Sign in to Peak Brain Path
Sign in to access your courses, books, and progress tracker. New here? Signing in creates your account automatically.
Want to explore courses first?
Browse courses and booksSign in to access your courses, books, and progress tracker. New here? Signing in creates your account automatically.
Want to explore courses first?
Browse courses and booksModule 8
Chapter 8 · 1.5 h · 8 quiz items · pass at 80%
BCIA Domain III treats artifact as a core competency because contaminated signal corrupts everything downstream. This module teaches the practitioner to recognize each artifact by its waveform and site, to reject it manually or automatically, and to understand the specific danger of training on un-gated signal. The quiz proves the learner can tell physiology from artifact and explain what un-rejected artifact does to a brain map and a training session.
Once you can see artifact, you cannot unsee it, and you can no longer trust a report from anyone who never learned to. This is the skill that separates a practitioner whose data means something from one whose confident maps are built on jaw muscle and eye blinks. The amplifier records every voltage reaching the scalp and has no way to tell brain from non-brain. That judgment is yours. Get it wrong and everything after it inherits the error: a database comparison inflated by muscle, a Z-score driven by a blink, a protocol aimed at a pattern that was never in the cortex. Get it right and you are doing real clinical work instead of generating impressive-looking noise.
Artifact is any signal in the recording that does not come from cortical neurons. This chapter takes the major types one at a time by their waveform signature, frequency content, and spatial distribution, so you can name what you are seeing: muscle, eye, movement, electrode, line noise, and cardiac. Then it covers how to reject artifact (manual visual editing versus automated methods, and the adaptive filtering that can subtract some artifacts rather than discard the data), and finally the part most practitioners never think through, what artifact does specifically to live neurofeedback training, where there is no chance to clean the data afterward and a contaminated signal means you are reinforcing the wrong thing in real time. The flagship treatment of artifacting, with its full library of traces and the field and phase-reversal analysis, lives in the companion QEEG Field Guide. This chapter gives the BCN candidate the recognition skills and the rejection logic.
The reason artifact detection matters more than any other single technical skill is that artifact looks like brain activity to mathematics. The analysis software does not know that the 25 Hz energy at a temporal electrode is a clenched jaw rather than cortical beta; it computes the power, places it in the beta band, paints it on the topographic map, and compares it to a normative database of relaxed subjects who were not clenching. Out comes a Z-score of +2.5 and a report sentence about "excessive frontal beta suggesting hyperarousal," recommending neurofeedback to reduce something that was never a brain pattern. The client is then referred to train down their jaw muscle. This is not a rare failure. It is the most common serious problem in the field, and it is why a single discipline runs underneath this whole chapter: never interpret a processed map without inspecting the raw EEG that produced it.
A subtle corollary follows. Artifact carries information even though it is not brain activity, and you can use that information and still avoid the cardinal error. Sustained forehead tension may genuinely tell you a client is anxious; frequent movement may reflect real motor restlessness; excessive blinking may signal fatigue or dry eyes. You may note "the client showed sustained muscle tension consistent with anxiety" in your record. What you may not do is also count the muscle-driven beta as a brain finding and write "elevated beta indicates hyperarousal," because the beta elevation is the muscle. The artifact is evidence about the person's state. It is not evidence about their EEG spectrum. Holding those two apart is the discipline.
Muscle is the most common contaminant after the eyes, and the most dangerous because it lands squarely in the bands neurofeedback cares about. Contracting muscle generates electrical activity (electromyographic, or EMG, activity) at high frequencies, characteristically 20 to 300 Hz, which overlaps and overwhelms the beta and gamma bands.
The signature. On the raw trace, muscle reads as high-frequency "fuzz" or "grass": dense, irregular, fast activity riding on top of the background, quite unlike the slower, more rhythmic and organized look of true brain beta. In the frequency domain it shows as a broadband elevation across the high frequencies rather than a clean spectral peak, which is one of the most reliable ways to distinguish it: brain rhythms make peaks, muscle makes a broad hump.
Where it appears. Spatial distribution is the other great tell, because each muscle sits under particular electrodes. The temporalis and masseter (jaw) muscles dominate the temporal sites T3/T4 (T7/T8) and reach the lateral frontal sites F7/F8, so jaw clenching or chewing produces maximal high-frequency fuzz at the temporal electrodes that drops off rapidly toward the midline. The frontalis (forehead) muscle, tensed by anxiety, concentration, squinting, or eyebrow-raising, contaminates the frontal sites Fp1/Fp2, F7/F8, and F3/F4, often bilaterally and often sustained for minutes in an anxious client. The neck muscles (trapezius, splenius) tense with poor posture or an unsupported head and contaminate the posterior sites O1/O2, P3/P4, and the posterior temporal electrodes, which is insidious because it corrupts the occipital alpha assessment in a region that is usually clean. Swallowing produces a brief, large, diffuse burst that is easy to recognize and easy to exclude.
