Reading Your Brain: A Guide to How the EEG Works
- John S. Anderson
- 52 minutes ago
- 22 min read
John S. Anderson's post summarizes important lessons from Nunez, Nunez, and Srinivasan (2016), “Electroencephalography (EEG): Neurophysics, Experimental Methods, and Signal Processing.”
Nunez, Srinivasan, and their colleagues have written some of the clearest descriptions of EEG and its analysis available. The summary presented here covers fundamental issues that shape neurofeedback and qEEG work. It may read as technical in places, and that is expected for this material. The goal is to familiarize practitioners with the scientific methods of electroencephalography.
What Is EEG, and Why Does It Matter?
Electroencephalography (EEG) is one of neuroscience’s most powerful tools. By placing electrodes on the scalp, researchers detect the brain’s electrical activity with millisecond precision, a speed that technologies such as functional magnetic resonance imaging (fMRI) cannot match.
From diagnosing epilepsy to powering brain-computer interfaces, the EEG has shaped more than a century of neuroscience and clinical medicine. The central questions are what we are actually measuring and how we make sense of the signal.
Where the Signal Comes From
The EEG does not directly record individual neurons firing. Instead, it picks up the cumulative effect of millions of synaptic currents, the coordinated electrical exchanges between brain cells that unfold over roughly 10 milliseconds or more. A patch of cortex just 3 to 5 mm across contains about 1 million neurons and 10 billion synapses. All of them contribute to the electrical field that eventually reaches the scalp.
The EEG is as much a measure of neural synchrony as of neural activity. When large populations of neurons fire in coordinated patterns across the cortex, their signals add together and produce detectable scalp potentials.
When neurons fire randomly, their signals cancel out. A small, synchronized patch of cortex can therefore be more visible to EEG than a large, disorganized one.
The Brain’s Frequency Bands
EEG signals are typically analyzed by separating them into frequency bands, each associated with different cognitive states. The table below summarizes the conventional bands and their associations.
The alpha rhythm is perhaps the most recognizable of these. It dominates when the eyes close and the person relaxes. It is strongest over the back of the head, where the visual cortex resides.
The Challenge of Getting a Clean Signal
Recording the EEG sounds simple, since the procedure involves placing electrodes on the head and measuring the signals. In practice, it is far more complicated.
The skull is a poor electrical conductor, so it blurs and spatially smears the signal before it reaches the electrodes. Beyond that, the EEG systems pick up many unwanted electrical signals unrelated to the brain.
The Sources of Artifact
Artifacts in the EEG fall into two broad categories, biological and environmental. Biological artifacts are by far the more troublesome.
Muscle activity, known technically as electromyographic (EMG) artifact, originates from the face, jaw, neck, and even distant muscles throughout the body. Near the ears, eyes, and neck, EMG signals can be as much as 200 times stronger than the cortical signals underneath.
This large amplitude difference means that even a small amount of muscle contamination can swamp the brain signal entirely. Common triggers include eye blinks, lateral eye movements, jaw clenching, and head movements, each producing its own distinctive spatio-temporal fingerprint on the scalp.
The rhythmic pulse of arteries in the temples and neck produces slow artifacts, as does the movement of the electrically polarized eyeballs. The heart’s electrical activity also bleeds into distant scalp recordings.
Environmental artifacts include power line interference, which appears at 60 Hz in North America and 50 Hz in Europe and Asia. They also include capacitive coupling from electrical equipment and transient potential shifts caused by electrode movement.
Some researchers argue against aggressively filtering out power line noise because its presence at a given electrode is a useful diagnostic indicator. If 60-Hz noise appears, that electrode likely has poor skin contact, and its data may be unreliable.
A particularly stubborn problem is that muscle artifact and genuine brain signals occupy overlapping frequency ranges. EMG has a broadband spectrum with substantial power above 15 Hz, which means it bleeds directly into the beta band and beyond.
This overlap makes it essentially impossible to separate brain from muscle using simple frequency filters alone. The overlap also has serious implications for any study that makes inferences about beta- or gamma-band activity.
