Artifact Correction and Rejection

Artifact Correction and Rejection are preprocessing responses applied to evidence that has been identified or reasonably suspected to contain artifact-related contamination. Artifact rejection excludes, masks, or withholds affected evidence from a declared use, effectively removing it from the analysis. In contrast, artifact correction attempts to estimate, attenuate, separate, replace, or otherwise reduce the artifact contribution while retaining some portion of the scientifically relevant target evidence. Artifact detection, artifact source attribution, artifact correction, artifact rejection, missing-data handling, and signal denoising are related but distinct activities, each serving a specific role in managing contaminated evidence without conflating their purposes or outcomes.

Meaning of Artifact Correction and Rejection

Artifact rejection is the decision to exclude contaminated samples, frames, events, intervals, channels, components, trials, streams, or other evidential supports from a specified analysis or representation. Rejection does not repair or restore the rejected evidence and often reduces temporal coverage, event coverage, participant coverage, or multichannel completeness.

Artifact correction is a transformation that attempts to reduce the contribution of an artifact while preserving scientifically relevant target information. Correction methods include subtraction, regression, projection, decomposition, component removal, reference-based cancellation, local replacement, spatial reconstruction, or model-based estimation. The corrected signal is an estimate based on assumptions and should not be described automatically as the uncontaminated original.

Scientific Goals and Decision Criteria

The scientific goal of artifact handling is to prevent artifact-related structure from creating false events, false condition differences, biased amplitudes, altered timing, misleading morphology, spurious spatial patterns, incorrect source attribution, or false cross-signal relationships, while preserving as much valid target evidence as the scientific use requires. The objective is evidential validity rather than merely visual cleanliness.

Decisions between correction and rejection represent trade-offs among factors such as artifact severity, artifact support extent, target overlap, separability, available reference information, expected target information loss, amount of evidence retained, uncertainty in artifact attribution, and the requirements of the intended scientific claim. There is no universal rule requiring correction or rejection for every artifact type.

Preservation priority dictates that evidence should be rejected or modified only to the extent justified by the quality limitation and scientific purpose. A mild artifact outside of features of interest can be less consequential than a small artifact overlapping a critical onset, peak, transition, or cross-source relationship. Artifact amplitude alone is therefore insufficient for choosing a response.

Uncertainty-aware decisions acknowledge that when artifact presence, source, extent, or separability is uncertain, the handling decision can preserve graded confidence, masks, alternative analyses, or unresolved status rather than forcing definitive correction or rejection. Uncertain artifact attribution should not be converted into certain signal modification without justification.

Rejection, Masking, and Exclusion

Artifact handling can exclude individual observations, intervals, events, trials, channels, spatial regions, components, streams, or participant-specific supports depending on where contamination is localized. Rejecting a larger support than necessary discards valid information, while rejecting too little retains contaminated evidence.

Masking differs from deletion in that a mask preserves the existence, location, and status of excluded evidence while preventing selected use; deletion removes both values and their explicit quality status if provenance is not retained. Prefer representations that preserve why evidence was excluded and over which support whenever that information affects interpretation.

Threshold-based rejection uses decision rules based on amplitude, slope, variance, morphology, spectral content, sensor state, missingness, timing, or composite quality indicators. Such thresholds are not natural boundaries between valid and invalid behavior and should be justified against the consequences of false retention and false rejection.

Rejection bias arises because artifact occurrence can depend on movement, task difficulty, participant state, condition, group, environment, device fit, or behavior. Therefore, rejected evidence need not be a random subset of the original observations. Differential rejection can change condition balance, participant comparability, event frequencies, state representation, or temporal coverage.

In multichannel evidence, rejecting one channel alters spatial coverage, reference relationships, source estimation, redundancy, cross-channel comparisons, or the validity of derived multichannel representations. A rejected channel should not be treated as though it contributed valid evidence to later joint calculations.

Artifact Correction Strategies

Reference-based subtraction and regression use a reference measurement carrying information about an artifact source. A model estimates the artifact contribution associated with that reference and subtracts it from the contaminated signal. Correction quality depends on the validity of the reference relationship; if the reference also carries target information or shares target-correlated structure, regression can remove valid signal.

Projection and spatial-filter approaches use multichannel structure to define directions, subspaces, or spatial patterns associated with artifact-related activity and attenuate those contributions. These methods depend on spatial separability and can distort target information when target and artifact occupy overlapping subspaces.

Component-decomposition approaches such as principal-component analysis (PCA), independent-component analysis (ICA), canonical-correlation analysis (CCA), or related blind-source methods express observed mixtures in components with different statistical or structural properties. Component decomposition does not guarantee that each component corresponds to one physical or physiological source.

Here, x(t) is the observed multichannel signal vector, s(t) are latent source or component signals, and A is a mixing matrix. This is a representative instantaneous linear mixing model underlying some decomposition methods, but it is not a universal model of artifact generation. Correction conceptually reconstructs evidence after excluding or attenuating components judged artifact-related; however, model mismatch or components mixing target and artifact can cause loss of target information.

Component identification is a separate inferential problem after decomposition. Evidence supporting an artifact label includes component time course, spatial pattern, spectral properties, correlation with reference signals, event relation, morphology, and contextual evidence. Statistical independence, variance explained, or component order alone should not be treated as proof that a component is artifactual.

