ICA for EEG Artifact Removal: A Comprehensive Guide for Biomedical Researchers

Abstract

Independent Component Analysis (ICA) has become a cornerstone technique for cleaning electroencephalography (EEG) data of confounding artifacts, which is a critical preprocessing step in both neuroscience research and clinical drug development. This article provides a comprehensive resource for researchers and scientists, covering the foundational principles of ICA, detailed methodological pipelines for its application on common artifacts like ocular, cardiac, and muscle activity, and advanced strategies for optimizing performance in challenging scenarios like free-viewing and mobile experiments. It further synthesizes empirical evidence on the validation of ICA's efficacy, compares it with alternative artifact removal methods, and discusses its critical impact on the reliability of downstream analyses, such as EEG microstates and event-related potentials, thereby ensuring data integrity for robust scientific and clinical conclusions.

Understanding ICA and EEG Artifacts: Core Principles for Effective Cleaning

Demystifying the Blind Source Separation Technique

Independent Component Analysis (ICA) is a advanced signal processing technique designed to solve the "blind source separation" problem. Its core purpose is to recover a set of independent, non-Gaussian source signals from their observed linear mixtures, without prior knowledge of the mixing process or the sources themselves [1] [2]. This capability makes it a powerful tool across various fields, including audio processing, financial analysis, and particularly in biomedical signal analysis where it has revolutionized electroencephalography (EEG) artifact removal [3].

The classic analogy used to explain ICA is the "Cocktail Party Problem," where multiple microphones in a room, each picking up a different mixture of voices, are used to isolate the individual speech of each talker [1] [2]. Similarly, in EEG applications, ICA treats the signals recorded from each electrode as a linear mixture of underlying independent sources, which include both brain activity and various artifacts like eye blinks, muscle movement, and heart activity [3]. By identifying and separating these sources, ICA enables researchers to isolate and remove contaminants while preserving the neural signals of interest.

Mathematical Foundation and Core Principles

Theoretical Framework

The mathematical model of ICA assumes that the observed data matrix X (representing multichannel EEG recordings) is generated by a linear mixing of statistically independent source components S (neural and artifactual sources) via an unknown mixing matrix A [1] [2]. This relationship is expressed as:

X = A S

The goal of ICA is to find an "unmixing" matrix W that approximates the inverse of A, thereby recovering the independent sources:

S = W X

where S contains the estimated independent components, and W is the unmixing matrix that separates the sources [1] [3]. The columns of the inverse matrix W⁻¹ represent the scalp topographies of the components, providing crucial information about their physiological origins [3].

Key Assumptions and Requirements

For ICA to successfully separate sources, several critical assumptions must be met:

• Statistical Independence: The source signals must be statistically independent of each other, meaning that knowing the value of one source provides no information about the value of another [2].
• Non-Gaussian Distribution: The source signals must have non-Gaussian distributions since Gaussian distributions are symmetric and maximize entropy, making separation impossible [2].
• Linear Mixing: The sources are assumed to mix linearly at the sensors, and propagation delays from sources to electrodes are considered negligible [3].

These assumptions are generally reasonable for EEG data, making ICA particularly well-suited for neurophysiological applications [3].

ICA in EEG Artifact Removal: Mechanisms and Workflows

How ICA Separates EEG Artifacts

In EEG analysis, ICA operates as a spatial filter that decomposes multichannel scalp recordings into temporally independent and spatially fixed components [3]. Each resulting component consists of a time course of activation and an associated scalp map that shows its projection strength at each electrode [3]. Artifacts such as eye blinks, eye movements, muscle activity, and cardiac signals typically generate components with distinctive characteristics:

• Eye Blinks: Produce components with large, punctate activations that project strongly to frontal sites [3].
• Eye Movements: Generate components with low-frequency time courses that also project mainly to frontal electrodes [3].
• Muscle Activity: Creates components with high-frequency spectral content (above 20 Hz) that project to temporal sites [3].
• Cardiac Artifacts: Show periodic activations synchronized with the heartbeat [4].

Once identified, artifactual components can be "zeroed out" by setting their contributions to zero and reconstructing the EEG signals from only the neural components, effectively removing the artifacts without discarding valuable data epochs [1].

Advantages Over Traditional Methods

ICA offers significant advantages compared to conventional artifact removal techniques:

• No Data Loss: Unlike epoch rejection methods that discard contaminated data segments, ICA preserves all collected information [1] [3].
• No Reference Channels Required: Unlike regression techniques, ICA does not require separate reference channels (like EOG for eye artifacts), which themselves often contain brain signals that would be inadvertently removed [1] [3].
• Versatility: ICA can remove various artifact types without specialized reference signals, including muscle noise, electrode artifacts, and line noise that lack clear reference channels [3].

Table 1: Quantitative Performance of ICA in EEG Artifact Removal

Artifact Type	Removal Efficacy	Key Identifying Features	Clinical Validation
Eye Movements & Blinks	High efficacy in isolating and removing [3]	Frontal projection, low-frequency or punctate activations [3]	Evident clearing of signals with minimal spike distortion [4]
Muscle Artifacts	Effective separation from neural activity [3]	Temporal projection, spectral peak >20 Hz [3]	Successful removal while preserving interictal activity [4]
ECG (Heart) Artifacts	Can be effectively identified and removed [4]	Periodic activations synchronized with heartbeat	Correlation analysis shows minimal signal distortion [4]
Line Noise	Effective reduction of 60Hz and aliased frequencies [3]	Narrow frequency band components	Improved signal-to-noise ratio in corrected EEG [3]

Experimental Protocols and Implementation

Data Collection Requirements

Successful ICA application depends on appropriate experimental design and data collection practices. Recent research has quantified the relationship between data quantity and decomposition quality, providing evidence-based guidance for researchers [5].

• Data Quantity: Studies using the AMICA algorithm, considered a benchmark in the field, demonstrate that decomposition quality (measured by mutual information reduction and component near-dipolarity) generally increases with more data [5]. While common heuristic thresholds exist, benefits may continue to increase with additional data collection beyond these thresholds [5].
• Experimental Considerations: For locomotion studies where motion artifacts are prevalent, researchers should design tasks with careful consideration of cognitive and locomotive demands to balance ecological validity with data quality [6]. Low-intensity locomotion (e.g., slow walking) typically produces more analyzable data than high-intensity movements (e.g., running) due to reduced motion artifacts [6].

Step-by-Step ICA Implementation Protocol

The following workflow outlines a standardized protocol for implementing ICA in EEG artifact removal:

ICA Implementation Workflow for EEG Artifact Removal

Step 1: Data Preparation and Preprocessing

• Begin with raw multichannel EEG data that has been properly collected with appropriate sampling rates (typically 200-500 Hz for standard systems) [6].
• Apply bandpass filtering (e.g., 1-100 Hz) and re-referencing as needed for your specific research questions.
• Ensure data meets ICA assumptions: sufficient data length, more time points than channels, and continuous recording where possible.

Step 2: ICA Decomposition

• Select an appropriate ICA algorithm based on your data characteristics and computational resources. Common choices include:
  • JADE (Joint Approximate Diagonalization of Eigenmatrices): Used in clinical validation studies [4].
  • FastICA: Efficient and commonly implemented in toolboxes like EEGLAB [2].
  • AMICA (Adaptive Mixture ICA): Considered a benchmark algorithm that may provide higher quality decompositions [5].
• Apply the selected algorithm to the preprocessed data to obtain the unmixing matrix W and independent components S.

Step 3: Component Identification and Classification

• Analyze the resulting components using both their time courses and scalp topographies.
• Apply established heuristics for identifying common artifacts [3]:
  • Eye blinks: Frontal projection, large punctate activations
  • Eye movements: Frontal projection, low-frequency time course
  • Muscle artifacts: Temporal projection, high-frequency content
  • Cardiac artifacts: Periodic waveform synchronized with EKG
• Document the criteria used for classifying each component as neural or artifactual.

Step 4: Signal Reconstruction and Validation

• Reconstruct artifact-corrected EEG signals by projecting only the neural components back to sensor space: clean_data = W⁻¹(:,neural) * activations(neural,:) [3].
• Quantitatively validate the results by comparing original and corrected data using metrics like correlation analysis to ensure minimal distortion of neural signals [4].
• Visually inspect the corrected data to confirm artifact removal and preservation of neural activity.

Python Implementation Example

For researchers implementing ICA in Python, here is a basic framework using the FastICA algorithm from scikit-learn:

Advanced Applications and Recent Developments

Covariate-Integrated ICA

Recent advances in ICA methodology have introduced approaches that integrate behavioral or clinical covariates directly into the decomposition process. A 2025 study demonstrated that incorporating cognitive performance metrics from the Woodcock-Johnson Cognitive Abilities Test during ICA decomposition strengthened and stabilized the correlations between EEG connectivity measures and cognitive performance [7]. This augmented ICA approach provides a more powerful multivariate framework for uncovering brain-behavior relationships than conventional ICA followed by post-hoc correlation analysis [7].

Comparative Analysis with Other Techniques

Table 2: ICA vs. Alternative Artifact Removal Methods for EEG

Method	Mechanism	Advantages	Limitations	Best Use Cases
Independent Component Analysis (ICA)	Blind source separation based on statistical independence [2]	No reference channels needed, preserves neural data, handles multiple artifact types [1] [3]	Requires sufficient data, computationally intensive, component identification subjective [2]	Research studies with high-channel counts, multiple artifact types [3]
Regression Methods	Uses reference signals to estimate and subtract artifact contributions	Simpler implementation, computationally efficient	Requires clean reference channels, removes correlated brain signals [3]	Studies with clear reference signals available
Principal Component Analysis (PCA)	Separates components based on variance [2]	Computationally efficient, objective component ordering	Mixes neural and artifactual sources, not designed for biological signals [2]	Initial data exploration, dimensionality reduction [2]
Epoch Rejection	Discards contaminated data segments	Simple implementation, guarantees artifact removal	Significant data loss, reduces statistical power [3]	Studies with rare artifacts and abundant data

Essential Research Reagents and Tools

Table 3: Research Reagent Solutions for ICA-EEG Studies

Tool Category	Specific Examples	Function in ICA-EEG Research
ICA Algorithms	JADE, FastICA, AMICA, ExtendedICA [3] [4]	Perform the core decomposition of mixed signals into independent components
EEG Processing Toolboxes	EEGLAB, MNE-Python [1]	Provide integrated environments for implementing ICA and visualizing components
Quality Metrics	Mutual Information Reduction (MIR), Component Near-Dipolarity [5]	Quantify decomposition quality and guide data collection requirements
Visualization Tools	Topographic maps, component activations, power spectra [3]	Facilitate identification and classification of artifactual components
Validation Tools	Correlation analysis, spectral comparison, expert rating [4]	Quantify artifact removal efficacy and neural signal preservation

Independent Component Analysis represents a powerful approach for blind source separation that has proven particularly valuable in EEG artifact removal. By leveraging the statistical properties of underlying sources, ICA can effectively separate and remove contaminants while preserving neural signals of interest. The technique continues to evolve with advancements like covariate-integrated ICA offering new possibilities for exploring brain-behavior relationships [7]. When implemented with careful attention to data requirements and component identification protocols, ICA provides researchers with a robust tool for enhancing EEG data quality and enabling more accurate neuroscientific and clinical investigations.

Electroencephalography (EEG) is a fundamental tool in clinical neurology, neuroscience research, and increasingly in non-clinical domains such as brain-computer interfaces (BCIs), wellness tracking, and neuroergonomics [8]. However, the EEG signal is highly susceptible to contamination by various non-neural artifacts, which can obscure cerebral activity and lead to misinterpretation of data. Effective artifact management is therefore a critical prerequisite for valid EEG analysis, particularly in studies investigating neural correlates of cognition, behavior, or drug effects [9]. The challenge is especially pronounced in the context of wearable EEG systems, which utilize dry electrodes, reduced channel counts, and record in uncontrolled environments, making artifacts more frequent and pronounced [8].

This application note establishes a detailed taxonomy of the three primary biological artifacts—ocular, muscular, and cardiac—framed within the methodological context of Independent Component Analysis (ICA), a dominant approach for artifact removal. We provide a structured classification, quantitative summaries, experimentally-validated protocols, and practical toolkits to support researchers in developing robust EEG preprocessing pipelines.

A Detailed Taxonomy of Major EEG Artifacts

Artifacts in EEG signals are typically categorized based on their origin. The following sections detail the characteristics of ocular, muscle, and cardiac artifacts, which are the most common biological contaminants.

Ocular Artifacts

Ocular artifacts originate from eye movements and blinks. The primary mechanism is the movement of the corneo-retinal dipole, which creates an electrical field that propagates across the scalp [10].

• Eye Blinks: Characterized by high-amplitude, low-frequency deflections, typically in the 0.5–2 Hz range, with a broad, symmetric frontal scalp topography [11] [10].
• Saccadic Eye Movements: Generate sharp, lateralized potentials in the frontal and fronto-polar regions due to the rapid rotation of the eyeball.
• Slow Eye Movements: Produce slow, drifting potentials that can be mistaken for slow cortical oscillations.

These artifacts are particularly problematic for the analysis of event-related potentials (ERPs) and low-frequency brain signals.

Muscle Artifacts

Muscle artifacts, or electromyogenic (EMG) artifacts, are caused by the electrical activity of cranial, facial, and neck muscles [8].

• Spectral Signature: EMG artifacts have a broadband spectral profile that can overwhelm EEG activity across all frequencies, but is most prominent in the high-frequency range (>20 Hz).
• Topographical Distribution: Their topography depends on the muscle group activated (e.g., frontalis muscle tension affects frontal channels, while temporalis muscle activity affects temporal channels).
• Amplitude and Duration: They manifest as high-frequency, low-amplitude bursts or sustained tonic activity, making them challenging to separate from neural gamma-band activity.

Cardiac Artifacts

Cardiac artifacts, or electrocardiographic (ECG) artifacts, result from the electrical activity of the heart.

• Primary Manifestation: The most common signature is the QRS complex, which appears as a stereotyped, periodic spike in the EEG trace, synchronized with the heartbeat.
• Propagation Mechanisms:
  • Direct Propagation: Volume conduction from the heart, often visible in ear lobe or temporal electrodes when the subject is lying down.
  • Pulse Artifacts: Caused by the pulsation of nearby blood vessels against an electrode, leading to slow, rhythmic waves at the heart rate.

Table 1: Taxonomic Summary of Major EEG Artifacts

Artifact Category	Spectral Domain	Spatial Topography	Temporal Signature	Key Identifying Features
Ocular (EOG)	Low-frequency (0.5–2 Hz for blinks; up to 10 Hz for saccades) [11]	Bilateral, fronto-polar maxima [10]	High-amplitude, smooth deflections (blinks); sharp, lateralized potentials (saccades)	Corneo-retinal dipole; strongly correlated with EOG channel
Muscular (EMG)	Broadband, dominant in high-frequencies (>20 Hz) [8]	Focal over specific muscle groups (frontal, temporal)	High-frequency, irregular, burst-like or tonic	Non-stereotyped, spatially focal; ICs often have "spiky" spectra
Cardiac (ECG)	Pulse artifact: <1-3 Hz; QRS: broadband [12]	Widespread, but often maximal in temporal/earlobe electrodes	Stereotyped, periodic QRS complexes; slow pulse waves	Precise temporal locking to heartbeat; can be identified via ECG channel

ICA as a Core Strategy for Artifact Removal

Independent Component Analysis (ICA) is a blind source separation technique that decomposes multi-channel EEG data into maximally independent components (ICs) [10]. The underlying assumption is that the recorded EEG is a linear mixture of independent neural and non-neural sources. ICA solves the "unmixing" problem, allowing for the identification and removal of artifact-laden components before signal reconstruction [10].

The ICA Workflow for Artifact Removal

The standard ICA-based artifact removal pipeline involves several key stages, from data preparation to component rejection and signal reconstruction.

Advanced ICA Considerations and Emerging Methods

While standard ICA is powerful, several critical considerations and novel extensions have been developed to enhance its efficacy.

• Data Requirements: ICA requires a substantial amount of data for stable decomposition. It is recommended to use more trials than channels, and the data should be as clean as possible beforehand [10].
• The "Over-Cleaning" Pitfall: A critical, often overlooked issue is that ICs are rarely purely artifactual or neural. Subtracting entire components deemed "artifactual" can inadvertently remove neural signals, leading to artificially inflated effect sizes in ERP and connectivity analyses, and biasing source localization [9] [13].
• Targeted Artifact Reduction: To mitigate this, advanced methods like the RELAX pipeline perform targeted cleaning. Instead of subtracting entire components, it selectively removes artifact-dominated periods (for eye movements) or frequency bands (for muscle noise) within components, better preserving neural information [9] [13].
• Challenges in Wearable EEG: The effectiveness of ICA can be compromised in wearable EEG systems with low channel counts (often <16), as it limits the spatial resolution needed for optimal source separation [8].

Experimental Protocols for Artifact Management

Protocol 1: ICA-Based Ocular Artifact Removal with RELAX

This protocol details the steps for implementing the RELAX pipeline, an advanced ICA-based method for targeted ocular and muscle artifact reduction [9] [13].

• Data Acquisition: Record EEG using a standard high-density (e.g., 64-channel) system. Ensure synchronization with EOG and EMG channels if available.
• Preprocessing:
  • Apply a high-pass filter at 1 Hz and a low-pass filter at 80 Hz.
  • Remove bad channels via visual inspection or automated methods (e.g., high-frequency noise, flat-line signals).
  • Reject grossly contaminated data segments.
• ICA Decomposition: Perform ICA using the Infomax algorithm (runica in EEGLAB) with extended options to capture sub-Gaussian sources [10].
• Component Labeling: Use an automated classifier like ICLabel to obtain preliminary labels for each IC (e.g., "Brain," "Eye," "Muscle") [14].
• Targeted Cleaning with RELAX:
  • Install the RELAX plugin for EEGLAB from its public repository (https://github.com/NeilwBailey/RELAX).
  • Configure the cleaning parameters. For ocular artifacts, RELAX will identify eye-movement-related ICs and subtract only the blink and saccade periods from the component time course.
  • For muscle artifacts, RELAX will identify muscle-related ICs and apply a frequency-domain filter to remove the high-frequency artifact power while preserving lower-frequency neural activity within the same component.
• Data Reconstruction: Reconstruct the EEG signal from the modified ICs.
• Validation: Quantify the success of artifact reduction by comparing the correlation between EEG and EOG/EMG channels before and after processing. For ERPs, verify that expected components (e.g., N400, P300) are preserved without inflation [9].

Protocol 2: Motion Artifact Removal for Mobile EEG

This protocol is tailored for EEG recorded during locomotion (e.g., walking, running), where motion artifacts are severe [14].

• Setup: Use a mobile EEG system with active electrodes. If available, integrate inertial measurement units (IMUs) to track head motion.
• Preprocessing with Advanced Algorithms:
  • Option A: Artifact Subspace Reconstruction (ASR). Apply ASR (e.g., the clean_rawdata EEGLAB plugin) with a threshold parameter k typically set between 10-30. A lower k is more aggressive and suitable for high-motion scenarios like running [14].
  • Option B: iCanClean. If pseudo-reference or dedicated noise sensors are available, use iCanClean with a canonical correlation threshold (R²) of 0.65 and a 4-second sliding window, which has been shown to be effective for walking data [14].
• ICA Decomposition: Perform ICA on the preprocessed data.
• Component Evaluation: Assess the quality of the decomposition by calculating the dipolarity of the resulting components. A higher proportion of dipolar components suggests a successful separation of brain sources from non-physiological motion artifacts [14].
• Validation:
  • Examine the power spectrum at the gait frequency and its harmonics; effective cleaning should show a reduction in power at these frequencies.
  • For task-based studies, confirm that expected ERPs (e.g., P300 in a Flanker task) can be recovered with the correct latency and topography [14].

Table 2: Key Research Reagents and Resources for EEG Artifact Management

Resource Name	Type/Category	Primary Function in Research	Key Features & Applications
EEGLAB	Software Environment	Provides a comprehensive framework for EEG processing, including ICA.	Core platform for running ICA; hosts essential plugins like ICLabel, RELAX, and ASR [9] [10].
RELAX Pipeline	Software Plugin (for EEGLAB)	Implements targeted artifact reduction within ICA components.	Mitigates effect size inflation and source localization bias; superior to full-component rejection [9] [13].
ICLabel	Software Plugin (for EEGLAB)	Automates the classification of ICA components.	Uses a trained dataset to label components as Brain, Muscle, Eye, Heart, etc., streamlining the review process [14].
ICanClean	Software Algorithm	Reduces motion artifacts using reference noise signals.	Leverages canonical correlation analysis (CCA); effective for motion artifacts in mobile brain imaging [14].
EGI HydroCel Geodesic Sensor Net	Hardware (EEG Cap)	Standardized high-density EEG acquisition.	Ensures consistent electrode placement; critical for high-quality ICA and spatial analysis [15].
Child Mind Institute Healthy Brain Network (HBN) Biobank	Public Dataset	Provides normative EEG data for method development and validation.	Includes resting-state EEG (eyes open/closed) from a large pediatric cohort; useful for establishing baselines [15].
EEGOAR-Net	Deep Learning Model	Calibration-free removal of ocular artifacts.	Montage-independent U-Net model; does not require EOG channels or subject-specific calibration data [16].

A rigorous and nuanced understanding of EEG artifact taxonomy is indispensable for neuroscientific and clinical research. While ICA remains a cornerstone technique for managing these artifacts, researchers must be aware of its limitations, particularly the risk of neural signal removal when using simplistic component-subtraction approaches. The adoption of targeted cleaning methods like the RELAX pipeline, along with specialized tools for motion artifact correction, represents the current best practice. As EEG applications expand into real-world, mobile, and wearable domains, continued refinement of these protocols will be essential to ensure the validity and reliability of electrophysiological findings in studies of brain function, therapeutic interventions, and drug development.

Theoretical Foundation of ICA in EEG

Independent Component Analysis (ICA) has become a cornerstone technique in electroencephalography (EEG) signal processing due to its powerful ability to separate mixed signals into their underlying source components. The core principle of ICA rests on the assumption that the observed EEG signals represent a linear mixture of statistically independent source signals originating from the brain and various non-neural sources.

The mathematical model for ICA defines the observed EEG signal matrix ( X ) as a linear mixture of independent source signals ( S ) through a mixing matrix ( A ), such that: [ X = A S ] The goal of ICA is to estimate an unmixing matrix ( W ) that separates the observed signals into statistically independent components: [ S = W X ] where ( S ) represents the independent components [17]. ICA algorithms operate under the fundamental assumption that the source components are statistically independent and have non-Gaussian distributions (with the exception that at most one component can be Gaussian) [17] [18].

This mathematical framework makes ICA particularly well-suited for EEG analysis because it effectively addresses the blind source separation problem - the challenge of recovering original source signals without prior knowledge of the mixing process. The physiological basis for this approach lies in the recognition that EEG recordings capture superimposed electrical activity from multiple distinct generators, including cortical neurons, ocular movements, cardiac activity, and muscle contractions, which can reasonably be assumed to originate from statistically independent processes [4].

