[Eeglablist] FW: Beyond good and evil of ICA

[Robert Thatcher]

Marius,

    The groundtruth of the EEG and phase delays begins with the physics of the electricalfield potential (or theelectrostatic potential) defined as theamount of work needed to move a unit positive charge from a reference point toa specific point inside the field without producing any acceleration. Whenthe curl ∇ × E is zero then the electrical potential is determined bythe gradient of the electrical field.  EEGdifferential amplifiers are used to measure the difference or gradient of theelectrical field and therefore the electrical potential is the differencebetween the two inputs to the amplifier. If one measures the electrical field and then removes ICs and startswith a smaller number of EEG channels to create a larger number of time seriesto replace the original time series then this changes the gradients of theelectrical field produced by the brain and also distorts the time differencesbetween samples then the physics of the electrical field no longer applies.
The majority of neuroscientists agree that the physiologicalground truth for the genesis of the EEG as being summated synaptic potentialsin groups of neurons that are connected by axons.   The ground truth of phase differencesbetween separated scalp locations are physiological factors of the electricalsources and the connections between sources (i.e., networks) such as axonalconduction velocities, length of connecting axons, synaptic rise times, synapticdelays and synaptic integration times, etc.  This is well established science and it was quite surprising to readposts on this forum by individuals that claim that there is no ground truth ofthe EEG, EEG is all noise and therefore this justified altering phasedifferences using a small number of ICs to create different time series (i.e.,reconstruction) to thereby decoupling the brain from the subject and thephysics of electrical fields.   LikePedro Valdez I have also recorded EEG from the dura and pia surface andsimultaneously from the scalp.  I alsohave impaled neurons and correlated the intracellular EEG with field potentialsas a function of distance from the impaled neurons.  These recordings were when I was a post docat Albert Einstein College of Medicine   With E. Roy John we moved multiple micro-electrodesthrough wide regions of the cat brain while also recording EEG from thescalp.  It was during this time that Royand myself and others quantified how phase differences increase or decrease asa function of interelectrode distance. At the University of Maryland myself and colleaques, while stillcollaborating with Roy, did extensive studies that linked phase differences to thecortical white matter with short distance ‘U’ shaped fibers vs long distancefasciculi.  These measures helped PaulNunez to use scalp EEG phase differences to estimate white matter conductionvelocities.  Additional validationstudies of the network physiological ground truth of EEG phase differences waswhen I worked at NIH with patients before and after craniotomy.   Further validations also involvedcorrelations with MRI T2 relaxation time.  
Luckily ICA reconstruction was not invented in the 1960s because if ithad then today we would have no idea that there is a replicable and verifiable groundtruth of EEG nor know about the relationship between EEG phase differences andthe cerebral white matter.   
 

As for volume conduction and phase delays betweenseparated sensors I think that there may be some confusion between theinstantaneous physics of the electrical potential and the volume conductormodel used for the inverse solution.  Thevolume conductor model pertains to a single instant of time where the accuracyof the spatial localization of sources is affected by inhomogeneities of tissueconductivities that primarily affect amplitudes and only negligibly timedifferences which are near to the speed of light.  A good book on this topic is Malmuvo andPlonsey “Bioelectricity”, Oxford University, 1995 or a standard physics textbook, for example, Richard Feynman’s book. Here is a url to a study where we evaluated phase differences as afunction of the Euclidean distance between Brodmann areas that comprise thedefault network - see figure 6 at:

http://www.appliedneuroscience.com/Default_Network_LORETA_Phase_Reset-Thatcher_et_al.pdf

 

We have also evaluated the distortion of effectiveconnectivity by ICA reconstruction.  Forexample, we computed the phase slope index before and after ICA reconstructionand 100% of the effective connectivity between 171 electrode pairs for allfrequencies is adulterated.  This isanother example of decoupling the “ground truth” of the physiology of the brainfrom the subject.   We recently published a study that correlatedintelligence with effective connectivity where we found significant differencesbetween low and high I.Q. subjects.   Ifwe had used ICA reconstruction then because phase differences are scrambled anddistorted and thereby decoupled from the brain it is unlikely that there wouldhave been any significant differences.  

Here is a url to this study: http://www.appliedneuroscience.com/Intelligence%20&%20information%20flow-Thatcher%20et%20al%202016.pdf

 

For those interested, here is a url to a book that E.Roy John and myself published in 1977 that reviews the history of how theground truth of EEG was discovered.  The ICA replacement violates the electricity and physiology of the brain and contradicts this history:  http://www.appliedneuroscience.com/FuncNeuro_v1n.pdf

 

Best wishes,

 

Robert


On Friday, August 4, 2017, 1:10:52 PM EDT, Marius Klug <marius.s.klug at gmail.com> wrote:

Hi everyone!

