Quantitative EEG (qEEG)

The visually inspected EEG trace is a one-dimensional voltage record that the human reader parses with extraordinary pattern sensitivity but limited quantitative precision. Quantitative EEG (qEEG) is the family of methods that transform that trace into numbers: power at each frequency, ratios between bands, measures of how rhythms at distant electrodes relate to one another, the rate at which the spectrum falls off with frequency, and statistical comparisons against normative databases. Done well, qEEG extends the expert eye by detecting graded changes too subtle or too slow for visual review, by enabling group statistics, and by compressing hours of recording into compact trend displays a single clinician can monitor. Done badly, it manufactures false precision from artifact, mistakes statistical deviation for disease, and lends spurious authority to claims the raw signal never supported. The whole discipline lives on a knife edge between genuine quantitative power and seductive numerical illusion. This module builds the mathematical foundation, separates the established applications from the investigational and the frankly improper, and then dwells, deliberately and at length, on the caveats - because in qEEG the caveats are not footnotes, they are the subject.

From waveform to power spectrum

The central operation of qEEG is spectral decomposition. Any finite segment of EEG can be expressed as a sum of sinusoids of different frequencies, amplitudes, and phases; this is the Fourier representation, developed rigorously in the next module. The FFT (Fast Fourier Transform) computes that decomposition efficiently, returning a complex number at each frequency whose magnitude encodes amplitude and whose angle encodes phase. The squared magnitude, suitably normalized, forms the periodogram, which estimates the power spectral density (PSD): the distribution of signal variance across frequency, with units of microvolts squared per hertz. Intuitively the PSD answers a single question - how much of this segment's wiggle is slow and how much is fast - and assigns a number to every shade in between.

In practice the PSD is almost never computed from a single raw segment, because the periodogram of one segment is a wildly noisy estimator: its variance does not shrink as the segment lengthens, so a single transform produces a jagged spectrum that looks different every time you recompute it on adjacent data. The standard remedy is Welch's method. The epoch is divided into overlapping windows; each window is multiplied by a smoothing taper to control spectral leakage; the periodogram of each is computed; and the periodograms are averaged. Averaging trades frequency resolution for a dramatic reduction in the variance of the estimate, yielding a spectrum stable enough to compare across patients and across time. A more sophisticated alternative, the multitaper method, multiplies the same segment by a family of orthogonal tapers (the discrete prolate spheroidal sequences, or Slepian sequences) that are mathematically optimal at concentrating energy within a chosen frequency band, then averages the resulting spectra. Multitaper estimation achieves a controlled bias-variance trade-off with explicit error bars and is the preferred estimator when statistical rigor matters, particularly for short or noisy records. The crucial conceptual point is that every spectrum a clinician reads is an estimate with its own variance and bias, governed by choices - window length, overlap, number of tapers - that are rarely shown on the display but entirely determine what the numbers could and could not have resolved.

The PSD is the substrate for almost every downstream qEEG metric. Because EEG power is conventionally summarized over canonical bands - delta (roughly 1 to 4 Hz), theta (4 to 8 Hz), alpha (8 to 13 Hz), beta (13 to 30 Hz), and gamma (above 30 Hz) - the first reduction is to integrate the PSD within each band. These boundaries are pragmatic conventions, not laws of nature, and an individual peak frequency may straddle them; treating the bands as sharp biological categories is itself a subtle error. The widget below lets you see how these rhythms differ in frequency and morphology, which is exactly the structure the spectrum quantifies.

Absolute power, relative power, and the aperiodic spectrum

Absolute band power is the integral of the PSD over a band, expressed in microvolts squared. It is intuitive but problematic for comparison: absolute power depends on skull thickness, electrode impedance, the reference montage, age, scalp and electrode contact, and amplifier settings, so two healthy people can differ severalfold for purely non-neural reasons. Relative band power normalizes each band by the total power across all bands, yielding the fraction of the spectrum occupied by that band. Relative power is dimensionless and far more robust to global scaling factors, which is why most clinically oriented qEEG and most normative databases emphasize it. The cost is a statistical trap that recurs throughout qEEG: relative measures are compositional, meaning they are constrained to sum to one, so they are mathematically coupled. An apparent rise in relative theta may reflect a true increase in theta or merely a fall in alpha, because if one fraction grows another must shrink. Reporting that relative theta rose without noting what fell elsewhere is at best incomplete and at worst misleading.

