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 10
Chapter 10 · 2.5 h · 8 quiz items · pass at 80%
This module opens IQCB Domain V (QEEG), at 21% the largest domain on the exam. It is the methodological heart of the credential: the candidate must understand every step from raw signal to quantitative metric, including how coherence and source methods handle volume conduction. The quiz confirms the learner can describe the pipeline and distinguish the power, coherence, and source measures a report will present.
Domain V is the heaviest single domain on the IQCB exam: 21 percent of your score, more than EEG interpretation, more than technical acquisition, more than neuroscience. The three chapters in this part carry that weight. This first one covers the machinery: how a recording becomes a set of numbers, what those numbers measure, and where each transform can mislead you. If you understand the pipeline at the level of "what does this metric assume, and when does the assumption break," you will answer most of the Domain V items without memorizing a single formula. Memorize the formulas anyway. The exam asks for both.
A point to settle before the methods. QEEG is disciplined pattern recognition under explicit constraints, not a diagnostic test. Every step below adds resolution and adds assumptions at the same time. The skill the exam tests, and the skill the clinic demands, is holding both in mind: knowing what each transform buys you and what it costs.
Quantitative analysis sits on top of a clean recording. It does not rescue a bad one. The garbage-in problem is absolute: a Fourier transform of muscle artifact produces a precise, reproducible, completely meaningless beta spectrum. The pipeline below assumes you have already acquired data competently (Chapters 5 and 6) and that you will review it ruthlessly before any number leaves the screen.
The standard pipeline runs in six stages.
1. Montage and re-referencing. The raw recording was acquired against some physical reference (linked ears or a single physical reference electrode). Before analysis you select the montage the database requires, which usually means re-referencing in software. Re-referencing is not cosmetic. It changes every voltage at every site, because EEG is reference-dependent at its core: you are always measuring a difference between two points, never an absolute potential. A frontal theta value computed against linked ears differs from the same data against an average reference. The database was built with a specific reference, and your recording must match it, or the z-scores are computed against the wrong distribution. Section 10.5 of the recording chapter covers montage mechanics. The methodology point here is that re-referencing happens before the transform and propagates into every downstream metric.
2. Artifact review. You inspect the raw traces and mark or remove everything that is not cortical EEG: eye blinks, lateral eye movements, muscle (EMG), cardiac (ECG), movement, electrode pop, and 60 Hz (or 50 Hz) line noise. This is the stage where clinical judgment dominates and where two competent analysts most often diverge. Automated tools assist (Section 10.8 on ICA, and the artifact methods in the recording chapters), but a human verifies. The exam will test whether you know that artifact handling sets what "clean" means, and that database developers artifacted their normative samples by their own standards, so your standards should match theirs.
3. Epoch selection. You divide the artifact-free record into analysis segments (epochs), one to two seconds each, and select enough of them to produce a stable spectrum. Section 10.2 covers the count and the reasoning.
4. The transform. You convert the selected epochs from the time domain (voltage over time) to the frequency domain (power over frequency), almost always by the Fast Fourier Transform, sometimes by wavelet methods. Sections 10.3 and 10.4 cover both.
5. Normalization. You compare the individual's metrics to an age-matched normative database, producing z-scores. This is Chapter 11. Here it is enough to say it is a stage, not an afterthought, and that it sits downstream of every choice above.
6. Visualization. The numbers become topographic maps, spectral plots, coherence matrices, and z-score displays. Chapter 12 covers reading them. A map is a rendering of the metrics, not new information, and a striking red blob represents exactly the z-score that produced it, no more.
Hold the order in mind for the exam: montage, artifact, epoch, transform, normalize, visualize. A question that asks "at what stage does reference choice affect the result" wants the answer "before the transform, and it propagates to everything after."
An epoch is one analysis window. The Fourier transform operates on a finite chunk of signal, and you decide how long each chunk is and how many you keep.
Why minimum data matters. A spectrum estimated from a few seconds of EEG is noisy: the power values bounce from epoch to epoch because each short window catches a slightly different slice of the brain's ongoing fluctuation. Averaging across many epochs smooths that variance and converges on a stable estimate. The standard clinical floor is two minutes of artifact-free data per condition. Less than that and the spectrum is not stable enough for confident z-scoring, in the lower-power high-frequency bands where a handful of contaminated epochs swing the mean. Two minutes is a floor, not a target. More clean data is better, and conditions that are hard to keep clean (eyes open, restless clients) need a longer raw recording to yield two clean minutes.
