[Eeglablist] Poor ICA Decomposition/Strange Scalp Maps

Thu Jul 22 14:39:45 PDT 2021

Hi Makoto,

Thanks for developing on this. I agree with you.
I was just suggesting these options considering her poor/strange ICA results obtained. These were just suggestions to attempt to increase ICA performance and decompositions. While it can be fine to use the pca option as you explained, I thought elevating the high pass edge, not using pca before ICA, improving data cleaning, and not interpolating bad channels before ICA would very likely improve ICA results in this situation.
If she rejected for example 6 bad channels and interpolated them, using pca -7 and a low high-pass could significantly decrease ICA performance, don’t you think?

Thanks,

Cédric

On Tue, Jul 20, 2021 at 21:35, Makoto Miyakoshi via eeglablist <eeglablist at sccn.ucsd.edu> wrote:

> Dear Casey and Cedric,
>
> If the electrode locations are confused (e.g. electrode labels are
> shuffled), the result could be messy.
>
> There is also a concern of rank deficiency in the input data for ICA. See
> these Wiki pages.
> https://sccn.ucsd.edu/wiki/Makoto's_useful_EEGLAB_code#How_to_avoid_the_effect_of_rank-deficiency_in_applying_ICA_.2803.2F31.2F2021_added.29
> https://sccn.ucsd.edu/wiki/Shannon%27s_ICA_rank_project
>
> On a separate note,
>
>> Either use the pca option with the reduced rank in input when running ICA
> (not recommended)
>
> It makes me worried to see this kind of belief keep spreading. Let me
> clarify it here.
>
> Fiorenzo published the following paper in 2018.
> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.sciencedirect.com_science_article_pii_S1053811918302143-3Fvia-253Dihub&d=DwIFaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=yFV5hvDMEHEFmIjhiehKZiKcz0Mm14M5jqJ93ctR3II&s=C4F0k-pYChrdf-hPV-Cj7vBZttjf6dJLP5gE4cP7j7I&e=
> The conclusion is NOT that using PCA is unconditionally bad for ICA.
> Instead, he showed that unnecessarily excessive dimension reduction is bad.
> You also know how redundant the scalp-recorded EEG data is. For example,
> how many PCs you need to represent 95% of variance of the scalp data (and
> the lesser degree of dimension reduction, in other words data
> sparsification, still happens for ICA!) In order to make the input data
> full-ranked, using PCA is TOTALLY fine. You don't lose ANY information in
> terms of linear algebra.
>
> Scalp EEG data is VERY redundant, which is why ICA works well (very
> redundant == very much mixed == very much approaching to Gaussianity due
> to central limit theorem == very much ICA-decomposable because ICA is a
> process of non-Gausianity maximization. Karl Friston pointed it out in his
> 1996 commentary paper in TICS.) It means that PCA can very efficiently
> 'compress' the data--but only in terms of variance! Variance does not
> necessarily respect multivariate phase relations and mutual information
> etc. which is why you need to spend much more PCs to keep the original info
> which ICA uses (as I understand, not tested.)
>
> To see how redundant the (raw) EEG data, see also my slide
> https://sccn.ucsd.edu/wiki/Makoto%27s_preprocessing_pipeline#Electroencephalosophy_.28For_100.2C000_page_views.2C_09.2F06.2F2019_updated.29
>
> Anyway, to repeat, in order to make the input data full-ranked, using PCA
> is TOTALLY fine. You don't lose ANY information.
>
> Makoto
>
> On Tue, Jul 20, 2021 at 9:33 AM Cedric Cannard via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
>> Hi Casey,
>>
>> > 2. Re-Reference to Mastoid channels
>> Average referencing would be much better, especially on 128-channel data.
>>
>> > 3. Filter: highpass .03Hz, low pass 40Hz
>> ICA works much better with highpass ~1 Hz
>> If you really need to keep low frequencies. You can run ICA and reject
>> your bad components on data highpassed at 1 Hz, and transfer them
>> (EEG.icaweights, EEG.icasphere, EEG.icawinv) to your original data
>> highpassed at 0.03 Hz.
>>
>> > 5. Identify channels to interpolate and epochs to reject
>> Either use the pca option with the reduced rank in input when running ICA
>> (not recommended), or interpolate rejected channels after ICA and rejection
>> of bad ICs.Each interpolated channel reduces data rank, which reduces ICA
>> performance.
>>
>> Also, you didn't mention data cleaning. ICA performs better after cleaning
>> major artifacts first (e.g., manually or with ASR).
>>
>> Hope this helps,
>>
>> Cedric
>>
>>
>> ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
>>
>> On Monday, July 19th, 2021 at 12:32 PM, Nicastri, Casey via eeglablist <
>> eeglablist at sccn.ucsd.edu> wrote:
>>
>> > Hi all,
>> >
>> > I am wondering if anyone can help give me some more information on a
>> possible reason my ICA decomposition is so poor. A picture of the first 35
>> scalp maps is linked below. This data was collected on a 128 channel system
>> (BioSemi/ActiView System)
>> >
>> > Current Pipeline:
>> >
>> > 1. Import, resample at 256Hz (collected at 512)
>> > 2. Re-Reference to Mastoid channels
>> > 3. Filter: highpass .03Hz, low pass 40Hz
>> > 4. Epoch to 1.2s epochs
>> > 5. Identify channels to interpolate and epochs to reject
>> > 6. Runica, component rejection
>> >
>> > Picture:
>> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.dropbox.com_s_h7y0r2dshcta2cu_Screen-2520Shot-25202021-2D07-2D14-2520at-252011.13.30-2520AM.png-3Fdl-3D0&d=DwIGaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=UKJhquGPgJ-e9o2GfVHIU_WgNx1T9I_aX4zuqcDdQJE&s=2G80tUh3e0z2iW3K4iJipbqd6-0lGV4WYGNBSlG2U6Q&e=
>> >
>> > Are there any large errors here? I’ve never seen components like
>> this. The data is a bit noisy but not worse than other data previously
>> collected on this system. Any ideas?
>> >
>> > Thanks!
>> >
>> > Casey
>> >
>> > --
>> >
>> > Casey Nicastri
>> >
>> > Clinical Research Assistant
>> >
>> > Division of Cognitive and Behavioral Neurology
>> >
>> > Center for Brain/Mind Medicine, Brigham and Women’s Hospital,
>> Harvard Medical School
>> >
>> > The information in this e-mail is intended only for the person to
>> whom it is addressed. If you believe this e-mail was sent to you in error
>> and the e-mail contains patient information, please contact the Mass
>> General Brigham Compliance HelpLine at
>> https://urldefense.proofpoint.com/v2/url?u=http-3A__www.massgeneralbrigham.org_complianceline&d=DwIGaQ&c=-35OiAkTchMrZOngvJPOeA&r=kB5f6DjXkuOQpM1bq5OFA9kKiQyNm1p6x6e36h3EglE&m=UKJhquGPgJ-e9o2GfVHIU_WgNx1T9I_aX4zuqcDdQJE&s=IYflcTPf6JaacGCJ5h6-scSE_GsCC9ESWhlJx67dQIE&e=
>> . If the e-mail was sent to you in error but does not contain patient
>> information, please contact the sender and properly dispose of the e-mail.
>> >
>> > Please note that this e-mail is not secure (encrypted). If you do
>> not wish to continue communication over unencrypted e-mail, please notify
>> the sender of this message immediately. Continuing to send or respond to
>> e-mail after receiving this message means you understand and accept this
>> risk and wish to continue to communicate over unencrypted e-mail.
>> >
>> > 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