Bad channel detection in EEG recordings

License.

Milana Komosar*, Patrique Fiedler, and Jens Haueisen

https://doi.org/10.1515/cdbme-2022-1066

Abstract: Electroencephalography (EEG) is widely used in

clinical applications and basic research. Dry EEG opened the

application area to new fields like self-application during

gaming and neurofeedback. While recording, the signals are

always affected by artefacts. Manual detection of bad channels

is the gold standard in both gel-based and dry EEG but is time-

consuming. We propose a simple and robust method for

automatic bad channel detection in EEG. Our method is based

on the iterative calculation of standard deviations for each

channel. Statistical measures of these standard deviations

serve as indications for bad channel detection. We compare the

new method to the results obtained from the manually

identified bad channels for EEG recordings. We analysed EEG

signals during resting state with eyes closed and datasets with

head movement. The results showed an accuracy of 99.69 %

for both gel-based and dry EEG for resting state EEG. The

accuracy of our new method is 99.38 % for datasets with the

head movement for both setups. There was no significant

difference between the manual gold standard of bad channel

identification and our iterative standard deviation method.

Therefore, the proposed iterative standard deviation method

can be used for bad channel detection in resting state and

movement EEG recordings.

Keywords: Electroencephalography, dry electrode,

artefacts, head movements, brain-computer interfaces

1 Introduction

Electroencephalography (EEG) is a non-invasive technique

for recording neural electrical activity with the help of

electrodes placed on the scalp. This method is widely used in

clinical applications and basic research. Moreover, EEG can

be used in brain-computer interfaces or rehabilitation. The

latest development of electronics for EEG recordings and dry

electrodes enabled self-application and thus usage in out-of-

the-lab scenarios [1].

The EEG signal has low amplitude and as such is prone to

artefacts and unwanted noise. The reconstruction of the brain

activity from bad channels is sometimes not possible.

Therefore, it is very important to remove non-reliable channels

before applying any type of EEG signal analysis or source

reconstruction. The most common procedure for bad channel

detection still is manual inspection. This procedure is often

time-consuming and subjective, which can lead to different

results depending on the experts evaluating the EEG [2–4].

Dry EEG is more prone to movement artefacts compared to

gel-based EEG. Due to the lower channel reliability in dry

EEG, the manual identification of bad channels is even more

time-consuming.

Various semi- or fully automated approaches for bad

channel detection have been proposed for gel-based EEG

including a combination of different statistical, temporal, and

frequency features [5,6]. Our aim is the development of a

simple and robust method for automatic bad channel detection

in dry EEG recordings [7], which is at the same time also

suitable for gel-based EEG.

2 Methods

2.1 Measurements

In the study participated 5 healthy volunteers, 3 females and 2

males, with a mean age of 27 ± 10. We analysed EEG data

from two different segments: one segment of resting EEG with

closed eyes for 3 minutes and one segment while volunteers

were performing head movements. Head movements were not

executed rapidly, but rather slowly. The participants were

instructed to move their heads downwards to the chest and

back to a straight position when they hear a tone. The

movement was repeated every 4 seconds. In sum, 45

movement epochs for each participant are recorded.

We used 64-channel gel-based and dry EEG caps with an

equidistant layout (waveguard, ANT B.V., Hengelo, The

_____

*Corresponding author: Milana Komosar: Institute of Biomedical

Engineering and Informatics, Faculty of Computer Sciences and

utomation, Technische Universität Ilmenau, Ilmenau, Germany,

milana.komosar@tu-ilmenau.de

Patrique Fiedler, Jens Haueisen: Institute of Biomedical

Engineering and Informatics, Faculty of Computer Sciences and

utomation, Technische Universität Ilmenau, Ilmenau, Germany

cdbme_2022_8_2.pdf 257 8/29/2022 5:45:49 PM

257

DE GRUYTER

Current Directions in Biomedical Engineering 2022;8(2): 257-260

Netherlands). The reference electrode was placed on the right

mastoid. The sampling rate was 1024 samples/second.

2.2 Data Processing

The data were processed using MATLAB (The

MathWorks, Inc., Natick, United States). Each of the 15

datasets was analysed separately. Data were pre-processed

using a finite impulse response (FIR) bandpass filter

implemented in EEGLAB [8] with a low cut-off frequency of

1 Hz and a high cut-off frequency of 40 Hz, as one of the

standard frequency ranges for EEG analysis [9]. After

filtering, the newly proposed iterative method for automatic

bad channel detection was performed. For comparison, the

data were manually inspected and bad channels were

identified as well.

The proposed method is applied and tested on two

different window lengths for the resting state EEG. The first

studied case is a window length of 60 s after skipping the first

10 s of the signal to avoid filter artefacts. For the second

studied case, a 30 s window is chosen as commonly used in

the standard EEG signal analyses [9]. These windows have the

same lengths for all the datasets. However, for the datasets

with the head movement, the analysis windows were starting

2.3 s after the movement and had a length of 0.5 s. Thus,

transient movement artefacts were not considered for

identifying bad channels.

