[Eeglablist] Beyond good and evil of ICA

Makoto Miyakoshi mmiyakoshi at ucsd.edu
Tue Aug 1 12:19:45 PDT 2017


Dear Marius,

> It was also different for different algorithms, meaning the areas were
not Brodmann, but different per algorithm used and the boundaries were
based on their inaccuracies etc. Do you know more of this and if so what is
your thought here?

In my humble understanding, dipole model is very simple. Meanwhile,
distributed source model is complex i.e. requires more parameters to make
it look nice. SCCN has also developed Sparse, Compact and Smooth (SCS)
algorithm etc, which I criticize too; I don't think 'compact' assumption
always holds.

It seems the biggest limitation of ICA+dipole approach is that when ICA
returns non-point source approximation with a single dipole naturally fails.

I have no experience in using distributed model (I admit that I have been a
'blind believer' on this point so far). My former colleagues, now at Qusp,
were positive to the distributed source models, but I'm not so sure.
Because I can easily imagine far more assumptions are necessary to make the
scheme work. My hunch is that both methods are equally bad :-) so I would
choose a simpler one. A nihilist's opinion.

 I say this because this field is still very primitive. One example is that
so far an electric forward model did not even have blood vessels, studies
of which started just recently (Fiederer et al., 2016) though all
radiologists have been using MR angiography for a few decades.

After all, EEG is still a adolesc-i-ence. We have many things to do!

Makoto



On Thu, Jul 27, 2017 at 4:17 PM, Marius Klug <marius.s.klug at gmail.com>
wrote:

