[Eeglablist] How to Clean EEG Data with Large Amplitudes After High-Pass Filtering
Jiongjiong Li
jiongjiongfly at gmail.com
Thu Jun 25 00:38:09 PDT 2026
Hello Tomas,
Many thanks for your kind and helpful guidance!
I preprocessed the EEG data following:
1) Apply 1 Hz high-pass filter
2) Apply Zapline for line-noise removal
3) Run GEDAI with auto+ and *bad-channel rejection enabled*
GEDAI identified FP1 as a bad channel. Therefore, I used the following
pipeline:
1) *Remove FP1* from the raw data
2) Apply 1 Hz high-pass filter
3) Apply Zapline for line-noise removal
4) Run GEDAI with auto+ and *bad-channel rejection disabled* (because it
remove 3 channels otherwise).
Overall, GEDAI appears to have substantially improved the data quality:
1. Only 0.05% of samples have |EEG| > 100 µV.
2. Amplitude distribution appears approximately Gaussian.
3. Channel data is much cleaner.
4. Channel spectra show substantially reduced power.
5. ICA classification identifies more ICs as brain-related.
However, I still mainly have two concerns:
1*. SSI Silhouette Score is only 0.36*, which seems relatively low.
2.GEDAI-removed data contained 12/15 ICs labeled as Brain. This may
indicate GEDAI removed some neural activity, or that ICA classification is
not reliable in this context.
Details:
1. Percentage of samples with |EEG| > 100 µV per channel after each
processing stage:
----------------------------------------------------------------
Channel Highpass (1Hz) Zapline GEDAI
----------------------------------------------------------------
FP2 31.31% 27.22% 0.03%
C3 37.70% 35.13% 0.09%
C4 34.83% 31.82% 0.07%
P7 77.78% 77.61% 0.03%
P8 50.24% 49.89% 0.03%
O1 85.52% 85.28% 0.04%
O2 52.29% 51.35% 0.21%
F7 27.34% 24.77% 0.01%
F8 42.44% 40.04% 0.02%
F3 43.28% 42.94% 0.02%
F4 34.78% 34.07% 0.04%
T7 25.87% 23.31% 0.03%
T8 27.63% 24.45% 0.03%
P3 87.47% 86.89% 0.01%
P4 69.98% 69.90% 0.04%
Total 48.56% 46.98% 0.05%
2. Additional comparison figures after each stage:
2.1 Amplitude distribution:
https://urldefense.com/v3/__https://drive.google.com/file/d/1iGUhps4UtATY1c060NV4rSllSrZQ5ION/view?usp=sharing__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82QZo5XmM$
2.2 Channel data:
https://urldefense.com/v3/__https://drive.google.com/file/d/1wjqGVyjTJXtzu9cMzbdtJi0azVmUcK2H/view?usp=sharing__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82QMlq3-p$
2.3 Spectra:
https://urldefense.com/v3/__https://drive.google.com/file/d/1mxaHMrenmd3LiZoM6jnlSo5rEhpPpWaP/view?usp=drive_link__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82Qq9Qv0I$
2.4 ICA results before vs after GEDAI:
https://urldefense.com/v3/__https://drive.google.com/file/d/1sX8davJ3TK6ShjIy1GfgUHIwALuZ1PtT/view?usp=drive_link__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82Zp8RzBk$
• Before GEDAI: 11/15 ICs labeled as brain, 2 as other, 2 as line noise
• After GEDAI: 14/15 ICs labeled as brain, 1 as other
2.5 Data Removed by GEDAI (Identified as Artifacts)
2.5.1 Spectra of data before GEDAI, data removed by GEDAI (Identified as
artifacts), and remaining data (identified as clean):
https://urldefense.com/v3/__https://drive.google.com/file/d/1hK4H_hTEFm5D1Otm6G9z6k61GaSoftX3/view?usp=sharing__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82TrSE9ZD$
2.5.2 ICA results of data before GEDAI vs data removed by GEDAI
(Identified as Artifacts):
https://urldefense.com/v3/__https://drive.google.com/file/d/10S03EQmlTjTMM7NuCbPdEh2b-LvsAx0o/view?usp=sharing__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82e3h3kYi$
• Before GEDAI: 11/15 ICs labeled as brain, 2 as other, 2 as line noise
• Data removed by GEDAI: 12/15 ICs labeled as brain, 2 as other, 1 as line
noise
3. SENSAI visualization:
https://urldefense.com/v3/__https://drive.google.com/file/d/1ngKiMtEp0U3BvUnkiiL2Qtzz1rYYcUSo/view?usp=drive_link__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82Yi55Dqa$
• SSI Silhouette Score: 0.36
• SENSAI score: 39%
• ENOVA: 98%
• Mean SSSI Before Denoising: 0.51
• Mean SSSI After Denoising: 0.78
• Mean NSSI After Denoising: 0.50
Any comments or suggestions would be very welcome!