ICA Implementation Protocols for EEG Artifact Removal

Data Preparation and Preprocessing

Successful application of ICA for EEG artifact removal requires careful data preprocessing to meet the statistical assumptions of the algorithm:

• High-Pass Filtering: ICA is sensitive to low-frequency drifts, requiring data to be high-pass filtered prior to fitting. A cutoff frequency of 1 Hz is typically recommended to remove slow drifts while preserving neural signals of interest [19].

• Data Collection Requirements: The quality of ICA decomposition depends on adequate data quantity. For optimal results using advanced algorithms like AMICA, sufficient data must be collected, as benefits to decomposition quality may continue to increase with more data beyond common heuristic thresholds [5].

• Handling Mobile EEG: For experiments with participant movement, moderate data cleaning (5-10 iterations of sample rejection in AMICA) improves decomposition quality, particularly for datasets with significant motion artifacts [18].

• Baseline Considerations: When working with epoch data, perform high-pass filtering but avoid baseline correction, as this can negatively impact ICA performance [19].

ICA Decomposition Workflow

The following protocol outlines a standardized approach for ICA-based artifact removal in EEG research:

Protocol 1: Standardized ICA for EEG Artifact Removal

• Data Input Preparation: Format continuous or epoched EEG data with all channels included except reference electrodes.

• Algorithm Selection: Choose an appropriate ICA algorithm based on data characteristics:

  • FastICA: Default choice for general applications [19]
  • Infomax: Effective for EEG, with optional extended version for greater flexibility [19] [17]
  • Picard: Extended Infomax equivalent [19]
  • AMICA: Considered a benchmark algorithm for optimal decomposition quality [5] [18]

• Parameter Configuration:

  • Set n_components to specify dimensionality reduction
  • For float values (0-1), ICA selects components explaining cumulative variance > threshold [19]
  • Define max_iter (500 for Infomax/Picard, 1000 for FastICA) or use 'auto' setting [19]

• Model Fitting: Execute ICA decomposition on the preprocessed data.

• Component Classification: Identify artifactual components using validated methods:

  • find_bads_eog() for ocular artifacts [19]
  • find_bads_ecg() for cardiac artifacts [19]
  • find_bads_muscle() for muscular artifacts [19]

• Artifact Removal: Reconstruct signals excluding artifactual components using ICA.apply() [19].

Table 1: ICA Algorithm Selection Guide

Algorithm	Best Use Case	Key Parameters	Advantages
FastICA	Standard EEG recordings	ncomponents, maxiter	Computational efficiency [19]
Infomax/Picard	Noisy data or when extended independence required	extended=True (for Infomax)	Robustness to noise [19] [17]
AMICA	High-quality decomposition for research	Data quantity, cleaning iterations	Considered benchmark for quality [5] [18]
JADE	Short EEG samples with evident artifacts	Component count	Effective for clear artifact separation [4]

Specialized Protocol for TMS-Evoked Potentials

For TMS-EEG artifact removal, specific considerations apply:

Protocol 2: ICA for TMS-Evoked Potentials (TEPs)

• Data Characteristics: Recognize that TMS-induced artifacts typically mask early (0-30 ms) TEP components [20].

• Variability Assessment: Measure trial-to-trial artifact variability using ICA-derived components, as low variability can lead to unreliable cleaning and potential removal of brain-derived activity [20].

• Accuracy Validation: Estimate cleaning reliability by measuring artifact component variability, which predicts cleaning accuracy even without clean ground-truth data [20].

• Component Exclusion: Apply conservative exclusion criteria, particularly for components with low trial-to-trial variability, to minimize unintended removal of neural signals [20].

Advanced ICA Applications in Neuroscience Research

ICA with Integrated Covariates

A novel approach enhances ICA's utility for cognitive neuroscience by integrating behavioral measures directly into the decomposition process:

Protocol 3: Covariate-Integrated ICA for Brain-Behavior Relationships

• Data Preparation: Create a combined dataset incorporating both EEG connectivity measures and behavioral assessment scores (e.g., Woodcock-Johnson Cognitive Abilities Test) [17].

• Dual Decomposition Approach:

  • Method A: Perform conventional ICA on EEG connectivity data followed by correlation analysis with behavioral measures [17]
  • Method B: Apply augmented ICA incorporating both EEG connectivity and behavioral measures simultaneously [17]

• Validation: Compare correlation strength and robustness between extracted components and cognitive performance measures across independent test datasets [17].

• Interpretation: Identify components that show significant relationships with cognitive performance, potentially serving as biomarkers for clinical and cognitive deficits [17].

Table 2: Quantitative Metrics for ICA Quality Assessment

Metric	Calculation Method	Interpretation	Optimal Range
Mutual Information Reduction (MIR)	Measures reduction in signal mutual information after decomposition	Higher values indicate better component separation [5] [18]	Increasing values preferred, may not plateau [5]
Near Dipolarity	Measures how closely components match expected dipole projections	Higher values indicate more physiologically plausible sources [5]	Component-dependent, higher generally better
Residual Variance	Variance unexplained after component projection	Lower values indicate better model fit [18]	Study-dependent, lower generally better
Component Class Proportion	Ratio of brain:muscle:other components	Higher brain component ratio suggests better decomposition [18]	Application-dependent

Visualization of ICA Workflows

Figure 1: Comprehensive ICA workflow for EEG artifact removal

Figure 2: Mathematical principles of ICA for source separation

Research Reagent Solutions for ICA-EEG Studies

Table 3: Essential Research Materials for ICA-EEG Investigations

Resource Category	Specific Tools/Solutions	Research Function	Implementation Notes
Software Libraries	MNE-Python ICA module	Complete ICA implementation including fitting, component selection, and signal reconstruction [19]	Provides multiple algorithms (FastICA, Infomax, Picard) and artifact detection methods
EEG Systems	19+ channel systems with 10-20 placement	Data acquisition with sufficient spatial sampling for effective source separation [17]	Higher channel counts (e.g., 71 channels) enable better decomposition quality [5]
ICA Algorithms	AMICA, FastICA, Infomax, JADE	Core decomposition engines with different performance characteristics [19] [4] [18]	AMICA considered benchmark; choice depends on data quality and research goals [5] [18]
Artifact Detection	MNE's findbads* methods, SVM classification	Automated identification of artifactual components after decomposition [19] [21]	Can be combined with manual inspection for validation
Connectivity Metrics	swLORETA, lagged coherence	Functional connectivity analysis for covariate-integrated ICA approaches [17]	Lagged coherence reduces volume conduction effects [17]
Quality Metrics	Mutual Information Reduction, Near Dipolarity	Quantitative assessment of decomposition quality [5] [18]	Essential for method validation and optimization

Validation and Quality Control Protocols

Quantitative Performance Validation

Protocol 4: ICA Decomposition Quality Assessment

• Metric Calculation:

  • Compute Mutual Information Reduction (MIR) to evaluate component independence [5] [18]
  • Assess near dipolarity to identify physiologically plausible neural sources [5]
  • Calculate residual variance after component projection to assess model fit [18]

• Component Categorization:

  • Classify components as brain, muscle, or 'other' based on spatial and temporal characteristics [18]
  • Use multiple criteria including topography, frequency spectrum, and time-course properties

• Signal-to-Noise Assessment:

  • Compare SNR in conditions of interest before and after ICA cleaning [18]
  • Validate preservation of neural signals while removing artifacts

Limitations and Considerations

Despite its powerful capabilities, researchers must recognize important limitations of ICA:

• Trial-to-Trial Variability: When artifacts show minimal variability across trials (as in TMS-EEG), ICA may become unreliable and potentially remove brain-derived activity along with artifacts [20].

• Algorithm Selection: Different ICA algorithms may yield varying results, requiring careful selection based on specific research needs and data characteristics [19] [18].

• Data Requirements: Decomposition quality improves with increased data quantity, with benefits potentially continuing beyond common heuristic thresholds [5].

• Linearity Assumption: ICA assumes linear mixing of sources, which may not fully capture the complex volume conduction properties of head tissues, though it remains a effective approximation for most EEG applications [17].

Independent Component Analysis (ICA) has become a cornerstone technique in electroencephalogram (EEG) preprocessing, particularly for the critical task of artifact removal. Its efficacy, however, is not guaranteed and hinges on several foundational prerequisites. The broader thesis of ICA for EEG artifact removal research posits that the success of the method is intrinsically linked to the quality and structure of the input data. This application note details the core prerequisites—data quality, channel count, and stationarity assumptions—that researchers must satisfy to ensure reliable and interpretable results. Failure to adhere to these principles can lead to incomplete artifact separation, the unintended removal of neural signals, or fundamentally invalid decompositions [20]. We synthesize current research to provide structured quantitative data, experimental protocols, and practical tools to guide researchers in optimizing their ICA workflows.

Quantitative Prerequisites for ICA

The following tables summarize the key quantitative findings from the literature regarding data requirements and performance for ICA in EEG.

Table 1: Data Quantity and Channel Configuration Requirements

Factor	Key Finding	Quantitative Evidence	Source
Data Quantity	Benefits of increased data may extend beyond common heuristic thresholds; no clear plateau observed.	Mutual Information Reduction (MIR) and near-dipolarity continued to improve with more data beyond common benchmarks in a 71-channel study.	[5]
Channel Count	An intermediate number of channels is optimal; too few or too many degrades performance.	5–8 channels were identified as the optimal range for motor imagery BCI applications, balancing information and noise.	[22]
Artifact Removal Performance	ICA can effectively clean various artifacts with minimal distortion to neural signals.	A quantitative study reported minimal distortion of interictal activity (measured via correlation analysis) after removing EKG, eye movement, and muscle artifacts.	[4]

Table 2: ICA Performance and Error Metrics in Validation Studies

Study Context	Performance Metric	Result	Implications
General Artifact Classification	Mean Squared Error (MSE) vs. Expert Labeling	<10% MSE on reaction time data; 15% MSE on an auditory ERP paradigm.	Automated classification can perform on par with inter-expert disagreement levels.	[23]
TMS-Evoked Potentials (TEP) Cleaning	Cleaning Accuracy	ICA becomes unreliable and may remove brain signals when artifact trial-to-trial variability is small.	Highlights a critical violation of the independence assumption in specific paradigms.	[20]

Experimental Protocols for Prerequisite Validation

Protocol: Assessing Data Sufficiency and Channel Setup

This protocol is designed to empirically determine the optimal amount of data and channel configuration for a given experimental setup.

• Data Preparation: Begin with a raw, continuous EEG dataset that has been acquired with a high channel count (e.g., 64 or 128 channels) and is of substantial duration (e.g., 20+ minutes).
• Subsampling Data Quantity: Randomly subsample the data to create multiple datasets of varying lengths. The data length should be expressed as a ratio of the total number of data frames (samples) to the number of channels (F/C). Test a wide range of F/C ratios (e.g., from 100 to 10000) [5].
• Subsampling Channel Count: From the full channel set, create multiple channel subsets. These should include:
  • Very sparse sets (e.g., 3-5 channels over motor cortex).
  • Intermediate sets (e.g., 5-8 channels, selected automatically or based on literature [22]).
  • A high-density set (e.g., 32+ channels).
• ICA Decomposition: Run a chosen ICA algorithm (e.g., AMICA [5] or Infomax [10]) on all combinations of the data quantity and channel count subsets.
• Quality Assessment: Evaluate the quality of each resulting decomposition using metrics such as:
  • Mutual Information Reduction (MIR): Measures the independence of the components [5].
  • Near-Dipolarity: The proportion of components with a dipolar scalp topography, which is indicative of a compact cortical source [5].
  • Component Reliability: Use a framework like RAICAR (Ranking and Averaging Independent Component Analysis by Reproducibility) to assess the stability of components across multiple ICA runs with different initializations [24].
• Optimal Point Determination: Plot the quality metrics against both the F/C ratio and the number of channels. The point where quality metrics begin to asymptote or peak indicates the optimal, efficient data configuration for your specific paradigm.

Protocol: Evaluating Stationarity and Trial-to-Trial Variability

This protocol tests the critical ICA assumption of statistical independence between sources, which can be violated by highly stereotypical artifacts.

• Simulated Artifact Injection: Start with a clean, artifact-free TMS-Evoked Potential (TEP) dataset or similar event-related potential data. If such data is unavailable, use resting-state EEG from which major artifacts have been meticulously removed.
• Generate Artifacts: Create simulated artifacts (e.g., a TMS pulse artifact or eye blink template) with controlled levels of trial-to-trial variability.
  • Low-Variability Condition: Add the artifact to each trial with nearly identical waveform and amplitude.
  • High-Variability Condition: Add the artifact with significant random variations in amplitude, latency, or morphology across trials [20].
• ICA Processing: Apply ICA to both the low-variability and high-variability datasets.
• Accuracy Assessment: Compare the ICA-cleaned data to the original artifact-free ground truth. Calculate the accuracy of artifact removal, for example, by measuring the correlation or mean squared error between the cleaned data and the true neural signal [20].
• Measure Component Variability: For real data where the ground truth is unknown, calculate the trial-to-trial variability of the artifact-dominated independent component itself. This can be a predictor of cleaning reliability [20].
• Interpretation: If accuracy is low and the identified artifact component shows very low variability, the results should be treated with caution, as ICA may have inaccurately partitioned the neural signal.

The following diagram illustrates the logical workflow and decision points for setting up a successful ICA analysis based on the aforementioned protocols.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Software and Algorithmic Tools for ICA Research

Tool Name	Type	Primary Function in ICA Research	Key Reference / Source
EEGLAB	Software Environment	Provides a comprehensive framework for ICA decomposition, visualization, and component rejection.	[10]
AMICA Plugin	Algorithm/Plugin	A multimodal ICA algorithm considered a benchmark for decomposition quality; requires significant computation.	[5]
JADE Algorithm	Algorithm	An ICA algorithm based on Joint Approximate Diagonalization of Eigen-matrices; effective for artifact removal.	[4]
FastICA Algorithm	Algorithm	A widely-used, computationally efficient ICA algorithm based on fast fixed-point iteration.	[25]
Infomax Algorithm	Algorithm	A popular ICA algorithm that maximizes the mutual information between inputs and outputs.	[10]
RAICAR Framework	Methodology	A framework for ranking and averaging ICA components by reproducibility across multiple realizations.	[24]
CW_ICA Method	Methodology	A recently proposed method for automatically determining the optimal number of ICs.	[26]

The successful application of ICA for EEG artifact removal is a deliberate process grounded in rigorous data preparation. Researchers must treat it not as a one-size-fits-all filter but as a powerful tool with specific requirements. The evidence clearly demonstrates that success depends on: providing sufficient high-quality data per channel, configuring an optimal number of electrodes for the given task, and critically evaluating the stationarity assumptions of the underlying sources, particularly in paradigms with stereotypical artifacts like TMS. By adhering to the validated protocols and utilizing the tools outlined in this document, researchers can significantly enhance the reliability and interpretability of their ICA results, thereby strengthening the validity of their subsequent neuroscientific and clinical findings.

A Step-by-Step Guide to Implementing ICA for EEG Cleaning

Independent Component Analysis (ICA) has become a cornerstone technique in electroencephalography (EEG) research for separating neural activity from various artifacts. The quality of ICA decomposition, however, is profoundly influenced by specific preprocessing steps applied to the data beforehand. Proper preprocessing is not merely a procedural formality but a critical determinant of ICA performance, directly impacting the algorithm's ability to isolate biologically plausible neural sources and effectively remove artifacts such as ocular movements, cardiac signals, and muscle activity. This application note provides detailed protocols for the essential preprocessing steps of filtering and referencing, framed within the context of preparing EEG data for optimal ICA decomposition in artifact removal research.

Theoretical Foundations: Preprocessing for Optimal ICA

The effectiveness of ICA relies on certain statistical assumptions about the input data, primarily that the underlying sources are statistically independent and mixed linearly through volume conduction. Preprocessing aims to condition the recorded EEG signals to better satisfy these assumptions.

The Impact of Filtering on ICA

Filtering choices directly influence ICA performance by controlling the frequency content available for decomposition. High-pass filtering is particularly crucial because EEG data exhibits a 1/f power spectral density, meaning lower frequencies dominate the signal amplitude [27]. Since ICA is biased toward higher amplitude activity when working with finite data lengths, excessive low-frequency content can cause the algorithm to focus on slow drifts rather than neurologically relevant activity in higher bands. Furthermore, low-frequency signals below 1 Hz are often contaminated by physiological artifacts such as sweating, which introduces spatiotemporal non-stationarity that violates ICA assumptions [27].

The official MNE documentation explicitly warns that "ICA is sensitive to low-frequency drifts and therefore requires the data to be high-pass filtered prior to fitting. Typically, a cutoff frequency of 1 Hz is recommended" [19]. Empirical evidence supports this recommendation, with studies concluding that "high-pass filtered data at 1-2Hz works best for ICA" in terms of signal-to-noise ratio and component quality [27].

The Role of Reference Choice in ICA

The reference electrode problem is fundamental to EEG interpretation, as there is no electrically neutral point on the body or head—a concept known as the "no-Switzerland principle" [28]. The choice of reference scheme affects the spatial structure of EEG data, which in turn influences the ICA decomposition process.

Different reference techniques transform the ideal infinite reference recording (VInf) through specific transformation matrices [28]. The average reference (AVG), while popular, results in a zero-sum constraint across all channels, which affects the spatial topographies of independent components [28]. Research comparing reference techniques for ICA has found that the Reference Electrode Standardization Technique (REST), which approximates a reference at infinity, shows "overall superiority" for ICA analysis, particularly in preserving both temporal ERP characteristics and spatial topographies of components [28].

Table 1: Common EEG Reference Techniques and Their Properties

Reference Technique	Mathematical Formulation	Impact on ICA	Best Use Cases
Linked Mastoids/Ears (LM)	( V{LM} = (I - \tilde{T}{LM}) V_{Inf} ) [28]	Standard approach, may introduce bias from non-neutral sites	Clinical protocols, tradition-based pipelines
Average Reference (AVG)	( V{AVG} = (I - \tilde{T}{AVG}) V_{Inf} ) [28]	Imposes zero-sum constraint on component topographies	High-density arrays, studies with uniform scalp coverage
Reference Electrode Standardization Technique (REST)	( V{REST} = T{REST} V_{Inf} ) [28]	Approximates ideal reference, often superior for ICA	Source localization, studies aiming for optimal component separation

Experimental Protocols

Protocol 1: Data Preparation and Filtering for ICA

This protocol outlines the critical steps for preparing raw EEG data to ensure optimal ICA decomposition, with specific attention to filtering parameters.

Materials and Equipment

• Raw EEG data recorded according to experimental requirements
• Computing environment with EEGLAB or MNE-Python software
• Data storage with adequate space for processed data files

Step-by-Step Procedure

• Data Import and Integrity Check

  • Import raw EEG data into your preferred analysis environment (EEGLAB/MNE-Python).
  • Verify that the data is continuous without inherent segmentation that might disrupt stationarity.
  • Check for channels filled entirely with zeros or clearly biologically implausible values, as these will severely compromise ICA and should be removed before decomposition [27].

• Downsampling (Optional but Recommended)

  • Downsample data to approximately 250 Hz if the original sampling rate is significantly higher [27].
  • Rationale: Reduces computational load and data size while maintaining neurologically relevant frequency content. Modern downsampling algorithms automatically apply anti-aliasing filters.

• High-Pass Filtering

  • Apply a high-pass filter with a cutoff frequency of 1-2 Hz. In practice, a pass-band edge of 1 Hz (equivalent to a -6 dB cutoff at 0.5 Hz) is recommended [27].
  • Rationale: This critical step removes slow drifts that violate ICA's stationarity assumption and bias the decomposition toward high-amplitude, low-frequency artifacts [19] [27].
  • Filter type: Use a zero-phase filter (e.g., FIR) to prevent temporal distortion of event-related potentials.

• Low-Pass Filtering (Optional)

  • Apply a low-pass filter with a cutoff of 40-50 Hz to reduce high-frequency noise, including muscle artifacts and line noise interference [29].
  • For studies specifically investigating high-frequency activity (e.g., gamma oscillations), adjust this parameter accordingly or omit this step.

• Bad Channel Identification and Interpolation

  • Identify channels with excessive noise, flat signals, or consistent artifacts using automated algorithms and visual inspection.
  • Interpolate identified bad channels using spherical spline or other spatially appropriate methods.
  • Note: Interpolation should be performed before ICA to ensure a complete channel set, but the same bad channels should be noted and potentially excluded from the actual ICA decomposition to maintain data quality.

Quality Control and Verification

• Visual Inspection: Plot filtered data to verify the removal of slow drifts while maintaining physiological signal characteristics.
• Spectral Analysis: Confirm appropriate frequency content using power spectral density plots.
• Stationarity Check: Ensure the filtered data exhibits relatively constant variance over time, without systematic trends or abrupt jumps.

Protocol 2: Referencing and ICA Decomposition

This protocol details the application of different reference techniques and the subsequent execution of ICA.

Materials and Equipment

• EEG data preprocessed according to Protocol 1
• EEGLAB or MNE-Python with ICA functionality
• Head model information (if using REST reference)

Step-by-Step Procedure

• Reference Application

  • Apply your chosen reference scheme to the preprocessed and filtered data. Common options include:
    • Linked Mastoids (LM): Rereference to the mathematical average of A1 and A2 (earlobe electrodes) [30].
    • Average Reference (AVG): Rereference to the average of all recording electrodes [28].
    • REST: Calculate a virtual reference at infinity using head model information [28].
  • Note: Each method has theoretical advantages and limitations, with REST often providing superior results for ICA decomposition [28].

• Data Segmentation for ICA (Critical Step)

  • Select a stationary segment of data for ICA training. This segment should:
    • Contain representative artifacts (especially ocular movements if removing ocular artifacts).
    • Be free of large-amplitude, transient non-stationary artifacts (e.g., large muscle spikes, electrode pops) [29].
  • In practice, this can be a dedicated "artifact training period" where participants perform systematic eye blinks and movements, or a clean segment extracted from the experimental data.

• ICA Decomposition

  • Execute ICA on the segmented, referenced data using preferred algorithm (Infomax, FastICA, or Picard).
  • Set the n_components parameter based on your data: use a fixed number (e.g., 20-30 for standard EEG arrays) or a float (e.g., 0.999) to select components explaining 99.9% of variance [19].
  • For MNE users, the ICA object is fitted to the data, producing mixing_matrix_ and unmixing_matrix_ attributes [19].

• Component Application to Full Dataset

  • Apply the computed ICA weights from the training segment to the entire continuous dataset, including portions that may contain non-stationary artifacts [29].
  • This transfers the spatial filters learned from clean data to all experimental data.

Quality Control and Verification

• Component Topography: Examine component scalp maps for physiologically plausible patterns.
• Time Courses: Review component activations for correspondence with known artifacts (e.g., eye blinks, saccades, muscle bursts).
• Diagnostic Tools: Use automated artifact detection (e.g., find_bads_eog, find_bads_ecg in MNE) to supplement visual inspection [19].

