EEG artefacts: causes, prevention and removal

An EEG artefact is any signal that enters an EEG recording from a source outside the brain or from the measurement set-up itself. Scalp EEG is measured in microvolts, while a blink, a jaw muscle or the mains supply can be many times larger and can fall into the same frequency bands. An analysis that fails to recognise artefacts may report an eye movement as a frontal effect, muscle tension as a beta rhythm or a loose electrode as an interesting channel. This guide covers the types of artefact, how they look in a recording, how to prevent them while recording, and which methods remove them afterwards and at what cost, from the point of view of research methodology.

01What is an EEG artefact, and why does it matter?

Artefacts are usually divided into two families. Physiological artefacts come from the participant’s body: the eyes, face and neck muscles, the heart, the sweat glands and movement. Technical artefacts come from the environment and the measurement chain: the mains supply, the contact between electrode and skin, the cables and the amplifier. In some cases a signal from the brain can also count as an artefact, depending on the research question; in a study of stimulus responses, for example, the alpha rhythm that rises when the participant closes their eyes.

Artefacts matter not only because they add noise; some imitate features of the EEG itself. Goncharova and colleagues (2003) showed that even a weak contraction of the forehead and temple muscles can produce EMG activity across the whole scalp that obscures or resembles the alpha, mu or beta rhythms. If an artefact is linked to an experimental condition, for instance if participants blink more in difficult trials, the result is not merely noisy but systematically biased. Managing artefacts is therefore not the last step of an analysis but part of the experimental design.

02Physiological artefacts

Blinks and eye movements (EOG)

The eye behaves like an electrical dipole whose front (the cornea) is positive relative to its back (the retina). When the eye turns or the eyelid slides over the cornea, this field changes and is picked up most strongly by the frontal channels (Fp1, Fp2, F7, F8). A blink leaves a large deflection in the frontal channels, symmetrical on both sides and lasting a few hundred milliseconds. Horizontal eye movements (saccades) produce step-like deflections of opposite polarity at F7 and F8, while vertical movements are prominent over the frontal midline. Gratton and colleagues (1983) showed that the factor by which the ocular signal propagates into EEG channels decreases from frontal to parietal channels and is larger for saccades than for blinks.

Muscle activity (EMG)

Clenching the jaw, frowning, swallowing, talking and neck tension all produce muscle activity. EMG is an irregular, fast-oscillating signal that extends to much higher frequencies than EEG. In the measurements of Goncharova and colleagues, EMG spanned a broad band from 0 Hz to above 200 Hz, with amplitude greatest around 20–30 Hz over frontal sites and 40–80 Hz over temporal sites. Muscle artefact is strongest in the edge channels near the muscle, but even a weak contraction could be measured near the vertex at frequencies above 12 Hz. This overlap makes muscle artefact hard to separate by filtering alone: brain activity in the beta and gamma bands lies at the same frequencies. The tongue also behaves like a small electrical dipole, so talking and swallowing can add slow deflections as well.

Cardiac activity (ECG) and pulse

The heart’s electrical field can appear as sharp, regular spikes synchronised with the heartbeat, particularly in earlobe, mastoid and temporal channels. When an electrode sits over an artery in the scalp, a pulse artefact arises instead: the pulsing vessel moves the electrode slightly, and slow waves recur shortly after each heartbeat. Both artefacts repeat at the heart rate, so they are easy to recognise when a separate ECG channel is recorded.

Sweat and skin potentials

In a warm room or a long session, sweat gland activity changes the electrochemical balance between electrode and skin. The result is very slow, broad baseline fluctuations lasting several seconds; they can look similar across neighbouring electrodes, and salty sweat can even form a bridge between two electrodes. Sweat artefact is usually concentrated at very slow frequencies below 1 Hz and falls in the same range as slow ERP components.

Head and body movement

Turning the head, talking, walking or leaning the cap against something shifts the electrodes on the skin and sways the cables. The recording shows large, irregular, mostly low-frequency deflections in many channels at once. In mobile paradigms such as walking, sport or virtual reality this artefact cannot be avoided entirely; an accelerometer on the cap records when movement happens and makes the artefact easier to recognise.

03Technical artefacts

Line noise (50 Hz)

The mains frequency in Türkiye and Europe is 50 Hz (60 Hz in most of the Americas). Power cables, adapters, lighting and nearby devices add sinusoidal noise at this frequency, sometimes also at its harmonics at 100 and 150 Hz. High impedance, or impedance that differs between electrodes, increases the noise because it weakens the amplifier’s common-mode rejection. It appears in the spectrum as a narrow, sharp peak at 50 Hz.