One thing after the other...:


Georges, sure, feel free to use the example, I'm always happy to contribute :-) 

A comparison of high to low density ICA cleaning would definitely be interesting and my hypothesis is that the ICA decomposition might be significantly worse (meaning more changes in the posterior channels) in the 19-channels set. If I have some spare time I might do that and post the results. Will take a while, though, I won't have time until end of September I think...



Robert, if you read again my very first email on the topic, I have not changed my mind the slightest bit... I have said from the beginning that ICA does indeed change the data, which is rather obvious. It's only about the question where this change is a positive or a negative one. However, as I said, it might be that 19 channels is not a sufficiently high number for a decent ICA decomposition for cleaning. It sure is not for working on the source level, but by now I did indeed assume that the eye components are so strong that even 19 channels are enough to cleanly take them out. If that's not the case, it's an important information and everyone needs to decide for themselves if they want to use ICA and risk losing a slight bit of brain data or other cleaning methods which might prove to be better in the low-density segment. I would, however still HIGHLY encourage you to test whether or not the results with the ICA-based cleaning are significantly better or worse than without. As long as you have not tested this, you keep rejecting a method that might prove useful for you.

The as you call it false belief of continuous artifacts has in my opinion already been adressed enough and I guess you just must have special means of keeping eye and neck muscles and heart beats from continuous signals in the EEG, since in ordinary participants eye components seem to correlate highly with eye tracker signals and I can still see the heart beats of my subjects in their EEG. As I've stated elsewhere, there are many means of increasing signal-to-noise ratio, ICA just being one of them (one other would for example be the averaging of ERPs). Many other published sources just tell us that those other means work as well and that the signal had been strong enough for statistically significant results. It does not say there are no continuous artifacts, that's a logical fallacy.

As for the statistically independent sources in ICA and connected brain regions: This is an interesting thing to think about, that's true. One answer would be that ICA has no temporal knowledge while separating the sources, it just works with point clouds, but the brain connections are temporal. One other might be that ICA doesn't separate based on correlation but more complex measures of the data set. I must admit I didn't have the time yet to fully understand the different ICA algorithms, though. My last answer to this is the fact that ICA gets nicely dipolar components which can be projected back to the brain, and this also works in simulations where the true sources are known. So for me, at least with sufficiently high density (64 at least, better 128) ICA components that have a neat separation from other components are brain sources. This has to be taken with a grain of salt though, since one needs to be experienced in understanding inhowfar a component might still be mixed sources. For example residual variance of the component gives at least a hint about this question. Unless proven otherwise, I will keep my working definition of ICA components as sources, and especially dipolar sources as brain sources, since I work on the source level anyway... 

Be it as it may, the 18-19 bitcoin example still doesn't hold true, you can just replace "sources" with "components". You made another strawman argument which I answered because it's an interesting question, not because it's an actual counter argument against my case.

Also, I think exactly the central-limit theorem is being used in ICA. A singular source is always less gaussian than mixed sources, so a de-mixing matrix which maximizes non-gaussianity of components would maximize the separation of the components from each other. I think your argument here goes exactly where the basic ICA argument goes so I don't see your point. See http://arnauddelorme.com/ica_for_dummies/



Makoto, yes, the dipole model is definitely far from perfect. We also attempt to create a head model including the eyes and neck muscles, since the head movements are vitally important for us. Blood vessels is an interesting point though, we hadn't thought about this, as far as I know at least. This would definitely improve localization of the muscle artifacts and simplify the discrimination of eye and muscle, but of course muscles are complex in their electrical properties so it's going to be an interesting question if this is working as intended. The distributed source model does appeal to me in the sense that it's just more realistic than a single dipole creating the signal at the scalp, but I am not familiar with the assumptions that are held for this.

Another big issue to me is that we don't get the same components for each subject. Clustering is a really bad way of combining the components, even though it's the best we have. In fact, I've realized that the clustering solution varies to a more-than-slight degree even if you re-cluster with the exact same parameters. I've thought of a method to do repeated clustering and compare the results to have a kind of probabilistic model slightly similar to Nima's Measure Projection, but it has proven to be more work than I had time. I might come back to it, though! As you said, a long way to go to make EEG a smoothly working brain imaging technique.. but all the more interesting and it's fascinating to be part of this!



Georges, I don't think anyone ever doubted that the data changes after one takes out eye components. It's obvious and I don't recall emails that state otherwise. The whole discussion was about the fact that the change in the data would increase the quality and make it closer to the actual brain data, that was even in Arno's very first post. It's reflected in the wording, for example you and robert use "distortion" or "adulteration", while Arno, Makoto and others from this ist use simply "change", because this implies no negative impact.