The error to avoid. The recurring mistake is reading muscle beta as brain beta and then as "hyperarousal." A topographic map of "elevated temporal beta" with maxima at T3/T4 and nothing at the midline is almost always jaw muscle; verify by looking at the raw trace for the characteristic fuzz and by checking whether the elevation is a broadband hump rather than a clean peak. Real brain beta is lower in frequency (13 to 25 Hz), more rhythmic, and more broadly distributed; muscle is higher (25 to 100 Hz), irregular, and focal to the muscle.
Management. Prevent it by coaching the client to relax the jaw, settle the forehead, and let the head rest on support; catch it in real time and prompt relaxation when you see the fuzz developing; reject contaminated epochs in offline review; and when muscle is pervasive, state in the report that the affected band cannot be reliably assessed rather than reporting an artifact-inflated value.
The eye is an electrical dipole, the cornea positive relative to the retina, so every eye movement and blink sweeps a voltage across the frontal scalp. Ocular artifact is the single most common contaminant in EEG, and it dominates the frontal channels.
Blinks (vertical movement). A blink produces a sharp, brief deflection (a typical blink lasts 200 to 400 milliseconds) that is maximal at the frontal-pole electrodes Fp1 and Fp2, bilateral and synchronous (both poles move together because both eyes blink together), and falls off rapidly toward the back of the head, so Fp1/Fp2 are large, F3/F4 smaller, and the posterior sites barely touched. The waveform is stereotyped, the same shape on every blink, and its energy is broadband, spilling into delta and theta. That last fact drives the classic error: a record full of blinks shows elevated frontal theta on the spectral analysis, the Z-score flags it, and it gets read as the frontal-theta pattern of ADHD when it is simply the client blinking. Never interpret frontal theta without first checking for blink contamination.
Lateral eye movements (saccades). When the eyes move side to side, the dipole swings horizontally, producing a different signature: opposite polarity at the two frontal poles, so Fp1 deflects up while Fp2 deflects down (or the reverse), because the cornea swings toward one eye and away from the other. Saccades are rapid (20 to 200 milliseconds), come in series during visual scanning, affect F7/F8 as well, and should not appear in a proper eyes-closed recording. Their signature error is manufacturing an apparent frontal asymmetry: the opposite-polarity deflections look like a left-right brain difference but are pure eye artifact. Sustained reading produces a continuous train of saccades (regular left-right sweeps with a large return sweep at each line's end) that both adds frontal artifact and suppresses the occipital alpha you were trying to measure, which is why the eyes-open fixation target must be a blank dot and any readable text must be out of the visual field.
Slow rolling eye movements. A different ocular pattern, slow sinusoidal eye rolls around 0.5 to 1 Hz with large amplitude, signals not contamination but a state change: the client is becoming drowsy. These accompany alpha fragmentation and rising theta and mean the recording has captured the transition toward sleep rather than a waking baseline. The response is to rouse the client and note or re-record, because drowsiness theta is a real physiological state but not the waking pattern you intend to interpret.
Any head or body movement shifts the electrodes relative to the scalp, momentarily changing the electrode-skin contact and transmitting mechanical force through the cables. The signature is large, slow voltage swings (typically below 1 Hz), often several hundred microvolts, affecting multiple channels at once because a movement disturbs many electrodes together, with an irregular morphology unlike any rhythmic brain oscillation. Movement artifact usually coincides with visible movement and frequently drags muscle and electrode artifact along with it. Cable movement deserves special mention: because the reference cable feeds every channel, a moving reference lead injects correlated artifact across the entire montage, which is one reason cables are secured and active electrodes (which buffer the signal at the site) resist this better. Management is mostly prevention and rejection: a comfortable, supported posture reduces the need to move, real-time coaching catches fidgeting, and the affected epochs are excluded.