Local replacement and spatial reconstruction estimate contaminated channels, frames, pixels, or localized values from neighboring, redundant, geometric, or model-based evidence. Such replacement can preserve useful structure but creates derived evidence whose uncertainty depends on spatial density, redundancy, geometry, gap extent, and model assumptions.

Template, model-based, and learned correction estimate artifact structure from repeated patterns, mechanistic models, statistical models, or learned mappings and then remove or replace it. A good fit to an artifact template does not guarantee preservation of target information overlapping the modeled artifact.

Separability, Target Leakage, and Overcorrection

Separability refers to the degree to which artifact and target contributions can be distinguished using available temporal, spectral, spatial, statistical, morphological, reference, or contextual structure. Correction becomes intrinsically uncertain when artifact and target share the same observable structure or when retained evidence lacks sufficient independent information to distinguish them.

Target leakage occurs when scientifically relevant target information enters the component, predictor, subspace, template, or reference used to represent the artifact. Removing that representation can suppress genuine behavior or physiology alongside the artifact.

Undercorrection and residual artifact happen when a correction method leaves part of the contaminating contribution due to imperfect reference information, nonstationarity, model mismatch, nonlinear coupling, changing artifact morphology, source overlap, or conservative correction. Residual contamination should remain distinguishable from corrected target evidence.

Overcorrection is removal or deformation of genuine target information beyond what is justified by the artifact contribution. It can attenuate events, distort amplitudes or morphology, alter spectra, remove behavioral variability, change phase relations, reduce condition differences, or create artificially homogeneous signals.

Correction-induced artifacts may include discontinuities, edge effects, ringing, spatial smoothing, altered covariance, synthetic temporal structure, or other patterns absent from the original retained evidence. Correction should therefore be evaluated for newly introduced distortions as well as residual contamination.

Multichannel, Multistream, and Temporal Consequences

Artifact correction can alter multichannel structure by removing components or projecting data, changing covariance, spatial patterns, channel relationships, reference behavior, rank, and information available for source-sensitive or multichannel analyses. A correction that improves individual channel appearance can still change the joint evidence structure.

Rejection can change temporal and cross-source support. Excluding intervals independently across streams can reduce valid overlap, break event correspondence, alter synchrony estimates, or leave different participants and modalities with unequal observation opportunities. Joint analyses should use the support on which all required evidence remains valid.

Temporal smearing and boundary effects occur when correction uses filters, windows, neighboring observations, or decomposition over finite intervals. Artifact influence and correction influence can extend beyond visibly contaminated intervals, so retained samples immediately around a rejected or corrected event are not automatically unaffected.

Propagation of quality status means corrected, rejected, masked, interpolated, and directly observed evidence should remain distinguishable when they later contribute to event measures, descriptors, averages, multichannel summaries, or cross-stream relations. Derived analyses should not silently erase which observations were modified or excluded.

Evaluating Correction and Rejection Effectiveness

Evaluation involves two simultaneous goals: reduce artifact-related bias or contamination and preserve the target signal properties needed for the scientific claim. A method that maximally suppresses an artifact proxy can still be scientifically poor if it also removes target information.

Reference-based and controlled evaluation use known artifact events, auxiliary reference signals, higher-quality observations, controlled contaminations, simulated mixtures, repeated measurements, or experimentally manipulated artifact conditions when appropriate. Every reference has limitations, and synthetic contamination can fail to reproduce real source overlap, nonlinear coupling, nonstationarity, or participant-specific behavior.

Target-preservation evaluation inspects waveform morphology, event timing, amplitude, spectral structure, spatial pattern, covariance, cross-source relations, known responses, or other properties expected to remain stable under successful correction. Evaluation should focus on properties actually used by the scientific analysis rather than rely solely on artifact attenuation.

Rejection-effect evaluation considers retained-data fraction, valid-event count, participant-specific coverage, condition balance, temporal coverage, source availability, and differential rejection across relevant groups or contexts. A low artifact rate after rejection can be achieved by discarding large amounts of evidence and is not sufficient proof of a superior strategy.

Sensitivity analysis recognizes that plausible changes in thresholds, component labels, rejected support, model settings, reference choice, decomposition method, or correction strength can materially change the retained signal and scientific result. When conclusions depend strongly on reasonable handling choices, that dependence should be treated as analytical fragility rather than hidden.

Provenance and Scientific Interpretation

Artifact-handling provenance includes information required to reproduce and interpret correction or rejection decisions. This comprises the input signal version, artifact label and confidence, affected support, rejection masks, thresholds, reference channels, decomposition or model method, component selections, correction parameters, spatial or temporal reconstruction settings, operation order, software or implementation version, output quality checks, rejected-data fraction, retained coverage, and uncertainty.

Artifact Correction and Rejection matter in Behavioral Signal Processing because artifact handling can alter behavioral event counts, condition contrasts, physiological morphology, timing, spectral content, spatial structure, participant comparability, cross-signal relations, descriptors, and inference. A defensible handling strategy does not merely produce cleaner signals: it makes explicit what evidence was removed or modified, why that action was justified, which target properties were preserved, what uncertainty remains, and how much valid evidence was sacrificed.