Electrode pop

When the gel dries or the electrode contact changes suddenly, a single channel shows an abrupt, steep jump followed by a slow return. That it affects only one channel is the strongest sign that it does not come from the brain: a field from the brain also spreads to neighbouring electrodes.

High impedance and poor contact

When contact between electrode and skin is poor, the channel becomes noisier, picks up more line noise and shows slow drifts. Active electrodes amplify the signal close to the skin, which reduces their sensitivity to interference picked up along the cable; good contact is still needed with every electrode type.

Cable movement

With passive electrodes, swaying or rubbing electrode cables produce low-frequency deflections. Fixing the cables to the cap and the body, carrying the amplifier close to the head, or using wireless systems reduces the risk.

Bridging

When excess gel or sweat connects two neighbouring electrodes, the two channels are electrically short-circuited and record almost the same signal. The difference between the two channels approaches zero and their correlation is unusually high. It is more common in densely spaced, high-density caps.

Amplifier saturation

When the signal exceeds the amplifier’s input range, the recording is cut off as a flat line at a limit value (clipping). Large electrode drifts, strong movement or electrical stimulation are among the causes. The information in a clipped segment cannot be recovered; such segments are excluded from the analysis.

Other environmental noise

Lifts, mobile phones, monitors and motorised devices can add electromagnetic noise, and vibrations such as a door slamming can add brief jumps.

04Artefacts at a glance: cause, appearance, prevention and removal

The table below summarises the most common artefacts. The removal column lists the usual options; which one fits depends on the research question and on how much data there is.

05Prevention during recording

No cleaning method can bring back a signal that was never recorded; the most effective steps in managing artefacts are taken before and during the recording. A laboratory protocol should cover the following points:

Skin preparation and impedance

Prepare the skin as the electrode type requires and bring impedance down to the level the manufacturer recommends before recording; write the values into the participant record.

Cap fit

Use a cap of the right size, placed in the right position, that does not slip; fix the cables to the cap and the body.

Extra channels

Record vertical and horizontal EOG (VEOG, HEOG) for eye movements, and ECG and neck EMG where needed; in mobile paradigms, record accelerometer data too. These channels make it easier to recognise artefacts later and to remove them with regression or ICA.

Instructions and experimental design

Ask participants to keep their jaw relaxed, to look at the fixation point on the screen and to save their blinks for short breaks between trials. Trial timing should be planned to allow for these breaks.

Calibration recording

Before the experiment, record a short segment in which the participant blinks on purpose, moves their eyes left and right and up and down, and clenches their jaw. This segment shows how artefacts look in that participant and gives the correction methods examples to work from.

Environment

Switch off unneeded electrical devices, keep power cables away from electrode cables and keep the room cool enough to prevent sweating.

Monitoring and notes

Watch the raw signal throughout the recording; note events such as coughing, talking or adjusting the cap, with their time, in the session log or as event markers in the recording.

Hardware suited to the paradigm

In mobile paradigms, active electrodes and wireless systems reduce cable-related artefacts; the choice should follow the movement the paradigm requires.

06Cleaning after recording: methods and trade-offs

After recording, artefacts are handled in one of two ways: the affected segment or channel is removed (rejection), or the artefact’s contribution is estimated and subtracted from the signal (correction). The review by Urigüen and Garcia-Zapirain (2015) notes that the best method depends on the type of artefact, the EEG signal and the ratio between signal and artefact, and recommends combining more than one method in successive stages.

Filtering and notch filters

A high-pass filter reduces slow drift and sweat fluctuations, a low-pass filter reduces part of the high-frequency muscle activity, and a 50 Hz notch filter reduces line noise. Filters are the cheapest and most predictable tool, but they only work for artefacts whose frequencies differ from the EEG. Blinks overlap the delta band and muscle activity the beta and gamma bands; these artefacts cannot be separated by filtering without harming brain activity. The high-pass cut-off also needs care: in ERP studies a high cut-off can distort slow components, so low values such as 0.1 Hz are usually preferred. Instead of a notch filter, methods that estimate the line component as a sinusoid and subtract it are also used; they aim to leave brain activity around 50 Hz less affected.

Interpolating bad channels

A channel that stays noisy, flat or disconnected throughout a recording is flagged by criteria such as amplitude deviation, correlation with neighbouring channels and the share of line noise. It can then be recomputed from its neighbours with spherical splines. The trade-off: an interpolated channel carries no new information, being only a weighted combination of its neighbours, so the number of independent dimensions in the data (its rank) falls, and if many channels in one region are bad, the repair is not reliable. Bad channels should be repaired before switching to an average reference; otherwise their noise spreads to every channel through the reference.

Epoch rejection

In data cut into trials, epochs that exceed an amplitude threshold (for example ±100 µV) or a peak-to-peak value are excluded from the analysis. The method is simple and transparent, but it loses data: if many epochs are rejected, statistical power drops, conditions can end up with different numbers of trials, and if an artefact is more frequent in one condition the result can be biased. The number of rejected epochs should be reported for every participant and condition.

Independent component analysis (ICA)

ICA separates a multichannel recording into statistically independent components. Blinks, eye movements, heartbeats and some muscle sources emerge as separate components; the researcher recognises them by their scalp maps, time courses and spectra, removes them and reconstructs the recording from the remaining components. Data are thus kept without deleting the affected segments. The MNE-Python documentation recommends fitting ICA on a copy high-pass filtered at 1 Hz; because filtering is a linear operation, the solution found can be applied to data that is unfiltered or filtered with a lower cut-off. The ICLabel plugin, which comes with EEGLAB, automatically estimates whether each component is brain, eye, muscle, line noise and so on.

ICA has its costs too. A good decomposition needs enough data that is reasonably clean; very noisy segments should be removed first. An average reference or channel interpolation lowers the rank of the data, and the EEGLAB documentation states that the number of components should then be reduced accordingly. A component can carry brain and artefact activity together, and removing it takes part of the brain signal with it, so component selection should be a reported decision based on criteria set in advance. Urigüen and Garcia-Zapirain consider ICA-based correction the safest approach when nothing is known in advance about the recording or the artefacts.

Artefact subspace reconstruction (ASR)

ASR learns the normal statistics of the data from a clean part of the recording; it then finds components that deviate strongly from those statistics in short, sliding windows and reconstructs that segment according to the structure of the clean data. It is automatic, can run in real time and is effective against large, transient artefacts such as movement and bursts of muscle activity; it is used through EEGLAB’s clean_rawdata plugin. Chang and colleagues (2020) reported, in 20 EEG recordings made during simulated driving, that ASR reduced eye and muscle components more than brain components, improved the quality of a subsequent ICA decomposition, and that a cut-off parameter between 20 and 30 balanced removing artefact against keeping brain activity. The trade-off: a cut-off that is too low also removes brain activity, and the method needs a clean calibration segment to start from.

Regression with EOG channels

When separate EOG channels are recorded, how much of the ocular signal spreads into each EEG channel is estimated as a propagation factor, and that contribution is subtracted. The method of Gratton, Coles and Donchin (1983) computes separate factors for blinks and saccades and keeps every trial instead of discarding those with artefacts. The trade-off: the EOG channel also records nearby brain activity, so regression removes part of the brain signal from the frontal region along with the ocular signal. The method cannot be applied to artefacts other than those from the eyes.

The strengths and weaknesses of the methods can be summarised as follows:

07An example preprocessing order

The order of the methods affects the result. The sequence below is a common one and matches the path taken by the analysis lab under this article; it should be adapted to the question of the study.

1. Inspect the raw recording and its spectrum; compare them with the notes in the session log.
2. Flag bad channels using criteria set in advance.
3. Filter: a high-pass filter for slow drift, a notch filter or a line-removal method for line noise, and a low-pass filter if needed.
4. Interpolate the bad channels, then change the reference (for example to the average reference).
5. Fit ICA on a copy high-pass filtered at 1 Hz; identify the eye, heart and clear muscle components and remove them from the main data.
6. Cut the data into epochs, apply baseline correction and reject the remaining epochs with large artefacts by threshold.
7. Report, for every participant, the number of rejected channels, components and epochs, together with the parameters used.

There are reasons for this order: if the average reference is applied before a bad channel is repaired, its noise spreads to every channel; if ICA is fitted with unfiltered slow drifts, the decomposition suffers; and leaving epoch rejection to the end means trials that could have been corrected are not thrown away needlessly. A final visual check of the cleaned data is still needed. In clinical EEG, telling artefacts apart and interpreting the recording depend on a physician’s assessment; this guide is written for preprocessing research recordings.

Active electrodes, wireless EEG systems and analysis software with offline artefact-processing tools are in the catalogue. For a recording set-up suited to your paradigm, you can reach us through the quote form.