woman making eeg eye blink artifact
Go back to blog
All about EEG artifacts and filtering tools
One of the main concerns when dealing with electroencephalographic signals (EEG) is assuring that we record clean data with a high signal to noise ratio. The EEG signal amplitude is in the microvolts range and it is easily contaminated with noise, known as “artifacts”, which need to be filtered from the neural processes to keep the valuable information we need for our applications. We review in this post different EEG artifacts and the main tools and techniques to remove them.

What is an EEG artifact?

Our brains are continuously working. Biochemistry exchanges between cells produce small electrical activity when the neurons communicate among them. A single electric signal from neuron to neuron is not recordable but when millions of neurons synchronize, the electric field generated can be measured from the scalp. These electroencephalographic signals (EEG) are transmitted through tissue, bone, and hair before it is recorded, and by then its amplitude is very attenuated (Sörnmo & Laguna, 2005; Nunez & Srinivasan, 2006). 

We denote an “artifact” as any component of the EEG signal that is not directly produced by human brain activity (under some circumstances neural processes generated by the brain can themselves be artifacts, but we skip them from the text as they are restricted to very specific research contexts). Thus, an artifact is the noise registered by the system that contamines the neural EEG data (Urigüen & Garcia-Zapirain, 2015). 

Types of EEG artifacts 

The ability to recognize artifacts is the first step in removing them. EEG artifacts can be classified depending on their origin, which can be physiological or external to the human body (non-physiological). The most usual are (Sörnmo & Laguna, 2005; Urigüen & Garcia-Zapirain, 2015; Clark, 1998; Ismal et al., 2016):

Physiological artifacts

  • Ocular activity
  • Muscle activity
  • Cardiac activity
  • Perspiration
  • Respiration

Non-physiological / Technical artifacts

  • Electrode pop
  • Cable movement
  • Incorrect reference placement
  • AC electrical and electromagnetic interferences
  • Body movements

The EEG device used to create most of the figures showing the artifacts was collected with the Bitbrain EEG versatile 16ch system, band pass filtered between 0.5 and 30 Hz. Independent components were extracted using logistic infomax ICA algorithm (Bell & Sejnowski, 1995).

Physiological artifacts

Ocular activity:

  • Origin: The eye can be electrically modeled as a magnetic dipole and it distorts the electric field in the region when it moves.
  • Why it affects EEG:  This distortion is known as the EOG (Electrooculogram) signal and has an amplitude usually one order of magnitude larger than the EEG signal, reaching values around 100-200 microvolts.
  • Types of effects: Blinking, lateral movement, eye movements
  • Effect on time-domain: Blinking produces a quick change with high amplitude on the EEG signals in the electrodes of the frontal area, more pronounced in those closer to the eyes. Lateral movements of the eye affect also the frontal areas but are more significant the closer to the temples. In general the artifact amplitude of the artifact is almost proportional to the angle of gaze. 
  • Effect on frequency domain: Effect in low frequencies that can be confused with delta and theta bands.

eeg eye blink artifact
Lateral eye movement

Muscle activity:

  • Origin: Muscles produce electrical activity when they are contracted. This activity can be measured and the resulting signal is called electromyography (EMG).
  • Why it affects EEG: That electrical activity produced by the muscles can interfere with the actual EEG activity. We can observe these high frequency artifacts with the naked eye.
  • Types of effects: Clenching the jaw, neck and shoulder muscles tension, swallowing, chewing, talking, sucking, sniffing, grimacing, frowning or hiccupping
  • Effect on time domain: We can observe a high frequency signal that overlaps the EEG signal. The amplitude correlates with the strength of the muscle contraction.
  • Effect on frequency domain: Effect in high frequencies overlapping artifacts in beta and gamma EEG bands

Jaw clenching eeg artifact

Cardiac activity:

  • Origin: Electrical activity from the heart. This signal is called Electrocardiogram (ECG) but also referred as to pulse artifact.
  • Why it affects EEG: Although the amplitude of the ECG is low on the scalp, sometimes, depending on the placement of the electrode or the body shape of the participant we would see a rhythmic distortion on the EEG signals.
  • Types of effects: Cardiac activity, pulse
  • Effect on time domain: A rhythmic pattern, corresponding with the heartbeats that overlaps the EEG signal.
  • Effect on frequency domain: The frequency components of the ECG overlap EEG band frequencies so it is difficult to visualize them with the naked eye.