The electrodes themselves fail in characteristic ways. An electrode "pop" is a sudden, large, often biphasic voltage jump confined to a single channel, produced when that electrode's contact abruptly changes, and it is sharply focal, affecting only the derivations connected to the problem electrode. Impedance drift is the slower cousin: as gel dries over a long session or an electrode gradually loosens, its impedance creeps up, the channel grows noisier, and the signal may wander, sometimes producing slow waves that masquerade as low-frequency brain activity. A single high-impedance electrode shows excessive broadband noise on its own channel while its neighbors stay clean, which is the tell that the problem is the electrode and not the brain. Bridging is the opposite failure: too much gel connects two adjacent electrodes into a conductive bridge, so the two channels show nearly identical waveforms when they should be independent, collapsing the spatial resolution in that region. The detection logic for all of these is that they are channel-specific and physically, not physiologically, distributed: a problem isolated to one electrode (or two bridged ones) that does not respect the smooth spatial gradients real brain activity follows is an electrode problem, confirmed by an impedance check.
The electrical power system radiates an electromagnetic field at the mains frequency (60 Hz in North America, 50 Hz in much of the world), the EEG leads act as antennas, and the amplifier picks it up. The signature is unmistakable: a perfectly regular sinusoid at exactly the line frequency, showing in the spectrum as a sharp, narrow peak at 60 (or 50) Hz, typically with harmonics at 120, 180, and 240 Hz. When the source is environmental it appears on all channels. When it appears on one channel it usually points to a single bad electrode or connection. The sources are fluorescent lights (especially older ballasts), outlets and wiring, nearby monitors and electronics, and poor grounding, and the same impedance physics from Chapter 6 applies: imbalanced electrode impedance lets common-mode line interference through the differential amplifier, so good, balanced impedance is itself a line-noise control.
The right response is prevention, not filtering: clean up the electrical environment, turn off and unplug nearby electronics, prefer a battery-powered amplifier, and balance the impedances. A notch filter can remove the 60 Hz peak after the fact, but it carries real caveats you must understand. It removes not just the noise but any genuine brain activity sitting at 60 Hz; it can introduce ringing of its own; and it does nothing about the harmonics unless each is filtered separately. A notch filter also sits right at the boundary of the high-beta and gamma ranges, so leaning on it routinely distorts exactly the bands where muscle and high-frequency assessment already demand care. Prevent the noise at the source whenever you can, and reach for the notch knowing what it costs.
The heart is a large electrical generator, and its signal reaches the scalp, particularly through electrodes near the ears, neck, and major vessels. The signature is regular spikes at the heartbeat rate (roughly 60 to 100 per minute, about 1 to 1.7 Hz), with the sharp QRS morphology of the cardiac complex, appearing most often at the temporal and occipital sites and especially when an ear reference sits near a pulsing vessel. The defining feature is its timing: the artifact is locked to the pulse, so a simultaneous ECG channel (when available) confirms it by showing the time-locked relationship, and even without one the metronomic regularity at pulse rate gives it away. A related pulse artifact, a mechanical bump from a vessel pulsing under an electrode, lags the heartbeat slightly because the pulse wave takes time to travel. Cardiac artifact's clinical impact on routine QEEG is usually modest because the heart rate is too slow to land in the clinically central bands, but it matters for sub-delta analysis, it can be mistaken for a real rhythm if not recognized, and it is one of the artifacts that independent component analysis isolates and removes cleanly thanks to its stereotyped, regular pattern.
Two more patterns belong here because they are routinely mis-handled, though strictly they are real physiology rather than artifact. Drowsiness is a state, not a contaminant: it shows as alpha fragmenting and slowing (10 Hz drifting toward 9 then 8), rising frontal theta, slow rolling eye movements, and eventually sleep features (spindles at 12 to 15 Hz centrally, K-complexes, vertex sharp waves). The danger is reading drowsiness theta as a waking trait pattern. The distinguishing feature is that drowsiness evolves over the recording and reverses when you rouse the client, whereas a true trait pattern is stable across the session and present when the client is plainly alert. The mirror caution is not to over-reject genuine physiology: high-amplitude slow activity and a slower posterior rhythm are normal in children, mild theta and a slightly slow alpha are normal in the elderly, and the breach rhythm of faster, sharper activity near a skull defect is real EEG, not artifact. The skill is bidirectional, recognizing non-brain signal as artifact and refusing to discard real brain signal that merely looks unusual.
Recognizing artifact is half the job. Deciding what to do about it is the other half, and there are three broad strategies, each with a proper place.
The foundational method is a trained human scrolling through the raw record and marking what is artifact. Its strengths are pattern recognition, contextual judgment (is this borderline theta drowsiness or trait?), and the flexibility to weigh the clinical question, and it remains the standard against which automated methods are judged. Its weaknesses are that it is slow, somewhat subjective, dependent on the reviewer's expertise, and vulnerable to fatigue. A systematic workflow keeps it honest: an initial fast scan of the whole record for a quality impression, then an epoch-by-epoch review marking each segment clean or rejected and noting the artifact type, a cross-check in a second montage for borderline cases (a pattern that vanishes when you re-reference is suspect), a frequency-domain pass to confirm that "elevated" bands are not artifact-driven, and a final tally of how much clean data survived. Done properly this takes 10 to 15 minutes or more, and rushing it inviting contamination into the analysis.
