Electroencephalography - Artifact Identification and Removal

EEG Artifacts and Removal Techniques Introduction When recording EEG, your electrodes pick up much more than just brain activity. Any recorded signal that does not originate from the brain is called an artifact. Artifacts are a major challenge in EEG research because they can mask genuine brain signals, distort your data, and lead to incorrect conclusions. Understanding where artifacts come from and how to remove them is essential for producing clean, interpretable EEG recordings. The key insight is that artifacts come from many sources—eye movements, muscle contractions, heartbeats, electrical equipment, and electrode problems—and each requires a different removal strategy. Your job as an EEG researcher is to identify which artifacts are present in your data and choose the most appropriate removal technique. What Are Artifacts and Where Do They Come From? Artifacts are any signals recorded at the scalp electrodes that do not represent neural activity. Common sources include: Ocular (eye) movements and blinks – The eyes act like an electrical dipole; when they move, they generate electrical fields that spread across the entire scalp. Muscle activity – Contractions of facial, neck, and scalp muscles produce electrical noise. Cardiac activity – The heart's electrical field can couple into the EEG recording. Tongue movements – Create artifacts similar to muscle activity. Electrode problems – Poor contact between the electrode and skin, or high impedance, produces "pops" and spikes. Environmental noise – 50 Hz or 60 Hz line noise from power supplies and nearby equipment. External equipment – IV drips and other medical devices can introduce contamination. The challenge is that these artifacts often have amplitudes comparable to, or larger than, genuine brain signals, making them impossible to ignore. Ocular Artifacts: Eye Movements and Blinks Eye-movement artifacts are particularly important because eyes are constantly moving, and the signals they generate are large and widespread. Understanding the mechanism helps you recognize these artifacts in your data. The human eye acts as an electrical dipole—the cornea (front of the eye) is positively charged relative to the retina (back of the eye). This creates a potential difference of roughly 15–100 microvolts. When the eyes move in their sockets, this dipole rotates, and the electrical field it produces changes across all the electrodes on the scalp. This is why eye movements contaminate not just frontal electrodes, but channels across the entire head. A special type of eye artifact occurs during eye blinks—the eyelid generates a "sliding potential" as it closes and opens. This produces characteristic large-amplitude transients called blink artifacts or Kappa artifacts, typically most visible at frontal electrode locations (like Fp1 and Fp2). To address eye-movement artifacts, researchers use electrooculography (EOG) electrodes placed around the eyes—typically on the outer corners (horizontal EOG) and above and below one eye (vertical EOG). These electrodes record eye movement directly. The eye-movement signal can then be estimated and subtracted from the EEG channels using regression methods, which we'll discuss later. This EEG recording shows the appearance of multiple channels over time. Notice how blink artifacts appear as large deflections, particularly in frontal channels. Muscular Artifacts Muscle contractions produce broadband electrical noise that spans a very wide frequency range: from $2 \text{ Hz}$ all the way up to $300 \text{ Hz}$. This is problematic because it contaminates important frequency bands, especially the gamma band ($30$–$100 \text{ Hz}$), which is often of research interest. Unlike ocular artifacts, which are localized to specific channels, muscular artifacts appear across all scalp channels simultaneously, though with different amplitudes depending on the nearby muscles. The amount of muscle artifact varies with: Body region – Neck and jaw muscles are particularly problematic. Sex and individual factors – Some participants naturally have more muscle activity. Contraction intensity – Stronger contractions produce larger artifacts. Muscle artifact is difficult to remove algorithmically because it overlaps so much with normal brain activity in frequency space. The most effective approach is often prevention—instructing participants to relax and avoid clenching their jaw or tensing muscles—followed by rejection of epochs containing obvious muscle activity. Cardiac Artifacts The heart generates its own electrical field through the organized contraction of cardiac muscle. This field can couple into the EEG recording, appearing as a periodic wave at approximately the heartbeat frequency (roughly $1 \text{ Hz}$ for resting heart rate). Cardiac artifacts are typically smaller than blink artifacts, but they're regular and systematic, which means they can be removed. To remove cardiac artifacts, researchers record an electrocardiogram (ECG) simultaneously with the EEG. The ECG electrodes are placed on the chest and directly measure the heart's electrical activity. Once the ECG is recorded, the cardiac component can be estimated from the ECG signal and subtracted from the EEG channels using regression or other cancellation methods. Environmental and Technical Artifacts Line noise is electrical contamination at $50 \text{ Hz}$ (in countries with 50 Hz power systems) or $60 \text{ Hz}$ (in countries with 60 Hz power systems). It enters the EEG recording through poor grounding or electromagnetic coupling with nearby electrical equipment. It appears as a narrow-band oscillation at exactly one frequency, making it relatively easy to identify and remove. Electrode artifacts arise from technical problems at the electrode-skin interface. When electrode impedance is high (poor contact) or when impedance changes rapidly (electrode "popping"), the result is high-amplitude spikes and noise that contaminate the signal. Prevention is key here: checking impedance before recording and ensuring good electrode contact eliminates most of these problems. Artifact Removal Strategies: Three Approaches There are three complementary strategies for handling artifacts: prevention, rejection, and cancellation. In practice, you typically use all three. Prevention The best artifact is the one that never enters the recording in the first place. Prevention