Sincerely,
Jiongjiong
On Wed, Jun 24, 2026 at 11:14 AM Ros, Tomas <dr.t.ros at gmail.com> wrote:
> Hello Jiongjiong,
>
> Having looked at the figures you sent, allow me to chip in regarding GEDAI
> denoising.
>
> Overall, the data you have is very noisy with quite a low channel count
> (16 channels). Your initial parameters are pretty much appropriate for this
> type of data (auto+ with a 1 Hz low cut). To further improve the final
> result I would recommend:
>
> 1. Remove the line noise BEFORE GEDAI (using a notch or low pass filter
> <55 Hz, or even better using Zapline)
> 2. Enable bad channel rejection within GEDAI (with default ENOVA of 0.95).
>
> That should further remove the very noisy parts in the data. For a deeper
> dive, check out the GEDAI FAQ
> <https://urldefense.com/v3/__https://github.com/neurotuning/GEDAI-master/wiki/Frequently-Asked-Questions-(FAQ)__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82QD_2Mk8$ >
>
> Good luck and all the best,
> Tomas
>
> ▬▬▬
>
> *Tomas Ros, PhD*
>
> Lecturer, Department of Clinical Neurosciences
> CIBM EEG HUG-UNIGE Section
> University of Geneva, Switzerland
> https://urldefense.com/v3/__https://www.tomasros.com/__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82ZCkXpaA$
>
> [image: CIBM logo]
>
>
> *Clinical and Translational Neuroimaging*
>
> UNIGE Campus Biotech
>
> 9 Chemin des Mines
>
> CH - 1202 Genève, Switzerland
>
>
> (+41) 0766 158 863
>
>
>
> cibm.ch <https://urldefense.com/v3/__http://www.cibm.ch/__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82R25Xa6Y$ >
>
> *Follow us: *[image: Twitter] <https://urldefense.com/v3/__https://twitter.com/CIBM_ch__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82Q25Gek4$ > [image:
> LinkedIn]
> <https://urldefense.com/v3/__https://www.linkedin.com/company/cibm-center-for-biomedical-imaging/__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82VQoNZFD$ > [image:
> YouTube] <https://urldefense.com/v3/__https://www.youtube.com/channel/UCA8rMvimtkVzFwWzDDlaGMQ__;!!Mih3wA!CaIG6ShccYM_SNsw1TOgp23Oeh4OtDBhcfGnFD9cxrOkA6TYzg9HWnfz3TUyv-XotIh8RbLYojsWzQGua5B82coZJE-Z$ >
>
>
>
>
>
>
> On Wed, 24 Jun 2026 at 17:48, Jiongjiong Li via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
>> Hello Евгений, Ingmar and Eugen,
>>
>> I deeply appreciate all of your support and invaluable guidance!
>>
>> Since our lab has already collected 80 recordings under the same
>> condition,
>> my initial goal is to clean the existing dataset. Following Ingmar's
>> suggestion, I applied *GEDAI *(
>> https://urldefense.com/v3/__https://github.com/neurotuning/GEDAI-master__;!!Mih3wA!DP_6iMSxhZ5nGMQgcR6NRZFO5LvXiIRNlofhI1K7OEXLKjqDu5TDmv_VUkJro-DS5xoeAc3e1MK9JWLcSr9RAoNOuMWN$
>> ) to
>> preprocess the raw EEG data. After cleaning, the proportion of samples
>> with *absolute
>> EEG amplitude > 100 µV was reduced to 8.89%*, making further analysis
>> feasible.
>>
>> Below are the data-denoising results using GEDAI:
>> 1. EEG signal 1 Hz high-pass only vs. GEDAI:
>>
>> https://urldefense.com/v3/__https://drive.google.com/file/d/17dnk7SIsP3gMiFzAuPG4WTmkBTIF8QLw/view?usp=sharing__;!!Mih3wA!DP_6iMSxhZ5nGMQgcR6NRZFO5LvXiIRNlofhI1K7OEXLKjqDu5TDmv_VUkJro-DS5xoeAc3e1MK9JWLcSr9RAhhcH-Pk$
>>
>> 2. EEG spectra Hz high-pass only vs. GEDAI:
>>
>> https://urldefense.com/v3/__https://drive.google.com/file/d/1vnjrqhCCGH_K_oaU_eSwMGKwCh0ygWHO/view?usp=sharing__;!!Mih3wA!DP_6iMSxhZ5nGMQgcR6NRZFO5LvXiIRNlofhI1K7OEXLKjqDu5TDmv_VUkJro-DS5xoeAc3e1MK9JWLcSr9RAiTBeN8A$
>>
>> 3. ICA results after GEDAI (all ICs labeled as “brain”):
>>
>> https://urldefense.com/v3/__https://drive.google.com/file/d/10NaJlz2FK7TVuvCT1F6hf7vpq00zciVs/view?usp=drive_link__;!!Mih3wA!DP_6iMSxhZ5nGMQgcR6NRZFO5LvXiIRNlofhI1K7OEXLKjqDu5TDmv_VUkJro-DS5xoeAc3e1MK9JWLcSr9RAmoNmDhg$
>>
>> Note: GEDAI was performed using the auto-denoising option with auto+
>> strength and a 1 Hz high-pass filter.
