[Eeglablist] [External] Re: Source analysis trade-offs: ICA + ECD versus distributed methods

Tue Jan 12 21:24:21 PST 2021

Dear Justin,

> But, I think the nature of my question might not have been clear.

Yes, it was not very clear to me. But I appreciated your enthusiasm.

> what is lacking imho is an earnest comparison of different source methods
in terms of utility, interpretational capacity, ground truth comparisons
across all types of methods (e.g., a phantom study).

If you say you read Electric Field of the Brain (EFB), I hope you did not
miss this part: ''If one type of data transformation suggests genuine brain
properties, we should be able to find supplementary computer methods that
yield similar physiological or cognitive conclusions. Otherwise, we run the
well-known scientific risk that our conclusions are no more than the
summation of assumptions made in the process of choosing experimental and
analysis methods. For example, some physiological interpretations have
depended on the unwarranted assumption that a few isolated sources generate
EEG. Another example is the implicit assumption that the brain is a system
with only a few degrees of freedom when interpreting measures of chaotic
dynamics." (EBF p.354-355)

The principle of functional brain map tells us that each brain region is
specialized to some specific functions. However, to the eyes that saw too
many nice activation maps in neuroimaging studies, watching a movie of raw
ECoG signal activation rendered on the cortex is purely mind boggling (was
it not?) We become surprised to see it because we realize that the speed
and complexity of the genuine brain dynamics is beyond our imagination.

Here is an interesting story--My colleague once said to me 'Where are the
independent components in this ECoG movie?' I liked this question as a
humble expression of curiosity at that time, but now I like it for another
reason: This symbolic question, which resonated in my mind for years since
then, finally brought one clear answer to me--independent components are
not in the brain.

Getting back to your discussion, what does it mean to compare equivalent
current dipole models with distributed source models?
Let me tell you--it compares their assumptions. It does not compare their
performance in terms of anatomical accuracy.

I generally agree to your opinion that using fMRI-based ROI is useful for
analyzing EEG, despite the cautions raised by the authors of EFB. We may
want to use as much prior information as possible, it is worth trying. Our
colleague Lars Kai Hansen once proposed a similar idea, a key hole
hypothesis (https://urldefense.com/v3/__https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5492868/__;!!Mih3wA!VpfjbanSu4M3WT7FQBGr_NEBib4APVhNTRsprLjkJYEItfdFbvxBmluGWbSSvImYaQDABQ$ ) But
again, we want to be careful when we say 'localize' or 'decompose' the 'EEG
sources': if we are implicitly or explicitly looking for an isolated,
stationary, independent generator as an EEG source, we may be wrong from
the beginning. It is worth remembering that the authors of EFB did not
choose to localize EEG sources because of radical skepticism on this point.
The authors actually criticized the popular implicit assumption that EEG
source are localizable.

Finally, this is my recent solution--talk only about models, and be always
ready to explain the difference between the model and the reality. Even if
your ruler is bent, as long as you can explain to others how bent it is,
there is a chance that your measurement is still meaningful. If you meet a
person who speak of EEG's source localization accuracy as if it were some
kind of neuroimaging modality, carefully ask them about his assumptions on
source distribution and dynamics.

Makoto

On Mon, Jan 11, 2021 at 7:07 PM Fine, Justin Michael <justfine at iu.edu>
wrote:

> To Makoto (and eeglab peeps):
>
> I appreciate the thoughtful response alongside Scott’s.  But, I think the
> nature of my question might not have been clear.
>
> Just to build background: I have read Nuñez, cohen’s book, Steve luck’s,
> the whole gamut. My question wasn’t about the indeterminacy of inverting a
> forward model.  That issue is clear.  For example, I regularly build custom
> neurostimulation optimization routines.  So the issues of solving for the
> inverse was not the question neither the issues of field cancellation of
> dipoles.  I’m saying this to clarify what the question is not. There’s a
> lot of material on this, what is lacking imho is an earnest comparison of
> different source methods in terms of utility, interpretational capacity,
> ground truth comparisons across all types of methods (e.g., a phantom
> study).
>
> But first, you state well why a priori sources
> Matter. Presumably the confidence in solving an inverse problem is always
> bolstered by putting more functional (e.g., functional localizer) and
> physiological constraints on a problem.  Leading into my question, but
> first, given the advent of available meta-analysis from fmri (e.g.,
> neurosynth), it would seem a good idea for eeg source analysis should be to
> always make a meta-analysis map of a priori regions (for a given task or
> concept, e.g., response inhibition) from fMRi data.  This places a strong
> constraint on spatial locations of, say, ECDs that would dominate variance
> in scalp maps. What to make of this idea?
>
> My main question was more practical in nature, given a certain level of
> resources (ranging from a standard study with only scalp eeg to one with
> T1s and recorded electrode positions, and an fmri prior map of activations)
> what source methods are preferable? Specifically, when the goal is to a
> priori examine how certain anticipated sources, such as rifg  (confirmed in
> fmri with 100s of studies), which method offers a good trade off: at
> simplest ICA + field trip style dipfit, to Bayesian constrained dipole
> analysis,  to distributed methods like eLORETA or nonlinear ones like MSP?
> What are the tradeoffs in interpretation? Why prefer individual bem/fem
> meshes versus canonical meshes (a la spm).  For example, I think only 3-5
> sources really drive the dominant activity during the time window of
> interest in my data.  Thus, it seems like Bayesian ECD is best because it
> allows directly testing this in a statistical model framework (see VB-ECD
> in spm).
>
> So, which method should someone consider first given these types of
> resources at hand and question?
>
> A final point/question  I would like to bring up about model constraints:
> it seems the strongest also be model-based Temporal constraints. I have
> dynamic causal modeling in mind here. Thoughts on this, regardless of
> connectivity, but really about neural mass like models as a generator of
> predicted eeg sources as a means to decompose scalp activity.
>
> I look forward to discussing this more.
> Cheers,
> Justin Fine
>
>
> Justin M. Fine, PhD
>
> > On Jan 11, 2021, at 2:46 PM, Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> wrote:
> >
> > This message was sent from a non-IU address. Please exercise caution
> when clicking links or opening attachments from external sources.
> > -------
> >
> > Dear Justin,
> >
> >> why would someone prefer finding a priori known sources in EEG with a
> > given source methods?
> >
> > It is because the inverse problem of the head volume conductor does not
> > have a unique solution. Because of this fundamental limitation, there can
> > be multiple reasonable assumptions about the spatial distribution of
> > 'source of EEG' to reach the unique solution (to a reasonable degree,
> given
> > the assumption).
> >
> > In the 'Electric Field of the Brain' (2006) by Nunez and Srinivasan, they
> > advocate an alternative approach in which you do get a unique solution in
> > estimating EEG source distribution at the cost of resolution in sulci. In
> > other words, you accept the limitation that you will only analyze
> > continuous gyri with no sulci and you get an unique solution. The
> > justification for this trade-off is that apparently potentials cancel
> each
> > between the two cortices facing to each other at a sulcus. You can find
> > quantitative evaluation on this in their book.
> >
> > In sum, what you can give up determines what you can say for sure. But
> the