Cardiac activity eeg artifact

Perspiration:

  • Origin: Sweat glands of the skin 
  • Why it affects EEG: Small drops of sweat produced by the glands cause changes in the electrical baseline of the electrodes. In case of intense perspiration it could even create shorts between electrodes.
  • Types of effects: Sweat glands, skin potentials.
  • Effect on time domain: Slow waves overlapping the EEG signal.
  • Effect on frequency domain: Low frequency artifact that overlaps delta and theta bands principally.

perspiration-skin-eeg-artifact

Respiration:

  • Origin: Movement of chest and head when breathing (inhale / exhale)
  • Why it affects EEG: It is more common in sleep recordings as respiration-related movement modifies the contact between the electrodes and the scalp if the participant is lying on a bed.
  • Types of effects: Inhale, Exhale.
  • Effect on time domain: Slow waves synchronized with breathing rhythm that overlaps the EEG signals.
  • Effect on frequency domain: Low frequency artifact that overlaps delta and theta bands.

respiration eeg artifact

Non-physiological / Technical artifacts

Electrode pop:

  • Origin: Temporary failures in the contact between the sensor and the scalp produced by touching the sensor or by spontaneous changes in electrode-skin contact.
  • Why it affects EEG: It is due to changes in contact potential between the scalp and the electrode. 
  • Types of effects: Electrode pop.
  • Effect on time domain: Abrupt and usually high amplitude interference on the EEG signal usually localized in a single channel. 
  • Effect on time domain: The characterization of an electrode pop is difficult due to the wide range of possible distortions.

Electrode pop by touching eeg artifact

Cable movement:

  • Origin: Movement of the cables connecting the electrodes and the amplification system.
  • Why it affects: changes in the electromagnetic fields produce distortion in the signal recorded and also in the scalp-sensor contact.
  • Types of effects: cable movement, cable touch.
  • Effect on time domain:  It is very dependent on the type of cable movement. If the movement is rhythmic, distortions overlapping EEG signals will appear with the same rhythm that the cable movement.
  • Effect on frequency domain: It also depends on the type of movement. If movements are rhythmic we can find non-EEG related frequency peaks.

cable movement eeg artifact

Incorrect reference placement:

  • Origin: Reference channel not placed or bad contact on the reference channel.
  • Why it affects: the signal recorded is not EEG.
  • Types of effects: reference sensor not placed.
  • Effect on time domain: abrupt changes in all the channels with high amplitude. All channels will converge slowly (filtering effects) to actual EEG signals when the reference is placed properly.
  • Effect on frequency domain: very high power in all channels, and in non-eeg related eeg signals.

eeg artifact due to the bad placement of the eeg reference

AC electrical and electromagnetic interferences:

  • Origin: AC electrical lines and devices
  • Why it affects EEG: Due to insufficient or lack of wire shielding, the signal can be affected by surrounding electromagnetic fields like AC power sources and wires. 
  • Types of effects: 50 Hz. or 60 Hz. 
  • Effect on time domain: You can observe a high frequency noise continuously overlapping the EEG signal.
  • Effect on frequency domain: You will see a big spike around 50 Hz. or 60 Hz. depending on the AC frequency standard for the country you are in (50 Hz. or 60 Hz. artifact).

graph on the eeg artifact due to AC interferece

Body movements:

  • Origin: Body movements, principally affected by head movements.
  • Why it affects: When moving, although unintentionally, the contact between electrode and skin is affected and the EEG signal corrupted. 
  • Types of effects: Head movements, arm movements, walking, running.
  • Effect on time domain: temporary slow waves corresponding with the rhythm of the movement.
  • Effect on frequency domain:  Effect is localized in lower frequencies overlapping delta and theta bands.

graph on eeg artifacts fue to body movements

EEG artifact filtering techniques (by data analysis)

There are four main ways to deal with artifacts depending on the data analysis:

1. EEG artifact Rejection

The first approach is to select and reject EEG epochs with artifacts. The different techniques define a pattern (usually one of the above artifacts) to select EEG epochs to be removed. The pattern identification methods range from visual inspection by an EEG expert, to automated statistics in the time or frequency domain (Nolan et al., 2010). For example, in an ERPs protocol, you can define a statistical threshold to remove trials that have a significantly higher amplitude. 

eeg-artifacts-filtering-technique-rejection

Rejection is a very costly method as, while almost all artifacts can be removed,  all the valuable EEG information of the epoch is also eliminated. Typically you will be interested in retaining as much of the EEG data as possible, especially when the recordings are short.

2. Filtering

The goal of these techniques is to remove the artifacts while keeping as much EEG information as possible. This classification includes techniques such as a simple linear filter to remove certain frequency bands (Panych et al.,1989), regression methods to remove EOG or ECG signals from EEG using a reference signal (Wallstrom et al., 2004), adaptive filters with reference signal (Marque et al., 2005), Wiener filters (Sweeney et al., 2012)  or Bayes filters (Sameni et al., 2007). 

For example, we can use linear filters to remove the 50 Hz. or 60 Hz. AC electrical interference.  This will also remove the EEG information (brain waves), however such high frequencies are not usually the focus of EEG studies. Another example is the use of EOG signal as a reference channel to remove that info from the EEG contaminated signal by regression or adaptive filters. 

Regression methods assume the recorded EEG is a combination of real EEG and artifacts (EOG). The regression filter calculates the proportion of the references (EOG) that are present in a single EEG channel and subtracts it. 

eeg-artifacts-filtering-technique-filtering
3. Blind Source Separation

These are demixing techniques that attempt to decompose the EEG into a linear combination of signal sources based on different mathematical considerations, such as orthogonality or independence. The most popular and useful technique is the Independent Component Analysis (ICA) (Choi et al., 2005), which linearly demixes the EEG into mathematically independent components or sources. As noise is usually uncorrelated and independent from EEG sources, we can observe that some components include the artifact information.

For example, they can be used to eliminate EOG or EMG artifacts by marking the noise sources (manually or via machine learning techniques as MARA (Winkler et al., 2011) and then removing them and linearly reconstructing the clean EEG data from the rest of components. 

Blind-Source-Separation-eeg-filtering-technique-for-artifacts

The advantage of the Blind Source Separation method is that it does not need a reference channel or any previous information about the noise. The main limitation is that it uses the full EEG matrix instead of filtering by channel, and does not work as well when the number of channels is reduced or the EEG data available is small.

4. Source decomposition methods:

These methods decompose each individual channel into basic waveforms, eliminating the ones that contain an artifact, and then reconstructing the EEG clean channel. The main example of these methods is the Wavelet decomposition (Unser & Aldroubi, 1996), and some variants less investigated such as the Empirical Mode Decomposition (EMD) (Safieddine et al., 2012), or Non-linear Mode Decomposition (NMD) (Iatsenko et al., 2015). 

In wavelet decomposition, the signal of each channel is decomposed in coefficients for different scales and drifts of the selected wavelet (“mother”). To filter the signal, after decomposing it, some coefficients are thresholded and then the signal is reconstructed. 

Source decomposition methods for eeg artifacts

The main advantage of these methods is that we can retain EEG data at a channel level.  The main disadvantage is that we need to find a correct basic waveform (wavelets, intrinsic mode functions, nonlinear modes) to decompose the noise in order to be able to threshold coefficients that only remove the artifacts without removing EEG data. They are also more complex and still under research.