>>
>> For comparison, EEG spectra before and after ASR denoising are shown
>> below:
>>
>> https://urldefense.com/v3/__https://drive.google.com/file/d/1AEOCy-MxqGBQzaU-_1lgVlmNR7_evLFd/view?usp=sharing__;!!Mih3wA!DP_6iMSxhZ5nGMQgcR6NRZFO5LvXiIRNlofhI1K7OEXLKjqDu5TDmv_VUkJro-DS5xoeAc3e1MK9JWLcSr9RAnyBFCsN$
>>
>> The table below summarizes the percentage of samples with absolute EEG
>> amplitude > 100 µV after each denoising method:
>>
>> Channel Percent of abs(EEG) > 100uV
>> High Pass (1 HZ) ASR GEDAI
>> FP1 92.946 Removed 2.5073
>> FP2 31.312 Removed 3.0691
>> C3 37.702 0.97549 8.0592
>> C4 34.826 1.6984 7.4946
>> P7 77.775 53.236 17.479
>> P8 50.242 10.092 10.864
>> O1 85.518 73.37 22.544
>> O2 52.286 Removed 8.4524
>> F7 27.337 1.5796 3.816
>> F8 42.444 17.465 6.726
>> F3 43.282 3.4959 11.055
>> F4 34.779 0.29215 7.5504
>> T7 25.874 0.52488 2.4978
>> T8 27.634 Removed 2.6692
>> P3 87.468 71.226 16.299
>> P4 69.978 30.943 11.147
>> Total 51.34% 22.07% 8.89%
>> Note: ASR removed 70% samples.
>>
>> I would also appreciate your guidance on the following questions:
>>
>> *1. Existing EEG data cleaning*
>> 1.1 Do the data look acceptable after GEDAI cleaning? If not, are there
>> additional remediation approaches for the existing dataset?
>> 1.2 Is it acceptable to apply further preprocessing steps (e.g., notch
>> filtering) after GEDAI, given the presence of a 60 Hz peak in the
>> spectrum?
>>
>> *2. Improving future EEG recordings*
>> 2.1 Ingmar mentioned that “100 µV is not unusual with moving
>> participants.”
>> May I further confirm whether 100 µV is still expected in mobile EEG under
>> our experimental conditions? Specifically, each trial lasts more than 20
>> minutes, and participants are allowed to walk and speak to simulate a
>> natural environment.
>> 2.2 Could you recommend any books, videos, or resources for learning how
>> to
>> identify EEG artifacts? I would greatly like to strengthen my
>> understanding
>> in this area.
>>
>> Thank you very much again for your time and guidance!
>>
>> Sincerely,
>>
>> Jiongjiong Li
>> Department of Computer Science
>> Illinois Institute of Technology
>> 10 W 35th St, Chicago, IL 60616
>>
>>
>>
>> On Tue, Jun 23, 2026 at 2:19 AM Евгений Машеров <emasherov at yandex.ru>
>> wrote:
>>
>> > Frequency filtering can be helpful if the artifact is concentrated in a
>> > fairly narrow frequency band outside the region of primary interest.
>> > However, if the artifact's shape is anything other than sinusoidal, it
>> > cannot be completely eliminated, since in addition to the fundamental
>> > frequency, there are also harmonics, which typically fall within the
>> band
>> > of primary research interest. Frequency filtering is helpful, for
>> example,
>> > in removing artifacts associated with electrode polarization or galvanic
>> > skin response, at the cost of losing some of the delta range. It is also
>> > possible to attempt to eliminate muscle artifacts by losing some of the
>> > beta range. This type of filtering aids visual analysis, but is rather
>> > detrimental to mathematical processing.
>> > Independent component analysis has already been proposed here, and it
>> can
>> > be a lifesaver if the component caused by motion is correctly
>> identified.
>> > However, if there is no signal at all, nothing will help. And in the
>> > figure shown, at least one channel (FP1) appears extremely unreliable.