I think an important thing to keep in mind is that we can not really make any meaningful conclusions from sensor level activity as to where this activity originated from. Sensor-level connectivity will always have this drawback, plus the mixing in of artifactual sources, which is why Johanna has said a while ago that she performs only high-density EEG source-level connectivity analysis, nothing on the sensor-level at all. It obviously depends on your data set and application, but to me, 19 channel surface EEG is always to be treated with extra care as to any interpretations.



Tarik, yes, it seems to me that there is tons of method research to be done and paper to be written. We soon will have a synchronized data set containing high-density EEG with 28 as EMG on the neck, together with motion capture of the head, shoulders, hands and feet in position and orientation and eye tracking while moving around in virtual reality. This might provide useful for investigating the continuous artifacts, since we have eye movement, muscle activity, and (if we record this in addition) ECG all synchronized and we can look for it's connections to the EEG data. Well, the things to come! It all needs time and resources... Anyways, I would be glad to contribute to some of the papers if I can, since investigating the EEG methods is one of my main interests as a researcher.


All the best,
Marius

2017-07-28 20:23 GMT+02:00 Tarik S Bel-Bahar <tarikbelbahar at gmail.com>:

Greetings, it's great to see a more congenial future-focused discussion ​building here.​ We all have knowledge we consider as true, biases that impact our perceptions and expectations, and this all sometimes gets in the way of (needed) discussion on research topics and assumptions. It seems we often need less "fixed ideas and assumptions" and to be more open to "new published data-based findings addressing specific topics". Overall, kudos for keeping the ​human ​conversation going, it may eventually lead to positive impacts for the EEG field.
>From my perspective,​ ​argument​s​ or point of view, whether based on theory​, math,​ or experience,will ​remain ​un​satisfactory until th​ey are ​in ​1) multiple published new high-quality​ empirical​ ​papers ​that 2) examine a ​range of topics related to the discussion here about ICA, EEG, ​phase metrics, ​artifacts, etc... from a ​3) ​data-and assessment-focused point of view. ​It's important to have conversations, but it's new peer-reviewed publications that will truly clarify things.​
​I may be too optimistic, but more collaboration amongst the discussants here could generate a half-dozen papers over the next year or two focused on these issues. These are opportunities for both young flexible researchers, as well as more mature researchers and clinicians. 
One thing that has struck me is that the conversation has reflected ​1) ​a divide between low-density and high-density EEG systems users, ​2) the differences in researcher and clinician values, ​and ​that 3) there is enough data ​already ​out there to provide data-based publications addressing these topics. Further, the I think there is a lot of room for theoretical and methodological consolidation in the EEG field, but this will take time and effort, just as a clear understanding of the nature of true neural sources will take time and effort. 
Below, a resend of my earlier list of topics that future articles can consider​ or that we can add to/clarify​, and a list of articles (and authors) that ​can be consulted regarding current use/opinions of ICA, EEG, phase, and con​n​ectivity​ analyses​.


Possible dimensions/constraints to consider or compare?
Channel density and relative total-head coverage
​Low-density vs. moderate-density vs high-density​
With and without EOG and/or EMG recording ?​

Number/type of artifact​ual detected, number/type of​ ICs removed
​IC-based vs. non-IC based artifact removal​
Number/type of artifact visible in channel record​​

Clarity/robustness of artifact​ual ICs ​(e.g., ICs that are mixed vs. ICs that
are mixed (containing both artifact and neural info)

Channel-level vs. ICA-level vs. source​-estimated connectivity metrics

Length of epochs/trials

Type of source analysis

Type of reference (e.g., Chella et al., 2016)

Type of ICA/blind source separation (e.g., Bridwell et al., 2016;
Brain Topography)

Event-related or resting data

Signal quality (e.g., gel versus saline, very noisy vs. quite clean)

Reliability of particular metric

Type of connectivity metric (e.g., various kinds of phase measures,
and the plethora of other connectivity and graph theoretical measures)

MEG vs. EEG ?

Sampling rate?
Type of population ?​ Rate of artifacts per population/sample?
Individual differences in cleanness or clarity of EEG signal?




Sample articles that seem to use ICA in relation to connectivity
metrics are listed below.​ ​Moving forward, it may be beneficial to survey these authors and their findings.
de Pasquale, F., Della Penna, S., Sporns, O., Romani, G. L., &
Corbetta, M. (2015). A dynamic core network and global efficiency in
the resting human brain. Cerebral Cortex, bhv185.

Lai, M., Demuru, M., Hillebrand, A., & Fraschini, M. (2017). A
Comparison Between Scalp-And Source-Reconstructed EEG Networks.
bioRxiv, 121764.

Colclough, G. L., Woolrich, M. W., Tewarie, P. K., Brookes, M. J.,
Quinn, A. J., & Smith, S. M. (2016). How reliable are MEG
resting-state connectivity metrics?. NeuroImage, 138, 284-293.

Kuntzelman, K., & Miskovic, V. (2017). Reliability of graph metrics
derived from resting‐state human EEG. Psychophysiology, 54(1), 51-61.

Siems, M., Pape, A. A., Hipp, J. F., & Siegel, M. (2016). Measuring
the cortical correlation structure of spontaneous oscillatory activity
with EEG and MEG. NeuroImage, 129, 345-355.

Toppi, J., Astolfi, L., Poudel, G. R., Innes, C. R., Babiloni, F., &
Jones, R. D. (2016). Time-varying effective connectivity of the
cortical neuroelectric activity associated with behavioural
microsleeps. NeuroImage, 124, 421-432.

Farahibozorg, S. R., Henson, R. N., & Hauk, O. (2017). Adaptive
Cortical Parcellations for Source Reconstructed EEG/MEG Connectomes.
bioRxiv, 097774.

Rueda-Delgado, L. M., Solesio-Jofre, E., Mantini, D., Dupont, P.,
Daffertshofer, A., & Swinnen, S. P. (2016). Coordinative task
difficulty and behavioural errors are associated with increased
long-range beta band synchronization, NeuroImage.

Cooper, P. S., Wong, A. S., Fulham, W. R., Thienel, R., Mansfield, E.,
Michie, P. T., & Karayanidis, F. (2015). Theta frontoparietal
connectivity associated with proactive and reactive cognitive control
processes. Neuroimage, 108, 354-363.

Nayak, C. S., Bhowmik, A., Prasad, P. D., Pati, S., Choudhury, K. K.,
& Majumdar, K. K. (2017). Phase Synchronization Analysis of Natural
Wake and Sleep States in Healthy Individuals Using a Novel Ensemble
Phase Synchronization Measure. Journal of Clinical Neurophysiology,
34(1), 77-83.

Vecchio, F., Miraglia, F., Piludu, F., Granata, G., Romanello, R.,
Caulo, M., ... & Rossini, P. M. (2017). “Small World” architecture in
brain connectivity and hippocampal volume in Alzheimer’s disease: a
study via graph theory from EEG data. Brain imaging and behavior,
11(2), 473-485.

Ranzi, P., Freund, J. A., Thiel, C. M., & Herrmann, C. S. (2016).
Encephalography Connectivity on Sources in Male Nonsmokers after
Nicotine Administration during the Resting State. Neuropsychobiology,
74(1), 48-59.

Vecchio, F., Miraglia, F., Curcio, G., Della Marca, G., Vollono, C.,
Mazzucchi, E., ... & Rossini, P. M. (2015). Cortical connectivity in
fronto-temporal focal epilepsy from EEG analysis: a study via graph
theory. Clinical Neurophysiology, 126(6), 1108-1116.

Smit, D. J., de Geus, E. J., Boersma, M., Boomsma, D. I., & Stam, C.
J. (2016). Life-span development of brain network integration assessed
with phase lag index connectivity and minimum spanning tree graphs.
Brain connectivity, 6(4), 312-325.

Chung, Jae W., Edward Ofori, Gaurav Misra, Christopher W. Hess, and
David E. Vaillancourt. "Beta-band activity and connectivity in
sensorimotor and parietal cortex are important for accurate motor
performance." NeuroImage 144 (2017): 164-173.

Shou, G., & Ding, L. (2015). Detection of EEG
spatial–spectral–temporal signatures of errors: A comparative study of
ICA-based and channel-based methods. Brain topography, 28(1), 47-61.

Kline, J. E., Huang, H. J., Snyder, K. L., & Ferris, D. P. (2016).
Cortical Spectral Activity and Connectivity during Active and Viewed
Arm and Leg Movement. Frontiers in neuroscience, 10.

van Driel, J., Gunseli, E., Meeter, M., & Olivers, C. N. (2017). Local
and interregional alpha EEG dynamics dissociate between memory for
search and memory for recognition. NeuroImage, 149, 114-128.

Castellanos, N.P., Makarov, V.A., 2006. Recovering EEG brain signals:
Artifact suppression with wavelet enhanced independent component
analysis. J. Neurosci. Methods 158, 300–312.
doi:10.1016/j.jneumeth.2006. 05.033

Mehrkanoon, S., Breakspear, M., Britz, J., & Boonstra, T. W. (2014).
Intrinsic coupling modes in source-reconstructed
electroencephalography. Brain connectivity, 4(10), 812-825.
______________________________ _________________
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/20170804/f1e4e6ea/attachment.html>