Data Presentation and Analysis

Quantitative Comparison of Filtering Parameters

Table 2: Comparative Analysis of High-Pass Filter Cutoffs for ICA Decomposition

Filter Cutoff (Hz)	Effect on ICA Stability	Residual Ocular Artifacts	Use Case Recommendation
0.1 Hz	Suboptimal due to increased low-frequency drift	Poor removal	Studies specifically investigating infraslow activity
0.5-1.0 Hz	Optimal stability and component quality	Effective removal	Standard ERP and spectral analysis; recommended default
2.0 Hz	Good stability, may attenuate some neural signals	Effective removal	Studies focusing on beta/gamma bands or with strong drift

Table 3: Key Computational Tools and Functions for ICA Preprocessing

Tool/Resource	Function/Purpose	Implementation Example
Zero-Phase FIR Filter	Removes low-frequency drifts without temporal distortion	mne.filter.filter_data() or EEGLAB's pop_eegfiltnew()
ICA Algorithm (Infomax)	Core decomposition engine identifying independent sources	mne.preprocessing.ICA(method='infomax') or EEGLAB runica()
Spherical Spline Interpolation	Reconstructs bad channels using spatial neighborhood information	mne.channels.interpolate_bads_eeg() or EEGLAB eeg_interp()
REST Reference Toolbox	Converts data to approximate infinity reference	EEGLAB plugins or standalone REST implementation
Automated Artifact Detection	Identifies component correlates of biological artifacts	ICA.find_bads_eog() and ICA.find_bads_ecg() in MNE

Workflow Visualization

The following diagram illustrates the complete preprocessing pipeline for ICA, integrating both filtering and referencing steps:

Diagram 1: Complete ICA Preprocessing Workflow. Critical steps (red/green) require precise parameter selection, while yellow boxes indicate input/output stages.

Proper preprocessing is an essential prerequisite for successful ICA decomposition in EEG artifact removal. The protocols detailed in this application note provide evidence-based guidelines for the critical steps of filtering and referencing. Specifically, high-pass filtering at 1-2 Hz effectively removes slow drifts that bias ICA, while careful selection of reference technique (with REST offering theoretical advantages) optimizes the spatial structure of components. By implementing these standardized protocols, researchers in both basic neuroscience and drug development can enhance the reliability of ICA for isolating neural signals from contaminants, thereby improving the quality of electrophysiological biomarkers in clinical research.

Independent Component Analysis (ICA) has become a cornerstone technique in electroencephalography (EEG) preprocessing for isolating neural activity from various artifacts. ICA is a blind source separation method that decomposes multichannel EEG recordings into statistically independent components, each with a fixed scalp topography and a time course of activation [31]. Within neuroscience research, this capability is crucial for identifying and removing confounding signals from eye movements, muscle activity, and cardiac rhythms without discarding valuable data segments [9] [10]. The success of this decomposition, however, depends critically on the choice of algorithm and the careful execution of the protocol. This application note provides detailed methodologies for implementing two prominent ICA algorithms—Infomax and AMICA—within the context of EEG artifact removal, offering researchers a structured framework for obtaining optimal results.

Core ICA Algorithms for EEG: Infomax vs. AMICA

Algorithm Comparison and Selection

For EEG decomposition, the selection of an ICA algorithm involves trade-offs between computational efficiency, decomposition quality, and stability. The following table summarizes the key characteristics of Infomax and AMICA, the latter being widely considered a benchmark for performance [32].

Table 1: Key Characteristics of Infomax and AMICA Algorithms

Feature	Infomax ICA	AMICA
Core Principle	Maximizes information transfer (infomax) or minimizes mutual information between components [31].	Uses adaptive mixtures of generalized Gaussians to model source densities; can fit multiple models to stationary data subsets [32] [33].
Typical Performance	Robust and reliable for super-Gaussian sources; performance can be enhanced with the 'extended' option for sub-Gaussian sources like line noise [10].	Often outperforms other ICA algorithms on quantitative metrics like mutual information reduction and component dipolarity [32].
Computational Demand	Moderate; suitable for standard computing resources.	High; often requires substantial computation time and resources, potentially benefiting from compiled versions or high-performance computing [33].
Stability	Decompositions can vary slightly between runs due to random initial weight matrices [10].	Known for producing high-quality, stable decompositions.
Best Use Cases	Standard artifact removal tasks; a good general-purpose choice.	Research requiring the highest possible decomposition quality; analysis of complex source interactions.

Quantitative Data Requirements for Stable Decomposition

A critical factor for a successful ICA decomposition is having a sufficient amount of data. The required data quantity scales with the square of the number of EEG channels. A common heuristic is the $\kappa$ value, defined as:

$$\kappa = \frac{\text{number of data frames}}{(\text{number of channels})^2}$$

While a $\kappa$ value of 20 is often recommended heuristically, recent empirical investigations using AMICA suggest that decomposition quality, as measured by Mutual Information Reduction (MIR) and component dipolarity, continues to improve with more data beyond this threshold, showing no clear plateau [32]. The following table quantifies this relationship based on a study using a 71-channel EEG setup.

Table 2: Data Quantity and ICA Decomposition Quality (AMICA on 71 channels)

$\kappa$ Ratio	Impact on Mutual Information Reduction (MIR)	Impact on Near-Dipolar Components
Low $\kappa$	Lower MIR, indicating less effective separation of sources [32].	Fewer components with residual variance < 10% when fitted with an equivalent dipole [32].
Increasing $\kappa$	Asymptotic increase in MIR observed [32].	General increasing trend in the number of near-dipolar components [32].
$\kappa = 20$	Often used as a heuristic minimum, but may not represent a performance plateau [32].	Benefits of collecting additional data are shown to extend beyond this common threshold [32].

Experimental Protocols for ICA Decomposition

Preprocessing for ICA

The quality of the ICA decomposition is heavily dependent on proper data preprocessing. The procedure aims to prepare the data such that ICA can model the most meaningful sources, rather than trivial noise or artifacts [34].

• Data Loading and Channel Selection: Load the continuous or epoched dataset into the processing environment (e.g., EEGLAB). ICA is typically run on EEG channels only, as other bio-signals like EMG involve propagation delays that violate ICA's assumption of instantaneous mixing [10].
• Filtering: Apply a high-pass filter (e.g., 0.1 Hz or 1 Hz cutoff) to remove slow drifts that can adversely affect the ICA solution [34]. A low-pass filter can also be applied to reduce high-frequency noise.
• Bad Channel and Segment Rejection: Identify and remove or interpolate consistently noisy or flat channels. It is also crucial to remove sections of data with infrequent, large-amplitude artifacts (e.g., SQUID jumps in MEG, electrode pops) before ICA. These atypical artifacts can "use up" components that would otherwise model brain activity or common artifacts [34]. Tools like ft_databrowser in FieldTrip or EEGLAB's eegplot can be used for this manual rejection. The removed sections can be replaced with NaN (Not a Number) values to maintain the data's continuous structure [34].
• (Optional) Data Reduction: For high-density datasets or to improve computational efficiency, data can be downsampled. However, be aware that this can reduce the quality of separation for high-frequency artifacts like muscle activity [34].

Protocol A: Running Infomax ICA in EEGLAB

Infomax ICA is a standard algorithm available in EEGLAB and provides a robust, general-purpose decomposition [10].

• Launch the ICA Function: After preprocessing, select Tools → Decompose data by ICA → Run ICA.
• Algorithm Selection: In the pop-up window, select Infomax or runica from the algorithm dropdown menu.
• Parameter Configuration:
  • For data containing strong line noise or sub-Gaussian sources, use the 'extended' option to enable the algorithm to find both super- and sub-Gaussian sources [10].
  • To achieve a cleaner decomposition, particularly for high-density data, you can lower the stopping criterion using 'stop', 1e-7 [10].
• Execution: Click Ok to run the decomposition. The command window will display iterative output showing the learning rate and weight change until convergence is reached [10].
• Post-processing: Once finished, the ICA weights are stored in the EEG structure and can be visualized and analyzed.

Protocol B: Running AMICA

AMICA is a high-performance algorithm that can be run as a standalone program or from within MATLAB/EEGLAB via a plugin [33].

• Plugin Installation: Install the AMICA and postAmicaUtility plugins for EEGLAB from the official plugin repository [33].
• Data Preparation: Follow the standard preprocessing steps outlined in section 3.1. AMICA benefits from large amounts of high-quality data.
• Running AMICA: AMICA creates its own set of menus within EEGLAB. Navigate to the AMICA menu to configure and launch the decomposition.
• Model Configuration: A key feature of AMICA is its ability to estimate multiple ICA models for different stationary subsets of the data. The number of models and other parameters can be configured at this stage.
• Execution and Loading: The computation may take a significant amount of time. After completion, use the postAmicaUtility functions to load the computed models and components back into the EEGLAB environment for further inspection [33].

The following workflow diagram summarizes the key steps for preprocessing and running ICA.

Table 3: Key Software Tools and Resources for ICA Research

Tool/Resource	Function in ICA Research	Access/Reference
EEGLAB	A collaborative, open-source MATLAB environment providing a comprehensive framework for ICA analysis, visualization, and artifact removal [10].	https://sccn.ucsd.edu/eeglab/
AMICA Plugin	An EEGLAB plugin that implements the Adaptive Mixture ICA algorithm, often yielding superior decomposition quality [32] [33].	Available via the EEGLAB plugin manager.
RELAX Pipeline	An EEGLAB plugin for a targeted artifact reduction method that cleans artifact periods/frequencies, helping to avoid the artificial inflation of effect sizes [9].	GitHub: NeilwBailey/RELAX
FieldTrip Toolbox	An alternative open-source MATLAB toolbox for advanced EEG/MEG analysis that includes its own implementations of ICA and other preprocessing tools [34].	https://www.fieldtriptoolbox.org/
DIPFIT	An EEGLAB extension used to fit equivalent dipole models onto ICA component scalp maps, aiding in the validation of their biological plausibility [32].	Included in standard EEGLAB distributions.

The effective application of ICA for EEG artifact removal hinges on a meticulous experimental approach. Researchers must choose an algorithm aligned with their quality and resource constraints, with Infomax serving as a robust default and AMICA offering potentially superior results at a higher computational cost. Crucially, the data itself must be prepared with care, ensuring that sufficient, clean data is presented to the algorithm—a requirement quantified by the $\kappa$ parameter. By adhering to the detailed protocols and leveraging the tools outlined in this document, researchers can reliably harness ICA to isolate neural signals, thereby enhancing the validity and interpretability of their EEG findings in both basic neuroscience and applied drug development research.

The accurate separation of neural activity from non-neural artifacts is a fundamental prerequisite for electroencephalography (EEG) research. Independent Component Analysis (ICA) has emerged as a powerful blind source separation technique that decomposes multi-channel EEG data into maximally independent components (ICs). A critical challenge lies in reliably identifying which of these components represent brain activity and which reflect various artifacts. This application note provides a comprehensive framework for classifying artifactual ICs based on their distinct topographical, temporal, and spectral signatures, equipping researchers with standardized protocols for enhancing EEG data quality in basic and clinical research.

Theoretical Foundations of ICA for EEG Decomposition

ICA operates on the principle of separating mixed signals into statistically independent sources without prior knowledge of their nature or mixing process. In the context of EEG, the recorded signals from scalp electrodes represent linear mixtures of underlying neural, ocular, muscular, and technical sources. The ICA algorithm estimates a demixing matrix that separates these sources into independent components, each characterized by a fixed topography (spatial distribution across electrodes), a time course of activation, and a spectral profile.

The validity of ICA decomposition relies on key assumptions: that the source signals are statistically independent and non-Gaussian, that the mixing process is linear and instantaneous, and that the number of recorded channels equals or exceeds the number of independent sources. Following decomposition, each IC can be examined through multiple feature domains to determine its origin, enabling researchers to exclude artifactual components before reconstructing the cleaned EEG signal [35].

Signature Profiles of Major Artifact Classes

Systematic identification of artifactual components requires multidimensional assessment across spatial, temporal, and frequency domains. The tables below summarize the characteristic signatures of common artifact types encountered in EEG research.

Table 1: Topographical, Temporal, and Spectral Signatures of Physiological Artifacts

Artifact Type	Topographical Signature	Temporal Signature	Spectral Signature
Ocular (Blink)	Bifrontal focus (Fp1, Fp2); Symmetrical distribution [36]	High-amplitude, slow deflections (200-400 ms) time-locked to blink events [37]	Dominant in delta (0.5-4 Hz) and theta (4-8 Hz) bands [37]
Lateral Eye Movements	Asymmetric frontal distribution; Phase reversal between F7/F8 [36]	Sawtooth waveform with opposing polarities at F7 and F8 [36]	Broadband with low-frequency emphasis
Muscle (EMG)	Focal over temporal regions (temporalis); Diffuse for neck/shoulder muscles [37]	High-frequency, low-amplitude bursts; Irregular pattern [37]	Broadband with peak power in beta (13-30 Hz) and gamma (>30 Hz) [37]
Cardiac (ECG)	Maximum over left temporal/occipital regions; Often unilateral [36]	Stereotyped waveform repeating at heart rate (60-100 bpm) [37]	Multiple harmonics of heart rate frequency

Table 2: Signature Profiles of Technical and Motion Artifacts

Artifact Type	Topographical Signature	Temporal Signature	Spectral Signature
Electrode Pop	Highly localized to single electrode; No field distribution [36]	Sudden, high-amplitude transient with steep onset [37]	Broadband, non-stationary noise [37]
Line Noise	Global across all electrodes; May vary with electrode impedance [37]	Persistent 50/60 Hz oscillation; Constant amplitude [37]	Sharp peak at 50 Hz (Europe) or 60 Hz (North America) [37]
Head Movement	Widespread, non-stationary topography; Affects multiple electrodes [38]	High-amplitude, low-frequency drifts; Irregular bursts [38]	Dominant low-frequency content (<2 Hz)
Sweat Artifact	Widespread, shifting distribution; Often anterior emphasis [36]	Very slow drifts (<0.5 Hz); Non-stationary baseline [37]	Extreme dominance in delta band (0.1-0.5 Hz) [37]

Experimental Protocols for IC Identification

Standardized IC Classification Workflow

Protocol 1: Comprehensive IC Evaluation

Objective: Systematically classify ICs using multi-domain features.

Materials:

• ICA-decomposed EEG data
• Visualization software (EEGLAB, MNE-Python, FieldTrip)
• Computing workstation with sufficient RAM for data handling

Procedure:

• Generate Topographical Maps
  • Plot IC scalp distributions using spline interpolation
  • Note focal extremes and spatial patterns
  • Reference against canonical artifact topographies [36]

• Analyze Temporal Characteristics

  • Visualize IC time courses across entire recording
  • Identify stereotyped waveforms repeating at physiological frequencies
  • Detect irregular, high-amplitude transients

• Examine Spectral Properties

  • Compute power spectral density for each IC
  • Identify spectral peaks at characteristic frequencies
  • Note broadband versus narrowband properties

• Cross-Domain Correlation

  • Correlate spatial and temporal features
  • Confirm consistency across signature domains
  • Make final classification decision

Duration: Approximately 30-45 minutes per dataset depending on recording length and number of components.

Protocol 2: Ocular and Cardiac Artifact Specific Identification

Objective: Specifically identify and remove ocular and cardiac artifacts.

Materials:

• ICA-decomputed EEG data
• Simultaneously recorded EOG/ECG channels (if available)
• EOG/ECG template maps for correlation

Procedure:

• Template Matching
  • Correlate IC topographies with canonical ocular artifact templates [35]
  • Identify components with highest correlation coefficients (>0.7)

• Time-Course Analysis

  • Create EOG/ECG epochs synchronized to events (blinks, saccades, heartbeats)
  • Average IC activation locked to these events
  • Verify consistent temporal relationship

• Spectral Validation

  • Confirm peak at expected frequency ranges (delta for blinks, heart rate for ECG)
  • Check for harmonic patterns in cardiac components

• Selective Removal

  • Mark identified artifactual components for exclusion
  • Preserve components with ambiguous characteristics
  • Reconstruct data and verify artifact reduction [35]

Quality Control: Verify artifact reduction by comparing pre- and post- correction data using variance metrics and visual inspection.

The Researcher's Toolkit

Table 3: Essential Tools and Resources for ICA-Based Artifact Removal

Tool/Resource	Function	Implementation Notes
MNE-Python	Open-source Python package for EEG/MEG analysis	Provides comprehensive ICA implementation with multiple algorithms (FastICA, Picard, Infomax) [35]
EEGLAB	Interactive MATLAB toolbox for EEG processing	Offers extensive IC visualization tools and plugin architecture for component classification [39]
FieldTrip	MATLAB toolbox for advanced EEG/MEG analysis	Includes ICA preprocessing and component rejection pipelines [34]
ICLabel	Automated IC classifier	Provides probabilistic classification of components into brain, muscle, eye, heart, line noise, channel noise, and other categories [14]
ADJUST	Automated artifact detector	Specifically designed for identifying blink, horizontal eye movement, and generic discontinuities in event-related paradigms [40]

Special Considerations for Challenging Populations

Pediatric and Infant EEG

ICA application in infant EEG presents unique challenges due to limited recording durations and rapid developmental changes in brain activity. Recent evidence suggests that standard ICA approaches may require modification for pediatric populations [40]. The minimum data requirement for effective ICA decomposition follows the formula: $k \cdot N^2 / fs$, where $N$ is the number of channels, $fs$ is the sampling frequency, and $k$ is a multiplier (typically ≥30) [40]. For high-density infant EEG (128 channels, 500 Hz), this translates to approximately 16 minutes of clean data—often challenging to acquire with restless infants.

Comparative studies indicate that while ICA effectively corrects eye-movement artifacts in infant EEG (sensitivity = 0.89), it may distort clean signals more than alternative methods like Artifact Blocking (specificity = 0.72 vs. 0.81 for AB) [40]. Researchers working with pediatric populations should consider these trade-offs when selecting artifact removal strategies.

Mobile EEG and Motion Artifacts

Studies involving movement, such as locomotion or naturalistic behaviors, introduce high-amplitude motion artifacts that challenge conventional ICA. Recent advances in artifact removal algorithms like iCanClean and Artifact Subspace Reconstruction (ASR) show promise for mobile EEG applications [14].

iCanClean leverages reference noise signals and canonical correlation analysis to detect and correct motion artifact subspaces, particularly effective when using dual-layer electrodes [14]. When benchmarked against ASR, iCanClean demonstrated superior performance in recovering dipolar brain components (dipolarity index: 15.3% improvement over ASR) and restoring expected P300 event-related potential patterns during locomotion tasks [14].

Validation and Quality Control

Robust validation of artifact removal efficacy is essential for ensuring data integrity. Recommended quality control measures include:

• Dipolarity Assessment: Calculate the percentage of brain ICs with dipolar scalp distributions (>70% expected in clean data) [14]
• Spectral Validation: Verify reduction of artifact-specific spectral peaks (e.g., >3dB reduction at gait frequency harmonics in locomotion studies) [14]
• Topographical Stability: Ensure microstate topographies remain stable across different ICA preprocessing strategies [41]
• Statistical Power: Confirm that experimental effects maintain or increase statistical power after artifact removal [41]

Quantitative benchmarks should be established a priori and reported in methodology sections to enhance reproducibility and cross-study comparisons.

Systematic identification of artifactual components through their topographical, temporal, and spectral signatures provides a robust foundation for EEG data cleaning. The protocols and frameworks presented in this application note empower researchers to implement standardized, transparent artifact removal procedures that preserve neural signals of interest while effectively mitigating contamination. As ICA methodologies continue to evolve, particularly for challenging recording scenarios and special populations, the multi-dimensional signature approach offers a flexible yet principled framework for adapting to new developments in the field.

Independent Component Analysis (ICA) is a foundational blind source separation technique in electroencephalography (EEG) preprocessing. It operates on the principle that measured scalp signals represent linear mixtures of statistically independent underlying sources, both neural and artifactual [1]. The primary goal of applying ICA in EEG analysis is to isolate and remove pervasive artifacts—such as those from eye blinks, eye movements, muscle activity, and cardiac signals—without discarding valuable EEG data segments [10] [1]. This correction-based approach is crucial for preserving data structure and trial uniformity across subjects and conditions, which is a paramount concern in clinical research and drug development [29]. The efficacy of ICA stems from its ability to decompose multi-channel EEG data into a set of independent components (ICs), each characterized by a fixed scalp topography and a temporally independent activity time course [10] [34]. Following the identification of artifactual components, the signal is reconstructed, effectively subtracting the artifact contribution while retaining the underlying neural signals of interest.

Theoretical Foundations and Algorithm Selection

The mathematical model underlying ICA is expressed as ( X = AS ), where ( X ) is the matrix of recorded EEG data from all channels, ( S ) contains the independent source signals (components), and ( A ) is the linear mixing matrix that projects these sources to the sensors [1]. The ICA algorithm's objective is to compute an "unmixing" matrix ( W ) such that ( S = WX ), yielding maximally independent components [1]. Success in this decomposition hinges on the criteria of statistical independence and non-Gaussianity.

Several ICA algorithms are available, each with distinct strengths. The choice of algorithm can influence decomposition quality and should be considered a key parameter in the experimental protocol.

Table 1: Common ICA Algorithms for EEG Analysis

Algorithm	Key Characteristics	Applicable Context
Infomax ICA (e.g., runica)	Uses logistic distribution and natural gradient; can be extended for sub-Gaussian sources [10].	Default choice for many EEG applications; use extended option for data with strong line noise [10].
AMICA	Considered a benchmark algorithm; multimodal and often yields high-quality decompositions [5].	Resource-intensive; optimal when data quality and quantity are high [5].
FastICA	Based on fixed-point iteration to maximize non-Gaussianity [10].	Requires separate toolbox installation; computationally efficient.
SOBI	Exploses time-delayed correlations; effective for separating temporally coherent sources [10].	Suitable for artifacts with distinct temporal structure.
Jader	Uses joint approximate diagonalization [10].	Less commonly used than Infomax or AMICA.

A critical practical consideration is data stationarity. ICA assumes that the underlying sources mix instantaneously and linearly in a stationary manner [29]. Violations of this assumption, such as those caused by large-amplitude, transient artifacts, can severely compromise decomposition quality. Therefore, proper data preparation—including the selection of a stationary data segment for ICA training—is as important as the choice of algorithm itself [29].

Experimental Protocols and Workflows

Comprehensive Preprocessing and ICA Protocol

A robust, semi-automatic protocol ensures consistent removal of major artifacts while preserving neural signals. This protocol is designed for EEG recordings without electrooculography (EOG) channels and emphasizes step-by-step quality checking [29].