> > critical difference concerns the qualitative differences, such as whether
> > the solution is unique or non-unique. For non-unique solutions, you may
> be
> > able to take advantage of electrophysiological facts by incorporating
> > sparticy and/or smoothness of the 'EEG sources' (such as
> > sparse-compact-smooth constraint by Chen used in Zeynep's SCALE, not
> SCORE)
> > as well as dynamics (such as Stefan Haufe's method that can be used for
> > SIFT which Tim Mullen always talk highly of) to 'improve' the result, or
> > better to say, 'inform' the result. Hence the goodness of the result
> > depends on the goodness of the assumptions.
> >
> > I'm not super knowledgeable in this topic but do have a strong interest
> to
> > learn. I'd be happy to discuss this topic with you.
> >
> > Makoto
> >
> >
> >
> >> On Sat, Jan 9, 2021 at 9:46 AM Fine, Justin Michael <justfine at iu.edu>
> wrote:
> >>
> >> Dear list:
> >>
> >> I have a question that the literature does not really seem to answer
> >> regarding source analysis: why would someone prefer findings a priori
> known
> >> sources in EEG with a given source methods? Specifically, I am asking
> about
> >> the benefits and obvious trade-offs if (1) ICA + ECD ,(2) distributed
> (and
> >> sparse or group hierarchical methods) source methods (e.g., MSP in SPM)
> or
> >> (3) a Bayesian ECD approach which does not rely on fitting separate IC
> >> components but relies on specifying a prior source locations?
> >>
> >> Quick background, I have T1s and recorded electrode positions (64
> channel
> >> acticap) for all participants.  The main goal here is extracting
> >> time-frequency and evoked (ERP) from an rIFG, pre-SMA/ACC/MCC, and left
> M1
> >> source.The study was a standard stop signal task, of which the
> literature
> >> tends to prefer the method (1) of ICA + ECD. But I gather that might
> have
> >> something to do with researchers typically (1) not having electrode
> >> locations and (2) T1s?
> >>
> >> Any thoughts and feedback would be greatly appreciated.
> >>
> >> Thanks!
> >> Justin Fine
> >> Post-doctoral researcher
> >> Indiana University
> >> -----Original Message-----
> >> From: eeglablist <eeglablist-bounces at sccn.ucsd.edu> On Behalf Of Scott
> >> Makeig
> >> Sent: Friday, January 8, 2021 2:32 PM
> >> To: JULIANA CORLIER <corlier at g.ucla.edu>; Johanna Wagner <
> >> joa.wagn at gmail.com>
> >> Cc: eeglablist at sccn.ucsd.edu
> >> Subject: [External] Re: [Eeglablist] Analysis of TMS-induced harmonics
> in
> >> the EEG
> >>
> >> This message was sent from a non-IU address. Please exercise caution
> when
> >> clicking links or opening attachments from external sources.
> >> -------
> >>
> >> Juliana -
> >>
> >> These will also soon be a new toolbox, the 'Independent Modulator
> Analysis
> >> Toolbox' (IMAT), that can separate harmonic from non-harmonic activity.
> If
> >> you might like to test its use, write Johanna Wagner <
> joa.wagn at gmail.com>.
> >>
> >> Scott Makeig
> >>
> >>> On Thu, Jan 7, 2021 at 3:20 PM JULIANA CORLIER <corlier at g.ucla.edu>
> wrote:
> >>>
> >>> Dear list,
> >>>
> >>> I would like to get some expert advice on how to assess/quantify the
> >>> presence of harmonics in the EEG.
> >>> Notably, our lab is using EEG recordings during repetitive
> >>> transcranial magnetic stimulation (TMS) and we would like to assess
> >>> whether the stimulation at a certain frequency elicits an entrainment
> >>> at the stimulation frequency but also at other frequencies.
> >>>
> >>> My first approach was to check for ‘ratios of frequencies’ that show
> >>> activation in the time-frequecy domain post stimulation, but that
> >>> turned out to be a more tricky than I have aniticipated.
> >>> I was wondering if there is a proper approach to harmonics analysis?
> >>> One would think that engineers and signal processing experts outside
> >>> of neurosciences deal all the time with that.
> >>>
> >>> Any advice is much appreciated!
> >>>
> >>> Thank you!
> >>>
> >>> Juliana Corlier
> >>>
> >>>
> >>>
> >>> _______________________________________________
> >>> 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
> >>
> >>
> >>
> >> --
> >> Scott Makeig, Research Scientist and Director, Swartz Center for
> >> Computational Neuroscience, Institute for Neural Computation,
> University of
> >> California San Diego, La Jolla CA 92093-0559,
> http://sccn.ucsd.edu/~scott
> >> _______________________________________________
> >> 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
>