Software can flag or remove artifact at scale, and several approaches are in common use. Threshold-based rejection discards any epoch whose voltage exceeds a set limit (say 100 microvolts). It is fast and objective and removes gross contamination, but it misses subtle artifact, cannot tell artifact from a legitimately high-amplitude alpha rhythm, and does not distinguish artifact types. Template matching recognizes characteristic shapes such as blinks but misses variations. Independent component analysis (ICA) is more powerful and more interesting (treated separately below). Machine-learning classifiers, from support vector machines to deep neural networks, can learn complex artifact patterns and process huge datasets tirelessly, but they generalize poorly off the equipment and population they were trained on, only detect artifact types represented in their training data, cannot integrate clinical context, struggle with borderline cases, and are often opaque "black boxes." Many published machine-learning detectors report impressive accuracy that collapses in real deployment because they were trained and tested on the same subjects or a single recording system. The reasonable stance toward automated tools is to trust externally validated, transparent methods used as first-pass screening with human review of what they flag, and to distrust black-box commercial systems with no validation data and claims of "perfect" detection.
A distinct strategy does not reject contaminated data but mathematically removes the artifact while keeping the underlying brain signal, which is valuable when the artifact is frequent enough that epoch rejection would throw away too much. Independent component analysis is the workhorse here. ICA decomposes the multichannel recording into a set of statistically independent components, the practitioner identifies which components are artifact (a blink component has a frontal topography and the characteristic waveform; a cardiac component fires at pulse rate; a muscle component is temporal and high-frequency), those components are zeroed out, and the rest are recombined into a cleaned recording. ICA excels at blinks and cardiac artifact, both stereotyped and well-isolated, and handles frequent lateral eye movements well. EOG regression, an older adaptive approach, uses a recorded eye channel to estimate and subtract the ocular contribution from the EEG channels, on the same subtraction principle.
The critical pitfall, and it is a serious one, is that ICA can remove brain activity that resembles artifact. A dominant posterior alpha rhythm, with its high amplitude and rhythmic morphology, can be flagged by automated component-scoring as a large artifact and removed, silently cutting alpha power and producing a false "reduced alpha" finding that the cleaning process itself created. This is not hypothetical. It is a well-documented failure mode. ICA also needs adequate data (several minutes) and enough channels (19 or more) to decompose well, struggles to separate continuous muscle from brain beta, and degrades the signal if too many components are removed. The non-negotiable safeguard is verification: always compare the spectra before and after component removal, confirm that expected patterns (alpha, if present) still survive, and decline to remove any component you are not sure is artifact. ICA is a powerful tool that requires human judgment at the identification step, not blind trust in an automated artifact score.
The field's consensus is a hybrid workflow: automated pre-screening to handle the bulk efficiently, then expert human review of the flagged segments and decisions, selective ICA for specific artifacts with before/after verification, documentation of both the automated parameters and the manual judgments, and a final confirming scan. This combines the efficiency of automation with the contextual judgment of a human, and it is the gold standard for clinical artifacting. One practical reason to understand the named research preprocessing pipelines (PREP for adult resting-state, the HAPPE family for developmental and high-artifact data, ICLabel for automated component classification, and others) is that normative databases are increasingly built with them, so knowing what a pipeline did to the data helps you understand what "normal" means in that database, and knowing these validated options keeps you from improvising ad hoc cleaning with unknown properties (Bigdely-Shamlo et al., 2015; Gabard-Durnam et al., 2018; Pion-Tonachini et al., 2019).
Two distinct settings suffer from artifact, and the second is the one this book's audience most needs to internalize.
In offline QEEG, the damage is Z-score inflation and false findings. Because the normative databases were collected from relatively clean recordings, any artifact in your recording that the database subjects did not have pushes your numbers above theirs: a client clenching their jaw shows temporal beta at +3.0 standard deviations, the report reads "significantly elevated temporal beta," and the reality is muscle. The same logic produces spurious "frontal theta" from blinks, false "asymmetry" from saccades, and impressive but meaningless source-localization blobs when an algorithm tries to localize a blink as if it were cortex. The mitigations are to artifact your data to the same standard the database used, to view Z-scores in artifact-prone regions (frontal and temporal beta especially) with deliberate skepticism, and above all to check the raw data before trusting any processed map. The cost of skipping this is paid by clients who get the wrong protocol because a number went red.
