[Eeglablist] Let's test whether GEDAI is a post-ASR EEG artifact rejection champion

[Makoto Miyakoshi]

Thank you Vikram for sharing your experience and opinion.

> I've used ASR for several years now, and it was a core part of my
preprocessing steps.

May I ask parameters, particularly the k value (i.e., cutoff threshold in
SD)?

> IClabel to automatically label the components and filter out the
artifacts. I usually keep relatively conservative settings: only ICs with
brain probability > 0.5 are retained.

It is possible that after GEDAI you don't use IC rejection and you can
retain more (non-ICA-decomposed) brain signals, particularly those with low
amplitudes.

Castellanos NP, Makarov VA. (2006).  Recovering EEG brain signals: artifact
suppression with wavelet enhanced independent component analysis. J
Neurosci Methods. 2006 Dec 15; 158(2) 300-312
DOI: 10.1016/j.jneumeth.2006.05.033, PMID: 16828877

> Can anyone please suggest whether this pipeline looks okay overall or can
it be further optimized?

You should band-pass filter the data at the beginning of your process so
that you do not need to use ZapLine. It should also help to perform your
bad channel removal (unless your recording is already
well-high-pass-filtered at 0.5-1.0 Hz, which is unlikely anyway) My
suggestion is:

* bandpass filtering (0.5 Hz-50 Hz)
* Bad channel removal
* GEDAI on the continuous data
* Bad channel interpolation (to prevent effective rand deficiency at the
previous GEDAI stage; see my ICA's bug paper)
* ICA + * IClabel (NOT for component rejection)

> I have not tried custom BEM on our datasets yet, as I was satisfied with
the default model. If anyone can confirm whether custom lead fields help,
I'd give it a try (we already have MRI+3D digitized positions for some
participants).

In the case of REST, it is not (very) sensitive to the choice of the
forward model, if I remember correctly.
Even if you do your best with MRI-derived morphometry + digitized electrode
locations, there are (at least) two other invincible limitations: tissue
conductivity and presence of parietal (and possibly other) foramens.
https://urldefense.com/v3/__https://en.wikipedia.org/wiki/Parietal_foramina__;!!Mih3wA!DWnnv7lCFyiURCxsJ9d6iuSeSBc8ex9Vn0NKqDin52tAOevJn7ZJxJ_qcH-xoqKZuDO8dfH5CzPtz-hQkACt-98WZdc$  I think the issue of
foramen alone is sufficient to make me pessimistic about obtaining an
accurate EEG forward model, not only because it is technically difficult to
address but also because no one knows it.

Makoto


On Thu, Apr 2, 2026 at 1:03 PM Vikram Shenoy Handiru <vikkyshenoy at gmail.com>
wrote:

> Hello all,
>
> First of all, thank you Makoto for kickstarting this discussion about
> GEDAI :)
>
> I routinely use GEDAI in my analyses (includes TMS-EEG, EEG during
> walking, EEG during balance perturbation, etc.) and I must say
> GEDAI-preprocessed data looks pretty good.  For context - I've used ASR for
> several years now, and it was a core part of my preprocessing steps. Having
> compared the results of ASR and GEDAI, I'm inclined to say GEDAI appears to
> be better.
>
> This is my usual pipeline (that has worked for most datasets):
>
> Data acquisition: Brain products ActiChamp (64-ch)+ Brainsight
> Neuronavigation for 3D digitization of electrode positions+MRI (whenever
> available) - my research involves EEG source reconstruction.
>
> *Preprocessing steps:*
> * Bad channel removal + interpolation (pop_rejchan + pop_interp)
> * Zapline for line noise (60 Hz) removal  (I found this to be better than
> cleanline).
> * bandpass filtering (0.5 Hz-50 Hz)
> * GEDAI on the continuous data (usually, with auto or auto- setting,
> depending on the nature of data recording). Caution: enabling parallel
> processing option can be counterproductive for long recordings in my
> experience.
> * ICA (cudaica implementation of extended infomax)
> * IClabel to automatically label the components and filter out the
> artifacts. I usually keep relatively conservative settings: only ICs with
> brain probability > 0.5 are retained.
>
> I have not tried custom BEM on our datasets yet, as I was satisfied with
> the default model. If anyone can confirm whether custom lead fields help,
> I'd give it a try (we already have MRI+3D digitized positions for some
> participants).
>
> Can anyone please suggest whether this pipeline looks okay overall or can
> it be further optimized?
>
> Thanks and Regards,
>
> *Vikram Shenoy Handiru, PhD*
>
> Research Scientist and Director - NeuroMuscular and Electrophysiology
> Laboratory (NMEL)
> Center for Mobility and Rehabilitation Engineering Research
>
> Kessler Foundation  | 1199 Pleasant Valley Way, West Orange, NJ 07052
>
> *T.* +1.973.324.3578 |* F*. +1.973.324.3527 |
> VShenoy at KesslerFoundation.org
>
> *Affiliation:* Research Assistant Professor,
>
> Department of Physical Medicine & Rehabilitation | Rutgers, New Jersey
> Medical School
>
>
>
> On Mon, Mar 30, 2026, 18:09 Makoto Miyakoshi via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
>> Hi EEGLAB mailing list subscribers,
>>
>> I watched this Youtube video on GEDAI, a relatively new EEG artifact
>> rejection algorithm, and got very impressed. If you are interested in, or
>> looking for, an effective artifact rejection method, you definitely want
>> to
>> check it out.
>>
>> https://urldefense.com/v3/__https://www.youtube.com/watch?v=qSM5narynzc__;!!Mih3wA!CLlPjuH-5y0g4OtnG1F8KKjQoNScMgDneZVoRYZ1yoUGb_ySMhoAaf4jO0oLiDxE_IZRlBbTXuSFkSA4XgNw9TOfbZ8$
>> Do not miss the scene in which Tomas shows a comparison before and after
>> removing the TMS-induced artifact (about 20 min position). I've never seen
>> a more dramatic demo than this. It is only comparable to successful demo
>> of
>> gradient artifact removal from fMRI-EEG recording.
>>
>> Here is a suggestion to this community:
>> If you have been already using ASR, why don't you please use GEDAI as well
>> in parallel, compare the results, and share your impression with us? I'll
>> do so in my next project and report it here.
>>
>> The reason why I like GEDAI is that the concept of combining lead field +
>> GED is elegant. The idea of 'learning from data's own cleanest part' is
>> the
>> same as ASR, but GEDAI does it more explicitly and simply. GEDAI is like
>> ASR + REST (a reference method using a forward model).
>>
>> Makoto
>> _______________________________________________
>> To unsubscribe, send an empty email to
>> eeglablist-unsubscribe at sccn.ucsd.edu or visit
>> https://sccn.ucsd.edu/mailman/listinfo/eeglablist .
>>
>