> Dear Makoto,
>
> thank you for this thread, this is vital to keep investigations in actual
> problems of ICA and spatial filters active and on point!
>
> The spatial stationarity is especially a problem as soon as you
> investigate any kind of muscle, as you've shown once yourself (I believe?),
> which is important for MoBI research. But yes, also cortical traveling
> waves might pose a more than trivial problem... The way spatial filters
> are, it's not really solvable, is it?
>
> It also goes together with the problem of temporal instationarity. It
> would effectively mean that we need a new filter for each frame of the
> traveling wave (for max resolution), which would disable any meaningful
> interpretation the way its done right now, right? A hard nut to crack for
> sure... As for different models for different tasks: I think this is also a
> result in itself. It shows nicely that the brain has different working
> modes, or states, if it solves different tasks. We thought of taking
> multiple models of the AMICA into account, but stick with one for now until
> we have everything working smoothly. Interesting thing!
>
> And for the last point: I heard a talk a while ago where source
> localization wasn't done based on dipoles but on specific areas most likely
> to be the origin of a filter. It was also different for different
> algorithms, meaning the areas were not Brodmann, but different per
> algorithm used and the boundaries were based on their inaccuracies etc. Do
> you know more of this and if so what is your thought here? I'm afraid I
> cannot do more research at the moment since I don't have access to a PC for
> a while (I'm writing on my phone)...
>
>
> Robert,
>
> those quotes show that we all agree that the phase is different if data is
> taken out. This is no surprise at all, since, obviously, data was taken
> out. The claim is that the data that was taken out was artifact. As I've
> written, it might be that this claim does not hold true perfectly with 19
> channels BUT this is untested. It may be that the changed phase is the
> better, i.e. more "brainy" phase, and results for tests are more precise.
> It may be that they are not, because a sufficiently large amount of brain
> data has slipped into the eye component (not impossible with 19 channels).
> The latter would be a problem, and investigating this would be important
> (writing the same things on the eeglablist all the time not so much) and it
> would answer Andy's second question.
>
> The 19-18 bitcoins example does not hold true either. One does create the
> same amount of channels from fewer sources. Not from fewer channels. The
> latter would be interpolation and is also regularly used, by the way. The
> former takes specific sources out and simulates a recording without these
> sources. Both final data matrices have a reduced rank. It is more so that
> the EEG contains a few gold nuggets (brain sources) and many stones
> (artifact sources) crushed and mixed together and then placed in a few
> bowls (channels), and ICA tries to separate those and created the most
> accurate representation of gold nuggets and stones it can. The most obvious
> stones are then taken out and the remaining gold and a smaller amount of
> stone powder is then placed back into the bowls again. They are a little
> less filled then (rank reduction) and the quality of the powder in the
> bowls is different (phase has changed) because a stone is missing. If there
> are only 19 channels (and also with high-density, but less gravely so) the
> ICA cannot separate gold from stone perfectly so there is a risk of
> throwing away some gold. There are other algorithms (spatio-spectral
> decomposition, SSD, for example) that separate based on different criteria,
> but in the end all are the same train of thought: Unmixing the mix of gold
> and stone. If one does not think there is stone in the bowls one of course
> thinks all changes lower the value. This is an assumption at least as big
> as the assumptions underlying ICA...
>
> I hope this is a nice example and helps all readers understand the topic a
> little better! Andy, has this helped you with your first question? I cannot
> contribute to the second, but I do think it is important to check!
>
> Best,
> Marius
>
> Am 27.07.2017 20:53 schrieb "otte georges" <georges.otte at pandora.be>:
>
> Dear Andrew
>
>
>
> As a clinical neuropsychiatrist I completely share your concerns as they
> boil down to the question: “do we need to conserve the information
> contained in the raw EEG data recorded at the scalp  or will cleaned up
> (ICA reconstruction) data will provide a better view on the brain
> functioning?”
>
>
>
> I have tried to address this in a mail at the beginning of this month but
> it got probably lost. See below.
>
>
>
> Basically, for me (but that is a personal view) the time series of EEG
> recorded at the scalp contain information about network node communication
> in health and disease. This information is contained in the phase
> differences between the sources (brain sources). As we all know EEG can be
> contaminated by activity originating from non-brain sources or even
> technical (non biological) causes.  Eliminating parts of the EEG that are
> contaminated by those elements is a daily job (as is taking care on
> recording situations with patients so that artefacts are minimal).
>
>
>
> There is a large class of methods available for elimination of artefacts.
> The classical ones keep the phase content (phase lag differences) between
> the channels intact and some may not do that.
>
>
>
> Personally, I would avoid methods that alter this information and prefer
> methods that do not.
>
>
>
>
>
> Sincerely
>
>
>
> Georges
>
>
>