Table 2: Step-by-Step Preprocessing and ICA Protocol

Step	Procedure	Key Parameters & Rationale	Quality Check
1. Data Preparation	Load raw data and channel locations. Select a stationary data segment for ICA training that contains artifacts of interest (e.g., eye blinks) but excludes large, atypical noise [29] [34].	A dedicated segment where the participant performs eye blinks and movements is ideal. This ensures ICA can reliably identify ocular components.	Visually inspect data (e.g., using ft_databrowser [34]) to confirm the presence of artifacts and stationarity.
2. Bandpass Filtering	Apply a bandpass filter. A common choice is 1-2 Hz high-pass and 40-50 Hz low-pass.	A high-pass cutoff ≥1 Hz is critical for successful ICA decomposition [29]. It reduces slow drifts that violate stationarity.	Compare data before and after filtering to ensure biological artifacts (e.g., blinks) are preserved.
3. Bad Channel Interpolation	Identify and interpolate consistently noisy or dead channels.	Prevents bad channels from disproportionately influencing the decomposition and reduces the number of usable components.	Use data summary plots (e.g., ft_rejectvisual [34]) to identify channels with abnormal variance or kurtosis.
4. ICA Decomposition	Run ICA on the preprocessed, stationary segment. Use the Infomax or AMICA algorithm.	For a high number of channels (e.g., >32), ensure sufficient data quantity. A higher frames-to-channels ratio improves decomposition [10] [5].	Check command-line output for convergence. The algorithm should report a final weight change below the stopping criterion (e.g., <1e-6) [10].
5. Component Inspection & Labeling	Plot component scalp maps, activity time courses, power spectra, and, for epoched data, ERPimages [10].	Eye Artifacts: Map shows frontal focus; spectrum has low-frequency dominance; ERPimage shows blink events [10].Muscle Artifacts: Map shows peripheral focus; spectrum has high-frequency broadband activity.	Use tools like ft_topoplotIC and ft_databrowser [34] or EEGLAB's "Inspect/label components" menu [10] to manually label artifactual components.
6. Signal Reconstruction	Back-project the data excluding the labeled artifactual components. This is mathematically equivalent to subtracting the artifact's contribution from the original data [10] [34].	The reconstructed data is now cleaned of the isolated artifacts. The unit of the data remains in microvolts.	Overlay original and cleaned data to verify artifact removal and signal preservation.
7. (Optional) PCA for Large Transients	Apply Principal Component Analysis (PCA) to remove any remaining large-amplitude, transient artifacts (e.g., muscle spikes) [29].	PCA is placed after ICA to avoid distorting ocular artifacts, which are best handled by the established ICA method [29].	Inspect data for residual high-amplitude spikes.

Figure 1: Workflow for ICA-Based EEG Artifact Removal. This diagram outlines the major steps for cleaning EEG data using ICA, from data preparation to final reconstruction.

Protocol for Single-Channel EEG

For single-channel EEG systems, where traditional multi-channel ICA is not applicable, alternative data-driven methods must be employed. One advanced approach involves the Fixed Frequency Empirical Wavelet Transform (FF-EWT) combined with a Generalized Moreau Envelope Total Variation (GMETV) filter [11].

• Decomposition: The single-channel EEG signal is decomposed into six Intrinsic Mode Functions (IMFs) using FF-EWT, which adaptively isolates oscillatory modes within fixed frequency ranges associated with EOG artifacts [11].
• Identification: EOG-artifact-related IMFs are automatically identified using feature metrics such as kurtosis, dispersion entropy, and power spectral density [11].
• Filtering: The identified artifact components are processed using a cascaded, finely-tuned GMETV filter to suppress the artifact while preserving the underlying EEG [11].
• Reconstruction: The cleaned IMFs are reconstructed to produce the final artifact-free single-channel EEG signal. This method has been validated on both synthetic and real EEG data, showing substantial improvements in metrics like Correlation Coefficient and Signal-to-Artifact Ratio [11].

Quantitative Metrics and Data Requirements

The quality of an ICA decomposition is not guaranteed; it must be quantitatively assessed to ensure reliable results. Furthermore, providing the algorithm with sufficient, high-quality data is a prerequisite for success.

Metrics for Decomposition Quality and Component Rejection

Table 3: Key Metrics for Assessing ICA Quality and Components

Metric	Description	Interpretation
Mutual Information Reduction (MIR)	Measures the decrease in statistical dependence among output components. Higher MIR indicates a more successful separation of independent sources [5].	A higher value signifies a better decomposition. The metric tends to increase asymptotically with more data [5].
Near Dipolarity	Measures the extent to which a component's scalp map can be explained by a single equivalent dipole. This is a hallmark of a physiologically plausible brain source [5].	Components with high near-dipolarity are likely of neural origin. Artifacts (e.g., eye blinks, muscle) typically have non-dipolar topographies.
Power Spectrum	The frequency content of a component's time course [10].	Eye artifacts: smoothly decreasing, low-frequency dominated spectrum. Muscle artifacts: broadband, high-frequency dominated spectrum.
Kurtosis	A statistical measure of the "peakedness" or "heaviness" of the tails of a component's amplitude distribution.	Useful for automatically detecting artifactual components, which often have highly peaked distributions (e.g., eye blinks) [11].

Data Quantity Requirements

A critical and often overlooked factor in ICA is the amount of data required for a stable decomposition. The relationship between data quantity, channel count, and decomposition quality has been quantitatively explored [5].

Table 4: Data Requirements for High-Quality ICA Decomposition

Factor	Impact on ICA	Practical Recommendation
Frames-to-Channels Ratio	The number of data frames (samples) relative to the number of channels is a key determinant of stability and quality. Higher ratios lead to better decompositions [10] [5].	Heuristics suggest very large amounts of data. For a 32-channel dataset, several minutes of continuous data are recommended. Benefits of more data may continue beyond common thresholds [5].
Data Quality	ICA works best with "basically similar and mostly clean data" [10]. Large, atypical artifacts can "use up" components and lead to a suboptimal decomposition [34].	Remove infrequent, large-amplitude artifacts (e.g., SQUID jumps, electrode pops) from the data segment used for ICA training, prior to running the decomposition [34].
Channel Count	As the number of channels increases (e.g., to 64 or 128), the amount of data required for a good decomposition increases substantially [10].	For high-density arrays, if insufficient data is available, consider using PCA dimensionality reduction during ICA to find fewer components than channels [10].

A recent study using AMICA on a 71-channel dataset found that metrics like MIR and near-dipolarity showed a general increasing trend with more data, without a clear plateau. This suggests that the benefits of collecting additional EEG data may extend beyond common heuristic thresholds [5].

The Scientist's Toolkit

Table 5: Essential Research Reagents and Computational Tools

Tool/Resource	Function	Application Note
EEGLAB	An interactive MATLAB toolbox for processing EEG and other electrophysiological data. It provides a comprehensive GUI and scripting environment for ICA and other analyses [10].	The primary platform for many ICA-based workflows. Includes implementations of Infomax, Jader, SOBI, and allows integration of other algorithms like AMICA and FastICA [10].
FieldTrip	A MATLAB toolbox for advanced analysis of MEG, EEG, and other electrophysiological data. It is particularly strong in script-based, reproducible research [34].	Offers robust functions for ICA (ft_componentanalysis), artifact rejection (ft_rejectvisual), and data visualization (ft_databrowser) [34].
AMICA Plugin	A high-performance ICA algorithm plugin for EEGLAB, often considered to produce state-of-the-art decompositions [10] [5].	More computationally intensive than Infomax. Best deployed on computing clusters for large datasets [10].
RELICA Plugin	An EEGLAB plugin for assessing the reliability and stability of ICA decompositions through bootstrapping [10].	Used to evaluate which components (and their features) are stable across multiple decompositions, adding confidence to component rejection decisions.
MNE-Python	A Python package for exploring, visualizing, and analyzing human neurophysiological data.	Provides a full suite of tools for preprocessing, ICA, and machine learning, ideal for integration into modern Python-based data science pipelines [1].
Artifact Removal Transformer (ART)	A deep learning model based on transformer architecture for end-to-end EEG denoising [42].	Represents a cutting-edge, data-driven alternative that can remove multiple artifact types simultaneously. Trained on pseudo clean-noisy data pairs generated via ICA [42].

Successful component rejection and signal reconstruction hinge on a principled approach that spans experimental design, data preprocessing, decomposition, and validation. The following best practices are synthesized from current literature and protocols:

• Prioritize Data Quantity and Quality: Provide ICA with ample, clean, and stationary data. The removal of large, infrequent artifacts before decomposition is crucial for maximizing the number of components available for modeling neural sources [29] [34].
• Validate, Do Not Assume: The quality of ICA decompositions can vary. Employ quantitative metrics like MIR and near-dipolarity to assess decomposition quality and avoid interpreting unstable component features [10] [5].
• Adopt a Systematic Labeling Protocol: Component rejection should be based on multiple lines of evidence: scalp topography, time course, power spectrum, and, when available, ERPimage features [10]. Automated tools can assist, but expert manual inspection remains the gold standard for reliable artifact labeling.
• Consider Advanced and Emerging Methods: For non-standard setups like single-channel EEG, specialized techniques such as FF-EWT+GMETV offer viable solutions [11]. Furthermore, deep learning approaches like the Artifact Removal Transformer (ART) are emerging as powerful, end-to-end alternatives that may complement or augment traditional ICA methods in the future [42].

By adhering to these structured protocols and best practices, researchers can consistently execute ICA-based artifact removal, thereby enhancing the validity and reliability of their EEG findings in clinical and cognitive neuroscience research.

Simultaneous EEG-fMRI and TMS-EEG represent advanced multimodal neuroimaging approaches that combine the unique strengths of each technique to investigate human brain function with unprecedented detail. Simultaneous EEG-fMRI integrates the millisecond temporal resolution of electroencephalography (EEG) with the millimeter spatial resolution of functional magnetic resonance imaging (fMRI), enabling researchers to capture both rapid neural events and their precise anatomical origins [43]. The blood oxygenation level-dependent (BOLD) signal measured by fMRI serves as a correlate of neural activity through the mechanism of neurovascular coupling, though it is limited by a temporal delay of several seconds [43]. In contrast, TMS-EEG combines transcranial magnetic stimulation (TMS), which allows non-invasive perturbation of specific cortical areas, with EEG recording to monitor the immediate electrophysiological consequences of the stimulation [44]. This approach provides insights into cortical reactivity and effective connectivity at high spatiotemporal resolution.

The integration of all three methods—TMS-EEG-fMRI—has recently been demonstrated as technically feasible, opening exciting possibilities for non-invasive investigation of causal brain dynamics [45] [46] [47]. This trimodal approach enables researchers to perturb a brain network node with TMS while monitoring the propagation of activity throughout cortico-subcortical networks with fMRI and tracking rapid oscillatory states with EEG [47]. However, these advanced multimodal recordings present significant methodological challenges, particularly regarding artifact removal, which must be addressed to ensure data quality and validity.

Table 1: Comparison of Multimodal Neuroimaging Approaches

Method	Temporal Resolution	Spatial Resolution	Primary Applications	Key Challenges
EEG-fMRI	Millisecond range	Centimeter range (EEG); millimeter range (fMRI)	Mapping neural correlates of BOLD signals, epilepsy focus localization	Gradient and BCG artifact removal, safety concerns
TMS-EEG	Millisecond range	Centimeter range	Assessing cortical excitability, effective connectivity	TMS-induced artifact removal, sensory confounds
TMS-EEG-fMRI	Millisecond (EEG/TMS); seconds (fMRI)	Millimeter range (fMRI)	Causal brain network dynamics, state-dependent connectivity	Complex artifact interactions, hardware compatibility

Artifact Challenges in Specialized Recordings

EEG-fMRI Artifacts

Simultaneous EEG-fMRI recording introduces several significant artifacts that contaminate the EEG signal. The gradient artifact arises from the rapidly switching magnetic fields used for image acquisition and represents the most prominent source of interference [43] [48]. This artifact is characterized by its high amplitude and consistent timing relative to the image acquisition sequence. The ballistocardiographic (BCG) artifact originates from cardiac-related phenomena, including pulse-driven movements of the scalp, head rotation in the magnetic field, and the Hall effect of pulsatile blood flow [48]. Unlike gradient artifacts, BCG artifacts exhibit complex spatiotemporal dynamics and substantial shape variability across cardiac cycles, making them particularly challenging to remove.

Additional safety concerns emerge from the interaction between EEG equipment and the MR environment. The primary risk involves heating of conductive materials due to electromagnetic induction, where radiofrequency (RF) fields and switching gradients induce currents in EEG electrodes and lead wires [43] [49]. This risk can be mitigated through specialized hardware designs incorporating current-limiting resistors, carbon fiber leads, and careful configuration of electrode caps [43].

TMS-EEG Artifacts

TMS-EEG recordings face distinct artifact types that can obscure the neural responses of interest. The TMS pulse artifact manifests as an extremely high-amplitude signal saturation that can overwhelm EEG amplifiers, potentially persisting for several milliseconds after pulse delivery [44]. Additionally, TMS-induced muscle artifacts arise from activation of scalp and facial muscles, particularly when stimulating superficial cortical regions, producing high-frequency activity that can persist for 10-20 ms after the pulse [44].

Perhaps more insidiously, sensory confounds represent a significant challenge in TMS-EEG experiments. The characteristic "click" sound produced by TMS coil discharge generates auditory evoked potentials, while the sensation of scalp stimulation beneath the coil produces somatosensory evoked potentials [44]. These sensory responses can contaminate the TMS-evoked potentials (TEPs) and may be misinterpreted as cortical reactivity unless properly controlled through experimental design.

ICA Methodologies for Artifact Removal

ICA Fundamentals

Independent Component Analysis (ICA) is a multivariate statistical technique that separates mixed signals into statistically independent components based on higher-order statistics [50]. The fundamental model assumes that observed signals represent linear mixtures of underlying independent sources:

x = As

where x is the vector of observed signals, s contains the independent sources, and A is the mixing matrix. ICA estimates an unmixing matrix W that recovers the original sources: y = Wx [50]. For neuroimaging data, two predominant algorithmic approaches are commonly employed: Infomax and FastICA, both capable of identifying artifactual and neural sources based on their statistical properties [50] [51].

ICA for EEG-fMRI Artifacts

In simultaneous EEG-fMRI, ICA has proven particularly valuable for addressing the challenging BCG artifact. The conventional approach involves applying ICA to multichannel EEG data and identifying components representing BCG artifacts based on their temporal and spatial characteristics [50] [48]. These artifactual components are then removed before reconstructing the cleaned EEG signals.

Recent methodological advances have focused on hybrid approaches that combine ICA with other artifact removal techniques. Some researchers propose performing initial gradient artifact correction using average artifact subtraction (AAS), followed by ICA to address residual BCG artifacts [48]. Alternatively, a more sophisticated approach involves ICA decomposition first, followed by application of optimal basis set (OBS) methods to individual components before reconstruction, potentially improving artifact removal while preserving neural signals [48].

The adaptive Optimal Basis Set (aOBS) method represents a significant advancement in BCG artifact removal. This approach incorporates beat-to-beat estimation of the delay between cardiac activity and BCG occurrence, followed by principal component analysis (PCA) of accurately aligned BCG epochs [48]. A key innovation of aOBS is its automated selection of artifact-related components based on explained variance criteria, reducing the need for manual intervention. Studies demonstrate that aOBS achieves superior artifact reduction compared to traditional methods, with significantly lower BCG residuals (5.53% for aOBS versus 9.20-20.63% for other methods) and reduced cross-correlation with ECG signals [48].

ICA for TMS-EEG Artifacts

ICA application in TMS-EEG requires specialized considerations due to the unique nature of TMS-related artifacts. The approach typically involves identifying and removing components corresponding to TMS pulse artifacts, muscle activity, and sensory evoked potentials [44]. However, careful validation is essential to ensure that neural responses of interest are not inadvertently removed during this process.

Expert recommendations emphasize that ICA should be applied as part of a comprehensive preprocessing pipeline that may include additional techniques such as signal space projection (SSP), source-based reconstruction, and advanced filtering [44]. The precise sequencing of these steps and the criteria for component rejection remain active areas of methodological development in TMS-EEG research.

Diagram 1: ICA-based artifact removal workflow for simultaneous EEG-fMRI. The core ICA processing steps (red) identify and remove artifact components following initial gradient correction and BCG detection.

Quantitative Performance Comparison

Table 2: Performance Metrics of Artifact Removal Methods in Simultaneous EEG-fMRI

Method	BCG Residual (%)	Cross-Correlation with ECG	Signal Preservation	Computational Demand
AAS	12.51%	0.051	Moderate	Low
Standard ICA	20.63%	0.067	Variable	Medium
OBS	9.20%	0.042	Good	Medium
aOBS	5.53%	0.028	Excellent	High
ICA + OBS Hybrid	~7-8% (estimated)	~0.035 (estimated)	Very Good	High

The performance of different artifact removal methods has been quantitatively evaluated using metrics such as BCG residual intensity, cross-correlation with ECG, and signal-to-noise ratio (SNR) of event-related potentials [48]. The adaptive Optimal Basis Set (aOBS) method demonstrates superior performance across these metrics, achieving approximately 50% lower BCG residuals compared to traditional OBS and significantly better artifact reduction than standard ICA or AAS approaches [48].

For TMS-EEG data, the impact of different preprocessing pipelines extends beyond simple artifact reduction to influence fundamental neurophysiological measures. Different preprocessing approaches can significantly alter TMS-evoked potential (TEP) waveforms, particularly at latencies under 100 ms where correlation coefficients between pipelines can range from 0.2 to 0.9 [52]. Furthermore, the test-retest reliability of TEP measurements has been shown to depend critically on the chosen preprocessing pipeline [52].

Experimental Protocols

Simultaneous EEG-fMRI with ICA Artifact Removal

Equipment Preparation: Utilize MR-compatible EEG systems with carbon fiber leads and current-limiting resistors to minimize heating risks [43]. Ensure all equipment is non-ferrous and specifically rated for MR environment use.

Data Acquisition Parameters:

• EEG: Sampling rate ≥ 5000 Hz to adequately capture gradient artifacts [48]
• fMRI: TR = 2000 ms, TE = 30 ms, voxel size = 3×3×3 mm³ [47]
• Cardiac monitoring: Simultaneous ECG recording for BCG artifact removal

Step-by-Step Protocol:

• Subject Preparation: Apply EEG cap following standard preparation guidelines. Verify impedance values < 10 kΩ before entering scanner room.
• Hardware Setup: Position EEG system amplifier outside scanner room with leads properly secured to minimize movement.
• Initial Recording: Acquire 2-minute resting-state data without fMRI acquisition for baseline EEG quality assessment.
• Simultaneous Acquisition: Collect simultaneous EEG-fMRI data during experimental paradigm, ensuring synchronization of triggers between systems.
• Gradient Artifact Removal: Apply average artifact subtraction (AAS) using scanner slice triggers [48].
• BCG Artifact Removal: Implement aOBS method with the following sub-steps:
  • Detect QRS complexes in simultaneously recorded ECG
  • Identify BCG peaks in gradient-corrected EEG data
  • Epoch EEG data around BCG peaks with adaptive alignment
  • Apply PCA to epoched data and automatically select components based on explained variance
  • Reconstruct and subtract BCG artifact from original data
• Additional ICA: Perform ICA decomposition to address any residual artifacts [50].
• Signal Reconstruction: Reconstruct cleaned EEG data for subsequent analysis.

TMS-EEG-fMRI Trimodal Recording

Equipment Setup: Utilize TMS equipment specifically designed for MR environments, compatible EEG-fMRI systems, and appropriate neuronavigation for precise coil positioning [47].

Experimental Parameters:

• TMS: Triple-pulse stimulation at 55% maximum stimulator output targeting dorsal premotor cortex [47]
• EEG: Continuous recording with high-density cap (64+ channels)
• fMRI: Sparse imaging acquisition to minimize interference

Protocol Steps:

• Subject Screening: Exclude participants with contraindications to TMS or MRI.
• Anatomical Localization: Acquire high-resolution structural scan for neuronavigation.
• Target Identification: Use functional localizer task to identify participant-specific stimulation target [47].
• Coil Positioning: Navigate TMS coil to target coordinates using MRI-guided neuronavigation.
• Trimodal Recording: Implement interleaved TMS-EEG-fMRI acquisition with jittered intervals between TMS pulses [47].
• TMS Artifact Handling: Apply specialized preprocessing for TMS pulse artifacts, potentially including data interpolation around pulse delivery [44].
• Sensory Control Conditions: Incorporate control conditions with matched auditory and somatosensory stimulation without TMS to account for sensory confounds [44].
• Data Integration: Analyze relationship between pre-TMS oscillatory states (EEG), TMS perturbation, and network-level responses (fMRI) [47].

Diagram 2: Experimental workflow for trimodal TMS-EEG-fMRI recordings, highlighting critical steps for successful integration of the three techniques.

The Scientist's Toolkit

Table 3: Essential Research Reagents and Equipment for Multimodal Recordings

Item	Specification	Function/Purpose	Considerations
MR-Compatible EEG System	64+ channels with carbon fiber leads	Neural activity recording in MR environment	Must include current-limiting resistors; certified for MR safety
TMS System	MR-compatible model with liquid cooling	Focal perturbation of cortical activity	Must operate reliably in high magnetic field environment
Neuronavigation System	MRI-guided with infrared tracking	Precise TMS coil positioning	Requires high-resolution anatomical scans
Artifact Removal Software	ICA implementation (EEGLAB, FastICA)	Separation and removal of artifacts	Custom scripts often needed for specialized pipelines
ECG Recording System	MR-compatible with high sampling rate	Cardiac monitoring for BCG artifact removal	Essential for aOBS and related methods
Visual Stimulation Apparatus	MR-compatible goggles or projection system	Presentation of experimental paradigms	Synchronization with EEG/MRI triggers required

ICA methodologies have become indispensable tools for addressing the complex artifact challenges inherent in simultaneous EEG-fMRI and TMS-EEG recordings. The continuous refinement of these approaches, including the development of hybrid methods like aOBS and specialized ICA pipelines for TMS-related artifacts, has significantly enhanced our ability to extract meaningful neural signals from contaminated recordings. The trimodal integration of TMS-EEG-fMRI represents the cutting edge of non-invasive human neuroscience, enabling unprecedented investigation of causal brain network dynamics. As these methodologies continue to evolve, standardization of acquisition parameters and processing pipelines will be crucial for enhancing reproducibility and comparability across research sites. Future developments will likely focus on real-time artifact removal to enable closed-loop brain state-dependent stimulation and more sophisticated integration of multimodal data streams to unravel the complex temporal and spatial dynamics of human brain function.

Optimizing ICA Performance: From Common Pitfalls to Advanced Protocols

In the realm of electroencephalographic (EEG) analysis, Independent Component Analysis (ICA) has established itself as a fundamental method for separating neural activity from various artifacts. Successful ICA decomposition, however, relies heavily on appropriate data pre-processing, with filtering being one of the most critical parameters. The strategic application of high-pass and low-pass filters directly influences the independence of components, the algorithm's convergence, and the ultimate efficacy of artifact removal. This application note synthesizes current research to provide evidence-based protocols for filtering EEG data to optimize ICA decomposition, thereby enhancing the reliability of EEG analysis for research and clinical applications.

The Critical Role of Filtering in ICA Decomposition

ICA is a blind source separation technique that decomposes multi-channel EEG data into statistically independent components (ICs). The core assumption is that artifacts (e.g., from eyes, heart, or muscles) and neural signals originate from spatially fixed, temporally independent sources. Filtering prepares the data for this decomposition by ensuring that the underlying sources meet these statistical assumptions as closely as possible.

The central challenge lies in the spectral characteristics of both neural signals and common artifacts. Insufficient high-pass filtering can allow slow drifts, often caused by sweating or skin potentials, to dominate the signal variance, thereby hindering ICA's ability to find optimal component separations [53]. Conversely, excessive high-pass filtering can distort or remove genuine neural activity of interest, such as event-related potentials. Similarly, low-pass filtering must be carefully calibrated to suppress high-frequency noise, including muscular artifacts, without eliminating relevant neural oscillations [54]. Research indicates that for the three most common ICA methods (extended Infomax, FastICA, and TDSEP), adequately high-pass filtering is very important, and compared to its effect, differences between the decomposition methods themselves are small [54].