Epoch length sets frequency resolution. This is the single most testable fact in spectral QEEG, and Section 10.3 develops it. The length of each epoch determines the spacing of the frequency bins: a longer epoch resolves finer frequency detail. A one-second epoch gives 1 Hz resolution; a two-second epoch gives 0.5 Hz resolution; a four-second epoch gives 0.25 Hz resolution. Practitioners trade off here: longer epochs resolve frequency better but are harder to keep artifact-free (a two-second window contaminated by a single blink is lost entirely), and they reduce the number of independent epochs available from a fixed recording.
Selection criteria. Keep epochs that are free of all artifact across all channels, recorded in a stable behavioral state (alert for eyes-open and eyes-closed resting, not drifting toward drowsiness), and representative of the condition. Discard transitional segments where the client is settling, micro-sleeping, or shifting position. The selection is partly subjective, which is why documentation matters: a defensible QEEG record states how much data was retained and on what criteria.
The Fourier transform decomposes a time-domain signal into the set of sinusoids that, summed together, reproduce it. EEG is a messy waveform of voltage over time. The FFT tells you how much power sits at each frequency. The "fast" in FFT is an algorithm (Cooley-Tukey and descendants) that computes this efficiently on sampled data; the underlying idea is the discrete Fourier transform.
What it produces. For each epoch, the FFT returns power (or amplitude) at a series of discrete frequency bins. Power is amplitude squared, reported in microvolts squared (μV²) or microvolts squared per hertz (μV²/Hz) for power spectral density. You then average power across all retained epochs to get the spectral estimate for that channel, and you sum or average across the bins inside each band (delta, theta, alpha, beta) to get band power.
Frequency resolution, restated as a rule. The bin spacing equals the reciprocal of the epoch length: resolution = 1 / T, where T is the epoch duration in seconds. A 1-second epoch yields 1 Hz bins; a 4-second epoch yields 0.25 Hz bins. You cannot resolve a 10.2 Hz alpha peak from a 10.0 Hz peak with 1 Hz bins. When individual alpha peak frequency matters (and it does, because a slow-alpha individual's alpha spills into the fixed theta band), finer resolution helps.
Windowing. The FFT assumes the epoch is one cycle of an infinitely repeating signal. Real EEG epochs do not start and end at the same value, so the abrupt cut at each edge introduces spurious high-frequency energy, an artifact called spectral leakage. To suppress it, you multiply the epoch by a window function that tapers smoothly to zero at the edges. The Hanning (Hann) window is the common clinical default. Hamming, Blackman, and others trade off differently between leakage suppression and frequency smearing. The exam-relevant point: windowing is not optional decoration. Without it, leakage contaminates the spectrum, especially the lower-amplitude high frequencies. With it, you accept slightly reduced frequency resolution in exchange for cleaner band estimates.
Sampling and the Nyquist limit. The transform can only represent frequencies up to half the sampling rate (the Nyquist frequency). A 256 Hz sampling rate resolves up to 128 Hz, which comfortably covers the clinical bands through gamma. Energy above Nyquist that was not filtered out before digitization aliases down into the analyzed range as a false low-frequency signal. This is why the anti-aliasing low-pass filter at acquisition (Chapter 5) is non-negotiable: it protects the spectrum the FFT will later compute.
The FFT answers "how much power at each frequency, averaged over this whole epoch." It does not tell you when within the epoch the power occurred. For stationary resting EEG, where the spectrum is assumed roughly stable across the recording, that limitation is acceptable. For transient or non-stationary events, it is not.
Wavelet analysis trades some frequency precision for time precision. Instead of correlating the signal against infinite sinusoids, it correlates against short, localized wavelets (scaled and shifted copies of a mother wavelet, the Morlet being common in EEG). The result is a time-frequency representation: power as a joint function of frequency and time, so you can see a burst of gamma appear at 340 ms and vanish by 480 ms.
The trade-off. You cannot have arbitrary precision in both time and frequency at once. This is a hard limit (the time-frequency uncertainty principle, analogous in form to the Heisenberg relation). A wavelet tuned for sharp time localization smears frequency, and one tuned for sharp frequency localization smears time. Wavelets handle this gracefully by using shorter windows at high frequencies (good time resolution where events are fast) and longer windows at low frequencies (good frequency resolution where rhythms are slow).