> There is a major evolution in modern neuropsychiatry that aims at linking
> clinical symptoms to brain network dysfunctions. While this approach was
> successful in grounding neurological symptoms to structural pathologic
> alternations in brain networks, in psychiatry the main momentum was not
> structural but functional network dysfunctions. While fMRI was the pioneer,
> the much better time resolution of MEG and EEG made them the preferred
> tools. Their output ( time series) is but a means to further construct a
> functional image of the networks involved where phase dynamics teach us the
> directionality of the information flow in the network nodes and allows us
> by comparison with a database of normal values what functional
> abnormalities can be detected. For me phase integrity in the data is thus
> very important to be able to construct valuable graph theory models of
> those networks be them dysfunctional or compensatory. Much work has been
> devoted on this topic since many decades by DrThatcher but also by many
> other authors such as Vinod Menon ( Stanford) linking psychiatric symptoms
> to specific network dysfunctions. For us, clinicians this introduces a new
> approach to neuroscientific psychiatry that links psychiatry back to it's
> neurobiological roots and can hopefully one day send the DSM categorization
> to the museum of the history of psychiatry.
>
> As phase is IMHO a most important parameter in order to establish the
> network internode information flow, it should not come as a surprise to
> hear that some find phase unimportant as contaminated by continuous
> artefact or hear about ICA’ s a signal reconstruction method that presents
> the danger of changing the phase dynamics in the original time series
> especially in low channel (19ch) recordings with perhaps more prominent
> effect due to overcompleteness.
>
> If in a 19 ch. EEG a clinician rejects (nulls out the rows of the mixing
> matrix ) ICAas components for blinks EMG, pletysmo and ecg ( 4 ) and then
> does a "reconstruction"  ( creating 19 channels  out of 15 ??) what we then
> get might look nice but is IMHO  not a valid base for a graph theoretical
> model of the underlying brain network.
>
> I think this is the reason this discussion is important and certainly not
> a trivial pro or contra ICA pugilism.
>
>
>
> Sincerely
>
>
>
> Georges
>
>
>
> *From:* Andrew Smart [mailto:andrew.johnsmart at gmail.com]
> *Sent:* Thursday, July 27, 2017 12:43 AM
> *To:* mmiyakoshi at ucsd.edu
> *Cc**:* Robert Thatcher <rwthatcher2 at yahoo.com>; EEGLAB List <
> eeglablist at sccn.ucsd.edu>; Georges Otte <georges.otte at telenet.be>
> *Subject:* Re: [Eeglablist] Beyond good and evil of ICA
>
>
>
> Hi all,
>
>
>
> I am somewhat of an outsider in this discussion so forgive my limited
> understanding as I haven't worked on EEG and ICA for a few years, but I am
> fascinated by this debate and many thanks for the clear and reasoned
> arguments from all sides. I have since worked in clinical science and FDA
> regulated areas with sensors and sensor data - and so have some familiarity
> with the validation required for example to use sensor data as a clinical
> endpoint.
>
>
>
> I have two questions regarding this discussion that I am not understanding
> entirely:
>
>
>
> 1) The idea of phase distortion as opposed to "true" brain phase. I would
> like to understand better what the arguments are for saying that the phase
> of the raw channel data is the ground truth (for lack of a better phrase)
> and that ICA distorts this "true" phase (this is my understanding of one
> side of the debate). It seems all agree that ICA changes the relative phase
> of the channel data - but the debate is about whether this is in fact a
> distortion? I.e., is the raw channel data somehow a better representation
> of the "true" electrical activity of the brain? It seems like the crux of
> the debate is whether the raw EEG is "truer" than the ICA cleaned data -
> from my perspective it seems like the ICA reconstructed time series is
> closer to whatever "true" underlying brain signals are contributing to the
> scalp recording.
>
> Another way to ask maybe: I don't understand what we're supposed to be
> comparing ICA phase to and why it's a distortion? A distortion of what? One
> way of looking at it is that ICA is actually correcting the phase by
> removing artifacts, not distorting it - is that fair?
>
>
>
> 2) Has anyone filed a 510k to FDA for example using ICA on EEG data for a
> medical purpose? I.e., where the intended use of the ICA results is to
> diagnose or treat neurological disease? My question is really - what is the
> clinical relevance of this discussion?
>
>
>
> Many thanks,
>
> Andy
>
>
>
> On Wed, Jul 26, 2017 at 2:40 PM, Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> wrote:
>
> Dear Robert,
>
>
>
> I want to know all publications that makes a clear claim that 'ICA
> distorts phase'. I will include all of them for our clarification paper. So
> far I know Montefusco-Siegmunt et al. (2013) is the only paper that makes
> this invalid claim. If you know other papers, please let me know.
>
>
>
> Again, you are calling the difference between 18 Bitcoins and 19 Bitcoins
> 'distortion'. It's a due change. See the pages below.
>
> https://sccn.ucsd.edu/wiki/How_phase_is_calculated_in_linear_decomposition
>
> https://sccn.ucsd.edu/wiki/ICA_phase_distortion
>
>
>
> Georges, Ramon told me that all the posts were published on the list. If
> otherwise, please let us know. Sorry for the trouble.
>
>
>
> Makoto
>
>
>
> On Wed, Jul 26, 2017 at 12:06 PM, Robert Thatcher <rwthatcher2 at yahoo.com>
> wrote:
>
> Dear Makoto,
>
>    I think your criticisms are important and note that there are traveling
> waves in the EEG and also there is nonlinearity in the form of wave
> dispersion as noted by Nunez, 1981 and demonstrated in the paper that can
> be downloaded at this url (see Table IV):
>
>
>
> http://www.appliedneuroscience.com/TWO-COMPARTMENTAL_MODEL_
> EEG_COHERENCE.pdf
>
>
>
> It seems that your 3rd criticism does not recognize that ICA
> reconstruction of a new time series violates the "Reciprocity" theorem of
> Helmoltz and the "Lead Field" necessary for a valid inverse solution.
>
>
>
> You mentioned a recent criticism on ICA that you stated is "technically
> invalid".   I doubt that you are referring to the criticism about ICA
> reconstruction adulterating phase differences between EEG channels?   The
> issue of ICA reconstruction and phase alteration is a settled issue based
> on math (not the separation of mixtures of phase or frequencies but rather
> the cross-spectrum at the same frequency at different locations) as well as
> multiple empirical demonstrations and tutorial demonstrations that anyone
> can verify for themselves.  Also, I am copying from the Eelablist
> statements by yourself and five others agreeing that ICA reconstruction
> alters phase differences.
>
> “If you remove IC and reconstruct channel EEG by back projecting the
> remaining ICs, of course it changes channel EEG phase!” (Makoto Miyakoshi,
> Eeglablist ICA and signal phase content, Sept. 16, 2014)
>
>
>
> “The EEG reconstruction after removing bad components/sources MAY change
> the phase value of the signal at any electrode.” (M. Rezazadeh Eeglablist
> ICA and signal phase content, Sept. 18, 2014).
>
>
>
> “The reconstructed data after removing spurious ICA components differs
> from the original time series, and because of that there are phase
> differences.” (Arnaud Delorme, Eeglablist ICA misinformation, June 10,
> 2017).
>
>
>
>   “I first noticed the problem with phase distortion more than a decade
> ago” (Robert Lawson, Eeglablist ICA misinformation, June 14, 2017).
>
>
>
> “I think Bob is right that the relative phase will be changed by deleting
> 1 or 2 artifact components.” (Ramesh Srinivasan, Eeglablist ICA
> misinformation, June 14, 2017).
>
>
>
> “We found phase distortions in the 8-10 Hz alfa band (greatest near the
> source of artefact) but also on more remote electrodes such as occipital
> and also in artefact free strokes of EEG.” (Georges Otte, Eeglablist ICA
> misinformation, June 15, 2017).
>
> Best regards,
>
>
>
> Robert
>
>
>
> On Wednesday, July 26, 2017, 2:01:59 PM EDT, Makoto Miyakoshi <
> mmiyakoshi at ucsd.edu> wrote:
>
>
>
>
>
> Dear List,
>
>
>
> Recently there was a criticism against ICA on the list. Unfortunately it
> is technically invalid so I remained unsatisfied. Let me share real
> problems of the ICA model (Onton and Makeig 2006) to re-do it. This is a
> continued discussion from the one titled 'How phase is calculated in linear
> decomposition' and now this is my turn to criticize ICA!
>
>
>
> As far as I know, there are three known limitations in ICA model.
>
>    1. Spatial stationarity. I have seen a nice traveling waves in ECoG
>    grid data during Joaquin Repela's presentation at SCCN. This clearly
>    violates the assumptions of spatial stationarity in ICA.
>    2. Temporal stationarity. Shawn Hsu at SCCN presented time-series data
>    of ICA model likelihood during drowsy driving task. Also, Jason Palmer's
>    AMICA also demonstrated temporal changes in model likelihood. So one model
>    per data does not fit the truth (unless the task has a strong control over
>    a subject's cognitive and behavioral states).
>    3. Dipolar source model. Although most of ICA results are fit with
>    dipole models, it seems ICA also returns (probably) non-point sources. When
>    one fits a dipole model to such a non-point source, the location tend to
>    end up with physiologically invalid depth (this is the most annoying thing
>    about ICA today)
>
> I'd like to hear detailed criticism about these points. Note I saw these
> critical counterevidence in SCCN; we are not a boring ICA cult who have
> blind belief in it.
>
>
>
> Nonetheless, ICA model has a critical merit. I named it *Independence-Dipolarity
> Identity (I-D Identity, or IDId)*. I-D Identity means that when ICA
> solves temporally problem, it also solves spatial problem at the same time *without
> using ANY spatial constraint*. Dipolarity can be thought of, in short, *biophysical
> origin-ness*. Hence I believe that this is evidence that ICA hits *some *physiological
> truth of EEG generation.
>
>
>
> There could be multiple criticisms against the limitations of ICA model,
> but at the same time any criticism, at least so far, was NOT strong enough
> to deny *I-D Identity *of the ICA model*. *After all, because of this *I-D
> Identity*, I still advocate ICA (but similar dipolarity can be achieved
> by using very different approach, such as SOBI... so independence is not
> the only requirement to reach the biophysical validity. It's still a
> mystery to me.)
>
>
>
> All models are wrong, but some are useful... but I want to go beyond this
> statement to reach the ground truth of EEG!
>
>
>
> --
>
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
>
>
>
>
>
> --
>
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
>
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.uc
> sd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>
>
>
>
>
> --
>
> Andrew Smart
>
>
>
> Author
>
> OR Books <http://www.orbooks.com/andrew-smart/>
>
>
>
> News Editor
>
> Idler <http://idler.co.uk/>
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.uc
> sd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>
>
>


-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170801/dbb19ba8/attachment.html>


More information about the eeglablist mailing list