Quantitative Filtering Parameters from Empirical Research

Systematic evaluations have yielded specific, quantitative recommendations for filter settings that promote high-quality ICA decompositions. The tables below summarize key findings from recent studies.

Table 1: Recommended High-Pass Filter Parameters for ICA

Recommended Cutoff	Research Context	Impact on ICA & Signal Quality
1-2 Hz [55]	Auditory oddball task (21 participants); ICA with MARA classifier	Consistently produced good results in terms of SNR, single-trial classification accuracy, and percentage of 'near-dipolar' components.
1 Hz [56]	General EEG pre-processing recommendation (EEGLAB tutorial)	Recommended to obtain good quality ICA decompositions and to remove linear trends.
2 Hz (for muscle artifacts) [54]	Self-paced foot movements (18 participants); evaluation of multiple ICA methods	Adequate high-pass filtering was crucial for the reduction of muscle artifacts by all three ICA methods tested (extended Infomax, FastICA, TDSEP).

Table 2: Recommended Low-Pass Filter Parameters for ICA

Recommended Cutoff	Research Context	Impact on ICA & Signal Quality
Adjustable parameter [57]	Free viewing & sentence reading; optimization of ocular artifact removal	Identified as one of four key parameters in the ICA pipeline; optimal value can depend on the specific artifacts and neural signals of interest.
Included in broadband 2-45 Hz filter [54]	Self-paced foot movements; removal of muscular artifacts	Data were broadband-filtered (2-45 Hz) before decomposition; low-pass filtering at 45 Hz helped limit high-frequency muscle contamination.

Detailed Experimental Protocols

Protocol 1: Optimizing ICA for General Artifact Removal

This protocol is adapted from Winkler et al. (2015) and standard EEGLAB practices, suitable for standard cognitive paradigms like the auditory oddball task [55] [56].

Step-by-Step Methodology:

• Data Import and Channel Setup: Load the continuous EEG data. Assign channel locations and remove non-EEG channels (e.g., EOG, EMG) from the data matrix to be decomposed.
• Filter Application:
  • High-Pass Filter: Apply a zero-phase shift FIR high-pass filter with a cutoff frequency of 1 Hz. A transition band as wide as possible is recommended to avoid artifacts [56].
  • Low-Pass Filter: Apply a zero-phase shift FIR low-pass filter with a cutoff frequency of 40-45 Hz to attenuate high-frequency noise and muscle activity [54].
• Data Segmentation (Optional): For studies with event-related potentials, the continuous data can be segmented into epochs around the events of interest. Note that filtering is often recommended on continuous data to minimize edge artifacts.
• Bad Channel Removal and Interpolation: Identify and remove channels with excessive noise. These can be interpolated after ICA decomposition.
• ICA Decomposition: Run an ICA algorithm (e.g., extended Infomax) on the filtered data. The data matrix should be of full rank.
• Component Classification and Removal: Use automated classifiers (e.g., MARA, ICLabel) or manual inspection to identify and flag artifactual components [55] [54].
• Signal Reconstruction: Reconstruct the artifact-corrected EEG signal by projecting all components back to the sensor space, excluding the artifactual ones.

Protocol 2: Advanced Protocol for Ocular Artifacts in Free-Viewing Experiments

This protocol, derived from Dimigen (2020), is optimized for challenging scenarios with abundant eye movements, such as visual search or reading experiments [57].

Step-by-Step Methodology:

• Data Preparation: Combine EEG with eye-tracking recordings to objectively quantify correction quality.
• Optimized Filtering:
  • High-Pass Filter: The study identified the high-pass filter cutoff as a key parameter. While an exact value is dataset-dependent, a cutoff of 1-2 Hz is a reasonable starting point for optimization.
  • Low-Pass Filter: Similarly, the low-pass filter cutoff must be optimized. The author found that training ICA on optimally filtered data was critical.
• Overweighting Artifactual Data: To improve the separation of ocular components, massively overweight the proportion of training data containing myogenic saccadic spike potentials (SPs). This teaches the algorithm to better isolate these artifacts.
• ICA Training: Run the ICA decomposition on this optimally filtered and weighted dataset.
• Component Rejection with Eye Tracker: Use the synchronized eye-tracking data to set an objective threshold for eye tracker-based component rejection. This helps minimize both undercorrection (residual artifacts) and overcorrection (removal of neurogenic activity).
• Validation: Quantify the correction quality by inspecting the removal of the saccadic spike potential and its associated spectral broadband artifact.

The following diagram illustrates the logical workflow for optimizing filter parameters as part of the ICA pre-processing pipeline.

Diagram 1: Workflow for Optimizing Filter Parameters in ICA

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Tools for ICA-based EEG Pre-processing

Tool / Material	Function in Research	Exemplary Solution / Implementation
FIR Filtering Toolbox	Applies zero-phase high-pass and low-pass filtering to prepare data for ICA.	EEGLAB's "Basic FIR filter" [56], BrainVision Analyzer.
ICA Algorithm	Performs blind source separation to decompose EEG signals into independent components.	Extended Infomax (EEGLAB) [54] [31], AMICA [18], FastICA.
Automated Component Classifier	Objectively identifies artifactual components based on features like topography, spectrum, etc.	MARA [55], ICLabel.
Synchronized Eye-Tracker	Provides objective data to identify and weight ocular artifacts for optimal removal.	Used in free-viewing paradigms to guide ICA training and validation [57].
Artifact Subspace Reconstruction (ASR)	An alternative/complementary method for removing large-amplitude, transient artifacts before ICA.	clean_rawdata plugin in EEGLAB [18].

The selection of high-pass and low-pass filter cutoffs is not merely a routine pre-processing step but a key parameter determining the success of subsequent ICA decomposition for EEG artifact removal. Empirical evidence consistently supports the use of high-pass filters with cutoffs between 1 and 2 Hz to facilitate a robust separation of neural and artifactual sources. Low-pass filtering, while equally vital, must be tailored to the specific experimental context and the frequency profile of the artifacts. By adhering to the detailed protocols and quantitative guidelines outlined in this application note, researchers and drug development professionals can significantly enhance the signal quality of their EEG data, leading to more reliable and interpretable results in both basic and applied neuroscience.

Independent Component Analysis (ICA) has become a cornerstone technique for isolating and removing artifacts from electroencephalography (EEG) data. The quality of the ICA decomposition is fundamentally dependent on the preprocessing steps applied to the data beforehand. Imperfect component separation can lead to the unintended removal of neural signals alongside artifacts, potentially inflating event-related potential effect sizes, biasing connectivity measures, and distorting source localization estimates [13]. This application note synthesizes current evidence and provides detailed protocols for optimizing data cleaning and sample rejection strategies to enhance the reliability of ICA decomposition in EEG research, with particular relevance for clinical trials and drug development studies where signal integrity is paramount.

The Impact of Data Cleaning on ICA Decomposition Quality

Empirical Evidence for Cleaning Efficacy

Recent studies have systematically evaluated how data cleaning protocols influence ICA decomposition outcomes. The relationship between movement intensity, sample rejection strength, and decomposition quality has been quantitatively assessed across multiple experimental paradigms.

Table 1: Effects of Cleaning Parameters on ICA Decomposition Quality

Parameter	Effect on Decomposition	Optimal Range	Evidence Source
AMICA Sample Rejection Iterations	Significantly improves component quality, though effect smaller than expected [18].	5 to 10 iterations [18].	Analysis of 8 open-access datasets with varying motion intensity [18].
Rejection Threshold (Standard Deviations)	Removes samples the algorithm cannot easily account for, improving model fit [18].	~3 SDs from mean log-likelihood [18].	Model-driven rejection based on log-likelihood during AMICA computation [18].
Data Volume for ICA	Critical for successful decomposition; more data yields more stable components [10].	Maximum available clean data; PCA reduction if insufficient data [10].	ICA algorithm requirements for estimating unmixing matrix [10].
Targeted Artifact Reduction	Reduces false positive effects and source localization biases compared to full component subtraction [13].	Clean artifact periods (eye) and frequencies (muscle) within components [13].	Testing across different EEG systems and cognitive tasks (Go/No-go, N400) [13].

Consequences of Inadequate Cleaning

Insufficient attention to data cleaning prior to ICA can introduce significant confounds in analysis. When artifacts systematically differ between experimental conditions, they can create spurious effects that mimic genuine neural activity [58]. For example, eyeblink artifacts that occur more frequently in one condition can create differences in ERP waveforms that might be misinterpreted as neural effects [58]. Furthermore, residual artifacts increase uncontrolled variance in the data, reducing statistical power and potentially obscuring genuine effects [58]. The decomposition process itself can be compromised when substantial artifacts remain in the data, leading to components that represent mixed neural and artifactual sources [13] [10].

Experimental Protocols for Data Cleaning

Comprehensive Pre-ICA Cleaning Protocol

This protocol outlines a systematic approach for preparing EEG data for ICA decomposition, integrating robust methods from recent literature.

Materials and Reagents

• Software Requirements: EEGLAB with RELAX plugin [13] and/or FieldTrip toolbox [59]
• Computing Resources: MATLAB environment with sufficient RAM for large datasets
• EEG System: High-density EEG recording system (≥58 channels recommended) [18]

Procedure

• Initial Data Assessment
  • Visualize continuous data using a data browser (e.g., ft_databrowser in FieldTrip) to identify grossly abnormal channels and obvious artifacts [59].
  • Plot channel locations and verify proper positioning against standard templates [60] [10].

• Bad Channel Identification and Interpolation

  • Manually or automatically identify channels with consistently poor signal quality throughout the recording [59].
  • Use interpolation methods (e.g., ft_channelrepair in FieldTrip) to reconstruct bad channels based on neighboring channels [59].
  • Apply weighted average interpolation using predefined channel neighbor templates [59].

• Temporally-Local Artifact Rejection

  • Identify and mark segments containing large, transient artifacts (e.g., electrode shifts, cable sway) [60] [59].
  • Apply artifact padding (e.g., 100ms) around identified artifacts to ensure complete removal [59].
  • Remove marked segments or use inpainting techniques to replace artifactual periods [59].

• Robust Detrending

  • Apply robust detrending algorithms (e.g., nt_detrend from Noisetools) to remove slow drifts without being affected by outliers [59].
  • Implement algorithms that can exclude artifactual periods during the detrending process [59].

• Re-referencing

  • Apply robust re-referencing after initial cleaning steps [59].
  • Common average referencing is widely used, though reference selection should align with research objectives [59].

The following diagram illustrates the sequential workflow for the comprehensive pre-ICA cleaning protocol:

AMICA Sample Rejection Protocol

The Adaptive Mixture ICA (AMICA) algorithm includes an integrated sample rejection function that automatically removes problematic samples during the decomposition process. This protocol details its optimal configuration.

Materials and Reagents

• Software: EEGLAB with AMICA plugin (v1.6.1 or higher) [18]
• Hardware: Substantial computing resources as AMICA is computationally intensive

Procedure

• Data Preparation
  • Complete minimal preprocessing including high-pass filtering (typically >1Hz) and bad channel removal [18] [59].
  • Ensure sufficient data quantity—ICA requires ample clean data for stable decomposition [10].

• AMICA Parameter Configuration

  • Enable sample rejection in the AMICA settings (disabled by default in command line) [18].
  • Set rejection iterations to 5-10 cycles for optimal results [18].
  • Configure rejection threshold to 3 standard deviations from the mean log-likelihood [18].
  • Adjust the number of model computation steps between rejection iterations (default varies by interface) [18].

• Decomposition and Validation

  • Run AMICA decomposition with the configured sample rejection parameters.
  • Assess decomposition quality using mutual information measures between components [18].
  • Evaluate the proportion of brain, muscle, and 'other' components identified [18].
  • Check residual variance and signal-to-noise ratio in representative conditions [18].

Targeted Artifact Reduction Protocol (RELAX Method)

The RELAX method represents a recent advancement in artifact handling that targets specific artifact characteristics within components rather than subtracting entire components.

Materials and Reagents

• Software: EEGLAB with RELAX plugin [13]
• Data Requirements: Previously computed ICA decomposition

Procedure

• ICA Decomposition
  • Perform standard ICA decomposition using preferred algorithm (Infomax, AMICA, etc.) [13] [10].
  • Identify artifactual components related to eye movements and muscle activity [10].

• Component Analysis

  • For eye movement components: identify artifact periods based on component time course [13].
  • For muscle components: identify artifact frequencies based on spectral characteristics [13].

• Targeted Cleaning

  • Apply cleaning specifically to identified artifact periods and frequencies within components [13].
  • Preserve neural activity in non-artifact periods and frequencies [13].
  • Reconstruct data using the modified components [13].

• Validation

  • Compare effect sizes before and after cleaning to check for artificial inflation [13].
  • Evaluate source localization biases using standardized protocols [13].

The following diagram illustrates the targeted artifact reduction workflow:

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Tools for EEG Data Cleaning and ICA

Tool Name	Function	Application Context	Implementation
RELAX Pipeline	Targeted artifact reduction within ICA components [13].	Protection against false positive effects in ERP and connectivity studies [13].	EEGLAB plugin, freely available [13].
AMICA Sample Rejection	Model-driven rejection of problematic samples during decomposition [18].	Mobile EEG studies with increased motion artifacts [18].	Integrated within AMICA algorithm in EEGLAB [18].
ICLabel	Automated classification of ICA components into neural and artifactual sources [60].	Standardized component selection across researchers and studies [60].	EEGLAB plugin component classification tool [60].
Clean Rawdata/ASR	Automated artifact removal using Artifact Subspace Reconstruction [60].	Initial cleaning before ICA; particularly effective for large transient artifacts [60].	EEGLAB plugin with adjustable thresholds [60].
FieldTrip Preprocessing	Modular pipeline for robust data cleaning and artifact rejection [59].	Resting-state EEG and task-based studies requiring reproducible workflows [59].	Open-source MATLAB toolbox [59].
Noisetools	Robust detrending and outlier detection algorithms [59].	Handling data with substantial drifts or intermittent artifacts [59].	MATLAB functions by de Cheveigné & Arzounian [59].

Discussion and Future Directions

The strategic implementation of data cleaning and sample rejection protocols significantly enhances the quality of ICA decomposition for EEG artifact removal. Current evidence supports a balanced approach that includes both traditional preprocessing methods and algorithm-specific sample rejection functions. The development of targeted artifact reduction methods, such as the RELAX pipeline, represents a promising direction that moves beyond binary component rejection toward more nuanced artifact handling [13].

For researchers in drug development and clinical trials, where signal integrity directly impacts outcome measures, adopting these optimized protocols can reduce false positive findings and improve measurement reliability. Future methodological developments will likely focus on adaptive cleaning approaches that automatically adjust parameters based on data quality metrics and specific research contexts. The integration of machine learning approaches for artifact identification and removal presents another promising avenue for enhancing ICA decomposition quality in increasingly mobile and real-world EEG applications.

Electroencephalography (EEG) recorded in naturalistic, mobile settings—such as free-viewing experiments and Mobile Brain/Body Imaging (MoBI)—captures brain activity during active exploration and whole-body movement. These paradigms are highly susceptible to complex artifacts that can overshadow neural signals of interest. Movement artifacts, muscle activity from head movements, eye blinks, cable sway, and electrode shifts introduce non-brain signals that are often higher in amplitude and broader in spectrum than neural activity [61] [18]. Independent Component Analysis (ICA) has emerged as a powerful method for separating and removing these artifacts while preserving neural data. This application note provides detailed protocols and quantitative data for optimizing ICA-based artifact removal in these challenging research scenarios.

Table 1: Characteristics of Challenging EEG Recording Scenarios

Scenario	Primary Sources of Artifact	Impact on EEG Signal	Recommended ICA Approach
Free-Viewing	Unconstrained eye movements (saccades, blinks), neck muscle tension	Ocular artifacts dominant (frontal), spectrum overlaps with neural signals	Informax or Extended Infomax ICA; focus on ocular & muscle component identification
Mobile EEG	Head movements, cable sway, muscle activity (neck, jaw), transient electrode pops	Increased high-frequency broadband power, large transient spikes, slow oscillatory artifacts	AMICA with moderate sample rejection; comprehensive feature set for classification
MoBI	Whole-body movement (gait), muscle activity (limbs), heavy sweating, cable motion	Complex mixture of motion, EMG, and mechanical artifacts; deep brain sources of interest	High-density EEG (>100 channels); AMICA with strong sample rejection; source localization

Experimental Protocols for Data Acquisition and Preprocessing

Data Acquisition Specifications for Mobile Paradigms

Successful artifact removal begins with optimized data acquisition. For MoBI studies, high-density EEG systems with 108 or more electrodes are recommended to provide sufficient spatial information for ICA to separate sources effectively [62]. Electrode impedances should be maintained below 5 kΩ to minimize motion-induced disruptions. Concurrent recording of auxiliary signals is crucial: surface electromyography (EMG) from relevant muscles (e.g., tibialis anterior during walking), electrooculography (EOG) for eye movements, foot switches or accelerometers for gait event timing, and goniometers for joint angles [62]. These signals provide ground truth for validating artifact components and understanding movement-related brain dynamics. All data streams must be synchronized via a common hardware clock.

Preprocessing Workflow for Motion-Rich EEG Data

A rigorous preprocessing pipeline is essential before ICA decomposition to handle severe motion artifacts.

Table 2: Preprocessing Steps for Motion-Contaminated EEG

Processing Step	Key Parameters	Rationale & Implementation Notes
Frequency Filtering	0.5–100 Hz bandpass (4th-order Butterworth); 50/60 Hz notch filter	Removes DC drift, high-frequency noise, and powerline interference [63]
Bad Channel Removal	>4-5 SD from mean channel variance; correlation with neighbors <0.8	Identifies chronically noisy or disconnected electrodes; requires interpolation later
Flatline/Saturation Removal	Identify isoelectric flatlines or constant saturated values; remove ±10s around event	Removes severe system failures or electrode disconnections affecting all channels [63]
Abnormal Peak Removal	Threshold: ±5 mV; check for coincidence in frontal/occipital/pole electrodes	Identifies large motion artifacts; remove if occurring in geometrically distributed channels [63]
Data Segmentation	10-minute segments (or shorter if necessary); minimum 10s duration	Creates manageable data portions for ICA; shorter segments for highly non-stationary data [63]
Channel Interpolation	Spherical spline interpolation for removed bad channels	Reconstructs missing channels to maintain full sensor array for ICA

Figure 1: Preprocessing workflow for motion-contaminated EEG. This pipeline prepares data for optimal ICA decomposition by removing severe artifacts while preserving data integrity.

ICA Decomposition and Optimization Protocols

Algorithm Selection and Configuration

The Adaptive Mixture ICA (AMICA) algorithm currently demonstrates superior performance for decomposing mobile EEG data compared to other algorithms like Infomax or FastICA [18]. AMICA's ability to model multiple data distributions makes it particularly suitable for handling the non-stationary characteristics of motion-contaminated recordings. For high-density EEG systems (≥64 channels), compute the ICA decomposition on the full channel set to maximize spatial resolution. The data should be mean-centered but not otherwise scaled before ICA. For exceptionally long recordings (>30 minutes), consider segmenting the data into smaller epochs (10-20 minutes) while ensuring each segment contains sufficient data points (samples > 20 × channels²) for stable decomposition.

Optimizing AMICA Sample Rejection

AMICA includes an integrated sample rejection feature that automatically identifies and removes time points poorly accounted for by the current model during decomposition. This function is particularly valuable for mobile EEG containing transient motion artifacts.

Table 3: AMICA Sample Rejection Parameters for Different Scenarios

Mobility Level	Recommended Iterations	Standard Deviation Threshold	Model Steps Between Rejection
Stationary/Sedentary	3-5 iterations	3 SD	3 steps
Light Movement	5-7 iterations	3 SD	2 steps
Whole-Body MoBI	8-10 iterations	2.5-3 SD	1-2 steps

Empirical evidence indicates that moderate cleaning (5-10 iterations) significantly improves decomposition quality across various movement intensities [18]. The sample rejection is model-driven, selectively removing artifacts that negatively impact ICA convergence while retaining informative data. This approach is more effective than amplitude-based cleaning methods that might remove biologically valid neural signals.

Automated Component Classification and Artifact Removal

Feature Extraction for Component Classification

Accurate classification of independent components as neural or artifactual requires a multi-domain feature set. The EPIC dataset, containing 77,426 expert-annotated components, provides a valuable resource for developing and validating classifiers [63]. The following feature categories should be extracted for each component:

• Spatial Features: Scalp topography maps, current density norms, range within pattern, equivalent current dipole fits with residual variance [63] [23]
• Spectral Features: Power spectrum density characteristics, mean local skewness, deviation from 1/f power law, relative power in specific bands (theta, alpha, beta, gamma) [23]
• Temporal Features: Autocorrelation function, kurtosis of activations, temporal predictability, relationship with external triggers (e.g., gait events) [23]

Machine Learning Classification Approaches

Both linear and nonlinear classifiers have been successfully employed for automated component classification. Linear classifiers using an optimized subset of 6-10 features can perform on par with inter-expert disagreement rates (<10% Mean Squared Error) when trained on sufficiently large annotated datasets [23]. For the EPIC dataset, classifiers achieved 70.44% brain components in training sets and 70.02% in testing sets, demonstrating consistency across data partitions [63]. Bayesian classifiers have also shown promising results, identifying EEG components with 87.6% sensitivity and 70.2% specificity in ictal recordings [64]. The choice between linear and nonlinear approaches involves a trade-off between interpretability and potential performance gains.

Figure 2: Automated component classification workflow. Features from multiple domains are extracted and processed by a classifier to separate neural from artifactual components.

Validation and Signal Quality Assessment

After artifact removal, comprehensive validation is essential to ensure neural signal preservation. Calculate the signal-to-noise ratio in task-relevant event-related potentials or oscillatory activities before and after cleaning. For MoBI paradigms, verify that movement-related spectral perturbations persist in physiologically plausible frequency bands. Assess the retention of known neurophysiological patterns, such as alpha power decreases during eye opening in visual areas or gait-related modulation in sensorimotor rhythms [61] [62]. When possible, correlate cleaned EEG signals with concurrently recorded physiological data (EMG, kinematics) to confirm that removed components predominantly reflect non-neural sources.

The Scientist's Toolkit

Table 4: Essential Research Reagents and Solutions

Tool/Category	Specific Examples	Function & Application Notes
ICA Algorithms	AMICA, Infomax (runica), FastICA, SOBI	Blind source separation; AMICA recommended for mobile data [18] [10]
Annotation Datasets	EPIC Dataset (77,426 components) [63]	Training automated classifiers; contains IC time-series, PSD, topography
Preprocessing Tools	EEGLAB Clean_Rawdata (ASR), AMICA sample rejection	Automatic artifact removal; ASR requires careful threshold tuning [18]
Mobile EEG Systems	108-channel high-density systems (MoBI) [62]	High spatial resolution for source separation; mastoid reference/ground
Auxiliary Sensors	EMG, goniometers, pressure sensors, accelerometers [62]	Provide timing correlates for validation; record true biological artifacts
Validation Metrics	Residual variance, mutual information, SNR	Quantify decomposition quality and artifact removal efficacy [18]

A central challenge in electroencephalography (EEG) research is the removal of biological artifacts without distorting or eliminating the underlying neurogenic signals of interest. Independent Component Analysis (ICA) has emerged as a powerful tool for this purpose, capable of separating mixed signals into their constituent sources. However, an overly aggressive approach to artifact removal can inadvertently discard or alter cerebral activity, a problem known as overcorrection. This application note details structured protocols and quantitative insights to help researchers navigate this challenge, ensuring the integrity of neurogenic activity in EEG data for downstream analysis in clinical and drug development contexts.