When to prefer wavelet over FFT. Use wavelet methods when timing matters: event-related spectral changes, transient events embedded in ongoing activity, phase-amplitude coupling where you track how a fast rhythm's amplitude rides on a slow rhythm's phase, and any analysis where the spectrum is expected to change within the window. Use FFT for standard resting-state spectral profiling, where the assumption of stationarity holds well enough and the goal is a stable average spectrum for database comparison. Most clinical QEEG reports are FFT-based. Wavelet methods appear in research and in the more advanced connectivity and cross-frequency measures.
Absolute power is the raw quantity: how many microvolts squared of activity sit in a given band at a given electrode. Frontal theta absolute power, occipital alpha absolute power, and so on.
Units and transformation. Absolute power is reported in μV² (or μV²/Hz). Its raw distribution across a population is right-skewed and lognormal: most people cluster at moderate values, with a long tail of high-power individuals. Z-scoring assumes a normal distribution, so databases log-transform absolute power before computing means and standard deviations. The exam may ask why log transformation is applied to absolute power. The answer is that it pulls the skewed lognormal distribution toward Gaussian so the z-score is meaningful.
What it is sensitive to. Absolute power reflects real cortical activity, but it is also sensitive to non-cortical factors that scale the whole signal: skull thickness (thicker skull attenuates more, lowering absolute power), electrode impedance, and the global amplitude trait of low-voltage EEG (a heritable variant that reduces amplitude everywhere). A client with a naturally low-voltage map shows low absolute power across all bands and sites. That is constitution, not pathology. Absolute power has good test-retest reliability (intraclass correlations 0.7 to 0.9 for standard bands), higher at posterior sites where the alpha generators are stable, which is why it anchors most phenotype detection.
Relative power expresses each band as a proportion of total power: theta relative power is theta divided by the sum of all bands at that site. It is a percentage, bounded between zero and one.
The normalization advantage. Because relative power divides out the total, it cancels the global scaling factors that contaminate absolute power. Skull thickness, impedance, and the low-voltage trait shift every band's absolute power together, so they wash out of the ratio. A low-voltage individual whose absolute power is low everywhere can have entirely normal relative power: the shape of the spectrum is preserved even when its overall amplitude is reduced. For comparing the distribution of activity across bands, relative power is the more stable metric.
The pitfall. Relative power introduces denominator instability, and this is the testable catch. Because every band's relative value depends on the total, a change in one band changes all the others mechanically, even if those others did not move. A burst of muscle-contaminated high beta inflates the denominator, which deflates the relative power of every other band, making alpha and theta look artifactually reduced when nothing happened to them. The classic trap: you see "reduced relative alpha" and infer an alpha deficit, when in fact the relative alpha fell only because contaminated beta inflated the total. Always check whether a relative-power finding is driven by a real change in that band or by a change somewhere else in the spectrum. Relative power has lower test-retest reliability than absolute power (0.6 to 0.85) for exactly this reason.
The practical rule the exam rewards: absolute and relative power answer different questions, and a complete read uses both. Absolute power asks "how much activity is here?" Relative power asks "what fraction of this brain's output is in this band?" A finding that shows up in both, in the same direction, is far stronger than one that appears in only one.
A power ratio divides one band's power by another's at the same site. The theta/beta ratio (TBR) is the famous one. Theta/alpha and other ratios appear in specific contexts.
Why ratios are seductive. A ratio compresses two metrics into one number that tracks a clinically intuitive contrast: TBR rises when slow activity dominates fast activity, which maps loosely onto an underarousal-versus-engagement axis. The number is compact, and it correlates with group-level findings.
Why ratios exaggerate. A ratio amplifies change because the numerator and denominator can move in opposite directions at once. If theta rises ten percent and beta falls ten percent, the ratio jumps more than either band did alone. That sensitivity cuts both ways: it makes ratios responsive, and it makes them volatile. A ratio inherits the state-sensitivity of both its components, so TBR is heavily affected by drowsiness (which raises theta) and by anything that shifts beta. Test-retest reliability of TBR is moderate at best (0.5 to 0.75) and collapses when drowsiness is not screened, because the single most common confound in clinical QEEG, low vigilance, inflates exactly the numerator.