>> The
>> > presence of power line interference in one channel often indicates a
>> > complete lack of contact with the scalp. Power line interference occurs
>> due
>> > to capacitive coupling between the electrical network and the electrode,
>> > and since the area of the "capacitor plates" (EEG electrode and power
>> > cable) is small and the distance between them is quite large, the
>> > capacitance is on the order of tenths of a picofarad or less, while the
>> > resistance at 50 or 60 Hz is tens or hundreds of gigaohms. If the
>> electrode
>> > and scalp contact are sufficiently good (resistance 10 kiloohms or
>> less),
>> > the resulting voltage divider attenuates the power line interference by
>> > several million times, and it is not detected. If contact with the
>> scalp is
>> > broken, the attenuation is equal to the ratio of the amplifier's input
>> > impedance to the resistance of the capacitive coupling to the electrical
>> > cables. If interference appears on one of the electrodes, it should be
>> > adjusted.
>> > Motion artifacts may be due to a cap that is too loose. Only after
>> > correcting all the causes of artifacts appearing during recording does
>> it
>> > make sense to use mathematical methods for their elimination, primarily
>> ICA.
>> >
>> > Your truly
>> >
>> > Eugen Masherov
>> >
>> > > Hello,
>> > > I'm working with EEG data collected using a Cyton + Daisy setup. The
>> > > experiment allows participants to move freely throughout the
>> recording,
>> > so
>> > > I expected a significant amount of motion-related noise.
>> > >
>> > > I applied a 1 Hz high-pass filter, but the data still shows a very
>> large
>> > > proportion of high-amplitude values. Overall, about *55% of samples
>> have
>> > > absolute values greater than 100 µV*, which seems too high for usable
>> > EEG.
>> > >
>> > > My question is: *how to clean this dataset*, since standard filtering
>> > does
>> > > not seem sufficient?
>> > >
>> > > 1. Percent of samples with |amplitude| > 100 µV per channel:
>> > > Channel PercentAbs > 100uV
>> > > FP1 92.946
>> > > FP2 31.312
>> > > C3 37.702
>> > > C4 34.826
>> > > P7 77.775
>> > > P8 50.242
>> > > O1 85.518
>> > > O2 52.286
>> > > F7 27.337
>> > > F8 42.444
>> > > F3 43.282
>> > > F4 34.779
>> > > T7 25.874
>> > > T8 27.634
>> > > P3 87.468
>> > > P4 69.978
>> > >
>> > > 2. Additional channel statistics:
>> > > Channel Min Max Mean Std P5 P95
>> > > FP1 -1.55E+05 3.98E+05 -0.33667 18173 -18654 14161
>> > > FP2 -3312.7 3268.7 -0.056876 243.58 -324.27 353.39
>> > > C3 -2650.7 4902.2 -0.064448 330.28 -465.32 490.06
>> > > C4 -3082.8 3528.5 -0.052471 269.78 -389.84 426.57
>> > > P7 -5018 17516 -0.018368 880.09 -1263.1 1431.4
>> > > P8 -2085.9 9144.2 0.11526 404.97 -523.25 638.35
>> > > O1 -11920 70997 0.43565 1346 -1509 1743.5
>> > > O2 -2705.5 9071.8 -0.014948 293.52 -395.37 447.87
>> > > F7 -2894.7 2943.4 0.024095 193.53 -234.4 255.67
>> > > F8 -2973.9 2639.3 0.054757 225.84 -314.76 324.82
>> > > F3 -4752.6 5627.2 0.08169 517.18 -635.01 672.64
>> > > F4 -6333.3 3600.2 0.07269 515.99 -392.74 630.2
>> > > T7 -2638.4 2738 0.015915 177.67 -229.59 253.66
>> > > T8 -3213.5 3242.7 0.0090257 223.3 -298.77 327.22
>> > > P3 -27780 43690 0.80651 2552.2 -3342.1 3681.6
>> > > P4 -14668 46330 0.17228 1622.3 -1771.8 1753.2
>> > > 3. Channel Data Screenshot:
>> > >
>> >
>> https://urldefense.com/v3/__https://drive.google.com/file/d/1FwlMrvYE1e7zn0mQOYnwG7WtYk1Q3SAh/view?usp=sharing__;!!Mih3wA!CrjzKaf8yxIzQ7S7-hNrGlTX3i4E6i-3GXMBj9uYu71UDEIrHRdtyhGKF0ShjK6kP-r-OnCRDllzIN6CQI-UPeoMk6Wj$
>> > >
>> > > Thank you so much for your guidance!
>> > >
>> > > Best Regards,
>> > > Jiongjiong
>> > > _______________________________________________
>> > > 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 .
>
>
More information about the eeglablist
mailing list