Variability is a property of features, not of EEG
The paper synthesizes 87 empirical studies that explicitly quantified or modeled EEG variability, spanning resting state, event-related potentials (ERPs), and task-related or brain-computer-interface paradigms such as motor imagery and SSVEP, in healthy and clinical cohorts. Its organizing conclusion is that variability is pervasive, structured, and consequential, but that it attaches to specific features and specific protocols rather than to EEG as a whole.
That framing matters because it replaces a single question — is EEG reliable? — with a narrower and more answerable one: which measurement, under which protocol, is stable enough for the intended inference? The review answers by grouping features into those that support individual-level inference and those that do not, and by tracing a substantial share of reported neural variability back to electrode placement, referencing, pipeline choices, and evaluation design.
Between-person differences usually exceed within-person drift
Across paradigms, inter-subject differences are typically larger than within-subject fluctuations. The review defines inter-subject variability as between-person differences in EEG features under comparable experimental contexts, and intra-subject variability as within-person fluctuations across trials, sessions, or tasks. Both, the authors stress, affect inference and model generalization — the larger between-person term does not make the smaller within-person term harmless.
The two terms also behave differently in practice. In the motor-imagery evidence summarized in the review, one study with 25 subjects over 5 days reported decoding accuracy falling from 68.8% to 53.7% before adaptation restored it to 78.9%, illustrating cross-session drift within participants. A separate comparison of an inter-subject and an intra-subject experiment concluded that cross-subject and cross-session patterns are not the same thing, so a pipeline validated one way is not automatically validated the other.
Large variability can coexist with high reliability
The review separates variability from reliability, and this distinction governs how every result should be read. Variability describes the magnitude and structure of dispersion — within a person, between people, across contexts. Reliability concerns repeatability: whether observed differences reproduce under repeated assessment. A feature can show large between-person spread yet remain highly reliable if individuals keep a consistent rank ordering across sessions; conversely, a feature can look stable at group level yet be unreliable if repeated measurements fluctuate unpredictably within individuals.
Because these are different quantities, the review treats the common metrics as answering different questions rather than as interchangeable scores. The intraclass correlation coefficient (ICC) estimates the between-subject share of total variance. The coefficient of variation expresses dispersion relative to the mean and becomes unstable when the mean approaches zero. Signal-to-noise ratio compares signal power to noise power, and high SNR does not imply high reliability when systematic shifts are present. Generalizability theory splits variance across design facets such as subject, session, and item. Test–retest correlation captures rank-order stability but ignores mean shifts and absolute agreement.
Alpha survives individual-level use; high frequencies often do not
Stability is feature-dependent, and the contrast is sharp. Alpha-band measures and individual alpha peak frequency are often relatively reliable, and the review's discussion singles out eyes-closed posterior alpha and individual alpha frequency as supporting individual-level inference. Higher-frequency power and many connectivity-derived metrics show more heterogeneous reliability. Single-edge connectivity is named among the measures that do not support individual-level inference, and the review notes that single edges are often noisy while pipeline choices can dominate connectivity outcomes.
ERP reliability varies by component rather than by paradigm. P300 measures frequently show moderate-to-good stability, and the review's evidence table records that P300 latency variability affects BCI use and that electrode choice affects P300 stability. Microstate findings follow the same split: spatial topographies can be relatively stable across sessions, whereas temporal parameters such as duration, coverage, and occurrence rate show only moderate reliability, with frequency of occurrence generally more stable than duration.
Four sources of variability, only two under investigator control
The review organizes variability into four interacting families. Biological and state-related sources are properties of the participant and must be modeled or stratified. Technical and analytical sources are properties of the protocol and pipeline and can be reduced through standardization and transparent reporting. The asymmetry is the practical point: an investigator cannot remove a participant's biology, but can change referencing, electrode placement, and analysis choices.
The evidence table gives concrete magnitudes for the controllable side. One study reported between-subject electrode placement variability of 3.94–7.17 mm across axes, though it had no test–retest design. Another decomposed variance and attributed roughly 32% to subjects, 9% to systems, and 1% to sessions — a small four-participant study whose proportions should not be generalized. The review also notes that analytic flexibility can substantially alter outcomes, particularly for connectivity, so variability estimates must be interpreted alongside pipeline choices.
Decoding accuracy hides variability as a generalization gap
For BCI and machine-learning readers, the review reframes variability as a measurable gap rather than an abstract nuisance. Validation schemes — within-session, cross-session, cross-subject, cross-dataset — operationalize variability as the accuracy drop between regimes, and improper splits can inflate accuracy. Domain-shift measures such as maximum mean discrepancy and classifier-based distances quantify the mismatch between training and test distributions across sessions or datasets.
Explicit models take a different route: instead of discarding variance, they partition it. Mixed-effects models combine fixed population effects with random subject and session effects; trait–state decomposition writes an observation as trait plus state plus error; Bayesian hierarchical models add uncertainty-aware variance partitioning. Domain adaptation aligns source and target distributions to mitigate cross-session and cross-subject drops, but may require target calibration data and carries a risk of negative transfer. The review also reports that variability can be partly predictable: one multimodal fNIRS-plus-EEG study predicted inter- and intra-subject performance with R² = 0.942, and alpha power predicted motor-imagery performance across subjects and sessions.
The evidence base is uneven where robustness claims are made
The review is explicit that its synthesis cannot support all the claims the field would like. It did not appraise methodological quality or risk of bias, so a four-participant study and a hundred-participant multi-session study carry equal weight, and cell counts in its figures are not evidence strength. Estimates were not pooled, and the reliability bands are an ordinal reading of the reviewed narratives rather than meta-analytic intervals. Screening and extraction were not independently duplicated, and the review was not prospectively registered.
The corpus itself has gaps that bear directly on generalization. Paradigm and metric remain confounded, so motor-imagery and resting-state findings rarely produce comparable numbers for the same feature; 31 of 64 paradigm × metric combinations contain no study at all; and only two of the reviewed studies evaluate cross-dataset or multi-site designs — precisely the evidence that robustness claims require. Nine of the 87 records are preprints, posted content, a dissertation, or a dataset descriptor and have not undergone peer review.
Report variance components as routinely as means
The review's practical recommendation is uniform across cognitive neuroscience, BCI, and clinical use: treat variability as a design variable rather than residual noise, report reliability and variance components as routinely as means, and evaluate generalization under the shift that deployment will actually impose. The consequences differ by field — attenuated effect sizes and failed replications in cognitive neuroscience, calibration burden and cross-session drift in BCI, constrained individual-level decisions in clinical use — but the remedy is the same.
The most decision-relevant split the paper offers is between features that can carry individual-level inference and those that cannot. Eyes-closed posterior alpha and individual alpha peak frequency sit on the usable side; high-frequency power, single-edge connectivity, and microstate temporal parameters sit on the other, and P300 measures land in a moderate-to-good middle band that depends on component and electrode choice. Closing the remaining gap, the authors argue, depends less on new metrics than on harmonized acquisition, transparent reporting of analytical degrees of freedom, and large multi-session and multi-site datasets that make realistic domain shift measurable rather than assumed.
Sources used in this answer
Inter- and Intra-Subject Variability in EEG: A Systematic Survey
Across 87 studies, inter-subject differences typically exceed within-subject fluctuations, but stability is feature-dependent: alpha-band measures and individual alpha peak frequency are often reliable, while high-frequency power, single-edge connectivity, and microstate temporal parameters are not.