The TBR cautionary tale. The theta/beta ratio was treated for years as a diagnostic marker for ADHD. The group-level difference is real: ADHD samples show elevated theta on average. The error was inferring the individual diagnosis from the group finding. Recent reappraisals (Poil et al., 2024; the Strzelczyk et al., 2026 multiverse analysis across hundreds of analytic specifications) found that TBR has no diagnostic value for ADHD at the individual level, and that previously reported group differences are highly contingent on analytic choices and partly reflect aperiodic spectral slope and individual alpha frequency rather than oscillatory differences. The American Academy of Neurology advises clinicians not to rely on TBR to confirm or refute an ADHD diagnosis. TBR retains utility for treatment stratification, selecting a starting protocol from a baseline profile, but not as a diagnostic test. For the full evidence trajectory and the reverse-inference logic behind the failure, see Chapter 12 and the Field Guide treatment of the same case. The exam-level takeaway: ratios are convenient and exaggerate; treat an extreme ratio as a flag to investigate, never as a conclusion.
Power metrics describe activity at single sites. Coherence describes the relationship between two sites: how consistently their activity is coupled at a given frequency. It is the workhorse connectivity measure in clinical QEEG, and the exam expects you to distinguish its variants, because they handle the central artifact problem (volume conduction) differently.
The core idea. Coherence at a frequency quantifies how stable the phase and amplitude relationship is between two channels across the recording. High coherence means the two sites rise and fall together consistently at that frequency; low coherence means their relationship is variable. Coherence is bounded between zero and one. Functionally, it is read as a proxy for communication or coupling between the underlying regions, with the heavy caveat that scalp coherence is an imperfect proxy for actual cortical connectivity.
Magnitude squared coherence. The standard form. It measures the consistency of the relationship between two signals at a frequency, regardless of the size of the lag between them. Computed from the cross-spectrum normalized by the two auto-spectra, it is essentially a frequency-resolved correlation. Its weakness is that it is contaminated by volume conduction: a single source spreading to two nearby electrodes through the conductive tissue of the head produces high apparent coherence between those electrodes even though there is no genuine coupling between two regions. The two electrodes are seeing one source, not two communicating sources.
Phase coherence and the phase lag index. To address volume conduction, phase-based measures focus on the timing relationship and discard the part of the signal that volume conduction produces. Volume conduction is effectively instantaneous: a source spreads to multiple electrodes with zero time lag (zero phase difference). The phase lag index (PLI) measures the consistency of the sign of the phase difference between two signals and is constructed to be insensitive to zero-lag coupling. Because genuine inter-regional communication takes time and therefore produces a nonzero phase lag, while volume conduction produces zero lag, PLI suppresses the volume-conduction artifact and keeps the true coupling. The cost is that PLI also discards any genuine zero-lag synchrony, which does occur biologically. No measure escapes the trade-off entirely. Phase coherence in general is insensitive to amplitude, keying only on timing.
Lagged coherence. A related solution. Lagged coherence (in the Pascual-Marqui formulation associated with the LORETA family) explicitly removes the zero-lag (instantaneous) component of the relationship before computing coherence, leaving only the lagged component that reflects time-delayed coupling. Like PLI, it reduces volume-conduction artifact by excluding zero-lag contributions. The exam may contrast magnitude-squared coherence (sensitive to volume conduction) with lagged coherence and PLI (designed to reduce it). The distinction is the zero-lag handling.
Reliability and reading. Coherence has moderate test-retest reliability (0.5 to 0.7), lower than spectral power, and short-range coherence between adjacent electrodes is more reliable than long-range coherence. Isolated coherence findings on a single recording deserve caution. Coherence interpretation, over-coherence versus under-coherence and what each predicts, is the subject of Chapter 12. The methodology point here is knowing what each coherence variant measures and how it treats volume conduction.
Phase is the timing of an oscillation: where in its cycle the wave sits at a given instant. Phase-based metrics extend connectivity analysis beyond coherence and underpin several modern measures.
Phase lag index, restated as a phase metric. PLI (Section 10.8) is at root a phase measure: it asks whether site A consistently leads or lags site B, ignoring the magnitude of the lag and ignoring zero-lag relationships. Its value rises when the phase-difference sign is consistent across the recording.
Inter-site phase clustering. Where PLI summarizes the relationship between two sites, inter-site phase clustering (also discussed in the literature as phase-locking value and related measures) quantifies how tightly the phase relationship between two sites clusters around a consistent value across epochs or trials. A tight cluster (phase differences that recur at nearly the same value every cycle) indicates strong phase coupling. A diffuse cluster (phase differences scattered around the circle) indicates weak coupling. The metric is computed with circular statistics, because phase is a circular quantity (it wraps from 2π back to 0), and ordinary linear means and standard deviations are undefined on a circle. This is a recurring methodological point: any phase metric requires circular statistics, not Euclidean ones.