The signal quality of each channel is manually evaluated

for the chosen windows [1,10]. The same signal windows were

used for the automatic bad channel detection. All datasets were

visually inspected by one annotator and the bad channels were

annotated similar to the criteria described in [11]. Predefined

bad channel characteristics are: exhibiting either a saturated

signal, an isoelectric line, or a predominantly artefactual EEG

recording.

2.3 Automatic ISD Method

We propose an Iterative Standard Deviation (ISD) method for

automatic bad channel detection in both dry and gel-based

EEG recordings. The standard deviation is calculated for each

of the 64 channels over the whole analysed data window.

First, the standard deviation of the signal for the j-th

sensor over the whole analysis window of length N is

calculated (see eq 1).

ࡿࡰ

࢐

ට

૚

ࡺି૚

หࢂ

(࢏,࢐)

െࢂ

ഥ

࢐

ห

૛

ࡺ

࢏ୀ૚

(1)

௜

is the i-th out of N samples for the j-th electrode and ܸ

ത

௝

mean voltage for the j-th electrode. The result are 64 values of

standard deviations. They are observed further as one

population from which outliers have to be identified. Four

criteria are established to detect the bad channels (see eq 2-5).

The criteria used to eliminate the outliers from the population

of SDs are the median and the 75th percentile range, the

standard deviation of channels lower than 10

-4

μV and higher

than 100 μV.

หࡿࡰ

(࢐,࢑)

െࡹ

࢑

ห>ૠ૞࢚ࢎ ࢖ࢋ࢘ࢉࢋ࢔࢚࢏࢒ࢋ (2)

ࡿࡰ

(

࢐,࢑

)

<૚૙

ି૝

μ܄ (3)

ࡿࡰ

(࢐,࢑)

>૚૙૙ μ܄ (4)

An additional criterion for ending the iterations is:

ࡿࡰ

࢖(࢑)

>૞ (5)

ܵܦ

(௝,௞)

is standard deviation of the j-th channel at the k-th

iteration. M is the median of the population of standard

deviations in the k-th iteration. SD

is the standard deviation of

all individual channel standard deviations in the k-th iteration.

2.4 Statistical Analysis

Statistical analysis was performed in MATLAB R2021a. The

aim was to compare the identification of bad channels between

the two methods. The manual selection of bad channels is

taken as ground truth. The performance of the proposed ISD

method is evaluated with the confusion matrix, sensitivity,

specificity, and accuracy calculation. In addition, Fisher’s

exact test at an alpha level of 0.05 is used to check if the two

methods are providing significantly different outputs.

3 Results

The main results of the analysis are shown in Table 1 for all

320 analysed channels from all 5 participants. For the 30 s long

analysis window and gel-based electrodes, the number of bad

channels manually identified is 11 while the ISD method

identified 10 bad channels. In the dry dataset 25 channels are

manually selected as bad, while the ISD method identified 24

of them. For the datasets with the head movements, the

manually identified number of bad channels is 12 and 30 for

gel-based and dry recordings, respectively. In this case, the

ISD method identified 10 and 32 bad channels in gel-based

and dry recordings, respectively. The accuracy of the ISD

method for the gel-based and dry alpha EEG recordings with

the chosen 30 s window is 99.69 %. The accuracy of the ISD

cdbme_2022_8_2.pdf 258 8/29/2022 5:45:49 PM

258

method is 99.38 % for both types of recordings and the 0.5 s

fixed window in the head movement datasets.

Figure 1 shows the method’s performance for the 10 s

exemplary alpha EEG dataset. In both cases (gel-based and

dry), both methods (manual and ISD) identified the same

channels as bad (colored red in Figure 1). For the gel-based

electrodes these are 3RD and 3LD and for the dry ones 3RD,

3LD, and 5LB. All bad channels in the datasets were detected

in the first iteration.

4 Conclusion

We found no significant differences in the channels labelled as

bad for the manual and ISD detection of bad channels for both

resting state and head movement EEG. The ISD method is

simple and robust and can be applied for bad channel

identification in both dry and gel-based recordings, replacing

manual evaluation. The detection criteria of the ISD method

can be adapted to other datasets. As this method is working

iteratively and does neither depend on pre-defined hard

thresholds nor standardized values, it has the potential to

automatize the process of bad channel detection. After the

application of the ISD method, processing steps can be applied

to further clean the data. However, the ISD method can be also

used to detect artefacts, such as e.g. movement artefacts.

We can notice a higher overall number of bad channels

detected in the datasets with the head movement for dry

recordings. As Debnath et al. reported, existing bad channel

detection methods may not perform well if the recordings have

a high number of bad channels as e.g. in EEG recordings with

movements [3]. The advantage of our proposed ISD method is

to overcome this limitation. The next phase of method testing

would consider a broader frequency range of the EEG signals.

Figure 1: Alpha EEG dataset for all 64 channels and 10 seconds a) gel-based and b) dry. The red color indicates bad channels that are

detected by the manual and ISD methods.

cdbme_2022_8_2.pdf 259 8/29/2022 5:45:49 PM

259

Future work will address the main limitation of this study

which is the low number of subjects.

Table 1: Statistical results of the ISD method for gel-based and dry

EEG. Number of channels identified as bad from a total of 320

channels for all 5 volunteers. P-values of Fisher’s exact test. The

difference between identified bad channels from manual and ISD

methods was compared at a significance level of alpha = 0.05.

Dataset Alpha Head

movement

Performance

parameter

30 s

window

60 s

window

0.5 s

window

Gel-

based

Sensitivity 90.91 84.62 83.33

Specificity 100 100 100

ccuracy 99.69 99.38 99.38

Manual 11 13 12

ISD 10 11 10

p-value 0.04*10

-16

0.01*10

-17

0.02*10

-17

Dry Sensitivity 96.00 100 100

Specificity 100 100 99.31

ccuracy 99.69 100 99.38

Manual 25 24 30

ISD 24 24 32

p-value 0.03*10

-33

0.01*10

-34

0.04*10

-38

Author Statement

Research funding: This research was supported by the

European Union’s Horizon 2020 Framework, Marie

Sklodowska-Curie Actions H2020-MSCA-ITN-2018 under

grant agreement No. 813843 and in part by the Free State of

Thuringia (2018 IZN 004), co-financed by the European

Union under the European Regional Development Fund

(ERDF).

Conflict of interest: Authors state no conflict of interest.

Informed consent: Informed consent has been obtained from

all individuals included in this study. Ethical approval: The

research related to human use complies with all the relevant

national regulations, institutional policies and was performed

in accordance with the tenets of the Helsinki Declaration, and

has been approved by the authors’ institutional review board

or equivalent committee.

References

[1] Di Fronso S, Fiedler P, Tamburro G, Haueisen J, Bertollo M,

Comani S. Dry EEG in Sports Sciences: A Fast and Reliable

Tool to Assess Individual Alpha Peak Frequency Changes

Induced by Physical Effort. Front Neurosci 2019;13:982.

[2] Daly I, Scherer R, Billinger M, Müller-Putz G. FORCe: Fully

Online and Automated Artifact Removal for Brain-Computer

Interfacing. IEEE Trans Neural Syst Rehabil Eng

2015;23(5):725–36.

[3] Debnath R, Buzzell GA, Morales S, Bowers ME, Leach SC,

Fox NA. The Maryland analysis of developmental EEG

(MADE) pipeline. Psychophysiology 2020;57(6):e13580.

[4] Hartmann MM, Schindler K, Gebbink TA, Gritsch G, Kluge T.

PureEEG: automatic EEG artifact removal for epilepsy

monitoring. Neurophysiol Clin 2014;44(5):479–90.

[5] Desjardins JA, van Noordt S, Huberty S, Segalowitz SJ,

Elsabbagh M. EEG Integrated Platform Lossless (EEG-IP-L)

pre-processing pipeline for objective signal quality

assessment incorporating data annotation and blind source

separation. J Neurosci Methods 2021;347:108961.

[6] Viana PF, Remvig LS, Duun-Henriksen J, Glasstetter M,

Dümpelmann M, Nurse ES, Martins IP, Schulze-Bonhage A,

Freestone DR, Brinkmann BH, Kjaer TW, Richardson MP.

Signal quality and power spectrum analysis of remote ultra

long-term subcutaneous EEG. Epilepsia 2021;62(8):1820–8.

[7] Hinrichs H, Scholz M, Baum AK, Kam JWY, Knight RT,

Heinze H-J. Comparison between a wireless dry electrode

EEG system with a conventional wired wet electrode EEG

system for clinical applications. Sci Rep 2020;10(1):5218.

[8] Delorme A, Makeig S. EEGLAB: an open source toolbox for

analysis of single-trial EEG dynamics including independent

component analysis. J Neurosci Methods 2004;134(1):9–21.

[9] Haueisen J, Fiedler P, Bernhardt A, Gonçalves R, Fonseca

C. Novel dry electrode EEG headbands for home use:

Comparing performance and comfort. Current Directions in

Biomedical Engineering 2020;6(3):139–42.

[10] Fiedler P, Fonseca C, Supriyanto E, Zanow F, Haueisen J. A

high-density 256-channel cap for dry

electroencephalography. Hum Brain Mapp 2022;43(4):1295–

308.

[11] Witteveen J, Pradhapan P, Mihajlovic V. Comparison of a

Pragmatic and Regression Approach for Wearable EEG

Signal Quality Assessment. IEEE J Biomed Health Inform

2020;24(3):735–46.

cdbme_2022_8_2.pdf 260 8/29/2022 5:45:49 PM

260