A development that has reshaped how spectra are read is the explicit separation of the periodic and aperiodic components of the PSD. The EEG power spectrum is not flat; it falls off roughly as a power law, so that power decreases with increasing frequency along a sloping background, and the rhythmic peaks (the alpha bump, a beta bump) sit on top of that slope like hills on a hillside. The sloping background is the aperiodic component, often summarized by an offset and an exponent (the steepness of the falloff, related to the so-called one-over-f slope). Modern parameterization methods fit the background and the peaks jointly, which matters enormously because the aperiodic slope is not noise to be ignored - it shifts systematically with age, arousal, anesthesia, and the balance of cortical excitation and inhibition. The clinical and conceptual hazard is that a change in the aperiodic slope can masquerade as a change in band power: if the whole background tilts, the integral within a fixed band moves even though no oscillation changed. Band-power metrics that do not first remove the aperiodic background can therefore attribute to a rhythm what was actually a broadband shift. This is one of the most important refinements in contemporary spectral analysis and a frequent source of irreproducible findings in older literature that integrated bands without separating the slope.

Beyond band power, several scalar summaries compress the whole spectrum. The spectral edge frequency (commonly SEF95) is the frequency below which 95 percent of the total power lies; it tracks the overall shift of the spectrum toward slower or faster activity and is widely used in depth-of-anesthesia and intensive-care monitors. The median frequency and the peak alpha frequency are sensitive markers of diffuse slowing in encephalopathy and the dementias, and peak alpha frequency in particular tends to fall with neurodegeneration and with age. The theta-to-beta ratio has been studied extensively in attention disorders, and its history is itself a cautionary tale. An early device using this ratio was cleared by the United States regulator as an aid in the evaluation of attention deficit hyperactivity disorder, yet larger and better-controlled studies failed to reproduce its discriminative power at the level of the individual patient, and the metric is now regarded as, at best, weakly supportive rather than diagnostic. The lesson generalizes throughout this module - a number that separates groups in one cohort frequently fails to classify the next individual who walks through the door.

Connectivity: coherence, phase-locking, and what they conflate

Single-channel spectra describe local rhythm but say nothing about how regions interact. Functional connectivity metrics quantify statistical dependence between signals at two electrodes, and they have become enormously popular because the language of brain networks is appealing. That popularity has badly outrun the rigor with which the metrics are usually applied. The classical measure is coherence, the frequency-resolved analogue of correlation. At each frequency it is computed from the cross-spectrum - the product of one channel's Fourier coefficient with the complex conjugate of the other - normalized by the two autospectra and averaged over segments. Coherence ranges from zero to one and conflates two distinct phenomena: the consistency of the phase difference between the channels and the covariation of their amplitudes. A high coherence at alpha between occipital and parietal electrodes suggests those regions share a rhythm with a stable relationship across the epoch, but it does not, by itself, establish that one region drives the other or even that they are genuinely interacting at all.

Because amplitude covariation can be driven by a shared external drive rather than by true interaction, many investigators prefer phase-only measures. The phase-locking value (PLV) discards amplitude entirely: for each segment it extracts the instantaneous phase difference between the two signals - via the Hilbert transform or wavelet phase - and measures how tightly those differences cluster around a constant value across segments. A PLV near one means the two oscillations maintain a fixed phase relationship, a candidate hallmark of genuine coupling, while a PLV near zero means the phase difference is random. A separate and physiologically distinct phenomenon is cross-frequency coupling, of which phase-amplitude coupling is the best studied: the phase of a slow rhythm modulating the amplitude of a faster one, as theta phase organizing gamma bursts. Cross-frequency coupling is a vibrant research area in memory and cognition, but it is notoriously vulnerable to spurious detection from non-sinusoidal waveform shape - a sharp, non-sinusoidal slow wave generates harmonics that mimic coupling - so positive findings demand careful controls and remain investigational rather than clinical.

Distinct from these symmetric, undirected measures is the family of directed or effective connectivity metrics, which attempt to infer the direction of influence between regions. Granger causality and its spectral relatives (such as the directed transfer function and partial directed coherence) ask whether the past of one signal helps predict the future of another beyond what the second signal's own past predicts. These methods can be informative but rest on strong assumptions - linearity, stationarity, and the absence of unmeasured common drivers - and they are acutely sensitive to the same volume-conduction problem that plagues the undirected metrics, sometimes assigning a spurious direction to what is merely a propagation delay through tissue. Directed connectivity from scalp EEG should be read as a hypothesis-generating tool, not as a measurement of literal information flow between brain regions.