Artifact Removal: Methods, Benefits, and Controversies
The field has developed several strategies for handling artifacts, each with meaningful strengths and significant limitations. No method is universally accepted, and several remain the subject of active scientific debate. The sections below walk through the main options in the order a careful analyst would consider them.
Prevention and Precleaning
Before any algorithm is applied, the most powerful intervention is simply good recording practice. Subjects should be instructed to remain still and to minimize jaw clenching, and the electrode cap should be fitted snugly but comfortably. After recording, a manual or semi-automated precleaning pass is typically performed.
The continuous EEG is segmented into short epochs or recording periods, typically 1 to 3 seconds, and segments containing gross movement artifacts are rejected before further processing.
These gross artifacts are identifiable by their unusually high variance compared with neighboring trials.
This precleaning step is not optional. Downstream algorithms such as ICA perform poorly when the data include occasional catastrophic artifacts. Those outlier segments distort the statistical structure on which the algorithm relies.
Ocular Regression
The oldest and most straightforward approach to artifact removal targets eye movements and blinks. Additional electrodes placed around the eyes record electrooculographic (EOG) signals, the corneo-retinal standing potential, with eye movements changing the orientation of that dipole.
Using linear regression, researchers estimate how much each EOG signal leaks into each EEG channel. They then subtract that estimated contribution from the recording.
Ocular regression is fast, mathematically transparent, and requires no subjective judgment about which components to remove, and it works well for clearly defined, repeating artifacts such as blinks.
Its main weakness is that it assumes a purely linear relationship between ocular and EEG signals, which is only an approximation. It cannot address muscle artifacts. It also requires dedicated EOG electrode placements that some experimental setups cannot accommodate.
Independent Component Analysis (ICA)
Independent component analysis (ICA) has become the dominant artifact removal tool in modern EEG research. The core idea is that the recorded EEG at each electrode is a linear mixture of multiple underlying sources, some from the brain, some from muscles, some from eye movements, and some from electrical noise.
ICA algorithms mathematically unmix these signals, decomposing the data into a set of statistically independent components. The analyst then evaluates each component to decide what it represents.
The two most widely used implementations are FastICA, which maximizes statistical non-Gaussianity, and InfoMax ICA, which maximizes information transfer through a neural network model. Both tend to yield components whose distributions contain outliers, which are extreme values. That property distinguishes structured artifacts such as eye blinks from the more Gaussian brain signals. Once the decomposition is complete, researchers inspect each component’s scalp topography, power spectrum, and time course.
Artifact components carry telltale signatures. An eye-blink component shows weights concentrated near the eyes and a low-frequency power spectrum. A muscle component shows highly focal, peripheral topography with power concentrated at high frequencies. An electrical discontinuity appears as a single isolated spike in the time course.
ICA’s main strength is breadth, since it can address eye blinks, eye movements, cardiac artifacts, and some muscle activity in a single pass. It can also identify artifacts that no simple filter could remove because they overlap in frequency with genuine brain signals.
Its weaknesses, however, are substantial and widely debated. The first and most fundamental problem is that component identification is inherently subjective.
Deciding whether a given component represents artifact or genuine brain activity requires expert judgment, and different researchers examining the same data may reach different conclusions. Eye-blink and eye-movement components are usually easy to identify. Muscle artifact components are far more ambiguous, since they can resemble certain patterns of cortical activity. The boundary between muscle and a brain-muscle mixture is often unclear.
A second concern is data rank reduction. Every component removed from the data reduces the recording's mathematical rank, permanently removing one dimension of variance from the dataset. If a researcher removes 10components, the remaining data has 10 fewer degrees of freedom. That loss can subtly yet meaningfully affect subsequent analyses, particularly those sensitive to the data's covariance structure.
A third concern, and the most contentious, is the lack of empirical evidence that ICA reliably isolates task-related muscle artifact. Shackman et al. (2009) and McMenamin et al. (2011) argued that ICA-based EMG removal may inadvertently discard genuine cognitive signals, particularly in studies of emotional processing where facial muscle activity and frontal brain activity tend to co-occur.
Whitham et al. (2007) demonstrated through pharmacological paralysis experiments that EEG frequencies above 20 Hz are substantially contaminated by EMG in typical recordings, a finding that casts doubt on much gamma-band EEG research. These findings generated significant controversy.
Olbrich et al. (2011) published a commentary contesting the conclusions about ICA’s sensitivity and specificity, asking in effect what counts as muscle and what counts as brain. McMenamin et al. (2011) responded in turn, and the debate over ICA-based EMG correction remains unresolved. Given these concerns, Nunez and colleagues recommend a conservative stance. They advise retaining EEG-artifact mixtures in the data unless there is a strong, specific reason to remove them.
When the research question does not require clean gamma-band data, the costs of aggressive component removal may well outweigh the benefits.
Automated ICA and the ADJUST Algorithm
To reduce the subjectivity of manual component labeling, researchers have developed automated tools. ADJUST is one of the most widely adopted. It uses quantitative features of each component, including the spatial distribution of scalp weights, signal variance, and kurtosis, to automatically classify components.
Kurtosis is a measure of how heavy-tailed a distribution is, and ADJUST uses it to flag eye blinks, vertical eye movements, horizontal eye movements, and generic potential discontinuities.
Automation has clear benefits. It dramatically reduces analysis time and removes the inter-rater variability that plagues manual component selection, and it imposes a consistent, reproducible decision criterion that can be reported and audited.
The limitation is that automated methods are only as good as the features they use. ADJUST was designed and validated primarily for ocular and discontinuity artifacts, so it is less reliable for muscle artifacts, which is precisely the category where reliable automated removal is most needed.
Any automated classifier can also err in both directions, labeling an artifact as brain activity or labeling genuine brain activity as an artifact.
The Frequency Band Workaround
A pragmatic alternative to aggressive artifact correction is to restrict analysis to frequency bands that are inherently more resistant to contamination. Delta, theta, alpha, and mu activity have spectral peaks well below the range where muscle artifact dominates. By focusing analyses on these lower bands, researchers can sidestep the worst of the EMG contamination problem without relying on imperfect removal algorithms.
This approach has real costs. It forecloses research questions about beta- and gamma-band dynamics, which are increasingly recognized as important for understanding cognition. Even so, it is arguably more honest than claiming to have cleanly separated brain from muscle in the high-frequency range when reliable tools to do so do not yet exist.
The Bottom Line on Artifact Removal
No current technique is perfect, and the authors are candid about this point. They note that no known artifact correction technique fully removes muscle artifact, and that no EEG recording is completely immune to it.
The practical upshot is that the best EEG studies combine several complementary strategies. These include rigorous recording practices, conservative precleaning, targeted removal of clearly identifiable artifacts such as eye blinks, and honest acknowledgment of the residual contamination that almost certainly remains.
Readers encountering EEG research should scrutinize the artifact-handling methods carefully, especially in studies that make claims about high-frequency brain dynamics. An ounce of prevention is worth a pound of cure. Coaching the client in how to prevent or minimize artifacts before recording is essential.
Analyzing the Signal: Frequency Methods in Depth
Once artifact handling is addressed, spectral analysis is almost always the first step. It measures how much power exists at each frequency and how that distribution relates to brain state, behavior, or disease.
Spectral analysis is not a single method but a family of techniques with different strengths, tradeoffs, and assumptions. Choosing the right one matters more than many researchers appreciate.
The Core Problem: EEG Is Not Stationary
All spectral methods grapple with the same core tension. Classical Fast Fourier Transform (FFT) analysis assumes the signal is stationary, meaning its statistical properties do not change over time. This is a reasonable approximation for resting-state EEG collected over short windows, but it is violated in almost every cognitive experiment, where brain dynamics shift rapidly in response to stimuli, decisions, and motor responses.
Because brain signals are highly dynamic and non-stationary, features such as power spectra can change over time, which has driven the development of methods that go beyond the basic FFT.
The Fast Fourier Transform (FFT)
The FFT remains the most widely used tool in EEG spectral analysis and is the natural starting point for any investigation. It decomposes a time series into its constituent sinusoidal frequencies, yielding both amplitude and phase information for each frequency bin. Because it uses a trigonometric basis of sinusoidal waves, it is well-suited to analyzing periodic components. A critical practical decision is the epoch length.
The frequency resolution of an FFT is the inverse of the epoch length, so a 2-second epoch yields 0.5 Hz resolution while a 1-second epoch yields only 1 Hz resolution. More epochs of shorter duration give better statistical estimates of the underlying random process, while fewer, longer epochs give sharper frequency resolution. This tradeoff is fundamental, and the right balance depends on the research question and the dynamics of interest.
Separating two alpha peaks at 9.5 Hz and 10.5 Hz, for example, requires an epoch of at least 2 seconds and careful attention to frequency resolution.
The strengths of the FFT are speed, universality, and interpretability. It is computationally efficient, well-understood mathematically, and produces results that can be compared directly across laboratories and studies, making it the common language of EEG analysis.
Its limitations follow from its assumptions, since EEG dynamics are nonlinear and non-stationary, and FFT results can therefore mislead. The FFT also suffers from spectral leakage, the tendency for power at one frequency to bleed into neighboring bins when the signal does not fit neatly into the analysis window.
Window functions such as Hanning and Hamming are applied to reduce leakage, though they introduce their own tradeoffs in frequency resolution.
Welch’s Method and the Multitaper Approach
Two important refinements on the basic FFT improve its statistical reliability. Welch’s method computes the FFT over multiple overlapping data windows and averages the resulting power spectra. This sharply reduces the variance of the spectral estimate, essentially smoothing out the noise, at the cost of some temporal resolution. It is the standard approach for computing a stable power spectral density estimate from a longer EEG segment.
Multitaper analysis goes further by using multiple orthogonal window functions, called Slepian sequences or tapers, simultaneously. Each taper emphasizes a slightly different aspect of the data, and the resulting estimates are averaged together. This provides high frequency resolution and low variance, particularly for short data segments where standard methods perform poorly. The key tuning parameter is the time-bandwidth product, which governs the tradeoff between frequency resolution and variance reduction.
A higher time-bandwidth product uses more tapers and reduces noise, but it blurs closely spaced spectral peaks. In practice, it helps to understand the data's oscillation structure and to set the spectral resolution to a value smaller than the minimum distance between known oscillators.
Multitaper analysis has become prominent in sleep EEG research, enabling detailed spectral characterization of sleep stages and microevents, such as spindles, from relatively short segments. Its main risk is that setting the time-bandwidth product too high can smear together peaks such as the 9.5 Hz and 10.5 Hz alpha sub-peaks described earlier.
Autoregressive (AR) Models
Rather than decomposing a signal into sinusoids, autoregressive (AR) models fit a parametric model to the data, treating each sample as a linear combination of previous samples plus noise. The power spectrum is then derived analytically from the model parameters.
Autoregressive models offer better frequency resolution and smoother power spectra than FFT methods, and they are particularly advantageous for short data segments. This makes them attractive for event-related analyses, where the windows of interest may be only a few hundred milliseconds long, too short for reliable FFT-based estimation.
For spectral estimates of comparable statistical performance, autoregressive analysis appears smoother and easier to interpret than windowed periodogram analysis. The method does require selecting an order, which is the number of previous samples included in the prediction. Lower orders poorly represent the signal, while higher orders increase noise, so the wrong order can either miss important features or introduce spurious peaks.
Criteria such as the Akaike Information Criterion and the Bayesian Information Criterion help guide this choice, although the selection remains partly subjective and the models assume a linear generating process that may not hold for nonlinear brain dynamics.
The Wavelet Transform
The wavelet transform addresses the stationarity problem directly by replacing the fixed sinusoidal basis of the FFT with wavelets, which are short, localized oscillatory waveforms that are scaled and shifted across the signal. This provides simultaneous information about when a frequency occurs and how much power it carries, a time-frequency representation that the standard FFT cannot provide.
The Morlet wavelet, a Gaussian-windowed sinusoid, is the most commonly used variant in EEG research. Time-varying power and coherence, calculated using the Morlet wavelet, allow researchers to track how brain rhythms evolve moment-by-moment following a stimulus.
Unlike Fourier-based methods, the wavelet transform offers multiresolution analysis, with adaptive resolution in both the time and frequency domains. It naturally uses shorter windows at high frequencies, where rapid changes matter more, and longer windows at low frequencies, where frequency precision matters more. This makes it excellent for analyzing brief, transient neural events without any stationarity assumption.
Its limitations are that the choice of wavelet affects the results, and that the scale parameter forces a tradeoff between temporal and spectral precision, the signal-processing analog of Heisenberg’s uncertainty principle.
The Hilbert Transform
The Hilbert transform takes a different conceptual approach. Rather than decomposing a signal into frequency components, it converts a real-valued signal into an analytic signal, a complex-valued representation. From that representation, two quantities can be derived at every instant: the instantaneous amplitude, which is the signal’s envelope, and the instantaneous phase, which is where in its cycle the oscillation currently sits.
This is particularly powerful for studying phase synchronization between brain regions, a measure of how consistently two oscillators maintain a fixed phase relationship independent of their amplitudes.
The Hilbert transform is useful for nonstationary signals because it expresses frequency as the rate of change of phase, allowing frequency to vary with time. Rather than assuming a fixed frequency and asking how much power exists there, it asks what frequency the oscillation is exhibiting at each moment. In practice, the Hilbert transform is almost always applied after bandpass filtering the signal into a frequency band of interest. The filtered signal is then transformed to extract the instantaneous amplitude and phase within that band.
This workflow makes the method natural for studying how oscillatory dynamics evolve around specific events, such as how the alpha phase at stimulus onset predicts subsequent perception.
The chief benefit is direct access to instantaneous amplitude and phase, which supports the study of phase synchronization and phase-amplitude coupling. Features derived from the Hilbert transform can also improve the performance of machine learning models for EEG classification.
The main limitation is that the signal must be approximately single-component, containing primarily one oscillatory mode at a time, which is why bandpass filtering beforehand is essential rather than optional.
The Hilbert-Huang Transform (HHT)
The most sophisticated entry in this family is the Hilbert-Huang transform (HHT), which combines two steps. First, it applies empirical mode decomposition, and then it applies the Hilbert transform to each resulting component.
Empirical mode decomposition is a fully data-driven method that adaptively extracts the oscillatory modes actually present in the data, called intrinsic mode functions, without assuming their frequency, shape, or stationarity. Each intrinsic mode function behaves like a single-component oscillation, which makes it a suitable input for the Hilbert transform.
The result is a fully adaptive time-frequency representation that can track frequency modulation, amplitude modulation, and transient events with high fidelity.
Because it does not force the data into predefined bands such as delta, theta, and alpha, the Hilbert-Huang transform can find the frequency bands actually present in the recording. This is a potential advantage when the brain’s oscillatory structure does not align neatly with the canonical bands defined decades ago. The tradeoffs are computational cost and reproducibility.
Empirical mode decomposition is time-consuming for long recordings at high sampling rates. It also suffers from mode mixing, in which a single intrinsic mode function contains oscillations from multiple frequency ranges, or a single physical oscillation is split across several functions, especially when the signal is noisy. The data-driven nature that makes the method powerful also makes results harder to compare across subjects and studies, since the extracted functions may not map to the same bands across recordings.
Arrufat-Pié et al. (2021) compared FFT and Hilbert-Huang marginal spectra in healthy adults and recommended the Hilbert-Huang approach when nonlinearity or non-stationarity may be present.
Choosing the Right Method
No single method dominates across all use cases. The FFT and Welch’s method are well-suited to resting-state recordings and long, stable epochs, but they assume stationarity. Multitaper analysis gives precise spectral estimates from short epochs, at the risk of blurring closely spaced peaks. Autoregressive models excel with very short windows and produce smooth spectra, though they depend on order selection and a linearity assumption.
Wavelets and the Hilbert transform are the tools of choice for time-varying, event-related dynamics, with the Hilbert transform requiring prior bandpass filtering. The Hilbert-Huang transform offers the most flexibility for nonlinear, non-stationary signals, at the cost of interpretability and computation.
In practice, most studies combine methods, using the FFT or multitaper analysis for overall spectral characterization and wavelets or the Hilbert transform for event-related dynamics. Coherence, which can be computed with any of these methods, then supports connectivity analysis.
Connectivity: From Spectrum to Coherence
Coherence measures how consistently two electrode sites oscillate in phase, allowing researchers to probe functional connectivity between brain regions. It is defined as the squared magnitude of the cross-spectrum between two channels, normalized by the products of their individual power spectra, and ranges from 0 for a random phase relationship to 1 for a perfectly consistent one.
High coherence in the alpha band across distant scalp sites, for example, suggests large-scale coordination in the underlying cortex. Interpreting coherence, however, requires substantial caution.
The electrical smearing caused by the skull can artificially inflate coherence between nearby electrodes at all frequencies, an effect that is largely independent of frequency and can be mistaken for genuine neural connectivity.
A practical rule of thumb is to require an electrode separation of at least 10 cm before treating coherence as cortical connectivity rather than volume conduction.
The surface Laplacian transformation offers an alternative, spatially high-pass filtering the EEG to reduce volume conduction effects before coherence is calculated.
Non-Stationary Designs: ERPs and Steady-State Responses
For studies involving cognitive tasks, researchers often shift to event-related potentials, the average brain response locked to a repeated stimulus. By averaging across hundreds of trials, random neural noise cancels out, leaving the consistent response behind.
Classic components such as the P1, N1, P2, and N2 peaks have been mapped to specific stages of sensory and cognitive processing. Their scalp topographies reveal where in the cortex each stage is computed.
A related paradigm, the steady-state evoked potential, uses rhythmically flickering or modulated stimuli to drive sustained oscillatory responses at known frequencies. Because the brain’s response is concentrated in a narrow spectral band set by the stimulus frequency, these responses have an inherently high signal-to-noise ratio. That property makes them particularly robust to the broadband muscle artifact that plagues other paradigms.
Steady-state responses have been used to track attention, to classify individual differences in perception, and even to distinguish video game players from non-gamers by their neural response profiles.
Integrative Summary
Electroencephalography records the summed synaptic activity of large neural populations, so its signal reflects neural synchrony as much as neural activity. A small, well-coordinated patch of cortex can dominate the scalp recording while a larger but disorganized region remains nearly invisible.
The skull complicates this picture by smearing the signal spatially, which limits spatial resolution but also carries diagnostic information about what the electrodes are actually capturing.
These physical realities shape every downstream decision an analyst makes, from electrode placement to interpretation. Understanding them is the foundation for both neurofeedback practice and quantitative EEG analysis.
The central practical obstacle in EEG analysis is separating genuine brain activity from contamination. Biological artifacts, especially muscle activity, can exceed cortical signals many times over and overlap directly with the beta and gamma frequency ranges. No single removal method solves this problem, so the field relies on a layered strategy that begins with prevention and careful recording and proceeds through precleaning, ocular regression, and component-based methods such as independent component analysis.
Each technique trades one limitation for another, and component identification in particular depends on expert judgment that different analysts may apply differently. The most defensible conclusion is that residual contamination almost always remains, which is why high-frequency results warrant extra scrutiny and why coaching clients to minimize artifacts before recording is essential.
Spectral analysis translates the cleaned signal into a description of how power is distributed across frequencies and how that distribution relates to brain state and behavior. The fast Fourier transform remains the common language of the field, but its assumption of stationarity breaks down in most cognitive tasks, where brain dynamics shift moment to moment.
A family of alternatives addresses this gap, including Welch and multitaper methods for stable estimates, autoregressive models for very short windows, and wavelet, Hilbert, and Hilbert-Huang approaches for time-varying dynamics. Each method makes different assumptions and reveals different features, so the choice of method is itself a scientific decision rather than a technical default. Most rigorous studies combine several methods, matching the tool to the question and to the temporal structure of the signal.
Beyond single-channel spectra, coherence and related measures allow analysts to probe how distant brain regions coordinate, though volume conduction through the skull can inflate apparent connectivity and must be controlled for. Event-related and steady-state designs offer complementary windows into cognition, with steady-state responses providing unusually high signal-to-noise ratios that resist muscle contamination.
Taken together, these themes point to a consistent message for the practitioner. The EEG is a powerful but interpretively demanding tool whose conclusions depend heavily on recording quality, artifact handling, and analytic choices. Treating those choices transparently, favoring prevention over correction, and matching methods to questions are what separate trustworthy EEG work from results that merely reflect the analyst's assumptions.
Key Takeaways
1. EEG measures synchrony, not just activity. Coordinated neural populations produce the strongest signals, and a small synchronized patch can outshine a large disorganized one.
2. The skull is both a limitation and a feature. It blurs spatial resolution, yet understanding its filtering effects is essential for accurate signal interpretation.
3. Artifact removal is a continuum of tradeoffs, not a solved problem. Prevention beats correction, regression handles the eyes, ICA covers a broader range while introducing subjectivity and rank reduction, and automation improves consistency more than accuracy.
4. High-frequency EEG results deserve extra scrutiny. The overlap between muscle artifact and beta and gamma brain signals is substantial and largely unresolved, so the most defensible analyses focus on lower bands unless strong controls are in place.
5. Method choice is a scientific decision, not a technical default. No single spectral method is universally best, and the method shapes what the analyst can and cannot see, so defaulting to the FFT for cognitive EEG or skipping bandpass filtering before the Hilbert transform risks conclusions that reflect analytic choices rather than the brain.
Glossary
artifact: any electrical signal in an EEG recording that does not originate from the brain processes under study, whether biological, such as muscle, eye, and heart activity, or environmental, such as power line and equipment noise.
autoregressive (AR) model: a parametric spectral method that treats each EEG sample as a linear combination of previous samples plus noise and derives the power spectrum from the fitted model.
coherence: a frequency-domain measure of how consistently two electrode sites oscillate in phase, ranging from 0 for a random relationship to 1 for a perfectly consistent one, used to estimate functional connectivity.
electromyographic (EMG) artifact: contamination of the EEG by muscle activity from the face, jaw, neck, and body, with a broadband spectrum that overlaps the beta and gamma ranges.
electrooculographic (EOG) signal: the corneo-retinal standing potential, with eye movements changing the orientation of that dipole.
empirical mode decomposition (EMD): a data-driven technique that adaptively separates a signal into intrinsic mode functions without assuming a fixed set of basis functions.
event-related potential (ERP): the averaged EEG response time-locked to a repeated stimulus, which isolates a consistent neural response by canceling random noise across trials.
Fast Fourier transform (FFT): the standard algorithm for decomposing a time series into its sinusoidal frequency components, providing amplitude and phase for each frequency bin.
Hilbert transform: a method that converts a real-valued signal into an analytic signal, yielding instantaneous amplitude and phase, typically after bandpass filtering.
Hilbert-Huang transform (HHT): a two-step method that applies empirical mode decomposition and then the Hilbert transform to each component, producing a fully adaptive time-frequency representation.
independent component analysis (ICA): a technique that unmixes the recorded EEG into statistically independent components so that artifact components can be identified and removed.
intrinsic mode function (IMF): a single-component oscillation extracted by empirical mode decomposition that serves as input to the Hilbert transform.
multitaper analysis: a spectral method that averages estimates from several orthogonal taper functions to reduce variance, especially for short segments.
power spectral density (PSD): a measure of how the power of a signal is distributed across frequency.
spectral leakage: the bleeding of power from one frequency into neighboring bins when a signal does not fit neatly into the analysis window.
stationarity: the assumption that a signal’s statistical properties do not change over time, which classical Fourier analysis requires but cognitive EEG usually
violates.
steady-state evoked potential (SSEP): a sustained oscillatory response driven by a rhythmically modulated stimulus, concentrated in a narrow band and resistant to broadband muscle artifact.
surface Laplacian: a spatial high-pass transformation that reduces volume conduction effects before connectivity measures such as coherence are calculated.
volume conduction: the spread of electrical activity through the skull and scalp, which can inflate coherence between nearby electrodes and mimic genuine connectivity.
wavelet transform: a time-frequency method that uses short, scaled, and shifted oscillatory waveforms to track how power at different frequencies changes over time.
Welch’s method: an FFT refinement that averages spectra from multiple overlapping windows to produce a stable power spectral density estimate.
References
Arrufat-Pié, E., Estévez-Báez, M., Estévez-Carreras, J. M., Machado-Curbelo, C., Leisman, G., & Beltrán, C. (2021). Comparison between traditional fast Fourier transform and marginal spectra using the Hilbert–Huang transform method for the broadband spectral analysis of the electroencephalogram in healthy humans. Engineering Reports, 3(8), e12367. https://doi.org/10.1002/eng2.12367
McMenamin, B. W., Shackman, A. J., Greischar, L. L., & Davidson, R. J. (2011). Electromyogenic artifacts and electroencephalographic inferences revisited. NeuroImage, 54(1), 4–9. https://doi.org/10.1016/j.neuroimage.2010.07.057
Nunez, M. D., Nunez, P. L., & Srinivasan, R. (2016). Electroencephalography (EEG): Neurophysics, experimental methods, and signal processing. In H. Ombao, M. Lindquist, W. Thompson, & J. Aston (Eds.), Handbook of neuroimaging data analysis (pp. 175–197). Chapman & Hall/CRC.
Olbrich, S., Jödicke, J., Sander, C., Himmerich, H., & Hegerl, U. (2011). ICA-based muscle artefact correction of EEG data: What is muscle and what is brain? Comment on McMenamin et al. NeuroImage, 54(1), 1–3. https://doi.org/10.1016/j.neuroimage.2010.04.256
Shackman, A. J., McMenamin, B. W., Slagter, H. A., Maxwell, J. S., Greischar, L. L., & Davidson, R. J. (2009). Electromyogenic artifacts and electroencephalographic inferences. Brain Topography, 22(1), 7–12. https://doi.org/10.1007/s10548-009-0079-4
Whitham, E. M., Pope, K. J., Fitzgibbon, S. P., Lewis, T., Clark, C. R., Loveless, S., Broberg, M., Wallace, A., DeLosAngeles, D., Lillie, P., Hardy, A., Fronsko, R., Pulbrook, A., & Willoughby, J. O. (2007). Scalp electrical recording during paralysis: Quantitative evidence that EEG frequencies above 20 Hz are contaminated by EMG. Clinical Neurophysiology, 118(8), 1877–1888. https://doi.org/10.1016/j.clinph.2007.04.027
About the Author
John S. Anderson, MA, LADC, BCB, BCN, QEEGD, is a veteran neurofeedback practitioner and educator with over five decades of experience in biofeedback and neurofeedback, beginning his work in 1974. He holds a master's degree in psychology and is certified by the Biofeedback Certification International Alliance (BCIA) and the International QEEG Certification Board. As the founder of the Minnesota Neuro-Training Institute, Anderson provides clinical services, mentorship, and professional training in neurotherapy. His clientele includes individuals with ADHD, learning disorders, chronic pain, and addiction. He is also a recognized instructor, offering BCIA-approved courses and QEEG certification programs, and contributes to educational initiatives such as Biosource Software's "Seminars Without Borders." Anderson integrates holistic healing practices with contemporary neurophysiological research to develop effective neurofeedback protocols.
Support Our Friends