The principle of ICA is based on blind source separation, where recorded EEG data (X) is considered a linear mixture of underlying sources (S) via a mixing matrix (A), such that X = A × S [1]. The goal is to calculate an unmixing matrix (W) to isolate the independent components (ICs). The core challenge lies in accurately identifying which ICs represent artifacts and removing them without compromising brain-related signals [10].

Quantitative Effects of Preprocessing on Neurogenic Content

The choice of preprocessing strategy directly impacts the stability and interpretability of brain-derived signals. Research on EEG microstates—dynamic sequences of brain network activity—provides a quantitative framework for assessing this impact.

Table 1: Impact of ICA Preprocessing Strategy on Microstate Metrics and Neurogenic Content

Preprocessing Strategy	Effect on Microstate Topography	Effect on Microstate Feature Stability	Statistical Power for EO/EC Comparison
No ICA Preprocessing	Unstable and unreliable	Low stability	Greatly reduced [41]
Remove Ocular Artifacts Only	Stable and brain-related	High stability	High statistical power [41]
Remove All Identified Artifacts	Stable and brain-related	High stability	High statistical power [41]

The data demonstrates that while skipping artifact removal is detrimental, provided a high-quality recording exists and ocular artifacts are removed, microstate topographies and features are robust to the level of preprocessing [41]. This paves the way for automated pipelines that prioritize preservation of neurogenic signals.

Experimental Protocols for Controlled Artifact Removal

Protocol 1: A Tiered ICA Preprocessing Strategy for Microstate Analysis

This protocol is adapted from Artoni et al. (2024) and is designed to systematically evaluate and control for overcorrection [41].

• Aim: To test the reliability of microstate extraction and the stability of microstate features against different ICA-based artifact removal strategies.
• Materials: Normative resting-state EEG data with alternating eyes-open (EO) and eyes-closed (EC) conditions.
• Method:
  • Data Preparation: Ensure data is high-quality and contains channel location information.
  • Preprocessing Tiers: Process the same dataset through four distinct pipelines:
    • Tier I: No ICA preprocessing.
    • Tier II: ICA with removal of ocular artifact components only.
    • Tier III: ICA with removal of all reliably identified physiological (muscle, cardiac) and non-physiological artifacts.
    • Tier IV: ICA retaining only components reliably identified as originating from brain activity.
  • Microstate Analysis: For each processed dataset, perform microstate extraction to identify dominant topographies and analyze their chronological sequences.
  • Comparison: Evaluate microstate evaluation criteria, topography stability, and the statistical power of EO/EC comparisons across the four tiers.
• Key Outcome: The protocol validates that removing ocular artifacts is critical, but more aggressive preprocessing does not substantially alter core neurogenic features, providing a justification for a conservative approach.

Protocol 2: Optimizing AMICA with Integrated Sample Rejection

For data with significant motion artifacts, such as from mobile or clinical populations, the AMICA algorithm's integrated cleaning can be optimized as per [18].

• Aim: To improve ICA decomposition quality for data with varying motion intensity without excessive data loss.
• Materials: EEG datasets with varying degrees of subject mobility.
• Method:
  • Algorithm Selection: Use the Adaptive Mixture ICA (AMICA) algorithm.
  • Parameter Setup: Enable AMICA's built-in sample rejection function. This function rejects samples based on their log-likelihood, effectively removing data points the algorithm cannot easily model.
  • Intensity Calibration: Set the sample rejection to perform between 5 and 10 iterations. This represents a moderate cleaning strength that has been shown to improve decomposition without being overly aggressive [18].
  • Quality Assessment: Evaluate decomposition quality using metrics such as:
    • Mutual information between components.
    • Proportion of brain vs. non-brain components.
    • Residual variance of brain components.
    • Signal-to-noise ratio in key experimental conditions.
• Key Outcome: This model-driven cleaning approach robustly improves decomposition for noisy data while preserving data integrity, as it primarily targets samples that hinder the ICA model itself.

Protocol 3: Hybrid ICA-Regression for Ocular Artifact Removal

This protocol, based on the method from [65], offers a refined technique for removing ocular artifacts while maximizing the recovery of neuronal signals from the artifactual components.

• Aim: To automatically identify and eliminate ocular artifacts with minimal loss of neuronal activity.
• Materials: Contaminated EEG data and simultaneously recorded EOG signals (vertical and horizontal).
• Method:
  • Decomposition: Perform ICA on the contaminated EEG data to obtain Independent Components (ICs).
  • Automatic Identification: Calculate two statistical measures for each IC:
    • Composite Multi-Scale Entropy: Measures signal complexity; ocular artifacts typically have lower entropy.
    • Kurtosis: Measures the "peakedness" of the signal distribution; eye-blinks have high kurtosis.
  • Artifact Attenuation: Apply a Median Absolute Deviation (MAD) filter to the identified ocular ICs to remove high-magnitude ocular activity peaks.
  • Neuronal Signal Recovery: Process the artifact-attenuated ICs using a linear regression model (e.g., Extended Recursive Least Squares) trained on the EOG references. This step aims to remove any remaining ocular artifact and recover neuronal signals present within these components.
  • Reconstruction: Back-project all ICs (including the corrected artifactual ones and the untouched neuronal ones) to reconstruct the artifact-free EEG signal.
• Key Outcome: This hybrid approach has been shown to outperform standard ICA or regression alone, resulting in lower mean square error and higher mutual information between the cleaned and original, artifact-free EEG [65].

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Tools for ICA-based EEG Artifact Removal

Tool / Solution	Function in Research	Application Note
High-Density EEG System	Records brain electrical activity from multiple scalp sensors.	A sufficient number of channels (e.g., >58) is critical for successful ICA decomposition [18].
AMICA Plugin	An advanced ICA algorithm for source separation.	More robust to limited data cleaning; includes integrated, model-driven sample rejection [18].
EEGLAB Software	An interactive MATLAB toolbox for EEG processing.	Provides a standard platform for running ICA, component inspection, and implementing removal protocols [10].
RELICA Plugin	A bootstrap-based method for assessing ICA reliability.	Quantifies the stability of extracted components, helping to avoid misinterpretation of unreliable ICs [41].
Artifact Subspace Reconstruction (ASR)	An automated method for removing large-amplitude, transient artifacts.	Can be used as a preprocessing step before ICA; requires careful threshold calibration to avoid overcleaning [18].

Workflow Visualization for Preserving Neurogenic Signals

The following diagram illustrates a decision-making workflow designed to minimize overcorrection during ICA-based artifact cleaning, integrating the principles from the protocols above.

Diagram 1: ICA artifact removal workflow for neurogenic signal preservation.

Preserving neurogenic activity during EEG artifact removal is a balance between rigorous cleaning and conservative interpretation. The quantitative data and protocols outlined herein demonstrate that a tiered, hypothesis-driven approach is superior to a one-size-fits-all application of ICA. By selecting a preprocessing strategy that matches the data's noise profile, leveraging advanced algorithms like AMICA for challenging recordings, and employing hybrid methods for critical artifact types, researchers can significantly mitigate the risk of overcorrection. This ensures that the biological signals central to neuroscientific discovery and clinical application remain intact and interpretable.

Within electrophysiological research, particularly in electroencephalography (EEG) studies for drug development and clinical neuroscience, Independent Component Analysis (ICA) has become a cornerstone technique for isolating and removing artifacts from neural data. The quality of the ICA decomposition is paramount, as it directly impacts the reliability of subsequent neural signals analysis. This application note focuses on two powerful automated preprocessing tools: the integrated sample rejection within the Adaptive Mixture ICA (AMICA) algorithm and Artifact Subspace Reconstruction (ASR). We detail their operational principles, provide quantitative comparisons, and outline standardized protocols to enable researchers to leverage these tools effectively within a robust, automated preprocessing pipeline [18] [60].

AMICA's Integrated Sample Rejection

AMICA is recognized as one of the most powerful algorithms for EEG ICA decomposition [18]. A distinctive feature of AMICA is its model-driven, integrated sample rejection capability. Unlike amplitude-based cleaning methods that may remove physiologically relevant data, AMICA's approach is to iteratively reject samples based on their log-likelihood during model computation. Samples that the algorithm finds difficult to account for (i.e., those with a log-likelihood falling significantly below the mean) are considered detrimental to a clean decomposition and are excluded [18]. This process targets artifacts that negatively impact ICA specifically, while often preserving stereotyped artifacts like eye blinks that ICA itself can later separate and remove [18]. The key parameters controlling this process are:

• Number of Rejection Iterations: How many cleaning cycles are performed.
• Standard Deviation (SD) Threshold: The stringency of rejection (e.g., samples 3 SDs below the mean log-likelihood are rejected).
• Timing and Intervals: When rejection starts and the number of model computation steps between each rejection iteration [18].

Artifact Subspace Reconstruction (ASR)

ASR is a popular method for cleaning continuous EEG data, available in EEGLAB via the clean_rawdata plugin. It functions by identifying and reconstructing high-variance, multi-channel artifacts in near-real-time. The core principle involves:

• Calibration: Learning a calibration covariance matrix from a segment of clean, "baseline" data.
• Detection and Reconstruction: For a sliding window of data, ASR performs a Principal Component Analysis (PCA). If the variance in any principal component exceeds a predefined threshold relative to the calibration data, that component is flagged as artifactual.
• Subspace Repair: The artifactual component(s) are removed, and the data within the window is reconstructed from the remaining "brain-like" components [60].

ASR is particularly effective for removing large, transient artifacts such as those caused by electrode pops, cable sway, or sudden head movements [60]. However, its performance is highly sensitive to the chosen threshold and the quality of the baseline calibration data [18].

Quantitative Comparison of Cleaning Methods

The effectiveness of data cleaning strategies can be evaluated using several quantitative metrics. The following table summarizes key findings from a systematic evaluation of AMICA's sample rejection and other common approaches.

Table 1: Quantitative Evaluation of Pre-ICA Data Cleaning Methods

Method	Key Mechanism	Optimal Parameters / Conditions	Impact on ICA Decomposition Quality	Key Advantages
AMICA Sample Rejection [18]	Model-driven iterative rejection based on sample log-likelihood.	5-10 iterations with an SD threshold of 3.	Significant improvement in decomposition, measured by reduced mutual information between components and a higher proportion of brain components. Effect was robust across datasets with varying motion intensity.	Targets only artifacts harmful to ICA; preserves data ICA can handle; integrated into the decomposition process.
Artifact Subspace Reconstruction (ASR) [18] [60]	Statistical detection and reconstruction of artifact subspaces using PCA.	Performance is highly threshold-dependent. Requires a clean baseline segment for calibration.	Can improve decomposition but is less robust and more variable than AMICA's method, especially without a properly tuned threshold and baseline.	Effective for large, transient, non-stereotyped artifacts; works on continuous data.
AMICA without Sample Rejection [18]	N/A	N/A	Serves as a baseline. Decomposition quality is lower but demonstrates AMICA's inherent robustness.	Preserves all data for decomposition.
Automated Peak-to-Peak Thresholding (e.g., Autoreject) [66]	Cross-validation to find optimal peak-to-peak amplitude thresholds per sensor.	Data-specific thresholds are learned automatically.	Not directly compared in the same study, but is a state-of-the-art method for bad trial rejection and sensor repair, improving signal quality for analysis.	Fully automated, transparent, and minimizes data loss by repairing bad channels.

Another critical consideration is the impact of the chosen artifact removal strategy on downstream EEG analysis. Research on EEG microstates has shown that the level of preprocessing can affect the stability of results.

Table 2: Impact of ICA-Based Artifact Removal Strategy on EEG Microstate Features

ICA Preprocessing Strategy	Description	Impact on Microstate Topography & Features
No ICA Preprocessing [67]	Using raw data after band-pass filtering and bad channel/epoch rejection.	Not recommended; leads to unstable microstate evaluation criteria and reduced statistical power.
Remove Ocular ICs Only [67]	Removing only components identified as eye blinks and movements. Recommended minimum.	Ensures stability of microstate topographies and features, allowing for robust statistical comparisons.
Remove All Identifiable Artifacts [67]	Removing ICs related to eyes, heart, muscle, etc.	Microstate topographies and features remain stable and capture brain-related physiology.
Retain Only Brain ICs [67]	Most aggressive approach, keeping only components reliably identified as brain-related.	Similar stability as removing all artifacts; both aggressive strategies are viable after ocular artifact removal.

Experimental Protocols

Protocol 1: Utilizing AMICA's Integrated Sample Rejection

This protocol is designed for use within the EEGLAB environment with the AMICA plugin installed.

Research Reagent Solutions:

• Software: MATLAB, EEGLAB, AMICA plugin.
• Computing Environment: Multi-core workstation or compute cluster (AMICA supports OpenMP/MPI for parallel processing [68]).
• EEG Data: Continuous, high-density EEG data (recommended ≥58 channels [18]).

Procedure:

• Standard Preprocessing: Perform standard EEG preprocessing steps including data import, channel location assignment, high-pass filtering (e.g., 1 Hz), and bad channel removal/interpolation [60].
• Data Referencing: Re-reference the data to the average reference [60].
• Configure AMICA Parameters: From the EEGLAB menu, select Tools > Decompose data by ICA > Run AMICA. In the AMICA settings, open the rejection sub-interface to enable sample rejection. The default parameters are a suitable starting point [18]:
  • numrej: 5 (Number of rejection iterations)
  • rejsig: 3 (Standard deviation threshold for rejection)
  • Other parameters can typically be left at their defaults.
• Execute AMICA: Run the decomposition. AMICA will iteratively compute the model and reject bad samples based on the specified criteria.
• Apply Components: Once complete, the computed ICA weights can be applied to the original, unfiltered, and uncleaned continuous data to retain maximum data for subsequent analysis [18].

Protocol 2: A Combined ASR and AMICA Pipeline

This protocol leverages ASR for aggressive initial cleaning of large artifacts, followed by AMICA for fine decomposition and further model-driven cleaning.

Procedure:

• Initial Preprocessing: Import data, assign channel locations, and apply a high-pass filter (e.g., 1 Hz).
• Apply ASR:
  • In EEGLAB, select Tools > Reject data using Clean Rawdata and ASR.
  • Enable the ASR option. A stringent threshold (e.g., 10 to 20) is recommended for initial aggressive cleaning to remove major motion artifacts and spikes [60].
  • This step will create a new, cleaner dataset.
• Prepare for ICA: Re-reference the ASR-cleaned data to the average reference.
• Run AMICA with Sample Rejection: Decompose the pre-cleaned data using AMICA, as described in Protocol 1. The integrated sample rejection will now work on a dataset that has already had the most egregious artifacts removed, potentially leading to an even more robust decomposition.
• Component Classification and Removal:
  • Use an automated classifier like ICLabel (Tools > Classify components using IClabel) to obtain probabilities for each component being brain, eye, muscle, etc. [69].
  • Based on these labels, flag and remove artifact-related components.
  • Finally, project the cleaned components back to the sensor space to reconstruct artifact-free EEG.

The logical flow of this combined pipeline is illustrated below.

Decision Framework for Tool Selection

The choice between using AMICA's rejection alone or a combined ASR-AMICA pipeline depends on the nature of the EEG data. The following decision chart provides a guided approach.

For researchers in neuroscience and drug development, automating and standardizing EEG preprocessing is critical for scalability and reproducibility. AMICA's integrated sample rejection offers a robust, model-driven method for improving ICA decomposition with minimal manual intervention, proving effective across a range of data qualities. For studies involving significant participant movement, a pipeline that combines an initial aggressive cleaning with ASR followed by AMICA decomposition provides a powerful strategy for handling severe artifacts. By adopting the protocols and decision frameworks outlined in this note, research teams can enhance the reliability of their electrophysiological biomarkers and streamline their analytical workflows.

Validating ICA Cleaning: Metrics, Comparisons, and Impact on Analysis

Independent Component Analysis (ICA) has become a foundational technique in electroencephalography (EEG) preprocessing for isolating neural activity from various artifacts, including those originating from eye movements, heart activity, and muscle noise [4] [70] [32]. A critical challenge in employing ICA is objectively evaluating the success of the artifact removal process. Effective cleaning must achieve a dual objective: the thorough suppression of artifactual signals and the minimal distortion of underlying cerebral neural activity. This Application Note details the core quantitative metrics and standardized experimental protocols that researchers can employ to rigorously measure the accuracy and efficacy of ICA-based cleaning procedures.

Quantitative Metrics for Cleaning Accuracy

The evaluation of ICA performance rests on several key metrics that quantify different aspects of decomposition quality and artifact suppression. The following table summarizes the primary quantitative measures used in the field.

Table 1: Key Quantitative Metrics for ICA Cleaning Accuracy

Metric	Description	Interpretation	Application Context
Mutual Information Reduction (MIR) [32]	Measures the reduction in mutual information between components after ICA. Reported in kilobits/sec (kbps).	Higher MIR values indicate more successful separation into independent sources.	General evaluation of ICA decomposition quality.
Dipolarity [32]	Percentage of independent components that can be modeled by a single equivalent dipole with a residual variance (RV) below a threshold (e.g., <10% RV).	A higher percentage of near-dipolar components suggests physiologically plausible brain sources.	Assessing the biological plausibility of separated components.
Change in Power Spectral Density (ΔPSD) [70]	Calculates the change in signal power within standard EEG frequency bands (delta, theta, alpha, beta) before and after artifact suppression.	Lower ΔPSD values indicate less distortion of the underlying brain rhythm activity.	Quantifying signal distortion introduced by the cleaning process.
Normalized Correlation Coefficient [4]	Measures the similarity in morphology and topography of brain signals (e.g., spikes) before and after artifact removal.	A coefficient close to 1 indicates minimal change to the neural signal of interest.	Validating the preservation of specific neurogenic signals.

These metrics can be complemented by objective, eye-tracker-based quantification of residual artifacts (undercorrection) and the removal of neurogenic activity (overcorrection) in paradigms involving free eye movements [57].

Experimental Protocols for Metric Evaluation

Protocol for Evaluating Mutual Information Reduction and Dipolarity

This protocol is suited for a general assessment of ICA decomposition quality using a 71-channel EEG setup, as an example [32].

1. Data Acquisition and Preprocessing:

• Equipment: A 64-channel or higher EEG system (e.g., Compumedics Neuroscan) is recommended.
• Parameters: Sample data at 1 kHz with a bandwidth of 0-200 Hz [70]. For the specific evaluation of data quantity, ensure a sufficiently long recording.
• Data Quantity: The amount of data required is crucial. A common heuristic is the κ value, defined as κ = number of data frames / (number of channels)². Studies suggest benefits may continue beyond κ=20, but this serves as a common reference point [32].

2. ICA Decomposition:

• Algorithm Selection: Employ a robust ICA algorithm such as Adaptive Mixture ICA (AMICA), which has been shown to outperform others in quantitative metrics [32].
• Data Submission: Run the AMICA decomposition on the preprocessed, multi-channel EEG dataset.

3. Post-ICA Calculation of Metrics:

• Mutual Information Reduction (MIR): Calculate MIR using the formula derived from the unmixing matrix W and the marginal entropies of the original channels x and the independent components y [32]: MIR = log\|detW\| + [h(x₁) + ... + h(xₙ)] - [h(y₁) + ... + h(yₙ)] Report the result in kilobits per second (kbps).
• Dipolarity: a. Source Localization: For each independent component, compute a best-fitting single equivalent dipole using a head model (e.g., a four-shell spherical Boundary Element Method (BEM) model in DIPFIT (EEGLAB)) [32]. b. Residual Variance (RV) Calculation: Calculate the RV for each component's dipole model. c. Threshold Application: Count the percentage of components with an RV of less than 10%. This percentage is the dipolarity metric.

Protocol for Evaluating ΔPSD in Artifact Suppression

This protocol quantifies the distortion introduced into the brain signal after suppressing specific artifacts, such as cardiac or ocular artifacts [70].

1. Targeted Data Selection:

• Identify segments of EEG data contaminated by the target artifact (e.g., EKG artifact in central electrodes like CP2, or EOG artifact in frontal electrodes like FP1).
• Select clean, non-overlapping reference data for EOG and EKG signals, if available.

2. Artifact Suppression via a Reference-Based Method:

• Obtain Clean Reference: Process the raw EOG/EKG reference data using Ensemble Empirical Mode Decomposition (EEMD) to extract Intrinsic Mode Functions (IMFs). Apply an unsupervised technique like Principal Component Analysis (PCA) to these IMFs to capture and reconstruct the clean artifact signal [70].
• Regression: Suppress the artifact from the contaminated EEG by correlating the measured EEG with the clean EOG/EKG reference and subtracting the scaled reference signal [70].

3. Power Spectral Density (PSD) Calculation:

• Pre- and Post-Suppression Analysis: Calculate the PSD for the selected EEG segment before and after artifact suppression. Use Welch's method for robust spectral estimation [70].
• Band-Limited ΔPSD: Compute the change in PSD (ΔPSD) for standard EEG frequency bands:
  • Delta (δ): 0.5–4 Hz
  • Theta (θ): 4–8 Hz
  • Alpha (α): 8–12 Hz
  • Beta (β): 13–30 Hz The result is four ΔPSD values (ΔPSDδ, ΔPSDθ, ΔPSDα, ΔPSDβ) for each evaluated segment. Lower values indicate less distortion [70].

The workflow for this quantitative assessment is outlined in the diagram below.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Materials and Tools for ICA-EEG Research

Item	Function / Description	Example / Note
High-Density EEG System	Records scalp electrical potentials. Essential for providing sufficient spatial information for effective ICA.	64-channel systems (e.g., Compumedics Neuroscan) are common; more channels can be beneficial [70] [32].
Electrooculography (EOG) & Electrocardiography (EKG) Hardware	Provides dedicated recordings of ocular and cardiac activity. Used as reference signals for artifact suppression or validation [70].	--
ICA Algorithms	Software implementations for performing the blind source separation.	AMICA is a benchmark algorithm [32]. Infomax ICA is also widely used and can be optimized [57].
EEGLAB	An interactive MATLAB toolbox for processing EEG data. It provides a ecosystem for ICA and related analyses [32].	Includes plugins like DIPFIT for source localization and dipole modeling [32].
Ensemble Empirical Mode Decomposition (EEMD)	A data-driven technique for signal decomposition. Used for extracting clean artifact signatures from reference EOG/EKG data [70].	--

The rigorous application of quantitative metrics—including Mutual Information Reduction, dipolarity, and changes in Power Spectral Density—is paramount for validating the success of ICA in EEG artifact removal. By adhering to the detailed experimental protocols outlined in this document, researchers and drug development professionals can objectively benchmark the performance of their ICA pipelines, ensuring both effective artifact suppression and the faithful preservation of neural signals. This quantitative approach is fundamental for producing reliable, high-quality EEG data in both basic neuroscientific research and clinical applications.

Electroencephalography (EEG) is a vital tool in neuroscience and clinical diagnostics, but its signal quality is often compromised by artifacts. As research and applications expand into real-world settings using wearable EEG, the challenge of effective artifact removal has become increasingly critical [8]. This document provides a structured comparison and benchmarking of Independent Component Analysis (ICA) against other prominent artifact removal techniques—Regression, Principal Component Analysis (PCA), and Artifact Subspace Reconstruction (ASR). The content is framed within the context of a broader thesis on ICA for EEG artifact removal, offering detailed protocols and application notes for researchers and drug development professionals.

Performance Benchmarking: Quantitative Comparison of Artifact Removal Techniques

The table below summarizes key performance metrics for various artifact removal methods, based on recent experimental findings. These metrics provide a quantitative basis for method selection.

Table 1: Performance Benchmarking of Artifact Removal Techniques

Method	Best For / Key Feature	Reported Performance Metrics	Key Strengths	Key Limitations
ICA (Independent Component Analysis)	Ocular artifact removal; General purpose artifact identification [10] [71].	Effective clearing of signals; Minimal distortion of interictal activity (e.g., spikes) [4].	Separates neural and artifactual sources; Widely used and validated; Does not require reference signals [71].	Requires multiple channels; Component selection can be subjective; Performance degrades with high-motion artifacts [8] [72].
Motion-Net (Deep Learning)	Subject-specific motion artifact removal [73].	Artifact reduction (η): 86% ±4.13; SNR improvement: 20 ±4.47 dB; MAE: 0.20 ±0.16 [73].	High accuracy for motion artifacts; Incorporates visibility graph features for stability on smaller datasets [73].	Subject-specific training required; Computationally intensive [73].
FF-EWT + GMETV (Wavelet-Based)	Single-channel EOG artifact removal [11].	Lower RRMSE and higher CC on synthetic data; Improved SAR and MAE on real data [11].	Effective for single-channel setups; Automated identification of artifact components [11].	Primarily targets EOG artifacts; Performance on other artifact types less established [11].
iCanClean	Motion artifact removal during locomotion (e.g., running) [72].	Reduced power at gait frequency; Recovery of more dipolar brain ICs; Identified P300 congruency effect [72].	Uses pseudo-reference noise signals; Effective for motion artifacts during high-mobility tasks [72].	--
ASR (Artifact Subspace Reconstruction)	Motion and instrumental artifacts; General-purpose cleaning [8] [72].	Reduced power at gait frequency; Produced ERP components similar to standing task [72].	Popular for wearable EEG; Works well with ICA by improving subsequent decomposition quality [72].	--
Hybrid ICA–Regression	Automatic ocular artifact removal [71].	Lower MSE and MAE; Higher mutual information vs. standalone ICA or Regression [71].	Automatically identifies artifactual ICs; Preserves neuronal activity better than rejecting whole ICs [71].	More complex pipeline than standalone methods [71].
DWT (Discrete Wavelet Transform)	Ocular and muscular artifacts; Feature preservation [8] [74].	MAE: 4785.08, MSE: 309,690 (for ASD EEG); Robustness in preserving signal characteristics [74].	Effective for single-channel analysis; Good balance between denoising and feature preservation [74].	--
Butterworth Filter	Basic frequency-based noise removal [74].	Moderate results across metrics (SNR, MAE, MSE) [74].	Simple and fast; Flat frequency response in passband minimizes distortion [74].	Ineffective when artifact and brain signal frequencies overlap [73].

Experimental Protocols for Key Methods

Protocol: Standard ICA for Artifact Removal

This protocol outlines the steps for using ICA to remove artifacts, such as those from eye blinks, from multi-channel EEG data [10].

Workflow Diagram: Standard ICA for Artifact Removal

Detailed Procedure:

• Data Loading and Preprocessing:

  • Load the multi-channel EEG dataset. Ensure channel locations are defined [10].
  • Perform standard preprocessing: high-pass filter (e.g., 1-2 Hz cutoff) to remove slow drifts, and optionally remove bad channels and noisy data segments to improve ICA decomposition quality [10].

• ICA Decomposition:

  • Select Tools → Decompose data by ICA in EEGLAB.
  • Choose an ICA algorithm (e.g., runica for Infomax ICA is a common default). The algorithm decomposes the EEG data into a set of independent components (ICs). The number of ICs will equal the number of EEG channels used [10].

• Component Inspection and Identification:

  • Plot component scalp maps (Plot → Component maps → In 2-D).
  • Inspect the properties of each IC to identify artifacts. Click on individual components to plot:
    • Scalp Topography: Eye blink artifacts typically show strong frontal projections [10].
    • Activity Time Course: Look for patterns characteristic of blinks or movements [10].
    • Power Spectrum: Ocular artifacts often have a smoothly decreasing spectrum [10].

• Artifact Removal and Data Reconstruction:

  • Flag the ICs identified as artifactual for removal.
  • Reconstruct the EEG signal by back-projecting the remaining (neural) components. This step subtracts the artifactual activity from the data [10].

Protocol: Hybrid ICA-Regression for Ocular Artifacts

This advanced protocol combines the strengths of ICA and regression to automatically remove ocular artifacts while maximizing the preservation of neural information [71].

Workflow Diagram: Hybrid ICA-Regression Method

Detailed Procedure:

• ICA Decomposition:

  • Perform ICA on the preprocessed multi-channel EEG data as described in the standard protocol. This yields a set of Independent Components (ICs) [71].

• Automatic Component Classification:

  • Use statistical measures to automatically classify ICs as "neuronal" or "artifactual" without manual inspection.
  • Composite Multi-Scale Entropy (CMSaE) and Kurtosis are used as classification features. Artifactual components related to eye blinks have distinct signatures in these measures compared to neural components [71].

• Processing of Artifactual ICs:

  • High-Magnitude Artifact Removal: Apply a threshold based on the Median Absolute Deviation (MAD) to the artifactual ICs. This step removes the highest-amplitude ocular artifacts (e.g., large blinks) from these components [71].
  • Regression: Further clean the artifactual ICs by applying a linear regression model to remove any residual EOG-related activity. This step aims to recover and retain the neural signals that were mixed into the artifactual ICs [71].

• Data Reconstruction:

  • Combine the unaltered neuronal ICs with the processed artifactual ICs.
  • Back-project the full set of components to reconstruct the artifact-free EEG signal. This approach minimizes the loss of neural data compared to simply discarding artifactual ICs [71].

Protocol: Motion Artifact Removal for Mobile EEG

This protocol is designed for scenarios involving substantial participant movement, such as walking or running, which introduces complex motion artifacts [73] [72].

Workflow Diagram: Motion Artifact Removal for Mobile EEG

Detailed Procedure:

• Initial Pre-cleaning:

  • Apply a robust pre-cleaning method to handle large-amplitude motion artifacts that can disrupt subsequent ICA.
  • Artifact Subspace Reconstruction (ASR): This method identifies and removes high-variance signal segments exceeding a statistical threshold from the data [72].
  • iCanClean: An alternative method that uses pseudo-reference noise signals (e.g., from accelerometers) to guide the artifact removal process. Studies suggest it can be more effective than ASR for recovering valid neural components during running [72].

• ICA Decomposition:

  • Perform ICA on the pre-cleaned data. The initial pre-cleaning step results in a higher-quality decomposition with more "dipolar" components that are likely of neural origin [72].

• Component Assessment and Removal:

  • Assess the quality of the resulting ICs. A higher number of dipolar components indicates a successful pre-cleaning step [72].
  • Identify and remove any remaining artifactual components using standard inspection or automated tools.

• Validation:

  • Validate the cleaning process by checking for a reduction in power at the gait frequency and its harmonics [72].
  • For event-related potential (ERP) studies, confirm that expected components (e.g., the P300) are recovered with the correct properties [72].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Resources for EEG Artifact Removal Research

Category	Item / Reagent	Function / Application	Example / Note
Software & Libraries	EEGLAB	Interactive MATLAB environment for EEG processing; includes ICA implementations and visualization tools.	Core platform for implementing standard ICA and related methods [10].
	Picard / FastICA Plugins	Alternative ICA algorithms available as EEGLAB plugins.	Can offer different performance characteristics (speed, accuracy) [10].
	RELICA Plugin	EEGLAB plugin for assessing reliability and stability of ICA decompositions.	Important for validating ICA results, given the stochastic nature of some algorithms [10].
Data	Standard EEG Datasets	Publicly available datasets for method development and benchmarking.	Klados et al. (2011) dataset used for hybrid method validation [71].
	Custom Mobile EEG Datasets	Data recorded during motion, synchronized with accelerometer data.	Essential for developing and validating motion artifact removal techniques [73] [72].
Algorithms & Methods	Infomax ICA (runica)	Standard ICA algorithm for decomposing EEG into independent components.	Default algorithm in many EEGLAB tutorials [10].
	ASR (Artifact Subspace Reconstruction)	Method for robustly removing high-amplitude, transient artifacts.	Used as a pre-cleaning step before ICA in mobile EEG studies [72].
	iCanClean	A deep learning-based method for removing motion artifacts.	Uses pseudo-reference signals; shown effective for running data [72].
Hardware	Mobile EEG System	Wearable EEG system with dry or semi-dry electrodes for naturalistic recording.	Enables data collection in real-world settings where motion artifacts are prevalent [73] [8].
	Accelerometer	Auxiliary sensor to measure head movement.	Provides reference signal for motion artifact removal methods like iCanClean [73] [72].

Electroencephalography (EEG) and event-related potential (ERP) research provide unparalleled insight into neural dynamics with millisecond temporal resolution. A significant challenge in this field is the presence of biological and non-biological artifacts that can obscure genuine brain signals and compromise the integrity of downstream analyses. As a cornerstone technique for artifact removal, Independent Component Analysis (ICA) has proven particularly valuable for its ability to separate and remove artifacts without discarding large portions of data. However, the application of ICA must be carefully optimized to preserve the neurophysiological features of interest, especially the rapidly evolving brain network dynamics captured by EEG microstates and the time-locked cognitive processes reflected in ERPs. This application note synthesizes current methodological standards and protocols for ICA-based artifact removal, with a specific focus on safeguarding the integrity of these critical analytical endpoints for research and clinical drug development.

Theoretical Foundations: ICA, Microstates, and ERPs

Independent Component Analysis (ICA) in EEG Preprocessing

ICA is a blind source separation algorithm that decomposes multi-channel EEG data into statistically independent components (ICs) [75]. The core assumption is that the signals recorded at the scalp are linear mixtures of underlying brain and non-brain sources. ICA mathematically reverses this mixing process, allowing the isolation of artifacts embedded within the data, such as those from ocular, cardiac, and muscular activity [10]. A key advantage of ICA is its capacity to subtract artifactual components rather than simply rejecting contaminated data segments, thereby preserving the continuity of the neural signal, which is crucial for subsequent analysis [10].

EEG Microstates and ERPs as Critical Downstream Analyses

EEG Microstates are brief periods (typically 60-120 ms) of stable scalp potential field topography that represent the rapid dynamics of large-scale neural networks [76] [77]. They are considered the "atoms of thought," with specific topographies associated with canonical brain networks like the auditory, visual, salience, and attention networks [78]. The analysis of their temporal dynamics—duration, occurrence, and time coverage—provides a powerful biomarker for understanding brain function in health and disease [76] [78].

Event-Related Potentials (ERPs) are neural responses time-locked to sensory, cognitive, or motor events. Components like the N170, P300, and N400 are well-established indices of specific cognitive processes. The accuracy of ERP analysis is heavily dependent on a high signal-to-noise ratio, which can be drastically reduced by artifacts [79].

The central challenge is that aggressive or inappropriate artifact removal can distort or eliminate the very neural signals these analyses seek to quantify. Therefore, protocols must be designed to maximize artifact removal while preserving the topographic and temporal features essential for microstate and ERP analysis.

Quantitative Impact of Artifact Removal on Downstream Analysis

The choice of artifact removal strategy has a direct and quantifiable impact on the outcomes of microstate and ERP analyses. The following tables summarize key findings from recent investigations.

Table 1: Impact of Artifact Removal on ERP Decoding Performance

Artifact Handling Method	Effect on ERP Decoding Performance	Key Findings	Study Details
ICA Correction + Artifact Rejection	No significant improvement in most cases [79]	Combining these methods did not enhance decoding accuracy for SVM- and LDA-based classifiers across multiple ERP paradigms.	Paradigms: N170, MMN, N2pc, P3b, N400, LRP, ERN [79]
ICA Correction Alone	Recommended to minimize confounds [79]	Effectively removes artifacts that could otherwise artificially inflate decoding accuracy, ensuring results reflect neural signals.	Recommended as a standard pre-processing step prior to decoding analyses [79]

Table 2: Performance of Artifact Removal Methods in Challenging Conditions

Method	Condition/Context	Performance Metrics	Implication for Microstate/ERP Analysis
Generalized Eigen Decomposition (GED)	Amulatory EEG (walking, jogging); Low SNR (0.1-5) [80]	Correlation: 0.93; RMSE: 1.43 μV; Increased brain components by ~11 [80]	Enabled extraction of canonical microstates during motion; observed task-related modulation (e.g., increased Microstate A duration) [80]
ICA	Amulatory EEG [80]	Inferior to GED in very low SNR regimes [80]	Performance may be insufficient for mobile neuroimaging without additional steps.
Fixed Frequency EWT + GMETV Filter	Single-channel EEG; EOG artifacts [11]	Lower RRMSE, higher CC (synthetic data); Improved SAR, MAE (real data) [11]	Effective for portable systems; preserves low-frequency EEG information critical for microstates.

Experimental Protocols for ICA in Microstate and ERP Research

Protocol 1: Standardized ICA for EEG Microstate Preservation

This protocol is optimized for resting-state EEG studies where the goal is to perform microstate analysis.

Workflow Overview:

Step-by-Step Methodology:

• Data Collection & Import: Collect high-density resting-state EEG (at least 64 channels recommended for superior spatial resolution) [77]. Ensure consistent recording parameters across subjects and sites. Import data into a standardized analysis environment like EEGLAB.
• Pre-ICA Preprocessing: This is a critical step for ensuring the success of ICA.
  • High-Pass Filter: Apply a filter at 1-2 Hz to remove slow drifts.
  • Bad Channel Removal: Identify and interpolate consistently noisy or flat-line channels.
  • Data Integrity: While some artifact rejection can be performed post-ICA, the data used to train the ICA model should be as clean as possible to prevent artifacts from influencing the component separation [10].
• ICA Decomposition: Run ICA (e.g., using the runica algorithm in EEGLAB) on the continuous data. Use the default extended option to detect both super- and sub-gaussian sources. Provide ICA with as much clean data as possible for stable training [10].
• Component Classification & Labeling: Visually inspect and label ICs using established criteria.
  • Plot 2-D Scalp Maps: Use Plot > Component maps > In 2-D [10].
  • Inspect Component Properties: Use Tools > Inspect/label components by maps to evaluate the power spectrum, time course, and ERP image for each component [10].
  • Artifact Identification: Identify and flag components corresponding to blinks (smooth frontal topography, slow time course), eye movements (lateralized frontal topography), muscle noise (high-frequency, chaotic time course), and channel noise (focal, single-channel topography) [10].
• Artifactual Component Removal: Subtract the identified artifactual components from the data. This step creates a cleaned dataset with the artifact signals removed while preserving the neural data from the remaining components.
• Microstate Analysis: Proceed with the cleaned data using a standardized pipeline (e.g., MICROSTATELAB toolbox) [77].
  • Clustering: Identify individual microstate maps via topographic clustering (e.g., 4-7 classes).
  • Back-Fitting: Competitively fit the group-level template maps to the individual's preprocessed and ICA-cleaned EEG to obtain the temporal dynamics (duration, occurrence, coverage) of each microstate class.

Protocol 2: ICA for ERP Decoding and Analysis

This protocol is tailored for task-based paradigms where the goal is to analyze or decode ERPs.

Workflow Overview:

Step-by-Step Methodology:

• Data Collection & Epoching: Collect task-based EEG data with accurate event markers. Import the data and segment it into epochs time-locked to the event of interest (e.g., -200 ms to 800 ms).
• Pre-ICA Preprocessing: Apply a high-pass filter (e.g., 1 Hz) to the continuous or epoched data. ICA is typically trained on the continuous or concatenated epoched data to ensure sufficient data quantity for a stable decomposition.
• ICA Decomposition: Perform ICA decomposition on the preprocessed data. For datasets with a high number of channels, consider using the PCA option to reduce dimensionality if the amount of data is insufficient for the number of channels [10].
• Component Rejection: Following the same visual inspection procedure outlined in Protocol 1, identify and remove components representing blinks, eye movements, and muscle activity. Critically, as noted in Table 1, subsequent trial rejection may not be necessary after ICA correction for many decoding analyses, helping to preserve statistical power [79].
• Epoch Validation & Statistical Analysis: After ICA cleaning, baseline-correct the epochs and perform a final check for any residual artifacts. Then, proceed with the standard ERP analysis pipeline, including averaging across trials and quantifying component amplitudes and latencies, or use the single-trial data for multivariate pattern analysis (MVPA).

The Scientist's Toolkit: Essential Research Reagents and Solutions

Table 3: Key Software Tools and Analytical Resources

Tool/Resource	Function	Relevance to ICA, Microstates & ERPs
EEGLAB	MATLAB-based GUI environment for EEG processing [10]	Provides the standard framework for running ICA, with plugins for component classification and validation.
MICROSTATELAB	EEGLAB plugin for resting-state microstate analysis [77]	Offers a standardized pipeline for clustering, visualizing, and quantifying microstate dynamics on ICA-cleaned data.
Cartool	Software for functional mapping and microstate analysis [76]	An alternative software suite for microstate analysis, used in numerous clinical and cognitive studies.
RELICA Plugin	EEGLAB plugin for assessing ICA reliability [10]	Allows researchers to test the stability of their ICA decomposition through bootstrapping, crucial for ensuring reproducible results.
ICLabel Plugin	EEGLAB plugin for automated component classification [10]	Uses a trained classifier to automatically label ICs as brain, muscle, eye, heart, etc., providing an objective check for manual labeling.

The rigorous application of ICA is fundamental to modern EEG research, particularly when the analytical pathway leads to microstate or ERP quantification. The protocols outlined herein emphasize that successful analysis is not merely about removing artifacts but about doing so in a way that is optimized for the specific downstream application. For microstate analysis, this means preserving the topographic integrity and temporal dynamics of global brain states. For ERP research, it means enhancing the signal-to-noise ratio without sacrificing the trial count needed for robust statistical power. As EEG continues to expand into real-world, mobile, and clinical settings, the development and validation of artifact removal methods like GED that can handle extreme noise will be essential. By adhering to standardized, careful preprocessing protocols, researchers can ensure that their findings reflect genuine neural phenomena, thereby accelerating discovery in basic neuroscience and drug development.

Independent Component Analysis (ICA) has become a cornerstone technique in EEG preprocessing for separating neural signals from artifacts. However, its application is not universally successful, and understanding its failure modes is critical for obtaining valid neurophysiological data. This application note details specific scenarios where ICA performance degrades, provides protocols for assessing reliability, and offers tools to mitigate inaccuracies.

Core Principles and inherent Assumptions of ICA

ICA is a blind source separation technique that decomposes multi-channel EEG data into underlying components based on the principle of statistical independence [81]. The core model is represented as X = AS, where X is the measured data, A is the mixing matrix, and S contains the independent source signals [82] [1] [83]. The goal is to find an unmixing matrix W to recover the sources: S = WX [81] [83].

This model relies on several critical assumptions, the violation of which leads to unreliable outcomes [81]:

• Statistical Independence: The underlying sources must be statistically independent, meaning the value of one source provides no information about the value of another [82] [81].
• Non-Gaussianity: The source signals must have non-Gaussian (non-normal) distributions. The separation relies on maximizing non-Gaussianity, as Gaussian distributions are fully symmetric and contain no information about the directions of the columns of the mixing matrix [82] [81].
• Linear and Instantaneous Mixing: The signals are assumed to mix linearly at the electrodes, and propagation delays are considered negligible [81].

When ICA Fails: Key Limiting Scenarios and Experimental Evidence

Under specific but common experimental conditions, the assumptions of ICA are violated, leading to biased or inaccurate results. The table below summarizes the primary scenarios, their mechanisms, and consequences.

Table 1: Scenarios Leading to Unreliable ICA Performance

Scenario	Underlying Mechanism	Impact on ICA & EEG Data	Supporting Evidence
Low Trial-to-Trial Variability of Artifacts	Artifacts (e.g., TMS-pulse artifacts) are highly stereotyped, creating dependencies between components and breaking the independence assumption [82].	Biased cleaning that removes genuine brain signals; unreliable TMS-evoked potential (TEP) cleaning [82].	Systematic analysis showing ICA becomes unreliable when artifact variability is small [82].
High-Amplitude Motion Artifacts	Gross head movement during locomotion produces large, non-linear noise that overwhelms neural signals and reduces ICA decomposition quality [14].	Poor separation of brain and non-brain components; reduced number of dipolar brain components identified [14].	Comparison study during running showed motion artifacts degrade ICA; methods like iCanClean and ASR are preferred [14].
Imperfect Component Separation	Standard ICA subtraction can imperfectly separate neural and artifactual signals within a component [13].	Removal of neural signals alongside artifacts; artificial inflation of ERP/connectivity effect sizes; biased source localization [13].	Novel targeted cleaning methods protect against these false positives [13].
High Measurement Uncertainty (Low SNR)	Analog-to-digital converter (ADC) noise and other non-biological noise sources introduce Gaussian noise, violating the non-Gaussianity assumption [83].	Degradation in artifact identification performance; for eyeblink removal, SNR <15 dB leads to >5% performance drop [83].	Simulation studies with characterized hardware uncertainty (e.g., ADS1299 ADC) show performance degradation with lowering SNR [83].

The diagram below illustrates the logical pathway from an underlying data problem to its ultimate consequence on research findings.

Experimental Protocols for Quantifying ICA Reliability

To ensure the validity of ICA-based findings, researchers should implement experimental checks. The following protocol provides a methodology for assessing ICA cleaning accuracy, particularly in challenging contexts like TMS-EEG.

Protocol 1: Assessing ICA Cleaning Accuracy Using Simulated Artifacts

This protocol is adapted from studies investigating ICA reliability in the presence of artifacts with low trial-to-trial variability [82].

1. Objective: To quantitatively evaluate the success of ICA in separating and removing artifact components without distorting the underlying neural signal.

2. Materials and Reagents: Table 2: Research Reagent Solutions for ICA Reliability Testing

Item	Function/Description	Example
Clean, Artifact-Free EEG Data	Serves as the ground-truth neural signal for benchmarking.	Pre-recorded resting-state EEG from a controlled, stationary session [82].
Artefact Simulation Algorithm	Generates simulated artifacts with controllable properties (amplitude, waveform, variability).	Custom MATLAB/Python scripts to create TMS-like pulse artifacts or eyeblink waveforms [82] [83].
ICA Processing Software	The ICA implementation under test.	EEGLAB with runica (Infomax) [10], FastICA [83], or a commercial software's ICA module.
Signal Quality Metrics	Quantifies the difference between cleaned and ground-truth data.	Root Mean Square Error (RMSE), correlation coefficient, amplitude difference in key ERP components (e.g., P300).