The lag-sensitive metrics deserve a closer look because they are the standard defense against volume conduction. Imaginary coherence keeps only the imaginary part of the complex coherency, which is necessarily zero for instantaneously mixed signals and therefore cannot be inflated by pure volume conduction. The phase-lag index similarly measures the consistency of the sign of the phase difference, ignoring zero-lag relationships altogether. These metrics genuinely suppress the artifact, but they purchase that robustness at a real cost: they are blind to any true interaction that happens to occur at or near zero lag, and they can be biased by the very volume conduction they aim to remove when sources are numerous and overlapping. There is no connectivity metric that is simultaneously sensitive to all real interactions and immune to all spurious ones; the analyst chooses which error to risk, and an honest report names that choice.

Validated uses and the catalogue of misuse

Where, then, does qEEG earn its place? Several applications are genuinely well validated. Continuous quantitative trending lets a single clinician monitor hours of intensive-care or operating-room data for seizures, ischemia, and burst suppression, and quantitative trend review is now standard practice in many neuro-intensive-care units, where it triages which raw segments deserve full visual review rather than replacing that review. The two common trend displays are conceptually different and complementary: the color density spectral array is a true spectral display, an FFT-derived spectrogram time-compressed and color-coded so a rising seizure or sinking slowing is visible at a glance, whereas amplitude-integrated EEG is an amplitude-domain rather than a spectral measure - the EEG is band-pass filtered, rectified, smoothed, and plotted on a semi-logarithmic amplitude scale, which is insensitive to frequency changes but excellent for tracking overall amplitude and the sleep-wake or burst-suppression envelope. Conflating the two, or assuming amplitude-integrated EEG resolves frequency content, is a common misconception. Processed-EEG indices derived from spectral and entropy features help guide anesthetic titration, and the suppression ratio and spectral edge are routine in the operating room. Quantitative slowing correlates with the severity of metabolic and degenerative encephalopathies and can track recovery over time within a patient. In research, spectral, aperiodic, and connectivity measures are indispensable for characterizing sleep, cognition, pharmacology, and disease states at the group level. These uses share a defining feature: the quantitative output supplements expert interpretation of an artifact-screened signal and is read as a trend or a group comparison, not as a standalone diagnosis of an individual.

The misuses are equally instructive and, unfortunately, more common in the wild. Muscle activity contaminates the beta and gamma bands and inflates the spectral edge; eye movement, sweat, and electrode drift inflate delta; a single drowsy epoch can shift every band ratio; a loosening electrode can fabricate a focal abnormality that exists only at the scalp. If artifact is not rigorously identified and removed before computation, qEEG quantifies noise with impressive decimal precision, and the precision lends the noise a credibility the eye would never have granted it. The statistical pitfalls are just as severe. Normative-database comparisons compute deviation scores - typically z-scores - across many frequencies and many electrodes simultaneously, so without correction for multiple comparisons, some region will exceed the threshold in almost any normal subject by chance alone. With dozens of electrodes and several bands, the number of tests reaches the hundreds, and a scattering of false positives is the mathematically expected result, not a finding. Topographic z-score maps speckled with red and blue regions are routinely over-read as evidence of pathology when they are exactly what pure chance predicts.

The most consequential abuse is diagnostic overreach. Commercial services have marketed qEEG to diagnose attention deficit hyperactivity disorder, depression, anxiety, traumatic brain injury, and a range of psychiatric conditions, and to prescribe individualized neurofeedback protocols, on the strength of group-level associations and normative deviations that do not transfer to the individual. The reproducibility of these individual-level claims across independent laboratories and populations has been weak. Professional society reviews and practice assessments have repeatedly concluded that qEEG is a useful adjunct in defined contexts - intensive-care monitoring, anesthesia, and research - and is not a validated standalone diagnostic for psychiatric or developmental disorders, and several have cautioned explicitly against its promotion for those purposes. A particular danger arises in the medicolegal setting, where colorful qEEG maps have been offered as objective proof of brain injury after trauma; courts and reviewers have grown appropriately skeptical, because the same maps appear in uninjured people once multiple comparisons and artifact are taken into account. The throughline is simple: a quantitative output is only as trustworthy as the artifact screening, the statistical correction, and the validation that produced it, and a number printed to three decimal places is not evidence of any of those.