Phase-amplitude coupling. The best-studied cross-frequency interaction. The phase of a slow rhythm (typically theta) modulates the amplitude of a fast rhythm (typically gamma), so gamma bursts preferentially at a particular phase of the theta cycle. This is read as a mechanism for coordinating distributed processing: the slow rhythm opens timed windows of excitability during which fast local activity fires. Phase-amplitude coupling is computed with time-frequency methods (wavelets, Section 10.4) because it requires tracking the slow phase and the fast amplitude jointly over time. It is a research and advanced-analysis measure, not yet a standard line in clinical reports, but the IQCB blueprint expects you to recognize the concept and its theta-gamma canonical form.
Asymmetry metrics compare homologous sites across the two hemispheres: F3 against F4, T3 against T4, O1 against O2. They quantify lateralization, the degree to which one side carries more activity than the other in a band.
Amplitude asymmetry. The standard formulation is a log ratio of power at the two homologous sites: the natural log of right power minus the natural log of left power, or equivalently the log of the ratio. Taking the log makes the measure symmetric around zero (a given proportional excess on the left produces the same magnitude as the same proportional excess on the right) and improves its distribution for z-scoring. A value of zero means the two sites are balanced. A positive or negative value indicates which side dominates, by convention depending on how the ratio is written.
Frontal alpha asymmetry, the canonical case. The most studied asymmetry is frontal alpha asymmetry (FAA): the alpha balance between F3 and F4. Because alpha is an inverse marker of cortical activation (more alpha means a region is idling), more alpha on the left implies less left-frontal activation. In the Davidson approach-withdrawal model, reduced left-frontal activation is associated with withdrawal motivation and depressive presentations, and reduced right-frontal activation with approach. The clinical correlates and the model belong to Chapter 12 and the Field Guide phenotype atlas. The methodology cautions are specific and testable.
Reference and artifact cautions. Asymmetry is acutely reference-sensitive. A linked-ears reference can introduce an imbalance if the two ear electrodes have unequal impedance or pick up unequal temporal activity, and that imbalance maps directly onto the asymmetry score as a false lateralization. Average reference or current source density (Laplacian) is preferred for asymmetry work. Frontal sites are also vulnerable to eye-movement artifact, which spreads from the prefrontal electrodes (Fp1, Fp2) into F3 and F4 and biases the alpha estimate. Clean the eye artifact before computing frontal asymmetry. Finally, frontal alpha asymmetry has poor single-session reliability (intraclass correlations around 0.3 to 0.6). It reflects a trait disposition modulated by session-specific state, so a single recording is insufficient to characterize trait-level asymmetry. Multiple recordings, ideally three or more, are needed for confident interpretation.
Scalp EEG measures voltage at the surface. The activity that produced it arose in the cortex below, but the mapping from scalp to source is not one-to-one: a given scalp pattern could be produced by many different arrangements of underlying sources. This is the inverse problem, and it has no unique solution from the data alone. Source analysis methods solve it by adding mathematical constraints (assumptions about what kinds of source distributions are plausible) to pick the most likely source configuration consistent with the scalp data and a model of the head.
LORETA. Low-Resolution Electromagnetic Tomography (Pascual-Marqui, 1994) estimates the three-dimensional distribution of current density throughout the brain volume by assuming that neighboring neurons are active together, so the solution should be maximally smooth (spatially coherent). It analyzes all electrodes simultaneously against a standardized head model and returns current density at a grid of voxels. The "low resolution" is honest: with 19 electrodes, spatial resolution is on the order of one to several centimeters, and the smoothness assumption blurs point sources into distributions. LORETA can estimate source depth, distinguishing surface cortex from deeper midline structures like the cingulate and insula better than raw scalp topography can, but it cannot resolve fine structure and it will localize artifact as confidently as it localizes brain activity. Clean data first, always.
sLORETA. Standardized LORETA improves the localization by standardizing the current density estimate using its own variance, which yields zero localization error for a single point source under ideal (noise-free) conditions, a property LORETA itself does not have. sLORETA produces standardized units rather than raw current density. It is widely used and is the source-space basis for some normative databases. Its resolution is still limited by the underlying physics and electrode count. Standardization improves where the peak sits, not how sharply it is resolved.
eLORETA. Exact LORETA is a further refinement that achieves exact (zero-error) localization for point sources even in the presence of measurement noise and structured source configurations, while preserving the smooth distributed character of the solution. It is the most current member of the family and is increasingly the default where the LORETA approach is used. The progression to remember: LORETA (smooth distributed solution, nonzero localization error), sLORETA (standardized, zero error for a point source in the noise-free case), eLORETA (exact zero-error localization with noise). All three remain low-resolution in the practical sense; the refinements concern localization accuracy of the peak, not the centimeter-scale blur inherent to 19-channel source estimation.
