin has been removed when using only 17 or 18 components. 
The open question is whether the original relative phase or the ICA-corrected relative phase is the better estimate of the relative phase between the populations that contributed to each electrode.  It's not obvious to me to prefer the original relative phase with the artifact components. 
Part of the problem for me (and I do use EEGLAB's ica) about identifying components as artifact in the ICA is that I don't think they contain just the artifact, they also contain some genuine brain activity that we are removing.  This bothers me, but I don't know a better solution.  Even the case of the eye-movement artifact components is likely a mixture. 
I'd like to see this discussion move away from algorithm to this harder question about artifact removal. 

  On 06/14/2017 10:43 AM, Robert Thatcher wrote:
     Thank you for the information.  I could only find a power point attachment of a simulation in your post.  I did not find a scientific publication where you compared the phase differences changes between an  original EEG recording and a ICA reconstruction after removing one or more components.   Please re-send your study.  Also please give the citation to any of your publications or other’s publications where phase differences were compared between the original EEG recording and post ICA reconstruction.  It will be interesting to see if you found similar changes like in the study by Montefusco-Siegmund et al or by  Georges Otte or even in the example pre vs post data files that you can download from the internet.  I am assuming that you have downloaded the EEG data and then used a JTFA like the Hilbert transform or even the FFT  cross-spectrum to prove to yourself that the phase differences between the original and the ICA reconstruction have not been preserved.
As for the mathematics concerning reconstruction from a lower dimensional matrix to a higher dimensional matrix where there are no simple linear transforms I refer you to Taken’s theorem where “The reconstruction preserves the properties of the dynamical system that do not change under smooth coordinate changes, but it does not preserve the geometric shape of structures in phase space.”  Also, in standard differential geometry math courses the issue of lower dimensional manifold mapping to higher dimensional manifolds shows a loss of information in all cases.  Also, commonsense operates here where one tries to reconstruct 19 channels of EEG using only 15 or 16 or 17 ICA  components hence a loss of information.
Finally, the brain is not a total chaotic organ.   As demonstrated by many scientists (e.g., Nunez; Walter Freeman; Roberto-Pascual Marqui; E. Roy John; Joel Lubar; etc) coherence and phase differences are well behaved and highly reproducible within and  between subjects.  Coherence and phase are dependent on the number and strength of connections between groups of neurons.  Here is a URL to a study that tested Paul Nunez’s two-compartmental model of Coherence and Phase Differences and found that these measures vary as a function of distance and packing density:
Here is a url to a study that used EEG LORETA correlations to replicate Diffusion Tensor Imaging measures of connectivity in the brain:
Here is a url to a study that measured phase lock and phase shift duration from birth to about 16 years of age in 458 and where phase differences were stable and well behaved:  http://www.appliedneuroscience.com/PhaseresetDevelopment.pdf
If you do a search of the National Library of Medicine database (Pubmed) using the search terms “EEG coherence” you will find 2,874 citations.  There is huge consistency in this vast literature which would be impossible if the brain was totally chaotic.
Best regards,
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     
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170615/f7d4cef5/attachment.html>

More information about the eeglablist mailing list