[Eeglablist] Order of Channel Removal+Interpolation and ICA (removing noisy components)

Thu May 28 11:58:49 PDT 2026

Hi Marshid and Tim,

Thank you for your comments. It's my honor to go against Claude's advice!

In fact, I recognize Claude's workaround works fine for a specific purpose:
it uses ICA purely as an electrode signal cleaner at the cost of
IC-electrode correspondence i.e., your EEG.icawinv will be invalidated. As
a result, certain types of analyses become impossible, such as
envelope-topography (envtopo) at the group-level analysis, because matrix
dimensions do not match across datasets.

Here is the step by step examination.

1. We all agree that bad channels need to be excluded before ICA. Suppose
that we reject n channels here.

2. We run ICA on this channel-reduced data. As a result, the reduced number
of electrodes are registered to your ICA-generated matrices
(EEG.icaweights, EEG.icasphere, EEG.icawinv, EEG.chaninds).

3. After ICA, you interpolate rejected channels to recover your original
EEG.nbchan. Technically, you can do it. Probably you see no error message
in EEGLAB. However, this interpolation does not take care of ICA-generated
matrices (as far as I know). As a result, you can't perform IC rejection
for electrode signal cleaning because the number of channels of your scalp
recording and of your ICA-generated matrices do not match. You do see an
error message here.

4. The easiest way (i.e., without developing a reasonable workaround and
publishing it a dedicated technical paper to address this specific problem)
to avoid this problem is to perform channel interpolation BEFORE ICA so
that all the original electrodes are registered to ICA-generated matrices
so that you can perform IC rejection etc. after ICA. The drawback is that
you need to use PCA (or whatever) dimension reduction to run full-rank ICA
decomposition: see my ICA's bug paper for detail. Spline channel
interpolation is the most dangerous process for ICA because it does not
cause a clean rank deficiency due to its nonlinearity. If the smallest
eigenvalue is < 1E-6, ICA starts to generate 'ghost ICs', even though
Matlab's rank() function says 'the data are full ranked'! I credit Sven
Hoffmann for finding this threshold.

To conclude, my suggestion guarantees group-level data compatibility
between scalp and IC sources at the cost of the use of PCA dimension
reduction in ICA, and Claude's suggestion works fine only for channel data
analysis.

Makoto

On Thu, May 28, 2026 at 9:24 AM Tim Curran <tim.curran at colorado.edu> wrote:

> Hi Makoto,
> I have been playing with Claude Code, and using the latest Claude.md file
> with EEGLAB.
>
> Claude.md file says:
> "Typical pipeline position: clean_rawdata (removes bad channels) ->
> re-reference -> ICA -> ICLabel -> remove components -> **interpolate** ->
> re-reference (again, optional) -> epoch.”
>
> You do not agree with eeglab’s Claude.md file or maybe I am missing
> something?
>
> thanks
> Tim
>
>
>
> *From: *eeglablist <eeglablist-bounces at sccn.ucsd.edu> on behalf of Makoto
> Miyakoshi via eeglablist <eeglablist at sccn.ucsd.edu>
> *Date: *Wednesday, May 27, 2026 at 5:41 PM
> *To: *eeglablist at sccn.ucsd.edu <eeglablist at sccn.ucsd.edu>
> *Subject: *Re: [Eeglablist] Order of Channel Removal+Interpolation and
> ICA (removing noisy components)
>
> [External email - use caution]
>
>
> Hi Naviya,
>
>    - My main question is if I should perform interpolation before ICA
> or after ICA?
>
> Do it BEFORE ICA.
> If you perform channel interpolation after ICA, your EEG.icasphere does not
> have columns for the added channel.
>
> Makoto
>
> On Tue, May 26, 2026 at 2:19 PM Naviya Lall via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
> > Hello all,
> >
> > My name is Naviya and I work with EEG data in a lab in Delhi, India. I
> have
> > a question about EEG data channel interpolation and its order in
> > preprocessing pipelines.
> >
> >    - We record data with high density EEG (128 channels).
> >    - I tried to perform ICA without removing any channels and it was
> giving
> >    me noise heavy components (I use ICA to remove noisy components of
> eye,
> >    heart, muscle etc.)
> >    - I just want to remove 2-4 channels in some participants or certain
> >    sessions.
> >    - The GUI has very easy direct interpolation
> >    steps- Tools>Interpolate>select from data channels which skips the
> >    "Removal" step altogether.
> >    - In code I feel that it would be easier to "remove" the noisy
> >    channel(s) and use this -  original_chanlocs = EEG.chanlocs; (to save
> >    original locations) and then  EEG = pop_interp(EEG, original_chanlocs,
> >    'spherical');
> >    to interpolate the removed data- *Is that right?*
> >    - My main question is if I should perform interpolation before ICA or
> >    after ICA?
> >    - I read through some older exchanges on EEGLABLIST Archive from 2015,
> >    2017, 2023 and 2025 however I am still unsure of the ideal order of
> >    performing interpolation.
> >    - My logical thought is to remove + interpolate before running ICA so
> >    that the rank and number of components generated is not affected but
> > most
> >    people advise to remove channel, then ICA and then interpolate.
> >    - Please advise on the method of channel removal+interpolation and the
> >    order of channel interpolation and ICA?
> >
> >
> > Thank you so much.
> >
> >
> > Best regards,
> > Naviya
> >
> > --
> > Naviya Lall
> > Junior Research Fellow
> > Cognitive Science Lab
> > IIIT Delhi
> > naviyalalluni.wixsite.com <
> >
> https://urldefense.com/v3/__https://nam10.safelinks.protection.outlook.com/?url=https*3A*2F*2Furldefense.com*2Fv3*2F__https*3A*2F*2Fnaviyalalluni.wixsite.com*2Fnaviyalall__*3B!!Mih3wA!AK4IIr3bl7l5vstVMLfS9SvOU__IwE4KJfUqRptsvdt6mN4V6stqKqyZVfvP6h_-ebr4vwz6RSNGbePI1sed-hDQ*24&data=05*7C02*7Ctim.curran*40colorado.edu*7C5079845aad8347e770d908debc4968d2*7C3ded8b1b070d462982e4c0b019f46057*7C1*7C0*7C639155220649594641*7CUnknown*7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ*3D*3D*7C0*7C*7C*7C&sdata=QGd*2F*2BQg2J2n0RkrlOzb*2FhSvbVx*2Fn0bVEQOYCEDOyJBs*3D&reserved=0__;JSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUlJSU!!Mih3wA!Fy84dba60JaJD0fse1s0aw2M57j4HXXwDjTTVm3jvINiwNtWFC0hdHgfyQySOxrDj_DzOc-vYNGAVTei2QI_ZmXOfiM$ 
> <https://urldefense.com/v3/__https://naviyalalluni.wixsite.com/naviyalall__;!!Mih3wA!AK4IIr3bl7l5vstVMLfS9SvOU__IwE4KJfUqRptsvdt6mN4V6stqKqyZVfvP6h_-ebr4vwz6RSNGbePI1sed-hDQ$>
> > >
> > _______________________________________________
> > To unsubscribe, send an empty email to
> > eeglablist-unsubscribe at sccn.ucsd.edu or visit
> >
> https://urldefense.com/v3/__https://nam10.safelinks.protection.outlook.com/?url=https*3A*2F*2Fsccn.ucsd.edu*2Fmailman*2Flistinfo*2Feeglablist&data=05*7C02*7Ctim.curran*40colorado.edu*7C5079845aad8347e770d908debc4968d2*7C3ded8b1b070d462982e4c0b019f46057*7C1*7C0*7C639155220649658506*7CUnknown*7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ*3D*3D*7C0*7C*7C*7C&sdata=c006lGx*2Fi56iu2LNb9*2FuIWUnxG7dj6so24fzPPOjWsw*3D&reserved=0__;JSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUlJQ!!Mih3wA!Fy84dba60JaJD0fse1s0aw2M57j4HXXwDjTTVm3jvINiwNtWFC0hdHgfyQySOxrDj_DzOc-vYNGAVTei2QI_6lBRk_o$ 
> <https://sccn.ucsd.edu/mailman/listinfo/eeglablist >  .
> >
> _______________________________________________
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu or visit
> https://urldefense.com/v3/__https://nam10.safelinks.protection.outlook.com/?url=https*3A*2F*2Fsccn.ucsd.edu*2Fmailman*2Flistinfo*2Feeglablist&data=05*7C02*7Ctim.curran*40colorado.edu*7C5079845aad8347e770d908debc4968d2*7C3ded8b1b070d462982e4c0b019f46057*7C1*7C0*7C639155220649710224*7CUnknown*7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ*3D*3D*7C0*7C*7C*7C&sdata=IT5TcjXlhHoNDwoMewsn*2BHl5tWNo1LhIj23ioNFTfhk*3D&reserved=0__;JSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUl!!Mih3wA!Fy84dba60JaJD0fse1s0aw2M57j4HXXwDjTTVm3jvINiwNtWFC0hdHgfyQySOxrDj_DzOc-vYNGAVTei2QI_Gjg6iIE$ 
> <https://sccn.ucsd.edu/mailman/listinfo/eeglablist >.
>