3. Procedure:

• Data Preparation: Select a segment of high-quality, clean EEG data with a known event-related potential (ERP) structure. This dataset will be the ground-truth reference.
• Artifact Simulation: Generate a set of simulated artifacts (e.g., mimicking TMS pulses or eyeblinks) and add them to the clean EEG data. Systematically vary the trial-to-trial variability of the artifact, from completely deterministic (no variability) to highly stochastic.
• ICA Decomposition and Cleaning: Apply the chosen ICA algorithm to the contaminated dataset. Identify and remove the component(s) that correspond to the simulated artifact.
• Signal Reconstruction: Reconstruct the EEG signal from the remaining components.
• Accuracy Measurement: Compare the ICA-cleaned data to the original ground-truth data. Calculate accuracy metrics such as:
  • Waveform Similarity: The correlation between the cleaned data and the ground-truth data.
  • Amplitude Recovery: The difference in amplitude of key ERP features (e.g., N100, P300) between cleaned and ground-truth data [82].
• Variability Assessment (Optional but Recommended): To estimate reliability without ground truth, measure the trial-to-trial variability of the identified artifact component itself. Low variability in the component's waveform suggests a higher risk of inaccurate cleaning [82].

4. Data Interpretation:

• A high correlation and low amplitude difference indicate successful artifact removal.
• A systematic change in accuracy with decreasing artifact variability demonstrates a key limitation of ICA.
• This protocol provides a framework for researchers to validate their specific ICA pipeline before applying it to experimental data.

The Scientist's Toolkit: Mitigating ICA Limitations

When standard ICA is deemed unreliable, several advanced strategies and tools can be employed.

Table 3: Solutions and Tools for Overcoming ICA Limitations

Tool / Method	Primary Function	Application Context
RELAX Pipeline	Implements targeted artifact reduction, cleaning only artifact-dominated periods or frequencies within a component, thus preserving neural data [13].	Ideal for protecting against false positive effects in ERP and connectivity studies [13].
iCanClean	Uses canonical correlation analysis (CCA) with reference noise signals to detect and subtract motion artifact subspaces from EEG [14].	Superior for motion artifact removal during locomotion (walking, running); can use dedicated noise sensors or create pseudo-references from EEG [14].
Artifact Subspace Reconstruction (ASR)	A preprocessing method that uses sliding-window PCA to remove high-amplitude, non-stationary artifacts before ICA is run [14].	Improving ICA decomposition quality in data with large, sporadic artifacts (e.g., movement, electrode pops). A k parameter of 10-20 is recommended for locomotion data [14].
RELICA Plugin	Assesses the reliability and stability of ICA decompositions using bootstrapping and cluster analysis [10].	Determining if the extracted components are consistent and trustworthy, especially for high-density EEG systems.

The integration of these tools into a robust preprocessing workflow is key to managing ICA's limitations, as shown below.

ICA is a powerful but imperfect tool. Its reliability is compromised when core assumptions are violated, notably by stereotyped artifacts, high-amplitude motion, and significant measurement noise. Researchers must critically evaluate the conditions of their experiment and their data's properties before relying on standard ICA. By employing validation protocols like simulated artifact testing and integrating next-generation tools such as RELAX, iCanClean, and ASR, scientists can mitigate these limitations, safeguard their findings against false positives and biased conclusions, and advance the rigor of EEG research.

Electroencephalography (EEG) provides non-invasive, high-temporal-resolution measurement of brain activity, indispensable for cognitive psychology research and clinical diagnostics. A paramount concern in EEG analysis is the contamination of neural signals by extra-cerebral artifacts, among which ocular artifacts (OA) generated by eye blinks and movements are the most pervasive and severe. These artifacts introduce high-amplitude, low-frequency potentials that can obscure or mimic neural activity of interest, fundamentally compromising data integrity and interpretation. This case study examines the application of Independent Component Analysis (ICA) as a premier method for OA removal, framing its methodology, efficacy, and protocols within the broader thesis of advancing EEG artifact removal research.

The Problem of Ocular Artifacts in EEG

Ocular artifacts present a unique challenge for several reasons. First, their amplitude can be an order of magnitude larger than that of cortical potentials, effectively swamping neural signals. Second, the spectral content of OAs, primarily below 4 Hz, significantly overlaps with that of key neural signals like event-related potentials and delta waves, making simple frequency-based filtering ineffective. Third, eye movements and blinks are physiological activities that occur frequently and involuntarily—approximately 20 times per minute for blinks—making their complete avoidance during recordings impossible.

The traditional approach of rejecting contaminated EEG epochs results in a substantial and often unacceptable loss of data. This is particularly problematic in experimental paradigms with limited trials or clinical settings where data is precious. Regression-based methods using electrooculographic (EOG) channels, while historically used, are flawed because the reference EOG signals are themselves contaminated with EEG activity, leading to the removal of neural signals along with the artifact.

Independent Component Analysis (ICA): Principles and Applicability

ICA is a blind source separation (BSS) technique that addresses the artifact removal problem by decomposing the multi-channel EEG recording into a set of temporally independent and spatially fixed components. The core assumption is that the signals recorded at the scalp are linear mixtures of underlying cerebral and non-cerebral sources. ICA solves the linear equation X = AS, where X is the observed data, A is the mixing matrix, and S contains the independent sources. The goal is to find an unmixing matrix W that yields U = WX, where U contains the independent components (ICs).

For ICA to be effective, the data must meet certain assumptions: the underlying sources must be statistically independent, their mixture must be linear, and the propagation delays from sources to sensors must be negligible. These conditions are generally considered reasonable for EEG data.

The power of ICA for OA removal lies in its ability to isolate OAs into a small number of components based on their statistical properties, without requiring a reference EOG channel. This is a significant advantage over regression methods. The artifact-corrected signal is then reconstructed by projecting the remaining components back onto the scalp sensors.

Quantitative Evidence for ICA Efficacy in Ocular Artifact Removal

Multiple studies have quantitatively demonstrated the success of ICA in removing ocular artifacts while preserving neural data.

Table 1: Quantitative Outcomes of ICA for Artifact Removal

Study	Artifact Type	Key Metric	Result
Iriarte et al. (2003) [4]	EKG, Eye Movements, Muscle, Electrode	Normalized Correlation Coefficient	Minimal change in signal morphology; distortion of interictal activity was minimal
Viola et al. (2011) [84]	General Artifacts (incl. Ocular)	Mean Squared Error (MSE)	Performance on level with inter-expert disagreement (<10% MSE)
Improved wICA (2019) [85]	EOG	Accuracy in Time/Spectral Domain	Outperformed other component rejection and wavelet-based methods

A seminal study by Iriarte et al. (2003) applied ICA to 80 EEG samples with evident artifacts. The signal was reconstructed after excluding artifactual components, and a normalized correlation coefficient was used to measure changes. The study found that ICA produced a clear "clearing-up" of signals with the morphology and topography of spikes remaining very similar before and after artifact removal, and the rest of the signal did not change significantly [4].

Furthermore, an optimized automatic classifier for artifactual ICA components achieved a Mean Squared Error (MSE) of less than 10%, a performance on par with the level of disagreement between human experts [84]. This underscores that ICA, especially when augmented with automated classification, can achieve highly reliable artifact removal.

Experimental Protocols for ICA-Based Ocular Artifact Removal

A robust, semi-automatic protocol is essential for effective and reproducible OA removal. The following workflow integrates key steps from current best practices [29] [86].

Protocol Workflow

Detailed Protocol Steps

Step 1: Critical Data Preprocessing for ICA ICA decomposition is highly sensitive to data quality. Proper preprocessing is not optional but mandatory for success.

• High-Pass Filtering: A cutoff of 1-2 Hz is crucial to remove slow drifts that ICA cannot handle well. This step improves the decomposition quality, though it may sacrifice some neural information below 1 Hz [29] [19]. Some protocols suggest computing ICA weights on a high-pass filtered (e.g., 1 Hz) version of the data and then applying the weights to a less stringently filtered version (e.g., 0.1 Hz) to preserve low-frequency content.
• Bad Channel Handling: Identify and interpolate malfunctioning or noisy channels before running ICA.
• Artifact Rejection: Mark non-stereotypical, high-amplitude artifacts (e.g., muscle bursts, electrode pops) as "Bad Intervals." Crucially, do not mark blink intervals, as these are needed for identifying ocular components [86].
• Data Stationarity: ICA assumes data stationarity. If the dataset contains large, transient artifacts, it is advisable to select a clean, stationary segment (which still contains blinks and eye movements) to compute the ICA decomposition. The resulting weights can then be applied to the entire dataset [29].

Step 2: ICA Decomposition

• Algorithm Selection: Common algorithms include Infomax (runica), FastICA, and Adaptive Mixture ICA (AMICA). AMICA is currently one of the most powerful options and includes an integrated sample rejection function to iteratively reject bad samples during decomposition, improving robustness [18].
• Data Requirements: ICA requires a sufficient amount of data. A rule of thumb is to use at least N² × 20 data points (samples), where N is the number of channels [86]. The data must include a representative number of blinks and eye movements for the algorithm to isolate them effectively.

Step 3: Identification and Removal of Ocular Components This is the most critical step, requiring a combination of automated tools and expert verification.

• Automated Pre-Selection: Tools can pre-select components correlated with OAs.
  • Correlation with EOG: Calculate the correlation between component time courses and EOG channel activity.
  • Relative VEOG/HEOG Variance: Calculate each component's contribution to the variance in EOG channels.
  • Global Field Power (GFP): During blink intervals, calculate each component's contribution to the GFP of all channels [86].
• Manual Inspection and Verification: Automated selection must always be verified. Ocular components are characterized by:
  • Topography (Scalp Map): A strong, frontal projection typical of eye artifacts [10] [3].
  • Time Course: Large, low-frequency deflections for eye movements and sharp, high-amplitude peaks for blinks [86].
  • Power Spectrum: A smoothly decreasing (low-frequency dominated) spectrum [10].
• Component Removal and Reconstruction: After selecting ocular components, they are removed, and the data is reconstructed from the remaining components using the mixing matrix: clean_data = Winv(:,good_components) * activations(good_components,:) [3].

Advanced Techniques and Hybrid Methodologies

While standard ICA with component rejection is effective, a significant limitation is the potential loss of neural information present in the artifactual component. To overcome this, hybrid methods have been developed.

Wavelet-Enhanced ICA (wICA) is a notable advancement. Instead of rejecting entire components, it applies a wavelet transform to the artifact-laden independent components. This allows for the selective correction of only the data points contaminated by the EOG artifact within the component, leaving the rest of the component's neural information intact. Studies have shown this method outperforms both full component rejection and earlier wavelet-based methods in preserving signal integrity in both the time and frequency domains [85].

Table 2: The Researcher's Toolkit for ICA-based Ocular Correction

Tool/Reagent	Function/Role in OA Removal
High-Density EEG System	Provides sufficient spatial sampling for ICA to resolve independent sources effectively.
Infomax/FastICA/AMICA Algorithm	Core ICA engine for decomposing EEG data into independent components. AMICA is noted for its robustness.
Automated Component Classifier (e.g., ICLabel)	Machine-learning tool for preliminary, objective labeling of components as brain, ocular, muscle, etc.
Wavelet Toolbox	For implementing advanced hybrid correction methods (e.g., wICA) that selectively remove artifacts within components.
Semi-Automatic Tool (e.g., BrainVision Analyzer ICA)	Software providing a guided, criterion-based workflow for identifying and removing ocular components.

The following diagram illustrates the logic of choosing between standard and advanced ICA correction methods.

Limitations and Critical Considerations

Despite its power, ICA-based OA removal has limitations that researchers must acknowledge.

• Data Dependence and Stationarity: ICA performance depends on the amount and quality of input data. When artifacts have low trial-to-trial variability, it can create dependencies between components, leading ICA to inaccurately separate sources and potentially remove brain-derived activity alongside the artifact [20].
• Subjectivity in Component Selection: Although automated tools are improving, the final decision on which components to reject often requires expert judgment, introducing a degree of subjectivity. Studies show that the component selection criteria must be carefully applied, as over-rejection can remove neural signals, while under-rejection leaves artifacts [84].
• Algorithmic Uncertainty: Running the same ICA algorithm twice on the same data can produce slightly different decompositions due to random initialization. This inherent uncertainty should be considered when interpreting results [10].

The removal of ocular artifacts is a critical step in EEG preprocessing, without which the validity of neural signatures is questionable. ICA has established itself as a powerful, versatile, and theoretically sound method for this task, capable of isolating and removing ocular artifacts without the need for reference channels or excessive data loss. While standard ICA with component rejection is effective, advanced hybrid methods like wICA promise even greater fidelity by preserving neural information. Successful implementation hinges on rigorous data preprocessing, a well-defined protocol that combines automated and manual component selection, and a clear understanding of the method's assumptions and limitations. As EEG research continues to expand into mobile and clinical settings, robust and reliable ICA-based ocular artifact removal will remain a cornerstone of rigorous electrophysiological data analysis.

Conclusion

Independent Component Analysis stands as a powerful and robust method for EEG artifact removal, essential for ensuring the validity of neural signatures in both basic research and clinical trials. When implemented with careful preprocessing and optimization—particularly the crucial removal of ocular artifacts—ICA reliably preserves brain-related physiological signals and enhances the statistical power of subsequent analyses. Future directions should focus on the development of fully automated, standardized pipelines to improve reproducibility across labs, enhanced algorithms for real-time applications in brain-computer interfaces and neurofeedback, and continued refinement for complex, naturalistic experimental designs. For the biomedical research community, mastering ICA is not merely a technical skill but a fundamental requirement for deriving clean, interpretable, and clinically meaningful insights from electrophysiological data.

References

• 1. Removing Artifacts From EEG Data Using Independent ... - TMSi [https://www.tmsi.artinis.com/blog/removing-eeg-artifacts-with-independent-component-analysis-ica]
• 2. - ML - GeeksforGeeks Independent Component Analysis [https://www.geeksforgeeks.org/machine-learning/ml-independent-component-analysis/]
• 3. EEG artifact removal with Blind Source Separation [https://sccn.ucsd.edu/~jung/Site/EEG_artifact_removal.html]
• 4. Independent component analysis as a tool to eliminate ... [https://pubmed.ncbi.nlm.nih.gov/14530738/]
• 5. Quantifying Data Requirements for EEG Independent ... [https://arxiv.org/abs/2506.10156]
• 6. Experiment protocols for brain-body imaging of locomotion [https://pmc.ncbi.nlm.nih.gov/articles/PMC10014824/]
• 7. Independent Component Analysis (ICA) With Covariates ... [https://pubmed.ncbi.nlm.nih.gov/40698201/]
• 8. A Systematic Review of Techniques for Artifact Detection ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12473706/]
• 9. EEG is better when cleaning effectively targets artifacts [https://pubmed.ncbi.nlm.nih.gov/41092567/]
• 10. Independent Component Analysis for artifact removal [https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html]
• 11. Single-channel EOG artifact removal using fixed frequency ... [https://www.nature.com/articles/s41598-025-10276-8]
• 12. Connectivity Analysis in EEG Data: A Tutorial Review of the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10044923/]
• 13. EEG is better when cleaning effectively targets artifacts [https://www.sciencedirect.com/science/article/abs/pii/S1388245725012301]
• 14. A Comparison of Approaches for Motion Artifact Removal from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12349273/]
• 15. Comparison of comprehensive quantitative EEG metrics ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10659091/]
• 16. Calibration-free Ocular artifact reduction in EEG signals ... [https://www.sciencedirect.com/science/article/pii/S1746809425006585]
• 17. Independent Component Analysis (ICA) With Covariates ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12282553/]
• 18. Optimizing EEG ICA decomposition with data cleaning in ... [https://www.nature.com/articles/s41598-024-64919-3]
• 19. mne.preprocessing.ICA — MNE 1.11.0 documentation [https://mne.tools/stable/generated/mne.preprocessing.ICA.html]
• 20. Measuring the accuracy of ICA-based artifact removal from ... [https://www.sciencedirect.com/science/article/pii/S1935861X23019629]
• 21. A method for independent component graph analysis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5346515/]
• 22. How many channels are suitable for independent ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809419300175]
• 23. Automatic Classification of Artifactual ICA-Components for Artifact Removal in EEG Signals | Behavioral and Brain Functions [https://link.springer.com/article/10.1186/1744-9081-7-30]
• 24. Ranking and averaging independent component analysis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6870671/]
• 25. Independent component analysis [https://en.wikipedia.org/wiki/Independent_component_analysis]
• 26. CW_ICA: an efficient dimensionality determination method ... [https://www.nature.com/articles/s41598-023-49355-z]
• 27. Makoto's preprocessing pipeline - SCCN [https://eeglab.ucsd.edu/wiki/Makoto's_preprocessing_pipeline]
• 28. A Comparative Study of Different EEG Reference Choices ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6798171/]
• 29. Protocol for semi-automatic EEG preprocessing incorporating ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11930125/]
• 30. Application of Referencing Techniques in EEG-Based ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2020.559969/full]
• 31. ICA for dummies - Arnaud Delorme [https://arnauddelorme.com/ica_for_dummies/]
• 32. Quantifying Data Requirements for EEG Independent ... [https://arxiv.org/html/2506.10156v1]
• 33. AMICA [https://eeglab.org/plugins/amica/]
• 34. Cleaning artifacts using ICA [https://www.fieldtriptoolbox.org/tutorial/preproc/ica_artifact_cleaning/]
• 35. Repairing artifacts with ICA — MNE 1.11.0 documentation [https://mne.tools/stable/auto_tutorials/preprocessing/40_artifact_correction_ica.html]
• 36. Artifacts [https://www.learningeeg.com/artifacts]
• 37. EEG Artifacts: Types, Detection, and Removal Techniques ... [https://www.bitbrain.com/blog/eeg-artifacts]
• 38. Toward the Understanding of Topographical and Spectral ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7199478/]
• 39. Detection of EEG Spatial–Spectral–Temporal Signatures ... [https://link.springer.com/article/10.1007/s10548-014-0397-z]
• 40. Eye-movement artifact correction in infant EEG [https://www.sciencedirect.com/science/article/abs/pii/S0165027025000469]
• 41. How does Independent Component Analysis ... [https://pubmed.ncbi.nlm.nih.gov/39904902/]
• 42. Augmenting brain-computer interfaces with ART: An artifact ... [https://www.sciencedirect.com/science/article/pii/S1053811925001259]
• 43. Simultaneous EEG-fMRI: What Have We Learned and What ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8952790/]
• 44. TMS combined with EEG: Recommendations and open ... [https://www.sciencedirect.com/science/article/pii/S1935861X23016960]
• 45. On the feasibility of concurrent human TMS-EEG-fMRI ... [https://pubmed.ncbi.nlm.nih.gov/23221407/]
• 46. On the feasibility of concurrent human TMS-EEG-fMRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3569123/]
• 47. Concurrent human TMS-EEG-fMRI enables monitoring of ... [https://www.nature.com/articles/s42003-020-0764-0]
• 48. Adaptive optimal basis set for BCG artifact removal in ... [https://www.nature.com/articles/s41598-018-27187-6]
• 49. Simultaneous EEG-fMRI: What Have We Learned and ... [https://www.mdpi.com/1424-8220/22/6/2262]
• 50. Mining EEG-fMRI using independent component analysis - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC2693483/]
• 51. Independent Component Analysis (ICA) [https://www.brainvoyager.com/bvqx/doc/UsersGuide/ICA/IndependentComponentAnalysis.html]
• 52. The impact of artifact removal approaches on TMS–EEG ... [https://www.sciencedirect.com/science/article/pii/S1053811921005486]
• 53. Getting to know EEG artifacts and how to handle them in ... [https://pressrelease.brainproducts.com/eeg-artifacts-handling-in-analyzer/]
• 54. Removal of muscular artifacts in EEG signals - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC5893498/]
• 55. On the influence of high-pass filtering on ICA-based artifact ... [https://pubmed.ncbi.nlm.nih.gov/26737196/]
• 56. a. Filtering [https://eeglab.org/tutorials/05_Preprocess/Filtering.html]
• 57. Optimizing the ICA-based removal of ocular EEG artifacts ... [https://www.sciencedirect.com/science/article/pii/S1053811919307086]
• 58. Evaluating the effectiveness of artifact correction and rejection ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10516012/]
• 59. Cleaning and processing resting-state EEG [https://www.fieldtriptoolbox.org/workshop/madrid2019/tutorial_cleaning/]
• 60. A quick tutorial on ICA artifact rejection [https://eeglab.org/tutorials/misc/Quick_Tutorial_on_Rejection.html]
• 61. Mobile brain/body imaging of landmark‐based navigation with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9291975/]
• 62. High-density EEG mobile brain/body imaging data ... [https://www.nature.com/articles/s41597-019-0223-2]
• 63. EPIC: Annotated epileptic EEG independent components for artifact reduction | Scientific Data [https://www.nature.com/articles/s41597-022-01524-x]
• 64. A system for automatic artifact removal in ictal scalp EEG based on independent component analysis and Bayesian classification - PubMed [https://pubmed.ncbi.nlm.nih.gov/16458594/]
• 65. Hybrid ICA—Regression: Automatic Identification and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4853904/]
• 66. Autoreject: Automated artifact rejection for MEG and EEG data [https://pmc.ncbi.nlm.nih.gov/articles/PMC7243972/]
• 67. How does Independent Component Analysis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11794336/]
• 68. Amica [https://sccn.ucsd.edu/~jason/amica_web.html]
• 69. Repairing artifacts with ICA automatically using ICLabel ... [https://mne.tools/mne-icalabel/0.4/generated/examples/00_iclabel_automatic_artifact_correction_ica.html]
• 70. Common Methodology for Cardiac and Ocular Artifact ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5840220/]
• 71. Hybrid ICA—Regression: Automatic Identification and ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2016.00193/full]
• 72. A Comparison of Approaches for Motion Artifact Removal ... [https://ui.adsabs.harvard.edu/abs/2025Senso..25.4810L/abstract]
• 73. Motion Artifact Removal from EEG Signals Using the ... [https://www.sciencedirect.com/science/article/pii/S1746809425013886]
• 74. A comprehensive EEG dataset and performance ... [https://www.nature.com/articles/s41598-025-18934-7]
• 75. Independent Component Analysis with Functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10111109/]
• 76. Current State of EEG/ERP Microstate Research - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10884048/]
• 77. The EEGLAB Toolbox for Resting-State Microstate Analysis [https://link.springer.com/article/10.1007/s10548-023-01003-5]
• 78. EEG Microstates Signatures of rTMS Response Over the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12464126/]
• 79. and LDA-based decoding of EEG signals [https://www.sciencedirect.com/science/article/pii/S1053811925003076]
• 80. Contrast-based artifact removal enables microstate ... [https://pubmed.ncbi.nlm.nih.gov/41196780/]
• 81. Independent Component Analysis (ICA) – demystified [https://pressrelease.brainproducts.com/independent-component-analysis-demystified/]
• 82. Measuring the accuracy of ICA-based artifact removal from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11636693/]
• 83. Simulated Eyeblink Artifact Removal with ICA [https://arxiv.org/html/2410.03261v1]
• 84. Automatic Classification of Artifactual ICA-Components for ... [https://behavioralandbrainfunctions.biomedcentral.com/articles/10.1186/1744-9081-7-30]
• 85. Improved EOG Artifact Removal Using Wavelet Enhanced ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6956025/]
• 86. Ocular Correction ICA [https://www.brainproducts.com/support-resources/ocular-correction-ica/]