EEG artifact filtering techniques (by online/offline operation)

 An important aspect of these techniques is whether they operate offline or online  (Ismal et al., 2016). Offline methods are not automatic and require human intervention, and therefore can not be integrated in a system that runs autonomously. For instance, visual inspection to reject EEG epochs or visual selection of artifactual components/sources are offline methods that require the supervision of an expert.

Online methods can be fully automated and integrated in a system that runs autonomously. For example, methods with a reference signal as regression or adaptive filters can easily run online. Also, processes involving signal decomposition as Blind Source Separation ones or Source decomposition methods can be automated by establishing some thresholds or statistic thresholds from clean EEG data to automatically remove components. 

Software for EEG artifacting

There are several toolboxes and libraries available for EEG signal filtering. Here we are going to focus on a small subset of them that are probably the most used ones at the time this document was written. All of them are software libraries that can be used independently of the EEG system that acquires the data:

  1. EEGLAB (EEGLAB, EEGLAB Wiki): This is an interactive Matlab toolbox for processing continuous and event-related EEG, MEG and other electrophysiological data. It includes filtering techniques such as independent component analysis (ICA) or artifact rejection and several filtering plugins can be downloaded to increase the toolbox potential. It also incorporates time/frequency analysis, event-related statistics, and several modes of visualization of the averaged and single-trial data. EEGLAB runs on Windows, Mac OS X, Linux and Unix. 
  2. FieldTrip (FieldTrip toolbox): This is a MATLAB toolbox for MEG, EEG, iEEG and NIRS analysis. It offers preprocessing techniques and analysis methods, such as time-frequency analysis or source reconstruction using dipoles. It supports the data formats of all major MEG systems and of the most popular EEG, iEEG and NIRS systems and new data formats can be added easily. You can implement your own analysis protocols in a MATLAB script using FieldTrip high-level functions. FieldTrip  is an open source software under the GNU general public license.
  3. MNE (MNE — MNE 0.20.0 documentation): Open-source Python software for exploring, visualizing, and analyzing human neurophysiological data: MEG, EEG, sEEG, ECoG, and more. The software has a growing community behind and several python packages has been developed to add a graphical user interface, automatic bad channel detection and interpolation, independent component analysis (ICA), connectivity analysis, general-purpose statistical analysis of MEG/EEG signals or a python implementation of the Preprocessing Pipeline (PREP) for EEG data among others.

Conclusions

Several artifacts and methodologies to remove or reject them have been presented in this post. There is not yet a magical rule to deal with all the spectrum of possible artifacts at the same time. Depending on the experimental nature of the data collection, some artifacts are more prone to appear and some removal methodologies fit better. 

There is a lot of research literature about using a combination of methods to deal with artifacts (Akhtar et al., 2012; Mijovic et al. 2010). ICA-based algorithms seem to be the default first approach to EEG filtering if simpler methods as regression or rejection are not valid for your application (Urigüen & Garcia-Zapirain, 2015).  As a rule, it is important to remind participants of the nature of EEG and ask them to avoid movements or actions that might contaminate the signals as much as possible. Moreover, non-physiological artifacts should be minimized by the experimenter.

If artifacts are reduced to those that are unavoidable (eye-movements or small body movements), it will be easier to pick a correct tool to remove those specific artifacts and clean the EEG data (Sörnmo & Laguna, 2005; Urigüen & Garcia-Zapirain, 2015Ismal et al., 2016). However, this is not always feasible when the EEG monitoring is performed in natural conditions. 