Here is the section the QEEG-centric treatments underplay and the practicing neurofeedback clinician cannot afford to. In offline analysis you get a second chance: you can scroll back, mark the artifact, and exclude it. In live training there is no second chance. The feedback is computed and delivered in real time, instant by instant, from whatever signal is arriving at that moment, and if that signal is contaminated, the client is being rewarded or inhibited based on artifact while it happens. The question every practitioner must keep in front of them is blunt: if the training is not artifact-gated, what are you actually reinforcing?
Consider the concrete failure. You are running a theta/beta protocol that rewards low beta at Cz. The client, anxious or concentrating, tenses their forehead or clenches their jaw. That muscle activity lands in the beta band, the reward band's amplitude jumps, and the system delivers reward, a tone, a point, a satisfying advance in the game. The client's brain, doing exactly what operant conditioning trains it to do, registers that whatever it just did produced the reward and does more of it. But what it just did was tense a muscle, not produce cortical beta. Over a session, over a course, you can inadvertently shape muscle tension, teaching the client to clench in pursuit of a reward that was supposed to be cortical. The same trap waits on the inhibit side: if blinks or movement inflate a theta-inhibit band, the client gets penalized for artifact and may learn to suppress blinking or hold still in ways that have nothing to do with the cortical change you intended.
This is why every clinical neurofeedback system worth using applies artifact inhibits: thresholds on the artifact-prone bands (high-frequency EMG, the low-frequency and high-amplitude excursions of movement and blinks) that withhold reward whenever artifact is present, so the client can only earn the reward with clean signal. When EMG in the high beta range spikes, the artifact inhibit blocks the reward regardless of what the reward band is doing; when a blink or movement throws a large transient, the reward is gated off. Setting these inhibits is not an optional refinement. It is part of configuring the protocol correctly, and a session run without them is a session that may be training the wrong thing. Real-time signal hygiene matters as much as offline artifacting: you watch the live trace for developing muscle fuzz and coach the client to relax, you keep impedances good so the signal stays clean across the session, and you treat a feed full of artifact as a reason to pause and fix the setup rather than push on. Automated real-time cleaning methods exist, such as artifact subspace reconstruction (ASR), which identifies and reconstructs high-variance contaminated segments on the fly and is built for exactly this real-time context (Mullen et al., 2015), but ASR has its own limits (it needs a clean calibration segment, struggles with sustained muscle, and can over-correct), so it supplements rather than replaces good electrode technique and artifact inhibits. The bottom line for the chair is that an artifact-gated session trains the brain you intend to train, and an ungated one may be quietly reinforcing the client's jaw.
When you sit with a recording, the order of operations is fixed: inspect the raw trace before you trust any map, identify artifact by its three tells together (waveform morphology, frequency content, and spatial distribution), reject or correct what you find with the lightest tool that does the job, verify any ICA removal against the before-and-after spectra, and document what you rejected and why so the data quality is part of the record. When you run a live session, configure artifact inhibits before you start, watch the live signal for muscle and movement, coach the client to stay relaxed and still, and keep the impedances clean across the whole session, because in real time there is no cleaning afterward.
For the BCN exam, fix each artifact by its signature. Muscle (EMG): high-frequency broadband fuzz, 20 to 300 Hz, maximal at temporal sites for jaw and frontal sites for forehead, the great false source of "elevated beta." Ocular (EOG): blinks are sharp, brief, bilateral, frontal-pole maxima that spill into theta and fake "frontal theta"; lateral saccades show opposite polarity at Fp1 versus Fp2 and fake "asymmetry"; slow rolling eye movements signal drowsiness. Movement: large slow swings under 1 Hz across many channels at once. Electrode artifact: pop is a sudden focal spike on one channel, drift is rising impedance and wandering signal, bridging makes two adjacent channels identical. Line noise: a perfectly regular sinusoid at exactly 60 (or 50) Hz with harmonics, prevented at the source, removed by a notch filter only with full awareness of its costs. Cardiac: regular QRS spikes locked to the pulse, temporal and occipital, usually minor for routine bands. Know that manual visual editing is the standard and automated methods (threshold, ICA, machine learning) supplement it within a hybrid workflow; that ICA removes artifact components but can destroy a real alpha rhythm if applied without verification; that artifact inflates Z-scores and produces false QEEG findings, which is why you always check the raw data; and that in live training an ungated session can reinforce artifact, which is why artifact inhibits gate the reward so the client earns it only with clean cortical signal. See the artifact, and you protect both the map and the brain at the chair.