Beamforming. A different strategy. Where the LORETA family estimates current density across the whole brain at once, a beamformer (the linearly constrained minimum variance, or LCMV, beamformer is standard) builds a spatial filter tuned to one target location at a time and suppresses contributions from all other locations. This trades LORETA's whole-brain coverage for sharper specificity at the target and good temporal resolution, making it well suited to tracking fast dynamics in a defined region. Beamforming has been the mainstay of MEG source analysis and works well with high-density EEG (64 channels and up), where the spatial filter has enough sensors to do its job. With a standard 19-channel clinical montage, beamforming lacks the sensor density it needs, and LORETA variants remain the practical choice. The exam-relevant contrast: distributed inverse solution (LORETA family) versus adaptive spatial filter (beamforming).
Clinical use and the standing caveat. Source localization adds apparent precision and adds assumptions: a standardized head model that may not match the individual's anatomy, the smoothness or filter constraints, and total dependence on clean input. It is useful for estimating whether a finding sits anteriorly or posteriorly, superficially or deep, and for moving from "F3 shows excess theta" toward "the generator looks anterior cingulate." It is not a substitute for the structural precision of MRI, and source findings on a single recording carry the same reliability cautions as the surface metrics they derive from, and more, because the inverse solution adds model-dependent variance. Report source findings with the caveat that they are model-based estimates, not measured locations.
Independent Component Analysis (ICA) is a blind source separation method. It takes the multichannel recording (a mixture of overlapping signals at each electrode) and decomposes it into a set of maximally independent components, each with its own scalp topography and time course. The mathematical premise is that the recorded channels are linear mixtures of underlying sources that are statistically independent of one another, and ICA recovers those sources.
Artifact separation. The dominant clinical use. Eye blinks, lateral eye movements, cardiac signal, and isolated muscle artifacts often separate into distinct independent components with recognizable signatures: a blink component has a frontal topography and a characteristic deflection; an ECG component carries the heartbeat rhythm; an EMG component sits at edge electrodes with broadband high-frequency content. You can identify the artifact components and remove them, then reconstruct the EEG from the remaining clean components, preserving cortical activity that overlapped the artifact in time. This is more surgical than simply deleting contaminated epochs, because it removes the artifact while keeping the underlying brain activity in that segment. Automated component classifiers (the kind that label components as brain, eye, muscle, heart, line noise, or other) assist, but a human verifies the classification; ICA can misassign, and removing a genuine brain component degrades the data.
Source decomposition. Beyond artifact removal, ICA is used to separate overlapping cortical processes, which matters for event-related and connectivity analyses where one scalp channel reflects several distinct generators. In that role it complements rather than replaces the source-localization methods of Section 10.11: ICA separates statistically independent signals; LORETA estimates where in the brain a signal arises. They answer different questions and are often used together.
Limits. ICA assumes the sources are independent and the mixing is linear and stationary across the recording; reality bends all three. It needs enough data and enough channels to estimate the components reliably (more channels separate sources better). And component labeling, brain versus artifact, retains a human-judgment element that the exam expects you to recognize: ICA reduces the consistency gap between analysts and reduces manual labor, but it does not eliminate the need for artifact literacy.
Run the chain forward one more time, because the exam tests it as a sequence and the clinic lives it as one. You re-reference to match the database, review and clean artifact by hand and with ICA, select at least two minutes of artifact-free epochs at a length that gives you the frequency resolution you need, transform to the frequency domain with a windowed FFT (or wavelets when timing matters), and compute the metrics: absolute power for "how much," relative power for "what fraction," ratios as flags rather than conclusions, coherence and phase for coupling (with attention to how each variant treats volume conduction), asymmetry for lateralization (with attention to reference), and source estimates for depth and locus (with attention to the inverse problem's assumptions). Then you hand the numbers to the normative database, which is the next chapter.
Every metric in this chapter buys resolution and charges assumptions. The competent practitioner, and the candidate who passes Domain V, is the one who can state, for any number on the report, what it measures, what transform produced it, and the one condition under which it lies. Carry that habit into the database comparison ahead: a z-score is only as trustworthy as the metric and the recording beneath it.