includes: Proper electrode placement and secure mounting Checking electrode impedance before recording (impedance should typically be below $5\text{–}10 \text{ k}\Omega$) Using proper grounding and shielding to reduce environmental noise Instructing participants to minimize eye movements and muscle tension Checking for external sources of electrical noise (equipment placement, power cables) Prevention reduces the amount of cleaning needed later and improves data quality overall. Rejection The simplest removal strategy is to discard ("reject") any epoch of data contaminated by artifacts. You manually inspect the EEG recording or use automated algorithms to identify segments with large artifacts, then exclude them from further analysis. Advantages: Simple and guaranteed to remove the artifact completely. Disadvantages: You lose data, which reduces statistical power. If many epochs are contaminated, rejection can leave you with very little usable data. Cancellation Rather than discarding contaminated data, cancellation uses mathematical techniques to estimate the artifact contribution and subtract it while preserving the underlying neural signal. This is more sophisticated but also more powerful because you keep all your data. Algorithmic Approaches to Artifact Cancellation Several computational methods exist for canceling artifacts. Each has different strengths and is appropriate for different situations. Notch Filtering A notch filter is a very narrow bandpass filter that removes power at one specific frequency. It's highly effective for line noise because line noise is precisely $50 \text{ or } 60 \text{ Hz}$. How it works: The filter attenuates the signal only at the target frequency (and a narrow band around it), leaving the rest of the signal unchanged. Advantages: Simple, fast, and guaranteed to remove line noise. Disadvantages: Only works for narrowband contamination. If an artifact spans multiple frequencies (like muscle noise), a notch filter won't help much. Regression Methods Regression uses reference recordings (like EOG or ECG) to estimate how much artifact is present in each EEG channel, then subtracts that estimated artifact contribution. How it works: If you have an EOG recording, you assume the artifact in each EEG channel is proportional to the EOG signal. The algorithm finds the best-fit scaling factor for each channel: $$\text{EEG}{\text{clean}} = \text{EEG}{\text{raw}} - \beta \cdot \text{EOG}{\text{reference}}$$ where $\beta$ is the regression coefficient that scales the reference channel to best match the artifact component. Advantages: Uses actual reference measurements of the artifact source, so it's based on real information about what you're trying to remove. Disadvantages: Requires that you record reference channels (EOG, ECG), and it assumes a linear relationship between the reference and the artifact in your EEG channels. This assumption doesn't always hold. This EEG recording with frequency spectrum on the right shows contamination across multiple channels. Notice the peaks in the frequency domain (right side), which indicate dominant frequencies of contamination. Blind Source Separation: ICA and PCA The most powerful and flexible approaches are blind source separation methods, particularly Independent Component Analysis (ICA) and Principal Component Analysis (PCA). These methods don't require reference channels; instead, they mathematically decompose the entire recorded signal into independent or principal components. How it works: Imagine your 30-channel EEG recording is a mixture of different "sources"—some are brain activity, some are eye movement, some are muscle activity, and so on. ICA mathematically unmixes this combined signal, trying to find components that are as statistically independent as possible. PCA does something similar but prioritizes components that explain the most variance. Once you have these components, you inspect each one and identify which ones correspond to artifacts (eye movement components have a characteristic pattern, muscle components have high-frequency noise, etc.) and which ones are likely brain activity. You then remove the artifact components and reconstruct the EEG without them. This display shows multiple EEG components (left) with their frequency spectra (right). Each component is a potential source. Researchers identify which components are artifacts and remove them. Advantages: No reference channel needed Can separate complex, overlapping artifacts Very flexible and powerful Particularly good at isolating eye-movement artifacts and muscle artifacts Disadvantages: Requires expertise to identify which components are artifacts Can remove genuine brain activity if you're not careful in component selection Computationally intensive Results depend on the quality of the decomposition and your ability to interpret components <extrainfo> A Note on ICA Interpretation One common mistake is assuming that ICA components are neuroscientifically meaningful—that is, that each component corresponds to a specific neural source or brain region. This is not necessarily true. ICA simply finds components that are statistically independent; whether they correspond to meaningful brain activity or artifacts requires careful interpretation. Many researchers look at the spatial pattern (how the component's activity is distributed across the scalp) and temporal characteristics (frequency content, correlation with task events) to determine if a component is likely to be a brain signal or an artifact. </extrainfo> Summary: Choosing Your Artifact Removal Strategy In practice, a complete artifact-removal pipeline looks like this: Prevent artifacts during recording – Check electrodes, use proper shielding, instruct participants. Use notch filtering – Remove $50/60 \text{ Hz}$ line noise if present. Record reference channels – Capture EOG and ECG for later artifact estimation. Apply regression – Remove ocular and cardiac components using regression on reference channels. Use ICA or manual rejection – Identify and remove remaining artifacts, especially muscle activity, using blind source separation or by visual inspection and rejection. Different studies may emphasize different steps depending on the experimental design, the participant population, and the specific brain signals of interest. The key is to understand each method's strengths and weaknesses so you can make informed choices about your data.