[Eeglablist] Open collaboration project: Multi-model Anatomical Grouping of Independent Components (MMAGIC)

[Makoto Miyakoshi]

Hi Cedric,

I found that the ICA+CNN+Grad CAM used in Lee et al. (2024) to be an
interesting example of tensor decomposition. I once tried to do something
similar using PARAFAC and other tensor decomposition techniques explained
by Cichocki, but I could not make it work well. If it is often the case
that the differences we look for in our time-frequency data is rather
subtle and non-dominant. CNN+Grad CAM may be more suitable for the
channel/IC-frequency-time-epoch 4-D structure. I want to give it a try
someday. Thank you for sharing your experience and the paper.

This is a straightforward engineering projects to collect and integrate
existing solutions, more or less. It's basically Marjan's project. You are
welcome to join us!
I have some experience with e or s LORETA and MNE. I'm interested in this
project in the following points.

1. Testing the combination of ICA with distributed source model. Hristos
did one project with John back in 2017.
https://urldefense.com/v3/__https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2017.00180/full__;!!Mih3wA!CKgHMK3VIds16ckQCK9_b3B85zPBKbQerVDNllFzaC6ff7yfkBhdO40MWRW11v1rJY6RSbl7QpBUstpjE6kPNvjh3GU$ 
2. Comparing several distributed source models in one application. I'm
curious to see how sLORETA, eLORETA, MNE, and other things differ.

Nowadays, we can get help from Chat AI. This makes our lives much easier in
both mathematics and coding.

Makoto

On Tue, Sep 15, 2026 at 8:07 PM Cedric Cannard via eeglablist <
eeglablist at sccn.ucsd.edu> wrote:

> Hi Makoto,
>
> I have experienced the same thing recently. A ROI-based approach worked
> better for me than the more classic k-means clustering, where I would lose
> many subjects who did not meet the criteria for inclusion in the group
> analysis (in line with a recent question on this list). My context is
> heartbeat-evoked potentials (HEPs), where the components involved are not
> as obvious as in typical ERPs, since they are subtler responses to the
> heartbeat (and have to be separate from the cardiac field artifact
> component that has much greater amplitude and signal). The study aimed to
> track meditation (interoception) depth with them, which is not trivial and
> comes with its own phenomenological difficulties. Although better, the ROI
> approach still did not give successful results. Interoceptive regions can
> be deep and unreliable, so I also tried DMN regions to capture processes
> more related to mind wandering, which worked better but remained
> non-significant. I then tried a third approach that I call fun
>  ctional IC selection, inspired by this paper:
> https://urldefense.com/v3/__https://pubmed.ncbi.nlm.nih.gov/39159703/__;!!Mih3wA!C-b4A1m52zlfmgoSsSDmRbxcvQ9yFQy15BZxlffsKTjdYoQkaUrNfTEJ2pQLdE-oTrDml3CrEE_J-kOmnjrr0LEjtA$
>
> In brief, for each component we take the spectral power in the time window
> and frequency range associated with interoception in the literature (e.g.,
> 300 to 600 ms post R-peak, 4 to 20 Hz), then select the component (or the
> few top candidates) with the highest score, where the score combines that
> power with a surrogate control ensuring the activity is time-locked to the
> heartbeat. The group analysis is then run on the selected components. This
> is functionally individualized to each participant and is not circular,
> since the selection is independent of the condition of interest (meditation
> depth).
>
> The same logic could apply to classic ERP or task designs: any case where
> the component of interest has a well-established latency and frequency
> signature, but its topography or source location varies enough across
> participants that clustering either fails to converge or drops subjects.
> And because the criterion is functional rather than spatial, it can be
> defined in whichever domain best captures the process at hand, whether
> time, frequency, time-frequency, phase, or nonlinear measures, as long as
> it stays independent of the contrast being tested to avoid circularity.
>
> There are caveats, mainly around making the selection robust to non-brain
> variance and other biases, but I think it is an interesting approach.
> Feedback and criticism welcome.
>
> All that to say, I would be happy to take part in this new project. I also
> have experience with MNE, LORETA, and LCMV beamforming (roiconnect EEGLAB
> plugin), in case that is useful.
>
> Best,
> Cedric
>
> Sent with Proton Mail secure email.
>
> On Tuesday, September 15th, 2026 at 3:12 PM, Makoto Miyakoshi via
> eeglablist <eeglablist at sccn.ucsd.edu> wrote:
>
> > Hi Marjan and list,
> >
> > I'd like to propose a small open collaboration project.
> >
> > The idea is that we want to use anatomical ROI membership as a common
> > coordinate across subjects, instead of relying on conventional IC
> > clustering approach in EEGLAB STUDY. In the proposed approach, users
> > specify an anatomical ROI, optionally with a distance tolerance, and ICs
> > whose source models fall within that region are collected across
> subjects.
> > In this implementation, anatomical ROI membership replaces IC clustering
> as
> > the cross-subject correspondence rule.
> >
> > Here is a prototype I originally developed for a project with Ilaria
> > Berteletti, which uses standard current dipole models. Although we
> > ultimately did not use it in that study, I can tell you it worked better
> > than I thought.
> >
> >
> https://urldefense.com/v3/__https://github.com/MakotoMiyakoshi/MMAGIC__;!!Mih3wA!AoKTTZuSu2CVqhzgmxuMpdjqVNFyIVR-LQZYBeLZwVRfqta0reT9bYiTM3uHevgACv2_Bf8HE91bDX8kg7QX9GoBaAI$
> > <
> https://urldefense.com/v3/__https://github.com/MakotoMiyakoshi/AnatomicalROIGroupICA__;!!Mih3wA!AoKTTZuSu2CVqhzgmxuMpdjqVNFyIVR-LQZYBeLZwVRfqta0reT9bYiTM3uHevgACv2_Bf8HE91bDX8kg7QXcKIZKQg$
> >
> >
> > I cleaned the implementation to publish here and tested it on real
> > ICA/DIPFIT datasets; details are available in the repository.
> >
> > However, my recent work (under review) showed that more than 80% of the
> > pre-qualified brain ICs by IC Label were localized deeper than
> > physiologically plausible cortical source depths:
> >
> >
> https://urldefense.com/v3/__https://www.medrxiv.org/content/10.64898/2026.01.23.26344529v2__;!!Mih3wA!AoKTTZuSu2CVqhzgmxuMpdjqVNFyIVR-LQZYBeLZwVRfqta0reT9bYiTM3uHevgACv2_Bf8HE91bDX8kg7QXDaIe1jw$
> >
> > So it's time to try different cortex-constrained source estimates such as
> > MNE, the LORETA family, LCMV, or other methods. The results can then be
> > compared within a common anatomical ROI space. Hence MMAGIC: Multi-model
> > Anatomical Grouping of Independent Components.
> >
> > Marjan and I would like to develop this into a methodological/software
> > project. If anyone is interested in contributing a source-localization
> > method, atlas-mapping approach, validation with empirical data, or
> > methodological idea, you are welcome to join us. Please let us know.
> >
> > Makoto
> > _______________________________________________
> > To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu or visit
> https://sccn.ucsd.edu/mailman/listinfo/eeglablist   .
> >
> _______________________________________________
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu or visit
> https://sccn.ucsd.edu/mailman/listinfo/eeglablist  .
>