References

  • Akhtar, M. T., Mitsuhashi, W., & James, C. J. (2012). Employing spatially constrained ICA and wavelet denoising, for automatic removal of artifacts from multichannel EEG data. Signal processing, 92(2), 401-416.
  • Bell, A. J., & Sejnowski, T. J. (1995). An information-maximisation approach to blind separation and blind deconvolution. Neural Computation, (February 1995), 1004–1034.
  • Choi, S., Cichocki, A., Park, H. M., & Lee, S. Y. (2005). Blind source separation and independent component analysis: A review. Neural Information Processing-Letters and Reviews, 6(1), 1-57.
  • Clark, J. W. (1998). The origin of biopotentials. Medical instrumentation: application and design, 3, 121-182.
  • Correa, A. G., Laciar, E., Patiño, H. D., & Valentinuzzi, M. E. (2007). Artifact removal from EEG signals using adaptive filters in cascade. In Journal of Physics: Conference Series (Vol. 90, No. 1, p. 012081)
  • Iatsenko, D., McClintock, P. V., & Stefanovska, A. (2015). Nonlinear mode decomposition: a noise-robust, adaptive decomposition method. Physical Review E, 92(3), 032916.
  • Ismal, K., Rastegarnia, A., & Yang, Z. (2016). Methods for artifact detection and removal from scalp EEG: A review. Clinical Neurophysiology, 46, 287–305.
  • Marque, C., Bisch, C., & Dantas, R. (2005). Adaptive filtering for ECG rejection from surface EMG recordings. Journal of Electromyography and Kinesiology, 15, 310–315.
  • Mijovic, B., De Vos, M., Gligorijevic, I., Taelman, J., & Van Huffel, S. (2010). Source separation from single-channel recordings by combining empirical-mode decomposition and independent component analysis. IEEE transactions on biomedical engineering, 57(9), 2188-2196.
  • Nolan, H., Whelan, R., & Reilly, R. B. (2010). FASTER : Fully Automated Statistical Thresholding for EEG artifact Rejection. Journal of Neuroscience Methods, 192, 152–162.
  • Nunez, P. L., & Srinivasan, R. (2006). Electric fields of the brain: The neurophysics of EEG. Oxford: Oxford University Press.
  • Panych, L. P., Wada, J. A., & Beddoes, M. P. (1989). Practical digital filters for reducing EMG artefact in EEG seizure recordings. Electroencephalography and Clinical Neurophysiology, 72, 268–276.
  • Safieddine, D., Kachenoura, A., Albera, L., Birot, G., Karfoul, A., Pasnicu, A., Biraben, A., Wendling, F., Senhadji, L. & Merlet, I. (2012). Removal of muscle artifact from EEG data: comparison between stochastic (ICA and CCA) and deterministic (EMD and wavelet-based) approaches. EURASIP Journal on Advances in Signal Processing, 2012(1), 127.
  • Sameni, R., Shamsollahi, M. B., Jutten, C., & Clifford, G. D. (2007). A Nonlinear Bayesian Filtering Framework for ECG Denoising. IEEE Transactions on Biomedical Engineering, 54(12), 2172–2185.
  • Sörnmo, L., & Laguna, P. (2005). Bioelectrical Signal Processing in Cardiac and Neurological Applications (Vol. 1). Elsevier.
  • Sweeney, K. T., Member, S., Ward, E., Member, S., Mcloone, F., & Member, S. (2012). Artifact Removal in Physiological Signals — Practices and Possibilities. IEEE Transactions on Information Technology in Biomedicine, 16(3), 488–500.
  • Unser, M., & Aldroubi, A. (1996). A review of wavelets in biomedical applications. Proceedings of the IEEE, 84(4), 626-638.
  • Urigüen, J. A., & Garcia-Zapirain, B. (2015). EEG artifact removal — state-of-the-art and guidelines. Journal of Neural Engineering, 031001(12).
  • Wallstrom, G. L., Kass, R. E., Miller, A., Cohn, J. F., & Fox, N. A. (2004). Automatic correction of ocular artifacts in the EEG : a comparison of regression-based and component-based methods. International Journal of Psychophysiology, 53, 105–119. 
  • Winkler, I., Haufe, S., & Tangermann, M. (2011). Automatic Classification of Artifactual ICA-Components for Artifact Removal in EEG Signals. Behavioral and Brain Functions, 7(1), 30.

You might also be interested in: