ASR vs. iCanClean: A Performance Evaluation for Motion Artifact Removal in Mobile EEG

Matthew Cox Dec 02, 2025 438

Motion artifacts present a significant challenge for electroencephalography (EEG) in mobile and real-world settings, such as clinical trials and neuromonitoring.

ASR vs. iCanClean: A Performance Evaluation for Motion Artifact Removal in Mobile EEG

Abstract

Motion artifacts present a significant challenge for electroencephalography (EEG) in mobile and real-world settings, such as clinical trials and neuromonitoring. This article provides a comprehensive performance evaluation of two prominent motion artifact removal methods: Artifact Subspace Reconstruction (ASR) and iCanClean. We explore their foundational principles, methodological applications, and optimal parameter configurations based on the latest research. A direct comparison evaluates their efficacy in improving Independent Component Analysis (ICA) decomposition, preserving event-related potentials (ERPs), and enhancing data quality during dynamic activities like walking and running. This analysis offers researchers and drug development professionals evidence-based guidance for selecting and implementing robust EEG cleaning pipelines to ensure high-fidelity brain data in biomedical applications.

Understanding the Battle Against Noise: Foundations of Motion Artifacts in Mobile EEG

The Critical Challenge of Motion Artifacts in Mobile Brain Imaging

Electroencephalography (EEG) is the only brain imaging method lightweight enough with sufficient temporal precision to assess electrocortical dynamics during human locomotion, making it indispensable for studying the brain in naturalistic settings [1]. However, the very mobility that enables real-world brain imaging also introduces a critical challenge: motion artifacts. These artifacts originate from multiple sources including head motion, electrode displacement, cable sway, and muscle contractions, collectively contaminating the EEG signal and reducing the quality of independent component analysis (ICA) decompositions essential for source separation [1] [2]. During whole-body movements like running, motion produces broadband spectral power particularly at step frequency and its harmonics, which can obscure neural signals of interest [1]. The pursuit of effective motion artifact removal has become a central focus in mobile brain-body imaging (MoBI) research, with Artifact Subspace Reconstruction (ASR) and iCanClean emerging as two prominent preprocessing approaches. This comparison guide objectively evaluates their performance against other emerging methodologies, providing researchers with experimental data and implementation protocols to inform their motion artifact correction strategies.

Methodological Approaches: Core Algorithms and Technical Principles

Artifact Subspace Reconstruction (ASR)

ASR employs a sliding-window principal component analysis (PCA) approach to identify and remove high-amplitude artifacts in continuous EEG [1]. The algorithm first calculates the root mean square (RMS) of sliding 1-second data segments from continuous EEG, then uses a condensed Gaussian distribution to convert RMS values into z-scores [1]. Data segments with z-scores falling within -3.5 to 5.0 of the Gaussian distribution for at least 92.5% of electrodes are selected as reference calibration data, assumed to be artifact-free [1]. During processing, ASR performs PCA on non-reference data and identifies principal components as artifactual if their standard deviation of RMS exceeds a user-defined threshold ("k"), then reconstructs the time series based on the calibration data [1]. The "k" parameter critically determines cleaning aggressiveness, with lower values (e.g., 10-20) producing more extensive data manipulation [1]. Recent improvements to ASR algorithms include ASR-DBSCAN and ASR-GEV, which address limitations in calibration data identification during intense motor tasks like juggling [3].

iCanClean Algorithm

iCanClean utilizes canonical correlation analysis (CCA) with reference noise signals to detect and correct artifact subspaces in EEG data [1] [4]. The algorithm operates by identifying subspaces of scalp EEG that correlate with subspaces of noise based on a user-selected criterion (R²) of correlation between corrupt EEG and noise signals [1]. When dual-layer EEG sensors are available, iCanClean leverages mechanically coupled noise electrodes that only capture motion artifacts without brain signals [1] [4]. For standard EEG systems, iCanClean can generate "pseudo-reference" noise signals by applying temporary notch filters to identify noise within specific frequency bands (e.g., below 3 Hz) [1]. After noise components are projected back onto EEG channels using a least-squares solution, those exceeding the R² threshold are subtracted [1]. Key parameters requiring optimization include the R² cleaning aggressiveness threshold (typically 0.65) and the sliding time window for CCA (typically 4 seconds) [4].

Emerging Alternative Approaches

Beyond ASR and iCanClean, several promising motion artifact removal approaches have recently emerged:

Motion-Net represents a deep learning approach using a CNN-based U-Net architecture for subject-specific motion artifact removal [2]. This framework incorporates visibility graph (VG) features that provide structural information to improve performance with smaller datasets, achieving an average motion artifact reduction of 86% ±4.13 and SNR improvement of 20 ±4.47 dB [2]. Unlike batch-processing methods, Motion-Net is trained and tested separately per subject using real EEG recordings with ground-truth references [2].

IMU-Enhanced LaBraM utilizes fine-tuned large brain models with inertial measurement unit (IMU) reference signals to identify motion-related artifacts [5]. This transformer-based neural architecture processes EEG signals alongside 9-axis IMU data (accelerometer, gyroscope, magnetometer) using an attention mechanism to map spatial channel relationships and artifact contributions [5]. The model contains approximately 9.2 million parameters and leverages 5.9 hours of EEG-IMU recordings for training [5].

EEG-Cleanse provides a modular, fully automated preprocessing pipeline specifically designed for cleaning EEG during full-body movement without specialized hardware [6]. Implemented in Python and MATLAB using EEGLAB toolbox, it integrates motion-adaptive preprocessing methods with a hybrid strategy for labeling artifacts [6].

The table below summarizes the core methodological characteristics of these approaches:

Table 1: Fundamental Characteristics of Motion Artifact Removal Methods

Method Core Algorithm Reference Signals Implementation Key Parameters
ASR Principal Component Analysis None (uses clean calibration data) Online/Offline k threshold (10-30) [1]
iCanClean Canonical Correlation Analysis Dual-layer electrodes or pseudo-references Online/Offline R² threshold (0.65), window length (4s) [4]
Motion-Net CNN U-Net Architecture None (subject-specific training) Offline Training epochs, VG features [2]
IMU-LaBraM Transformer Network IMU sensors (9-axis) Online/Offline Attention mapping, feature alignment [5]

Experimental Comparisons: Performance Metrics and Benchmarking

Phantom Head Validation Studies

Controlled studies using electrical phantom heads with embedded brain source antennae provide ground-truth validation for artifact removal algorithms. In one comprehensive comparison evaluating multiple artifact types simultaneously (motion, muscle, eye, line-noise), iCanClean consistently outperformed other methods with a Data Quality Score improvement from 15.7% (before cleaning) to 55.9% after processing in the "Brain + All Artifacts" condition [7]. By comparison, ASR, Auto-CCA, and Adaptive Filtering only improved scores to 27.6%, 27.2%, and 32.9% respectively [7]. For context, uncontaminated "Brain only" condition scored 57.2% without cleaning, establishing a reasonable performance target [7].

Human Locomotion Studies

During overground running tasks incorporating Flanker paradigm assessments, both ASR and iCanClean with pseudo-reference noise signals led to recovery of more dipolar brain independent components [1] [8]. In these analyses, iCanClean demonstrated somewhat greater effectiveness than ASR [1]. Both methods significantly reduced power at gait frequency and harmonics, and produced ERP components similar in latency to those identified in stationary Flanker tasks [1]. However, only iCanClean successfully identified the expected greater P300 amplitude to incongruent flankers during dynamic jogging [1] [8].

For walking tasks across different populations (young adults, high-functioning older adults, low-functioning older adults), iCanClean with optimal parameters (4-second window, R²=0.65) improved the average number of "good" ICA components (dipolar, high brain probability) from 8.4 to 13.2 (+57%) [4]. Performance remained robust even with reduced noise channels, maintaining 12.7, 12.2, and 12.0 good components for 64, 32, and 16 noise channels respectively [4].

Real-World Motor Task Performance

During intensive whole-body motor tasks like three-ball juggling, next-generation ASR algorithms (ASR-DBSCAN and ASR-GEV) demonstrated superior performance compared to original ASR [3]. The improved methods found 42% and 24% of data usable for calibration respectively, compared to only 9% with ASR-original [3]. Subsequent ICA showed that data preprocessed with ASR-DBSCAN and ASR-GEV produced brain independent components accounting for more variance of the original data (30% and 29%) compared to ASR-original (26%) [3].

The table below summarizes quantitative performance metrics across experimental conditions:

Table 2: Performance Comparison Across Experimental Conditions

Method Experimental Condition Performance Metrics Comparative Results
iCanClean Phantom Head (All Artifacts) Data Quality Score: 55.9% [7] Outperformed ASR (27.6%), Auto-CCA (27.2%), Adaptive Filtering (32.9%) [7]
iCanClean Human Running (Flanker Task) P300 Congruency Effect Identified expected effect [1]
ASR Human Running (Flanker Task) P300 Congruency Effect Failed to identify effect [1]
iCanClean Multi-Population Walking Good ICA Components: 13.2 [4] 57% improvement over no cleaning (8.4 components) [4]
Motion-Net Real Motion Artifacts Artifact Reduction: 86% ±4.13 [2] SNR improvement of 20 ±4.47 dB [2]
ASR-DBSCAN Three-Ball Juggling Usable Calibration Data: 42% [3] Outperformed ASR-original (9%) [3]

Implementation Workflows and Parameter Optimization

iCanClean Implementation Protocol

Successful iCanClean implementation requires careful parameter selection and processing sequence. The optimal window length and cleaning aggressiveness established for human locomotion data are 4 seconds and R² = 0.65 respectively [4]. The algorithm can be effectively implemented with dual-layer electrodes (mechanically coupled noise sensors) or with pseudo-reference signals derived from temporarily notch-filtered EEG when dedicated noise sensors are unavailable [1]. The processing workflow begins with standard preprocessing including high-pass filtering (1 Hz cutoff) and average re-referencing, followed by channel rejection to remove excessively noisy channels [4]. iCanClean processing then proceeds with canonical correlation analysis comparing cortical electrode signals (brain + noise) with noise electrode signals (noise only) across sliding windows, removing components exceeding the R² correlation threshold [4].

G iCanClean Signal Processing Workflow RawEEG Raw EEG Data Preprocessing Preprocessing: High-pass filter (1Hz) Average re-reference RawEEG->Preprocessing ChannelRejection Channel Rejection: Remove outlier channels (>3x median amplitude) Preprocessing->ChannelRejection NoiseReference Noise Reference Creation: Dual-layer electrodes or pseudo-reference signals ChannelRejection->NoiseReference CCA Canonical Correlation Analysis: Sliding window (4s) Identify noise subspaces NoiseReference->CCA Threshold Component Subtraction: Remove components exceeding R² threshold (0.65) CCA->Threshold CleanedEEG Cleaned EEG Data Threshold->CleanedEEG ICA ICA Decomposition: Improved component quality and dipolarity CleanedEEG->ICA

ASR Implementation Protocol

Effective ASR implementation requires appropriate selection of the "k" threshold parameter and calibration data. For human locomotion data, a k threshold between 10-30 is recommended, balancing artifact removal against potential "over-cleaning" that inadvertently manipulates neural signals [1]. The calibration data can be supplied explicitly by recording clean resting-state data, or automatically extracted from contaminated data by identifying segments with RMS z-scores within -3.5 to 5.0 for at least 92.5% of electrodes [1]. Recent improvements through ASR-DBSCAN and ASR-GEV utilize point-by-point amplitude evaluation to eliminate collateral rejection of clean data, significantly increasing the amount of usable calibration data during intense motor tasks [3].

G ASR Signal Processing Workflow RawEEG Raw EEG Data Calibration Calibration Data Selection: Clean segments or automatic extraction RawEEG->Calibration RMS RMS Calculation: 1-second sliding windows Z-score conversion Calibration->RMS PCA1 PCA on Calibration Data: Establish reference component distribution RMS->PCA1 PCA2 PCA on Test Data: Sliding window analysis Identify outlier components PCA1->PCA2 Threshold Artifact Removal: Reconstruct data using k threshold (10-30) PCA2->Threshold CleanedEEG Cleaned EEG Data Threshold->CleanedEEG ICA ICA Decomposition: Improved source separation and localization CleanedEEG->ICA

Research Toolkit: Essential Materials and Experimental Setup

Table 3: Research Reagent Solutions for Mobile EEG Motion Artifact Studies

Tool Category Specific Examples Function/Purpose Implementation Notes
EEG Systems BrainAmp DC with actiCAP MOVE, 32-120+ channels [9] Mobile EEG acquisition with active electrodes High-density (120+ channels) preferred for source localization [4]
Noise Reference Sensors Dual-layer EEG electrodes [4], Head-mounted IMU (APDM Opal) [5] Provide reference noise signals for artifact subtraction Dual-layer: 120 scalp + 120 noise electrodes optimal [4]
Motion Platforms Robotic motion platforms [7], Treadmills with variable terrain [4] Standardized motion artifact induction Enable controlled validation studies [7]
Phantom Heads Electrically conductive head models with embedded sources [7] Ground-truth validation of artifact removal Contain 10+ brain sources and contaminating sources [7]
Software Toolboxes EEGLAB [1], BCILAB [7], Custom MATLAB/Python scripts Implementation of processing algorithms EEGLAB includes ASR by default [1]

The empirical evidence demonstrates that both iCanClean and ASR provide effective preprocessing for motion artifact removal during human locomotion, with iCanClean showing superior performance in preserving neural signals during running and complex motor tasks [1] [7]. iCanClean's use of reference noise signals enables targeted subtraction of artifact subspaces without excessive compromise of neural data, particularly valuable for event-related potential studies where cognitive components like the P300 must be preserved [1] [8]. ASR offers the advantage of not requiring additional hardware when clean calibration data is available, with improved versions (ASR-DBSCAN, ASR-GEV) showing enhanced performance during intense motor tasks [3].

Emerging approaches incorporating deep learning (Motion-Net) and multi-modal sensor fusion (IMU-LaBraM) represent promising research directions, potentially offering improved adaptability to individual subject characteristics and motion patterns [2] [5]. For researchers selecting motion artifact removal strategies, consideration should be given to experimental paradigm, available hardware, target neural signals, and computational resources. iCanClean is particularly recommended for studies requiring preservation of event-related potentials during high-motion conditions, while ASR variants provide robust performance for general locomotion studies without requiring specialized electrode configurations. As mobile brain imaging continues to advance toward more naturalistic and ecologically valid scenarios, ongoing refinement of these artifact removal approaches will remain critical for extracting clean neural signals from motion-contaminated EEG data.

Motion artifacts represent a significant challenge in electroencephalography (EEG), particularly as research expands into mobile brain-body imaging and real-world applications. These artifacts compromise signal quality, reduce the effectiveness of source separation algorithms like independent component analysis (ICA), and can obscure genuine neural signals. Understanding the nature and origin of these contaminants is fundamental to selecting appropriate artifact removal strategies. Within the context of performance evaluation between Artifact Subspace Reconstruction (ASR) and the iCanClean algorithm, this guide provides an objective comparison of their efficacy, supported by experimental data and detailed methodologies.

Motion artifacts in EEG are mechanical/non-physiological signals that originate from multiple sources related to physical movement. Unlike biological artifacts from muscle or eyes, motion artifacts are primarily external and can severely corrupt the EEG time series, especially during whole-body movement [1].

The primary mechanism involves physical displacement of EEG electrodes relative to the scalp. As cables sway through the air during movement, they interact with each other and background electromagnetic fields through inductive coupling and electromagnetic radiation [10] [7]. This is particularly problematic given the small voltage of electrocortical signals (approximately 20 µV). Furthermore, head motion during activities like running produces artifacts that are typically time-locked to the gait cycle, manifesting as increased spectral power at the step frequency and its harmonics [1] [8].

The table below summarizes the core characteristics of motion artifacts and other common contaminants.

Table 1: Types and Sources of Common EEG Artifacts

Artifact Type Origin Primary Characteristics Common Scenarios
Motion Artifact [1] [10] Electrode displacement, cable sway, head movement Time-locked to movement cycle (e.g., gait); broadband spectral power Running, walking, sports
Muscle Artifact (EMG) [10] Contractions of facial, neck, or jaw muscles High-frequency, non-stationary Talking, chewing, grimacing
Ocular Artifact [10] Eye blinks and saccades High-amplitude, low-frequency, frontal topography Any task involving visual activity
Line-Noise Artifact [10] Electrical interference from power lines 50/60 Hz steady-state oscillation All recordings, lab environments

Experimental Comparison of Removal Methodologies

Core Algorithmic Workflows

The fundamental difference between ASR and iCanClean lies in their approach to identifying noise. ASR is a blind source separation method that relies on statistical outliers in the data itself, while iCanClean uses a reference-based approach to directly identify and subtract noise subspaces.

G cluster_asr Artifact Subspace Reconstruction (ASR) cluster_ican iCanClean Algorithm ASR1 1. Calibration Phase: Identify clean EEG segments as reference ASR2 2. PCA Decomposition: Sliding-window PCA on incoming data ASR1->ASR2 ASR3 3. Artifact Identification: Flag components where RMS exceeds threshold (k) ASR2->ASR3 ASR4 4. Reconstruction: Reconstruct corrupted segments using reference data ASR3->ASR4 iCan1 1. Noise Signal Acquisition: Use dual-layer or create pseudo-reference signals iCan2 2. Canonical Correlation Analysis (CCA): Identify subspaces of EEG correlated with noise iCan1->iCan2 iCan3 3. Threshold Application: Remove components exceeding correlation criterion (R²) iCan2->iCan3 iCan4 4. Signal Reconstruction: Subtract noise components via least-squares solution iCan3->iCan4

Diagram: Core Workflows of ASR and iCanClean Algorithms

Key Experimental Protocols

To objectively evaluate the performance of ASR and iCanClean, researchers have employed rigorous experimental designs, including phantom head validation and human subject studies during locomotion.

Phantom Head Experiments

These studies utilize an electrically conductive phantom head with embedded brain source antennae to provide known ground-truth brain signals [10] [7]. The setup includes 10 simulated brain sources and 10 contaminating sources for artifacts. The protocol involves testing under multiple conditions: Brain only, Brain + Eyes, Brain + Neck Muscles, Brain + Facial Muscles, Brain + Walking Motion, and Brain + All Artifacts. The key metric is the Data Quality Score (0-100%), calculated from the average correlation between the known brain sources and the cleaned EEG channels [10].

Human Locomotion Studies

Human studies typically involve collecting high-density EEG during activities like standing, walking, and running [1] [4]. A common paradigm is an adapted Flanker task, where participants respond to congruent or incongruent stimuli while either standing still or jogging. This allows for the evaluation of artifact removal on the recovery of stimulus-locked event-related potential (ERP) components, such as the P300. Key evaluation metrics include:

  • ICA Component Dipolarity: The number of independent components well-localized as dipoles (residual variance < 15%) and with high brain probability [1] [4].
  • Spectral Power Reduction: Reduction in power at the gait frequency and its harmonics [1].
  • ERP Congruency Effect: The ability to capture expected neural responses, like the greater P300 amplitude to incongruent Flanker stimuli [1].

Quantitative Performance Comparison

The following tables summarize key experimental findings from direct comparisons between ASR and iCanClean.

Table 2: Performance Comparison in Phantom Head Study (Data Quality Score %) [10]

Condition Uncleaned ASR iCanClean Auto-CCA Adaptive Filtering
Brain + All Artifacts 15.7% 27.6% 55.9% 27.2% 32.9%
Brain + Walking Motion Data Not Shown Data Not Shown >55.9% Data Not Shown Data Not Shown
Brain (Clean Target) 57.2% - - - -

Table 3: Performance in Human Running Study [1] [8]

Evaluation Metric ASR Performance iCanClean Performance
Increase in Dipolar Brain Components Significant Improvement Somewhat More Effective than ASR
Power Reduction at Gait Frequency Significant Reduction Significant Reduction
Recovery of P300 ERP Component Similar Latency to Standing Task Similar Latency to Standing Task
Detection of P300 Congruency Effect Not Identified Successfully Identified

Table 4: iCanClean Parameter Optimization for Mobile EEG [4]

Parameter Function Recommended Value Impact of Variation
R² Threshold Cleaning aggressiveness 0.65 Lower R² = more aggressive cleaning
Window Length Temporal locality of CCA 4 seconds Shorter windows adapt faster to changing noise
Noise Channels Reference signal sources 16-64 channels More channels improve performance

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 5: Essential Materials for Mobile EEG Motion Artifact Research

Item Function/Application Example Use Case
Dual-Layer EEG Cap [4] Outer layer provides noise-only reference signals mechanically coupled to scalp electrodes. Provides optimal reference signals for iCanClean in human locomotion studies.
Electrical Phantom Head [10] [7] Provides ground-truth brain signals for validating artifact removal algorithms. Comparing algorithm performance against known sources in controlled artifact conditions.
High-Density EEG System (64+ channels) [4] Enables better source separation via ICA and provides more channels for noise reference. Essential for mobile brain imaging studies during whole-body movement.
Inertial Measurement Unit (IMU) [5] Measures acceleration and angular velocity to directly quantify motion. Used as a reference signal for adaptive filtering or modern deep learning approaches.
Robotic Motion Platform Induces controlled, repeatable motion for testing artifact removal methods. Isolating motion artifact effects in phantom head experiments [11].

The field of motion artifact removal is evolving toward multi-modal sensor fusion and advanced computational models. Research is exploring the integration of Inertial Measurement Units (IMUs) with deep learning. One approach involves fine-tuning large brain models (LaBraM) using attention mapping to correlate IMU signals with EEG motion artifacts, showing improved robustness over traditional ASR-ICA pipelines [5]. Furthermore, while iCanClean is highly effective with dual-layer electrodes, recent studies confirm it also performs well using pseudo-reference noise signals created from the raw EEG itself, making it more accessible for standard EEG systems [1]. These advancements point toward a future of more integrated, robust, and computationally efficient cleaning frameworks for real-world mobile brain imaging.

This guide provides an objective comparison of Artifact Subspace Reconstruction (ASR) and iCanClean, two prominent algorithms for cleaning motion artifacts from mobile electroencephalography (EEG) data. We detail their core principles, workflows, and performance based on current experimental research to aid in methodological selection.

ASR and iCanClean are both real-time capable processing techniques designed to handle artifacts in mobile EEG, but they are founded on different principles and input requirements.

The table below summarizes the fundamental characteristics of each method.

Feature Artifact Subspace Reconstruction (ASR) iCanClean
Core Principle Identifies and removes high-variance, artifactual components using Principal Component Analysis (PCA) [1] [12]. Uses Canonical Correlation Analysis (CCA) to identify and subtract noise subspaces correlated with reference signals [1] [7].
Primary Input Requires a segment of clean EEG data for calibration [1] [12]. Requires reference noise signals (e.g., from dual-layer EEG caps or generated as pseudo-references) [1] [4].
Core Operation Reconstructs corrupted data segments by replacing artifactual PCA components with data from the calibration period [12]. Subtracts noise components that are highly correlated with the reference signals beyond a set threshold (R²) [1] [4].
Key Parameter k: Standard deviation threshold for identifying artifact components; lower values lead to more aggressive cleaning [1]. R²: Correlation threshold that determines cleaning aggressiveness; lower values lead to more aggressive cleaning [1] [4].

Detailed Workflows

The Artifact Subspace Reconstruction (ASR) Workflow

ASR operates in two main phases: a calibration phase to learn the "clean" EEG space, and a processing phase to identify and correct artifacts in new data [12].

ASR_Workflow start Start ASR Workflow calib Calibration Phase start->calib clean_data Clean Calibration EEG Data calib->clean_data preprocess Preprocess Data (Detrend, Filter) clean_data->preprocess pca_calib Perform PCA (Compute Mixing Matrix Mr) preprocess->pca_calib calc_thresh Calculate Rejection Threshold (Γ) pca_calib->calc_thresh processing Processing Phase calc_thresh->processing new_data New Data Segment (Xt) processing->new_data pca_new Perform PCA on Xt new_data->pca_new compare Compare Components to Threshold (Γ) pca_new->compare reconstruct Reconstruct Cleaned Signal compare->reconstruct output Cleaned EEG Output reconstruct->output

Calibration Phase: This initial phase establishes a baseline of clean brain activity [12].

  • Input Clean Data: A segment of artifact-free EEG data is provided as a reference.
  • Preprocess: The calibration data is detrended (mean subtracted) and filtered to remove specific neural rhythms like alpha, preserving the data's covariance structure [12].
  • Compute PCA Space: Principal Component Analysis (PCA) is performed on this clean data to compute a mixing matrix (Mr) that defines the principal component space of clean EEG [12].
  • Set Threshold: A rejection threshold (Γ) is calculated based on the root-mean-square (RMS) of the principal components. This threshold is scaled by a user-defined parameter k (typically 5-30), where a lower k value sets a more aggressive, lower threshold [1] [12].

Processing Phase: This phase is applied to the continuous EEG data stream.

  • Process New Data: A short, sliding window of new EEG data is taken [12].
  • Project into PCA Space: The data window is projected into the PCA space defined during calibration.
  • Identify and Remove Artifacts: The principal components of the new data are compared to the rejection threshold (Γ). Any component with variance exceeding the threshold is identified as an artifact and removed (truncated) [12].
  • Reconstruct Signal: The retained "clean" components are projected back to the channel space using the calibration's mixing matrix (Mr), reconstructing the data with artifacts removed [12].

The iCanClean Workflow

iCanClean leverages reference noise signals to isolate and remove artifacts that are mixed with the brain signals in the scalp EEG.

iCanClean_Workflow start Start iCanClean Workflow input Input: Scalp EEG & Reference Noise Signals start->input segment Segment Data into Sliding Time Windows input->segment cca Perform Canonical Correlation Analysis (CCA) segment->cca measure_r2 Measure Correlation (R²) of CCA Components cca->measure_r2 apply_thresh Apply User R² Threshold measure_r2->apply_thresh subtract Subtract Noise Components from Scalp EEG apply_thresh->subtract output Cleaned EEG Output subtract->output

  • Input Signals: The algorithm requires two inputs: the scalp EEG signals (containing a mixture of brain activity and noise) and reference noise signals [1] [7].
  • Segment Data: The data is processed using a sliding time window. A 4-second window is often optimal [4].
  • Canonical Correlation Analysis (CCA): CCA is applied to find linear subspaces within the scalp EEG that are maximally correlated with subspaces in the reference noise signals [1] [7].
  • Apply Threshold: The correlation (R²) of these CCA components is measured. The user defines an R² threshold (e.g., 0.65); components with correlation exceeding this threshold are classified as noise [1] [4].
  • Subtract Noise: The identified noise components are projected back to the channel space and subtracted from the original scalp EEG, resulting in a cleaned signal [1].

Experimental Protocols and Performance Evaluation

Key Experimental Protocols

Recent studies have established rigorous protocols to evaluate ASR and iCanClean, often using mobile EEG recorded during whole-body movement.

Common Evaluation Paradigm:

  • Task: EEG is recorded during dynamic activities like overground running or treadmill walking, often while participants perform a cognitive task (e.g., a Flanker task) to elicit event-related potentials (ERPs) [1] [8].
  • Comparison Baseline: Performance is compared against data collected during static conditions (e.g., standing or seated) where motion artifacts are minimal [1].
  • Hardware: Studies frequently use high-density EEG systems (64+ channels). Evaluations of iCanClean often employ dual-layer EEG caps, where an outer layer of electrodes records only environmental and motion noise, providing ideal reference signals [1] [4].

Primary Performance Metrics:

  • ICA Dipolarity: The number of brain-like independent components (ICs) after decomposition that are well-explained by a single dipole (Residual Variance < 15%). This measures how well the cleaning process facilitates source separation [1] [4].
  • Spectral Power at Gait Frequency: The reduction in power at the step frequency and its harmonics after cleaning. Effective methods significantly reduce this motion-related power without compromising neural signals [1].
  • ERP Fidelity: The ability to recover expected neural components, such as the P300 ERP, during a cognitive task performed while moving. The amplitude and latency of the cleaned ERPs are compared to those from the static condition [1] [8].
  • Data Quality Score: In phantom head studies with known ground-truth brain signals, the correlation between the cleaned EEG and the original brain sources provides a direct measure of quality [7].

Comparative Performance Data

The following tables summarize quantitative results from recent, direct comparative studies.

Table 1: Performance in Phantom and Human Locomotion Studies

Evaluation Metric Artifact Subspace Reconstruction (ASR) iCanClean Experimental Context
Data Quality Score (Phantom) Improved from 15.7% to 27.6% [7] Improved from 15.7% to 55.9% [7] Phantom head with all artifacts (motion, muscle, eye, line-noise) present [7].
Good Brain ICs (Human) ~12-13 components [1] ~13 components [1] Mobile EEG during running; "good" ICs are dipolar and brain-like. iCanClean showed a 57% increase from baseline (8.4 to 13.2 components) [4].
P300 Recovery (Human) Produced ERP components similar to standing task [1] Produced ERP components similar to standing task; identified the expected greater P300 amplitude to incongruent flankers [1] Flanker task performed during jogging vs. standing [1] [8].
Power at Gait Frequency Significantly reduced [1] Significantly reduced [1] Mobile EEG during running [1].

Table 2: Optimal Parameters and Methodological Notes from Experimental Studies

Aspect Artifact Subspace Reconstruction (ASR) iCanClean
Recommended Parameters k = 20-30 for general use [1]; k = 10 for locomotion to avoid over-cleaning [1] Window length = 4s, R² = 0.65 [1] [4]
Strengths Does not require separate noise sensors; effective for large-amplitude artifacts [1] [12]. Highly effective at removing multiple artifact types while preserving brain activity; can use pseudo-reference signals if dedicated noise sensors are unavailable [1] [7].
Limitations / Notes Performance is dependent on the quality of the calibration data. Overly aggressive cleaning (low k) can remove brain activity [1]. Performance is optimal with true reference signals from a dual-layer setup, but still effective with pseudo-references [1] [4].

The Scientist's Toolkit: Essential Research Reagents and Materials

The following tools are critical for implementing and testing these artifact removal methods in experimental research.

Tool / Material Function in Research Example Use Case
Dual-Layer EEG System Provides mechanically coupled noise electrodes that record only environmental and motion artifacts, serving as ideal reference signals for iCanClean [1] [4]. Essential for obtaining optimal iCanClean performance in human locomotion studies [4].
Electrical Phantom Head Simulates brain electrical activity and artifacts (motion, muscle, ocular) in a controlled setting, providing ground-truth signals for algorithm validation [7]. Used for initial validation and parameter sweeps without the variability of human subjects [7].
Robotic Motion Platform Induces precise, repeatable motion artifacts in a phantom head, allowing for standardized testing of artifact removal algorithms [11]. Enables systematic evaluation of an algorithm's sensitivity to specific types of motion [11].
High-Density EEG Caps (64+ channels) Provides the high spatial resolution necessary for effective source separation techniques like ICA, which is used to evaluate the quality of cleaning [1] [4]. Standard in mobile brain imaging studies to ensure enough signal channels remain after processing and rejection [4].
IMU (Inertial Measurement Unit) Records accelerometer and gyroscope data synchronized with EEG, providing an alternative source of motion reference signals for artifact removal algorithms [13]. Can be used as an additional reference signal for motion artifact cancellation, though methods for integration are still being refined [13].

Understanding the Motion Artifact Challenge in Mobile EEG

Electroencephalography (EEG) is the only brain imaging method lightweight enough with sufficient temporal precision to assess electrocortical dynamics during human locomotion [1]. However, the promise of mobile brain imaging is hindered by a significant challenge: motion artifacts. When participants move their heads during whole-body movements like walking or running, these motions produce artifacts that contaminate the EEG signal and substantially reduce the quality of subsequent independent component analysis (ICA) decompositions [1].

Motion artifacts originate from multiple sources, including cable sway as EEG cables move through the air and interact with each other and background electromagnetic fields [10] [7]. Additional artifacts come from electrode displacement and mechanical stress on the recording system [1]. These artifacts are particularly problematic because they are typically time-locked to the gait cycle, creating rhythmic patterns that can mask genuine neural activity and hinder the identification of true brain sources [1]. The presence of large motion artifacts contaminates ICA's ability to identify maximally independent sources, necessitating effective preprocessing methods specifically designed to handle motion-related noise [1].

The iCanClean Algorithm: Core Principles and Mechanism

iCanClean is a novel generalized framework for removing EEG artifacts that leverages reference noise signals and canonical correlation analysis (CCA) to detect and correct noise-based subspaces in EEG data [1] [10]. The algorithm operates by comparing cortical electrode signals (which record mixtures of brain activity plus noise) with reference noise signals (which record only noise), enabling the identification and subtraction of noisy subspaces without removing underlying brain signals [4].

The core innovation of iCanClean lies in its use of canonical correlation analysis to identify subspaces of the scalp EEG that are correlated with subspaces of noise [1]. This process is based on a user-selected criterion of the correlation between scalp EEG subspaces and noise subspaces (R²) within specific sliding time windows [1]. After noise components are projected back onto the EEG channels using a least-squares solution, those components that exceed the R² threshold are systematically subtracted from the scalp EEG [1].

iCanClean can be implemented in two primary configurations:

  • Dual-layer electrode systems where specialized noise sensors are mechanically coupled to traditional EEG electrodes but only capture noise as they are not in contact with the scalp [1] [4]
  • Pseudo-reference systems where noise signals are created from the raw EEG by temporarily applying a user-selected notch filter to identify noise within the EEG [1]

G Start Start: Contaminated EEG Signal NoiseRef Obtain Reference Noise Signals Start->NoiseRef CCA Canonical Correlation Analysis (CCA) NoiseRef->CCA Threshold Apply R² Correlation Threshold CCA->Threshold Subtract Subtract Noise Components Threshold->Subtract End End: Cleaned EEG Signal Subtract->End

Figure 1: iCanClean Algorithm Workflow. The process begins with contaminated EEG signals, obtains reference noise signals, performs canonical correlation analysis, applies a correlation threshold, and subtracts identified noise components to produce cleaned EEG data.

Experimental Protocols and Performance Evaluation

Phantom Head Validation Studies

Researchers have rigorously evaluated iCanClean's performance using electrically conductive phantom heads with embedded brain source antennae, enabling comparison with known ground-truth brain signals [10] [7]. In one comprehensive study, investigators tested iCanClean under six different conditions: Brain only, Brain + Eyes, Brain + Neck Muscles, Brain + Facial Muscles, Brain + Walking Motion, and Brain + All Artifacts [10]. The performance was quantified using a Data Quality Score (0-100%), based on the average correlation between known brain sources and EEG channels [10].

The most striking results emerged from the "Brain + All Artifacts" condition, where iCanClean improved the Data Quality Score from 15.7% before cleaning to 55.9% after cleaning [10]. For context, the Brain-only condition scored 57.2% without cleaning, indicating that iCanClean restored data quality to near-pristine levels despite multiple simultaneous artifacts [10].

Human Locomotion Studies

In human studies during dynamic activities like running, researchers have employed adapted Flanker tasks to evaluate how different artifact removal approaches affect stimulus-locked event-related potential (ERP) components [1]. EEG is recorded during both dynamic (jogging) and static (standing) versions of the task, enabling direct comparison of ERP components known to appear in stationary conditions [1].

Performance metrics in these studies include:

  • ICA component dipolarity: The number of well-localized independent components with residual variance <15% and high brain probability [1] [4]
  • Spectral power reduction: Reduction in power at the gait frequency and its harmonics [1]
  • ERP component recovery: The ability to capture expected neural components like the P300 congruency effect [1]

G DataCollection Data Collection: • Dual-layer EEG during locomotion • Flanker task (static & dynamic) • Phantom head with known sources Preprocessing Preprocessing: • High-pass filter (1Hz) • Channel rejection • Average re-referencing DataCollection->Preprocessing Cleaning Artifact Removal: • Parameter sweep (R², window length) • Apply iCanClean/alternative methods Preprocessing->Cleaning Analysis Performance Analysis: • ICA decomposition • Dipole fitting (RV <15%) • ICLabel classification • Spectral power analysis • ERP component recovery Cleaning->Analysis Comparison Method Comparison: • Data Quality Score • Good components count • Statistical testing Analysis->Comparison

Figure 2: Experimental Evaluation Protocol. The comprehensive methodology for evaluating iCanClean performance includes data collection, preprocessing, artifact removal with parameter optimization, multiple analysis dimensions, and comparative statistical testing.

Performance Comparison with Alternative Methods

Quantitative Performance Metrics

iCanClean demonstrates superior performance across multiple metrics compared to other artifact removal methods.

Table 1: Performance Comparison Across Artifact Types (Phantom Head Data)

Method Brain + All Artifacts Brain + Walking Motion Brain + Muscle Brain + Eyes
iCanClean 55.9% 61.2% 63.5% 66.8%
ASR 27.6% 34.1% 38.3% 42.7%
Auto-CCA 27.2% 29.8% 35.1% 41.2%
Adaptive Filtering 32.9% 36.5% 40.2% 45.3%
No Cleaning 15.7% 18.9% 22.4% 25.6%

Data Quality Scores (0-100%) represent the average correlation between known brain sources and EEG channels after cleaning. Higher scores indicate better preservation of brain signals while removing artifacts [10] [7].

Table 2: Performance in Human Locomotion Studies

Method Good ICA Components Power Reduction at Gait Frequency P300 Recovery Computational Efficiency
iCanClean 13.2 (+57%) Significant reduction Expected P300 amplitude to incongruent flankers identified Real-time capable
ASR 10.3 (+23%) Significant reduction Similar latency to standing task Real-time capable
ICA Alone 8.4 (baseline) Minimal reduction Limited recovery Computationally intensive (5+ hours)
Adaptive Filtering 9.1 (+8%) Moderate reduction Partial recovery Real-time capable

Performance metrics for different artifact removal methods applied to human EEG data during running. Good ICA components are defined as those with residual variance <15% and ICLabel brain probability >50% [1] [4].

Key Performance Advantages

The experimental data reveals several distinct advantages for iCanClean:

  • Superior Artifact Removal: iCanClean consistently outperforms other methods regardless of the type or number of artifacts present [10]. When all artifacts were simultaneously present, iCanClean achieved a Data Quality Score improvement of 40.2 percentage points, compared to 11.9-17.2 percentage points for competing methods [10].

  • Enhanced ICA Decomposition: In human locomotion data, iCanClean preprocessing increased the average number of good ICA components from 8.4 to 13.2 (+57%), significantly improving source-level analysis capabilities [4].

  • Effective Neural Component Recovery: Preprocessing with iCanClean enabled identification of the expected P300 amplitude differences to incongruent flankers during running, demonstrating its ability to preserve cognitively relevant neural signals while removing motion artifacts [1].

  • Robust Performance with Reduced Sensors: Good performance could be maintained with reduced sets of noise channels (12.7, 12.2, and 12.0 good components for 64, 32, and 16 noise channels, respectively), enhancing practical implementation flexibility [4].

Implementation Considerations and Parameter Optimization

Optimal Parameter Settings

Through comprehensive parameter sweeps, researchers have identified optimal settings for iCanClean implementation. For mobile EEG data collection during walking, the optimal window length and cleaning aggressiveness were determined to be 4 seconds and R² = 0.65 [4]. These settings successfully balance aggressive artifact removal with preservation of underlying brain signals.

The R² threshold is particularly critical as it determines cleaning aggressiveness, with higher values (near 1) corresponding to less cleaning [4]. The 4-second window length appears optimal for capturing the temporal structure of gait-related artifacts while maintaining stationarity within analysis windows [1] [4].

Research Reagent Solutions

Table 3: Essential Research Materials for iCanClean Implementation

Research Tool Function Implementation Considerations
Dual-Layer EEG Cap Provides mechanical coupling between scalp electrodes and outward-facing noise electrodes 120+120 electrode configuration ideal; requires specialized hardware [4]
Pseudo-Reference System Creates noise references from existing EEG data when dedicated noise sensors unavailable Uses temporary notch filtering to identify noise subspaces [1]
High-Density EEG System Enables sufficient spatial sampling for effective source separation Minimum 64 channels; 100+ channels recommended [10] [4]
Electrical Phantom Head Validation with known ground-truth brain signals Contains embedded source antennae and contaminating sources [10] [7]
Motion Capture System Synchronizes neural data with movement kinematics Enables artifact correlation with specific movement phases [1]

iCanClean represents a significant advancement in motion artifact removal for mobile EEG research, consistently outperforming established methods like Artifact Subspace Reconstruction, Auto-CCA, and Adaptive Filtering across multiple performance metrics [1] [10] [4]. Its ability to improve ICA decomposition quality while preserving cognitively relevant neural signals makes it particularly valuable for studying brain dynamics during naturalistic movements like walking and running [1] [4].

The method's flexibility in implementation—working with both dedicated noise sensors and pseudo-reference signals—enhances its practical utility across different research environments and equipment configurations [1]. As mobile brain imaging continues to evolve toward more ecologically valid paradigms, iCanClean offers researchers a powerful tool for addressing the fundamental challenge of motion artifacts, potentially unlocking new possibilities for understanding brain function in real-world contexts [1] [10] [4].

Why Traditional ICA Struggles with Motion Artifacts

Electroencephalography (EEG) is a cornerstone technique for studying brain dynamics, but its utility in real-world, mobile settings has been historically limited by motion artifacts. These artifacts pose a significant challenge for traditional signal processing methods, particularly Independent Component Analysis (ICA). This guide examines the inherent limitations of ICA in handling motion artifacts and objectively compares the performance of two advanced solutions: Artifact Subspace Reconstruction (ASR) and the iCanClean algorithm, providing researchers with experimental data to inform their methodological choices.

The Fundamental Challenge: Why ICA Falters

Independent Component Analysis (ICA) is a blind source separation method that linearly decomposes multi-channel EEG data into maximally independent components (ICs). Its effectiveness relies on key assumptions, which are often violated in the presence of motion.

  • Non-Stationarity of Motion Artifacts: Motion artifacts are inherently non-stationary; their statistical properties change over time. ICA, however, assumes a stationary mixing process where the sources are statistically independent and mixed linearly. The transient and high-amplitude nature of motion artifacts corrupts this process, reducing the quality of the ICA decomposition and its ability to isolate brain sources effectively [1].
  • Overwhelming of Signal Space: The electrical signals generated by head movements, cable sway, and electrode displacement are often orders of magnitude larger than the microvolt-scale cortical signals. These artifacts dominate the input to ICA, consuming a disproportionate number of independent components and leaving fewer components to represent brain activity [1] [10].
  • Dependence on Clean Data and Expert Categorization: ICA does not remove artifacts by itself; it separates sources, leaving the arduous task of classifying each IC as a "brain" or "non-brain" source to the researcher. This process is subjective and time-consuming. While automated classifiers like ICLabel exist, they are typically trained on datasets from stationary participants and perform poorly on mobile EEG data contaminated with motion artifacts they have not been trained to recognize [1].

The diagram below illustrates this breakdown in the ICA process when motion is introduced.

G A Raw EEG Signal B ICA Assumptions A->B E Motion Artifact Violations A->E C Stationary Mixing Process B->C D Maximally Independent Sources B->D H Poor ICA Decomposition C->H D->H F Non-Stationary Signal E->F G Overwhelming Amplitude E->G F->H G->H I Corrupted & Non-Brain ICs H->I

Beyond ICA: A Performance Comparison of ASR vs. iCanClean

To address ICA's limitations, researchers have developed preprocessing algorithms designed specifically to handle motion artifacts before ICA is run. The table below summarizes the core principles and requirements of two leading approaches.

Feature Artifact Subspace Reconstruction (ASR) iCanClean Algorithm
Core Principle Identifies and removes high-variance, artifact-dominated components using Principal Component Analysis (PCA) [1] [12]. Uses Canonical Correlation Analysis (CCA) to find and remove data subspaces correlated with reference noise signals [1] [10].
Noise Reference Requires a segment of clean "calibration" EEG data from the same session to establish a baseline [1] [12]. Uses dedicated noise sensors or generates "pseudo-reference" signals from the raw EEG itself [1] [10].
Key Parameter Standard deviation threshold (k); lower values (e.g., 20-30) are more aggressive [1]. Correlation criterion (R²); e.g., 0.65 was effective for locomotion data [1].
Computational Efficiency Suitable for real-time application [12]. Computationally efficient and capable of real-time cleaning [10] [14].
Experimental Data and Performance Metrics

Recent studies using both phantom-head models (with known ground-truth signals) and human participants during locomotion provide quantitative data on the performance of these methods. The following table compiles key findings from controlled experiments.

Experiment Context & Metric Artifact Subspace Reconstruction (ASR) iCanClean Algorithm
Phantom Head with All Artifacts [10]Data Quality Score (0-100%) 27.6% 55.9%
Human Running EEG [1]ICA Component Dipolarity Improved Most Improved
Human Running EEG [1]Reduction in Gait-Frequency Power Significant Significant
Human Running EEG [1]Recovery of P300 ERP Effect No Yes
Key Experimental Protocols

The data in the table above are derived from several critical experimental designs:

  • Phantom Head Validation: A conductive phantom head with embedded "brain" sources and contaminating sources (eyes, neck muscles, facial muscles, motion) was used to obtain ground-truth data. The Data Quality Score was calculated as the average correlation between the known brain sources and the cleaned EEG channels, providing an objective measure of each algorithm's efficacy [10].
  • Human Locomotion EEG: Young adults performed a Flanker task (measuring P300 event-related potentials) during both static standing and overground jogging. Preprocessing with ASR (k=20) and iCanClean (pseudo-reference, R²=0.65) was compared based on:
    • ICA Dipolarity: The number of brain-like ICs with a single equivalent dipole, indicating a cleaner decomposition [1].
    • Spectral Power: Reduction of motion-induced power at the step frequency and its harmonics [1].
    • ERP Fidelity: The ability to preserve the expected P300 amplitude difference between congruent and incongruent Flanker stimuli, a key neural marker [1].

The workflow below contrasts the fundamental cleaning approaches of ASR and iCanClean.

G A1 Calibration: Clean EEG Segment A2 Learn Clean PCA Space & Threshold (k) A1->A2 A3 Process Data in Sliding Windows A2->A3 A4 Identify & Remove High-Variance PCs A3->A4 A5 Reconstruct Cleaned Data A4->A5 B1 Input: EEG & Reference Noise B2 Canonical Correlation Analysis (CCA) B1->B2 B3 Identify Correlated Subspaces (R²) B2->B3 B4 Subtract Noise Correlations B3->B4 B5 Output Cleaned EEG B4->B5

The Scientist's Toolkit: Key Research Reagents & Solutions

The following table details essential materials and software tools for implementing motion artifact correction in mobile brain imaging research.

Item Function in Research Example/Note
High-Density Mobile EEG System Records scalp potentials with sufficient spatial sampling for effective source separation. Systems with 64+ active electrodes; integrated amplifiers minimize cable sway [1] [10].
Dual-Layer Noise Sensors Provides optimal reference noise for iCanClean; mechanically coupled to EEG electrodes but not in contact with the scalp [1] [10]. Captures motion and environmental noise without brain signal.
Optical Motion Tracking System Quantifies head movement in six degrees of freedom, enabling motion-synchronized analysis and validation [15]. Used for prospective motion correction in MRI and to validate artifact timing in EEG [15].
Phantom Head Apparatus Provides ground-truth signals for controlled algorithm validation with known "brain" and "artifact" sources [10]. Critical for benchmarking performance of ASR, iCanClean, and other methods without biological variability [10].
iCanClean Software Implements the iCanClean algorithm for offline analysis or real-time cleaning [10] [14]. Can operate with dual-layer sensors or pseudo-reference signals derived from raw EEG.
EEGLAB with ASR Plug-in A standard EEG processing environment that includes an implementation of the ASR algorithm for offline use [1] [12]. Allows tuning of the k parameter and is widely accessible to the research community.

The evidence demonstrates that traditional ICA is an inadequate standalone solution for motion artifact removal due to its vulnerability to non-stationary, high-amplitude noise. Both ASR and iCanClean offer significant advancements as preprocessing steps. Quantitative results from phantom and human studies consistently show that iCanClean outperforms ASR in restoring data quality, improving ICA decompositions, and preserving nuanced neural signals like the P300. While ASR is a viable and accessible tool, iCanClean's superior performance, especially when using dual-layer noise sensors, makes it the more robust choice for demanding research applications involving intense motion, such as studies of running, sports, or rehabilitation. Researchers should select their methods based on the required fidelity of neural signals and the availability of hardware for noise reference.

From Theory to Practice: Implementing ASR and iCanClean in Your Research Pipeline

Step-by-Step Guide to Implementing the ASR Algorithm

Electroencephalography (EEG) is the only brain imaging method with the temporal precision and portability to assess electrocortical dynamics during human locomotion. However, head motion during whole-body movements produces substantial artifacts that contaminate EEG signals and reduce the quality of subsequent data analysis, particularly independent component analysis (ICA) decomposition [1]. As research increasingly focuses on neural dynamics in naturalistic environments, effective motion artifact removal has become a critical preprocessing step.

Two prominent approaches for addressing this challenge are Artifact Subspace Reconstruction (ASR) and iCanClean. This guide provides a detailed, step-by-step protocol for implementing the ASR algorithm, with objective performance comparisons to iCanClean based on recent experimental findings. The content is framed within the broader thesis of performance evaluation between these two methods for mobile brain imaging research, providing drug development professionals and neuroscientists with practical implementation guidance and evidence-based methodological selection criteria.

Theoretical Foundations of the ASR Algorithm

Artifact Subspace Reconstruction is an automated, component-based method for removing high-amplitude artifacts from continuous EEG data. The algorithm employs a sliding-window approach to identify and remove artifact components based on their statistical deviation from a calibration period of clean baseline data [1].

Core Mathematical Principles

ASR utilizes principal component analysis (PCA) to identify high-amplitude artifacts in continuous EEG. The method first calculates a covariance matrix from clean reference data, then uses this to identify components in new data that exceed a user-defined threshold. The key mathematical operation involves comparing the root mean square (RMS) of sliding-window principal components to the distribution of RMS values in the calibration data [1].

The algorithm identifies artifactual components when their standard deviation of RMS exceeds the threshold parameter "k," which determines the algorithm's aggressiveness. Lower k values (e.g., k=10) result in more aggressive cleaning, while higher values (e.g., k=20-30) are more conservative [1].

Step-by-Step Implementation Protocol

Prerequisite Data Collection and Preparation

Before implementing ASR, proper EEG data collection protocols must be established:

  • EEG System Setup: Use a high-density mobile EEG system (100+ channels recommended) with a sampling rate of at least 500 Hz [1] [7].
  • Calibration Recording: Collect 2-5 minutes of clean baseline EEG data during stationary resting state. This should be recorded with the participant in the same setup as the experimental task.
  • Experimental Recording: Proceed with the dynamic task recording (e.g., running, walking, or other movements).
ASR Parameter Configuration

Configure the core ASR parameters based on your specific research needs:

  • k threshold: Set between 10-30 based on desired aggressiveness (lower = more aggressive). For locomotion studies, k=10 is often effective while avoiding over-cleaning [1].
  • Sliding window size: Typically 500 ms to 1 second [1].
  • Reference criteria: Segments with z-scores within -3.5 to 5.0 of the Gaussian distribution for at least 92.5% of electrodes qualify as reference data [1].
Step-by-Step Computational Implementation

The following diagram illustrates the complete ASR workflow from data input to cleaned output:

G start Raw EEG Data calib Extract Calibration Data start->calib pca1 Calculate Reference PCA & Statistics calib->pca1 slide Sliding Window PCA on New Data pca1->slide detect Detect Components Exceeding k Threshold slide->detect recon Reconstruct Data Using Reference Components detect->recon output Cleaned EEG Data recon->output

Step 1: Extract Calibration Data
  • Identify clean segments from the baseline recording using statistical criteria
  • Calculate RMS values for 1-second sliding windows across all channels
  • Convert RMS values to z-scores using a condensed Gaussian distribution
  • Select segments with z-scores between -3.5 and 5.0 for at least 92.5% of electrodes as reference data [1]
Step 2: Calculate Reference Statistics
  • Perform PCA on the calibration data to establish a baseline component structure
  • Calculate the mean and standard deviation of RMS values across all components
  • Store the resulting covariance matrix and component statistics for comparison with new data [1]
Step 3: Process New Data with Sliding Window
  • Apply a sliding window (typically 500ms-1s) to the continuous EEG data
  • For each window, perform PCA and calculate RMS for each component
  • Compare component RMS to the reference distribution using the k threshold [1]
Step 4: Identify and Remove Artifactual Components
  • Flag components as artifactual if their RMS standard deviation exceeds the k threshold
  • Reconstruct the data using only the non-artifactual components
  • Interpolate removed components using the calibration data covariance matrix [1]
Step 5: Output Cleaned Data
  • The algorithm outputs continuous EEG with high-amplitude artifacts removed
  • Proceed with standard EEG preprocessing pipelines (filtering, ICA, etc.)

Experimental Comparison: ASR vs. iCanClean

Methodology for Performance Evaluation

To objectively compare ASR and iCanClean performance, researchers conducted a controlled study using:

  • Participants: 21 young adult athletes performing adapted Flanker tasks [1]
  • Conditions: Static standing vs. dynamic jogging during cognitive task
  • Evaluation Metrics:
    • ICA component dipolarity (measure of decomposition quality)
    • Power reduction at gait frequency and harmonics
    • Recovery of expected P300 event-related potential congruency effects [1]

The table below summarizes the key performance differences established in recent empirical studies:

Table 1: Performance Comparison of ASR and iCanClean for Motion Artifact Removal

Evaluation Metric ASR Performance iCanClean Performance Experimental Context
Data Quality Score 27.6% improvement from baseline 55.9% improvement from baseline Phantom head with all artifact types [7]
ICA Dipolarity Improved component quality Superior component quality, more dipolar brain sources Human running study [1]
Gait Frequency Power Reduction Significant reduction Significant reduction Overground running [1]
P300 ERP Recovery Produced ERP components similar to standing task Captured expected P300 amplitude differences Adapted Flanker task during jogging [1]
Computational Efficiency Suitable for real-time processing Suitable for real-time processing Both algorithms show real-time capability [7]
Implementation Requirements Comparison

The table below contrasts the technical requirements and implementation considerations for both algorithms:

Table 2: Implementation Requirements Comparison

Parameter ASR iCanClean
Reference Data Needed Clean calibration data Pseudo-reference or dedicated noise sensors [1]
Key Parameters k threshold (typically 10-30) R² threshold (typically 0.65), window length (typically 4s) [1]
Optimal Use Case General artifact removal, moderate motion environments High-motion environments, running studies [1]
Hardware Requirements Standard EEG systems Benefits from dual-layer EEG with dedicated noise sensors [1] [7]
Computational Load Moderate Moderate to high

The Researcher's Toolkit: Essential Materials and Reagents

Table 3: Essential Research Materials for Motion Artifact Removal Studies

Item Specification Research Function
Mobile EEG System High-density (64+ channels), sampling rate ≥500 Hz Captures neural signals during movement [1]
Active Electrodes Integrated amplification, dry or gel-based Reduces motion artifact at acquisition stage [7]
Motion Tracking System IMU sensors, optical motion capture Synchronizes motion data with EEG for analysis [1]
Dual-Layer EEG Cap Separate scalp and noise electrodes Enables reference noise recording for iCanClean [7]
Phantom Head Apparatus Electrically conductive, embedded sources Validates algorithms with known ground-truth signals [7]
Signal Processing Software EEGLAB, MATLAB, Python Implements ASR, iCanClean, and analysis pipelines [6]

Advanced Technical Considerations

Parameter Optimization Guidelines

Successful ASR implementation requires careful parameter tuning:

  • k Threshold Selection: For locomotion studies, start with k=10 and increase if over-cleaning is suspected (evidenced by loss of neural signals) [1]
  • Calibration Data Quality: Ensure reference data is truly artifact-free; manually inspect calibration segments
  • Window Size Adjustment: Adjust based on artifact characteristics - shorter windows for transient artifacts, longer for rhythmic movement artifacts
Integration with Downstream Processing

ASR should be strategically positioned within the preprocessing pipeline:

  • Apply before ICA to improve decomposition quality [1]
  • Follow with appropriate filtering (e.g., 1-40 Hz bandpass for ERP studies)
  • For source localization, verify that ASR doesn't distort spatial characteristics

The following diagram illustrates the recommended position of ASR within a comprehensive mobile EEG processing pipeline:

G raw Raw Mobile EEG import Data Import & Inspection raw->import asr ASR Processing import->asr filter Bandpass Filtering asr->filter ica ICA Decomposition filter->ica analysis Downstream Analysis (ERP, Spectral, Source) ica->analysis

Based on current evidence, ASR provides an effective approach for motion artifact removal in mobile EEG, particularly when configured with appropriate parameters. The algorithm demonstrates significant improvements in ICA decomposition quality and reduction of motion-related power interference.

However, comparative studies indicate that iCanClean generally outperforms ASR in demanding scenarios such as running, producing more dipolar brain components and better preserving expected neural responses like the P300 congruency effect [1]. For research involving high-impact movements or requiring optimal signal fidelity, iCanClean represents the current state-of-the-art, while ASR remains a valuable tool for moderate motion environments and research with computational constraints.

Implementation success depends on careful parameter optimization, appropriate calibration data collection, and integration within a comprehensive preprocessing pipeline. Researchers should validate their specific implementation using ground-truth paradigms where possible, such as known ERP components or phantom head setups.

Step-by-Step Guide to Implementing the iCanClean Algorithm

Electroencephalography (EEG) is a powerful, non-invasive tool for recording brain activity with high temporal resolution, making it ideal for studying neural dynamics during movement and in real-world settings [10]. However, a significant challenge in mobile EEG research is the contamination of signals by various artifacts, including those from motion, muscle activity, eye movements, and line noise [10] [7]. These artifacts can severely degrade data quality, hindering subsequent analysis such as independent component analysis (ICA) for source separation [4]. This guide focuses on two prominent computational approaches for mitigating these artifacts: the established Artifact Subspace Reconstruction (ASR) and the novel iCanClean algorithm. We will provide a detailed, step-by-step guide for implementing iCanClean, objectively compare its performance against ASR using published experimental data, and outline the essential resources required for implementation.

Core Algorithm Principles

iCanClean is a noise-canceling algorithm that uses canonical correlation analysis (CCA) to identify and remove subspaces of corrupted EEG data that are most strongly correlated with subspaces of reference noise recordings [10] [14]. Its name stands for "implementing Canonical correlation to Cancel Latent Electromagnetic Artifacts and Noise" [7]. A key advantage is its flexibility in using various reference signals, including dedicated noise sensors from a dual-layer EEG cap or internally generated "pseudo-reference" signals derived from the raw EEG itself [1] [4].

Artifact Subspace Reconstruction (ASR) is a method based on principal component analysis (PCA) that identifies and removes high-variance, artifact-containing components from the EEG signal by comparing them to a clean baseline calibration period [1] [12]. It functions by calculating the principal components of short, sliding windows of data and rejecting those components whose variance exceeds a user-defined threshold (k) relative to the calibration data [12].

Quantitative Performance Comparison

Extensive testing, particularly on a phantom head with known ground-truth brain signals, has demonstrated iCanClean's superior performance in removing a wide range of artifacts while preserving brain activity [10] [7].

Table 1: Data Quality Score Comparison Across Artifact Types (Phantom Head Study)

Artifact Condition No Cleaning iCanClean ASR Auto-CCA Adaptive Filtering
Brain + All Artifacts 15.7% 55.9% 27.6% 27.2% 32.9%
Brain (Target) 57.2% N/A N/A N/A N/A

Source: Adapted from [10] [7]

As shown in Table 1, in the most challenging condition with all artifacts present simultaneously, iCanClean improved data quality nearly to the level of the clean "Brain" target, outperforming other real-time-capable methods by a significant margin [10] [7].

In human studies during running, both ASR and iCanClean have been shown to improve the quality of subsequent ICA decompositions, measured by an increase in the number of "good" dipolar brain components [1].

Table 2: ICA Improvement in Human Locomotion Data (Running)

Preprocessing Method Average Number of "Good" Brain Components
No Cleaning (Basic Preprocessing) 8.4
iCanClean (with pseudo-reference) 13.2
Artifact Subspace Reconstruction (ASR) ~12 (estimated from trends)

Source: Data synthesized from [1] and [4]

Furthermore, only preprocessing with iCanClean successfully recovered the expected P300 event-related potential congruency effect during a running task, a key neural marker that ASR and other methods failed to fully isolate under high-motion conditions [1].

Step-by-Step Implementation Guide for iCanClean

Prerequisites and Data Collection
  • EEG System: A standard high-density EEG system (64+ channels) is recommended. For optimal performance, a dual-layer EEG cap with outwardly facing noise electrodes mechanically coupled to the scalp electrodes is ideal, as it provides direct reference noise recordings [4].
  • Data Recording: If a dual-layer cap is unavailable, iCanClean can generate "pseudo-reference" noise signals from the raw EEG data. In this case, ensure you record sufficient data; for mobile scenarios, collecting at least 30 minutes of high-density EEG at a sampling frequency of 500 Hz or higher is recommended [10].
Data Preprocessing

Before applying iCanClean, perform basic preprocessing on both the scalp EEG channels and the separate noise channels (if available) [4]:

  • High-Pass Filtering: Apply a 1 Hz high-pass filter to remove slow drifts.
  • Re-referencing: Average re-reference the scalp electrodes and the noise electrodes to their own respective averages.
  • Bad Channel Rejection: Identify and remove channels with abnormally high amplitude (e.g., standard deviation >3 times the median across channels).
Core iCanClean Workflow

The following diagram illustrates the core signal processing workflow of the iCanClean algorithm.

G Start Preprocessed EEG & Noise Data Step1 1. Identify Candidate Noise Components (Canonical Correlation Analysis) Start->Step1 Step2 2. Select Noise Components (User-defined R² Threshold) Step1->Step2 Step3 3. Project Noise Back to Channels (Least-Squares Solution) Step2->Step3 Step4 4. Subtract Noise from EEG Step3->Step4 Param Key Parameters: - Window Length - R² Threshold Param->Step2

  • Identify Candidate Noise Components: The algorithm performs CCA on the scalp EEG data (a mixture of brain signal and noise) and the reference noise data. This step finds linear combinations of each dataset that are maximally correlated with each other. These combinations are the "canonical components" [10] [7].
  • Select Noise Components: The user sets an R² threshold (aggressiveness parameter). Canonical component pairs with a correlation (R²) exceeding this threshold are marked as noise and selected for removal. A lower R² value results in more aggressive cleaning [4].
  • Project Noise Back to Channels: The selected noise components are projected back onto the original EEG channel space using a least-squares solution. This creates a channel-level estimate of the pure noise [10] [7].
  • Subtract Noise from EEG: The estimated noise is subtracted from the original, contaminated EEG data, resulting in the cleaned EEG output [10].
Parameter Tuning and Optimization

The two critical parameters for iCanClean are the window length and the R² threshold. Based on parameter sweeps in human locomotion studies [4]:

  • Window Length: A 4-second sliding window is recommended as the optimal setting for balancing local artifact removal with computational efficiency.
  • R² Threshold: An R² value of 0.65 has been identified as the optimal aggressiveness level for effectively removing motion artifacts during walking while preserving brain activity.

Essential Research Toolkit

Successfully implementing and comparing artifact removal methods requires specific hardware, software, and analytical tools.

Table 3: Research Reagent Solutions for EEG Artifact Removal

Item Function / Purpose Example / Specification
Dual-Layer EEG Cap Provides dedicated noise sensors mechanically coupled to scalp electrodes for optimal reference noise recording. 120 scalp + 120 noise electrodes [4]
High-Density EEG System Enables sufficient spatial sampling for effective source separation and ICA. 64+ channel mobile amplifier [10]
Accelerometer (IMU) Measures head motion; can be used as an alternative or supplementary reference signal. 9-axis IMU mounted on the forehead [5]
iCanClean Code The core algorithm for cleaning. Available via published papers and associated code repositories [10] [14]
EEGLAB with ICLabel Software environment for preprocessing, ICA decomposition, and automated component classification. EEGLAB plugin: ICLabel [4]
Artifact Subspace Reconstruction (ASR) Benchmark algorithm for performance comparison. Included in EEGLAB/BCILAB [12]

This guide provides a comprehensive framework for implementing and evaluating the iCanClean algorithm for EEG artifact removal. The strong experimental evidence from both phantom and human studies indicates that iCanClean offers a performance advantage, particularly in challenging scenarios with multiple, simultaneous artifacts like those encountered in mobile brain imaging research [10] [1]. By following the detailed protocols for implementation, parameter optimization, and comparative benchmarking with ASR outlined herein, researchers can make informed decisions to enhance the fidelity of their EEG data in real-world applications.

Electroencephalography (EEG) is the only brain imaging method lightweight enough with sufficient temporal precision to assess electrocortical dynamics during human locomotion and other real-world activities [1]. However, the presence of mechanical and non-physiological artifacts in the EEG time series due to head motion, electrode displacement, and cable sway presents a significant barrier to progress in understanding the neurophysiology of natural human behavior [1]. Motion artifacts contaminate EEG signals and reduce the quality of Independent Component Analysis (ICA) decomposition, hindering source-level analysis of mobile EEG data [4]. While Artifact Subspace Reconstruction (ASR) has been a commonly used approach for handling motion artifacts, iCanClean has emerged as a promising alternative that uses reference noise recordings to remove noisy EEG subspaces [1] [10]. This comparison guide examines the two primary configurations for implementing iCanClean—dual-layer hardware systems and pseudo-reference software approaches—to help researchers select the optimal configuration for their mobile brain imaging studies.

Understanding iCanClean: Core Algorithm and Mechanism

The iCanClean algorithm employs a novel noise-canceling approach that uses canonical correlation analysis (CCA) to identify and remove subspaces of corrupted data recordings that are most strongly correlated with subspaces of reference noise recordings [14]. The algorithm is computationally efficient, making it suitable for real-time applications such as brain-computer interfaces [14]. iCanClean operates by leveraging reference noise signals and CCA to detect and correct noise-based subspaces based on a user-selected criterion (R²) of the correlation between the subspaces of corrupt EEG and noise signals [1].

The fundamental principle behind iCanClean is that it compares cortical electrode signals (recording mixtures of brain activity plus noise) with noise electrodes or pseudo-reference signals (recording only noise) to remove noisy subspaces of EEG data without eliminating underlying neural signals [4]. After noise components are projected back onto the EEG channels using a least-squares solution, the noise components that exceed the R² threshold are subtracted from the scalp EEG [1].

G Raw EEG Signal Raw EEG Signal Noise Signal Acquisition Noise Signal Acquisition Raw EEG Signal->Noise Signal Acquisition Dual-Layer Hardware Dual-Layer Hardware Noise Signal Acquisition->Dual-Layer Hardware Pseudo-Reference Software Pseudo-Reference Software Noise Signal Acquisition->Pseudo-Reference Software Mechanically Coupled Noise Electrodes Mechanically Coupled Noise Electrodes Dual-Layer Hardware->Mechanically Coupled Noise Electrodes Notch-Filtered EEG Signals Notch-Filtered EEG Signals Pseudo-Reference Software->Notch-Filtered EEG Signals Canonical Correlation Analysis (CCA) Canonical Correlation Analysis (CCA) Mechanically Coupled Noise Electrodes->Canonical Correlation Analysis (CCA) Notch-Filtered EEG Signals->Canonical Correlation Analysis (CCA) Noise Subspace Identification Noise Subspace Identification Canonical Correlation Analysis (CCA)->Noise Subspace Identification Component Subtraction (R² Threshold) Component Subtraction (R² Threshold) Noise Subspace Identification->Component Subtraction (R² Threshold) Cleaned EEG Output Cleaned EEG Output Component Subtraction (R² Threshold)->Cleaned EEG Output User Parameters User Parameters User Parameters->Canonical Correlation Analysis (CCA) User Parameters->Component Subtraction (R² Threshold)

Figure 1: The iCanClean signal processing workflow demonstrates two parallel pathways for noise signal acquisition converging through canonical correlation analysis.

Experimental Approaches: Methodologies for Comparative Evaluation

Phantom Head Validation Studies

Researchers have employed electrically conductive phantom heads with embedded brain source antennae to quantitatively evaluate iCanClean's performance with known ground-truth brain signals [10]. These phantom systems typically incorporate multiple simulated brain sources and contaminating sources to replicate various artifact types, including ocular, muscle, motion, and line-noise artifacts [10]. Studies calculate a Data Quality Score (0-100%) based on the average correlation between known brain sources and EEG channels before and after cleaning, enabling direct comparison of different artifact removal methods [10].

Human Studies During Locomotion and Sports

Human studies during dynamic activities provide ecological validation of iCanClean's performance. Research designs typically involve participants completing tasks such as treadmill walking, running, or sports activities like table tennis while wearing high-density EEG systems [1] [16]. Evaluation metrics include ICA component dipolarity (residual variance < 15%), power spectral density changes at gait frequency and harmonics, and the ability to recover expected event-related potential components such as the P300 congruency effect [1].

Parameter Optimization Protocols

Systematic parameter sweeps are essential for determining optimal iCanClean settings. Studies typically vary two primary parameters: the R² threshold (cleaning aggressiveness from 0.05 to 1.0) and the window length for CCA (commonly 1s, 2s, 4s, or infinite) [4]. The optimal configuration is determined by measuring the number of high-quality brain components resulting from ICA decomposition, with components classified as 'good' based on dipole model fit (residual variance < 15%) and brain probability (ICLabel > 50%) [4].

Dual-Layer Configuration: Hardware-Based Noise Reference

The dual-layer EEG approach utilizes specialized hardware with noise electrodes mechanically coupled to traditional scalp electrodes but electrically isolated from them [16]. These noise electrodes are inverted and positioned to capture environmental and motion artifacts without recording brain activity, providing ideal reference signals for the iCanClean algorithm [4].

Implementation Requirements and Technical Specifications

G Dual-Layer EEG Cap Dual-Layer EEG Cap Scalp Electrodes Layer Scalp Electrodes Layer Dual-Layer EEG Cap->Scalp Electrodes Layer Noise Electrodes Layer Noise Electrodes Layer Dual-Layer EEG Cap->Noise Electrodes Layer Mixed Signal (Brain + Noise) Mixed Signal (Brain + Noise) Scalp Electrodes Layer->Mixed Signal (Brain + Noise) Reference Noise Only Reference Noise Only Noise Electrodes Layer->Reference Noise Only iCanClean Processing iCanClean Processing Mixed Signal (Brain + Noise)->iCanClean Processing Reference Noise Only->iCanClean Processing 3D-Printed Couplers 3D-Printed Couplers Mechanical Connection Mechanical Connection 3D-Printed Couplers->Mechanical Connection Mechanical Connection->Noise Electrodes Layer Cable Bundling Cable Bundling Identical Motion Exposure Identical Motion Exposure Cable Bundling->Identical Motion Exposure Identical Motion Exposure->Noise Electrodes Layer Clean Brain Activity Output Clean Brain Activity Output iCanClean Processing->Clean Brain Activity Output

Figure 2: Dual-layer EEG hardware configuration showing mechanical coupling and electrical isolation between layers.

Performance Metrics and Experimental Results

Dual-layer iCanClean demonstrates superior performance in multiple validation studies. In phantom head testing with all artifacts simultaneously present, dual-layer iCanClean improved the Data Quality Score from 15.7% before cleaning to 55.9% after cleaning, significantly outperforming ASR (27.6%), Auto-CCA (27.2%), and Adaptive Filtering (32.9%) [10]. For context, the brain-only condition scored 57.2% without cleaning, indicating that iCanClean nearly restored the data to its original quality despite massive contamination [10].

In human studies during walking, dual-layer iCanClean with optimal settings (4-second window, R²=0.65) improved the average number of high-quality brain components from 8.4 to 13.2 (+57%) across young adults, high-functioning older adults, and low-functioning older adults [4]. The approach maintained good performance even with reduced noise channels, yielding 12.7, 12.2, and 12.0 good components with 64, 32, and 16 noise channels respectively [4].

Pseudo-Reference Configuration: Software-Based Noise Estimation

The pseudo-reference approach provides an accessible alternative when specialized dual-layer hardware is unavailable. This method creates artificial noise references by applying a temporary notch filter to the raw EEG to isolate noise components, typically below 3 Hz for motion artifacts [1]. These software-generated noise signals enable iCanClean operation with standard EEG systems.

Implementation Workflow

G Raw EEG Data Raw EEG Data Temporary Notch Filter Application Temporary Notch Filter Application Raw EEG Data->Temporary Notch Filter Application Canonical Correlation Analysis Canonical Correlation Analysis Raw EEG Data->Canonical Correlation Analysis Pseudo-Noise Signal Generation Pseudo-Noise Signal Generation Temporary Notch Filter Application->Pseudo-Noise Signal Generation Pseudo-Noise Signal Generation->Canonical Correlation Analysis Noise Subspace Identification (R² Threshold) Noise Subspace Identification (R² Threshold) Canonical Correlation Analysis->Noise Subspace Identification (R² Threshold) Noise Component Subtraction Noise Component Subtraction Noise Subspace Identification (R² Threshold)->Noise Component Subtraction Cleaned EEG Output Cleaned EEG Output Noise Component Subtraction->Cleaned EEG Output User-Defined Parameters User-Defined Parameters User-Defined Parameters->Temporary Notch Filter Application User-Defined Parameters->Noise Subspace Identification (R² Threshold)

Figure 3: Pseudo-reference workflow showing software-based noise estimation from existing EEG signals.

Performance Assessment

Pseudo-reference iCanClean demonstrates substantial effectiveness, though generally less powerful than the dual-layer approach. In studies comparing motion artifact removal during running, preprocessing with iCanClean using pseudo-reference signals led to the recovery of more dipolar brain independent components and significantly reduced power at the gait frequency [1]. The approach produced ERP components similar in latency to those identified in stationary tasks and successfully captured the expected P300 amplitude enhancement to incongruent flankers [1].

Notably, iCanClean with pseudo-reference signals demonstrated somewhat greater effectiveness than ASR in recovering dipolar components and identifying expected P300 effects during running [1]. This positions pseudo-reference iCanClean as a valuable software upgrade for researchers with standard EEG equipment who need improved motion artifact handling.

Comparative Performance Analysis: Quantitative Results

Table 1: Performance comparison of iCanClean configurations and ASR for motion artifact removal

Method Configuration Data Quality Score (%) Good Brain Components Power Reduction at Gait Frequency ERP Component Recovery
iCanClean Dual-Layer (Optimal) 55.9 (from 15.7 baseline) 13.2 (from 8.4 baseline) Significant reduction Excellent
iCanClean Pseudo-Reference Not reported Significantly improved Significant reduction Good P300 recovery
ASR Standard (k=20-30) 27.6 (from 15.7 baseline) Moderate improvement Moderate reduction Moderate recovery
Adaptive Filtering IMU Reference 32.9 (from 15.7 baseline) Limited improvement Limited reduction Limited recovery

Table 2: Technical requirements and implementation considerations

Parameter Dual-Layer iCanClean Pseudo-Reference iCanClean ASR
Hardware Requirements Specialized dual-layer cap with noise electrodes Standard EEG system Standard EEG system
Optimal Parameters 4s window, R²=0.65 Similar to dual-layer, may require adjustment k=20-30 (standard), k=10 (conservative)
Calibration Needs None beyond standard setup None beyond standard setup Clean calibration data required
Computational Load Moderate Moderate Low to moderate
Best Application Context High-motion studies with budget for specialized equipment Standard motion studies with existing equipment Low to moderate motion with clean calibration segments

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key research materials and solutions for iCanClean implementation

Research Reagent Function/Application Implementation Considerations
Dual-Layer EEG Cap Provides hardware-based noise reference through mechanically coupled noise electrodes Requires specialized equipment; optimal for high-quality mobile brain imaging
Active Electrodes Amplifies EEG signals prior to transmission to reduce motion artifacts Compatible with both iCanClean configurations; improves baseline signal quality
3D-Printed Couplers Mechanically connects scalp and noise electrodes while maintaining electrical isolation Essential for dual-layer implementation; ensures noise electrodes experience identical motion
Conductive Fabric Creates artificial skin circuit for noise electrodes in dual-layer systems Bridges noise electrodes; improves noise reference quality
Canonical Correlation Analysis Algorithm Core computational method for identifying noise subspaces correlated with reference signals Central to iCanClean; requires parameter optimization (R², window length)
Pseudo-Noise Generation Algorithm Creates artificial noise references from existing EEG data via temporary notch filtering Enables iCanClean with standard EEG systems; typically filters below 3Hz for motion

The comparative analysis reveals that both iCanClean configurations offer superior motion artifact removal compared to traditional methods like ASR, with the dual-layer hardware approach providing the highest fidelity results and the pseudo-reference method offering an accessible and effective alternative for researchers with standard equipment.

For research requiring the highest signal quality during intense motion activities (e.g., running, sports), the dual-layer iCanClean configuration is recommended, with optimal parameters typically around a 4-second window and R² threshold of 0.65 [4]. For studies with limited access to specialized equipment or where moderate motion artifacts are expected, the pseudo-reference iCanClean approach provides significant improvement over ASR and other traditional methods [1].

Future research directions include optimizing iCanClean for non-cyclical movements, integrating complementary sensors such as IMUs, and developing automated parameter selection based on motion characteristics to enhance usability and effectiveness across diverse research applications [17] [18].

Electroencephalography (EEG) is the only brain imaging method with sufficient portability and temporal precision to assess electrocortical dynamics during human locomotion [1]. However, head motion during whole-body movements produces substantial artifacts that contaminate EEG signals and reduce the quality of subsequent data analysis, particularly independent component analysis (ICA) decomposition [1] [19]. Motion artifacts present a significant barrier to studying neural processes during ecologically valid walking and running tasks. Two prominent computational approaches have emerged to address this challenge: Artifact Subspace Reconstruction (ASR) and iCanClean [1] [10]. This guide objectively compares their performance across key metrics relevant to locomotion researchers, drawing from recent comparative studies in both walking and running paradigms.

Artifact Subspace Reconstruction (ASR)

ASR employs a sliding-window principal component analysis (PCA) to identify and remove high-amplitude artifacts from continuous EEG [1]. It first establishes a calibration period from reference data considered relatively artifact-free, then calculates the covariance matrix of this reference data. For new data segments, ASR identifies principal components that exceed a user-defined standard deviation threshold ("k") compared to the reference and reconstructs these artifactual components using the calibration data [1]. The "k" parameter controls cleaning aggressiveness, with lower values (e.g., k=10) producing more extensive cleaning but potentially risking "over-cleaning" and manipulation of neural signals [1].

iCanClean

iCanClean utilizes canonical correlation analysis (CCA) with reference noise signals to detect and remove noise subspaces from EEG data [1] [4]. The algorithm identifies subspaces of scalp EEG that correlate with noise subspaces based on a user-selected correlation criterion (R²) and subtracts these noise components [1]. iCanClean can operate with dedicated noise sensors (ideal) or with "pseudo-reference" noise signals derived from the EEG itself when dedicated sensors are unavailable [1] [19]. Key parameters include the R² threshold (cleaning aggressiveness) and the sliding window length for CCA computation [4].

G cluster_ASR Artifact Subspace Reconstruction (ASR) cluster_iCanClean iCanClean Algorithm ASR1 Establish calibration period from reference data ASR2 Calculate covariance matrix of reference data ASR1->ASR2 ASR3 Sliding-window PCA on new data segments ASR2->ASR3 ASR4 Identify components exceeding k threshold ASR3->ASR4 ASR5 Reconstruct artifact components using calibration data ASR4->ASR5 End Cleaned EEG ASR5->End IC1 Obtain reference noise signals (physical or pseudo-reference) IC2 Canonical Correlation Analysis (CCA) between EEG and noise signals IC1->IC2 IC3 Identify noise subspaces exceeding R² threshold IC2->IC3 IC4 Subtract noise components from EEG data IC3->IC4 IC4->End Start Motion-Corrupted EEG Start->ASR1 Start->IC1

Performance Comparison in Human Locomotion Studies

Quantitative Performance Metrics

Table 1: Comparative Performance of ASR and iCanClean in Running Research

Performance Metric ASR Performance iCanClean Performance Testing Paradigm Citation
ICA Component Dipolarity Improved dipolarity vs. no cleaning Superior improvement in dipolarity vs. ASR Overground running with Flanker task [1]
Power Reduction at Gait Frequency Significant reduction Significant reduction Overground running with Flanker task [1]
ERP Component Recovery Produced ERP components similar to standing task Produced ERP components similar to standing task + captured P300 congruency effect Overground running with Flanker task [1]
Data Quality Score (All Artifacts) 27.6% (from 15.7% baseline) 55.9% (from 15.7% baseline) Phantom head with multiple artifacts [10]
Optimal Parameters for Locomotion k=10-30 (aggressive but avoiding over-cleaning) R²=0.65 with 4-s window Parameter optimization studies [1] [4]

Table 2: iCanClean Performance Across Different Noise Channel Configurations During Walking

Number of Noise Channels Good Components After iCanClean Performance Improvement vs. Baseline Study Population
Baseline (No iCanClean) 8.4 components Reference Mixed ages and functioning levels
16 noise channels 12.0 components +42.9% Mixed ages and functioning levels
32 noise channels 12.2 components +45.2% Mixed ages and functioning levels
64 noise channels 12.7 components +51.2% Mixed ages and functioning levels
120 noise channels (full array) 13.2 components +57.1% Mixed ages and functioning levels

Key Comparative Findings

Recent research directly comparing ASR and iCanClean during overground running reveals important performance differences. A 2025 study examining both approaches during a dynamic Flanker task found that while both methods improved data quality, iCanClean demonstrated superior performance in several key areas [1]. Both approaches significantly increased ICA component dipolarity and reduced power at the gait frequency, but only iCanClean successfully captured the expected P300 congruency effect in event-related potentials during running [1].

In controlled phantom head experiments with known ground-truth signals, iCanClean consistently outperformed ASR across multiple artifact conditions [10]. When all artifacts were simultaneously present (motion, muscle, eye, and line-noise), iCanClean improved data quality scores from 15.7% to 55.9%, while ASR only achieved 27.6% [10]. This substantial performance gap suggests iCanClean may be more effective for complex real-world scenarios where multiple artifact types coexist.

For walking research, iCanClean has demonstrated robust performance across diverse populations. A comprehensive parameter sweep established optimal settings for walking data (R²=0.65 with 4-second windows), which increased the number of high-quality brain components from 8.4 to 13.2 (+57%) across young adults, high-functioning older adults, and low-functioning older adults [4]. The method maintained effectiveness even with reduced noise channel configurations, demonstrating practical utility for systems with varying sensor densities [4].

Experimental Protocols for Method Evaluation

Overground Running Flanker Task Protocol

Recent comparative studies have employed sophisticated paradigms to evaluate artifact removal methods during high-motion conditions. The overground running Flanker task involves participants jogging at a comfortable pace while responding to congruent or incongruent arrow stimuli presented on a display [1] [19]. This design enables simultaneous assessment of multiple performance metrics:

  • ICA Component Dipolarity: Components with residual variance <15% and high brain probability (>50% via ICLabel) are classified as 'good' brain components [1] [4].
  • Spectral Power at Gait Frequency: Power reduction at the fundamental gait frequency and harmonics quantifies motion artifact suppression [1].
  • ERP Component Recovery: Comparison with standing Flanker task ERPs establishes neural signal preservation [1].

G cluster_preprocessing Preprocessing cluster_cleaning Artifact Removal (Experimental Condition) Start EEG Recording During Overground Running P1 High-pass filter (1 Hz cutoff) Start->P1 P2 Channel rejection (amplitude >3x median) P1->P2 P3 Average re-referencing P2->P3 C1 Apply ASR or iCanClean P3->C1 C2 Parameter optimization (ASR k=10-30, iCanClean R²=0.65) C1->C2 E1 ICA Decomposition C2->E1 subcluster_evaluation subcluster_evaluation E2 Component dipolarity (RV<15%) E1->E2 E3 Spectral power at gait frequency E2->E3 E4 ERP component recovery E3->E4

Phantom Head Validation Protocol

Controlled phantom head studies provide ground-truth validation through simulated brain sources with known characteristics. The standard protocol involves:

  • Phantom Setup: Electrically conductive head phantom with embedded dipole sources (typically 10 brain sources and 10 contaminating sources) mimicking human electrical properties [10].
  • Artifact Introduction: Systematic introduction of motion, muscle, eye, and line-noise artifacts at controlled amplitudes.
  • Data Quality Assessment: Calculation of Data Quality Scores based on correlation between known brain sources and recovered EEG signals [10].

This approach enables precise quantification of artifact removal effectiveness while preserving neural signals, with iCanClean demonstrating 55.9% data quality recovery versus 27.6% for ASR under maximum artifact conditions [10].

Table 3: Research Reagent Solutions for Motion Artifact Removal Studies

Tool/Resource Function/Purpose Example Implementation
Dual-Layer EEG Systems Provides dedicated noise reference signals mechanically coupled to EEG electrodes 120+120 electrode configurations with outward-facing noise electrodes [4]
Electrical Phantom Head Ground-truth validation with known brain sources and controllable artifacts Ballistics gelatin models with embedded dipoles and broadcast antennas [10]
Mobile EEG Platforms Wireless EEG systems enabling natural locomotion studies 32-256 channel systems with head-mounted IMUs for motion tracking [5]
ICA Decomposition Algorithms Blind source separation for identifying neural and artifactual components AMICA, Infomax, FastICA implementations in EEGLAB [4]
ICLabel Classifier Automated component classification using deep learning EEGLAB plugin with convolutional neural network for component labeling [1]
Motion Tracking Systems Quantification of movement parameters for artifact correlation IMU sensors (accelerometers, gyroscopes) synchronized with EEG [5]

The comparative evidence from both walking and running studies indicates that while both ASR and iCanClean significantly improve mobile EEG data quality, iCanClean generally demonstrates superior performance across multiple metrics, particularly for challenging high-motion scenarios like running. iCanClean's ability to preserve neural signals while effectively removing complex artifacts makes it particularly valuable for studying cognitive processes during locomotion.

For researchers selecting between these approaches, consider iCanClean when studying high-amplitude movements (e.g., running), when multiple artifact types coexist, or when dedicated noise references are available. ASR remains a viable option for less demanding motion scenarios or when computational resources are constrained. Future directions include integration with IMU-based approaches [5] and the development of specialized electrode geometries [20] for further performance enhancement in mobile brain imaging research.

Integrating Cleaning Methods with ICA for Source-Level Analysis

Electroencephalography (EEG) is the only brain imaging method lightweight enough with sufficient temporal precision to assess electrocortical dynamics during human locomotion [8]. However, head motion during whole-body movements produces artifacts that contaminate EEG signals and reduce the quality of Independent Component Analysis (ICA) decomposition [8] [1]. ICA is a fundamental tool for source-level analysis that linearly decomposes mixed EEG data into maximally independent components (ICs) by detecting linear subspaces based on higher-order statistics [4]. Successful decomposition relies on the identification of components that are statistically independent and non-Gaussian, assumptions often violated in the presence of substantial motion artifacts [21].

The integration of artifact removal methods prior to ICA has become essential for mobile brain imaging research. This comparison guide objectively evaluates two prominent approaches—Artifact Subspace Reconstruction (ASR) and iCanClean—focusing on their performance in recovering neural signals during intense motion conditions and their effectiveness in facilitating subsequent ICA decomposition.

Artifact Subspace Reconstruction (ASR)

ASR employs a sliding-window principal component analysis (PCA) approach to identify and remove high-amplitude artifacts from continuous EEG data [1]. The method uses clean EEG segments for calibration to establish thresholds for artifact identification based on the distribution of signal variance [1]. Data segments are identified as potential reference data when their root mean square (RMS) values fall within -3.5 to 5.0 z-scores of a Gaussian distribution for at least 92.5% of electrodes [1]. A key user-defined parameter ("k") determines the threshold for identifying artifactual components, with lower values producing more aggressive cleaning [1].

iCanClean

iCanClean utilizes canonical correlation analysis (CCA) with reference noise signals to detect and correct artifact-contaminated subspaces in EEG data [10] [1]. The algorithm identifies subspaces of scalp EEG that correlate with noise subspaces based on a user-selected correlation criterion (R²) [1]. When dedicated noise sensors are unavailable, iCanClean can generate "pseudo-reference" noise signals by applying a temporary notch filter to identify noise within the EEG itself [1]. The method is particularly effective with dual-layer EEG systems where outward-facing noise electrodes mechanically coupled to traditional EEG electrodes provide reference recordings of artifacts without neural content [4].

Table 1: Key Technical Characteristics of ASR and iCanClean

Feature ASR iCanClean
Core Algorithm Principal Component Analysis (PCA) Canonical Correlation Analysis (CCA)
Reference Requirement Clean calibration data Noise sensors or pseudo-references
Primary Parameters "k" threshold (typically 10-30) R² threshold (typically 0.65), window length
Computational Demand Moderate Moderate to High
Real-time Capability Yes Yes

Experimental Comparisons and Performance Metrics

ICA Decomposition Quality

The quality of ICA decomposition can be quantified through component dipolarity, typically measured by residual variance (RV) in dipole fitting, with RV < 15% indicating well-localized brain components [4]. Studies comparing preprocessing methods during overground running found that both iCanClean and ASR improved ICA decomposition, but iCanClean demonstrated superior performance.

Table 2: ICA Decomposition Quality Metrics During Locomotion

Condition Good Components (Pre-cleaning) Good Components (Post-cleaning) Improvement
iCanClean (4-s window, R²=0.65) 8.4 13.2 +57% [4]
ASR (k=20) Not reported Moderate improvement Less than iCanClean [8]
No Cleaning 8.4 - Baseline

In research examining overground running during a Flanker task, preprocessing with iCanClean using pseudo-reference noise signals or ASR led to the recovery of more dipolar brain independent components, with iCanClean proving somewhat more effective than ASR in analyses [8] [1].

Motion Artifact Reduction

Power spectral analysis at the gait frequency and its harmonics provides a direct measure of motion artifact reduction. During running, power was significantly reduced at the gait frequency after preprocessing with both ASR and iCanClean [8] [1]. In phantom head studies with known ground-truth signals, iCanClean demonstrated superior artifact removal capabilities across multiple artifact types. When all artifacts were simultaneously present (motion, muscle, eye, line-noise), iCanClean improved Data Quality Scores from 15.7% to 55.9%, significantly outperforming ASR (27.6%), Auto-CCA (27.2%), and Adaptive Filtering (32.9%) [10] [7].

Neural Signal Preservation and Recovery

The ultimate validation of any artifact removal method lies in its ability to preserve or recover genuine neural signals. Studies employing event-related potentials (ERPs) during motion have demonstrated that both ASR and iCanClean can produce ERP components similar in latency to those identified in stationary conditions [8] [1]. In the P300 component during a Flanker task, the expected greater amplitude to incongruent flankers was successfully identified when preprocessing used iCanClean [8] [1]. This suggests that iCanClean may offer advantages for preserving stimulus-locked neural activity during motion-intensive paradigms.

Methodological Protocols for Performance Evaluation

Experimental Paradigms for Validation

Research evaluating motion artifact removal methods has employed several sophisticated experimental designs:

  • Dual-task paradigms during skateboarding, walking, and running incorporate auditory or visual stimuli to generate measurable neural responses (e.g., ERPs) alongside motion artifacts [22]. Machine learning classification of single-trial responses provides an objective measure of cleaning effectiveness.

  • Phantom head validation uses electrically conductive head models with embedded brain source antennae and contaminating sources, enabling ground-truth comparison of cleaning methods [10] [7].

  • Mobile brain imaging during treadmill walking or overground running across varying terrains and speeds systematically tests method robustness [4].

Pipeline Configurations and Processing Order

Studies have investigated different processing sequences for combining artifact removal with ICA:

  • ASR before ICA (ASRICA): This sequence has demonstrated superior performance in challenging motion environments [22]. ASR first removes non-stationary transient artifacts, enhancing subsequent ICA decomposition by addressing violations of spatiotemporal stationarity assumptions.

  • ICA before ASR: Less effective as large motion artifacts may compromise the initial ICA decomposition [22].

  • iCanClean before ICA: Optimal for maximizing brain component yield, particularly with dual-layer EEG systems [4].

G cluster_0 Alternative Pipelines RawEEG Raw EEG Data Preprocessing Basic Preprocessing (Filtering, Re-referencing) RawEEG->Preprocessing ASR ASR Cleaning Preprocessing->ASR ASRICA Pipeline iCanClean iCanClean Processing Preprocessing->iCanClean iCanClean Pipeline ICA_first ICA Only Preprocessing->ICA_first ASR_only ASR Only Preprocessing->ASR_only ICA ICA Decomposition ASR->ICA iCanClean->ICA Analysis Source-Level Analysis ICA->Analysis ICA_first->Analysis ASR_only->Analysis

Diagram 1: EEG Cleaning Pipeline Workflow. The ASRICA and iCanClean pipelines provide the most effective preprocessing for ICA decomposition during motion.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Research Materials and Solutions for Motion Artifact Studies

Research Reagent Function/Specification Application Context
Dual-Layer EEG Cap 120+120 electrodes; noise electrodes mechanically coupled to but electrically isolated from scalp electrodes [4] Provides reference noise signals for iCanClean
Phantom Head Apparatus Electrically conductive model with embedded brain sources and contaminating sources [10] [7] Method validation with known ground-truth signals
Inertial Measurement Units (IMUs) 9-axis sensors (accelerometer, gyroscope, magnetometer) sampled at 128Hz+ [5] Head motion tracking for reference-based artifact removal
AMICA Algorithm Adaptive Mixture ICA implementation; includes sample rejection based on log-likelihood [23] High-quality ICA decomposition for mobile EEG
ICLabel Convolutional neural network for automated component classification [4] Objective component categorization after ICA

Performance Optimization Guidelines

Parameter Selection

Optimal parameters vary by application but research suggests:

  • iCanClean: 4-second window with R² threshold of 0.65 provides optimal balance between artifact removal and brain signal preservation during walking and running [4]. Performance remains acceptable with reduced noise channels (12.0 good components with 16 noise channels vs. 13.2 with 120) [4].

  • ASR: "k" values of 10-30 are recommended, with lower values (more aggressive cleaning) potentially needed for high-motion scenarios but risking overcleaning [1]. For human locomotion, k=10 preserves dipolarity without excessive data manipulation [1].

Method Selection Framework

Research supports the following decision process:

  • For maximum brain component yield: iCanClean with dual-layer EEG systems is superior, particularly when dedicated noise sensors are available [4].

  • For standard EEG systems: ASR before ICA (ASRICA pipeline) provides substantial improvement over minimal cleaning or ICA alone [22].

  • For computational efficiency: Both methods offer real-time capability, though parameter tuning is essential [10].

  • For high-intensity motion: The combination of ASR and ICA has proven effective even during extreme activities like skateboarding on half-pipe ramps [22].

G Start Start: Motion Artifact Removal Method Selection DualLayer Dual-Layer EEG Available? Start->DualLayer iCanCleanRec Recommended: iCanClean with R²=0.65, 4s window DualLayer->iCanCleanRec Yes MotionLevel Motion Intensity Level? DualLayer->MotionLevel No Components Evaluate Component Quality (Dipolarity, ICLabel) iCanCleanRec->Components HighMotion High Intensity (Running, Sports) MotionLevel->HighMotion High LowMotion Low-Moderate Intensity (Walking, Standing) MotionLevel->LowMotion Low/Moderate ASRRec Recommended: ASR before ICA (ASRICA) with k=10-20 HighMotion->ASRRec LowMotion->ASRRec Pipeline Implement ASRICA Pipeline ASRRec->Pipeline Pipeline->Components

Diagram 2: Method Selection Decision Framework. The optimal artifact removal strategy depends on available hardware and motion intensity.

Integrating artifact removal methods with ICA is essential for valid source-level analysis in mobile EEG research. Both ASR and iCanClean significantly improve ICA decomposition quality during motion, but demonstrate distinct performance characteristics and optimal application contexts. iCanClean generally outperforms ASR in recovering brain components and preserving neural signals, particularly when implemented with dual-layer EEG systems. However, ASR remains a robust, accessible option that requires no specialized hardware and provides substantial improvements over minimal preprocessing. The methodological guidelines and performance data presented in this comparison guide empower researchers to select and implement appropriate artifact removal strategies for their specific mobile brain imaging applications.

Achieving Peak Performance: Parameter Tuning and Troubleshooting for ASR & iCanClean

In mobile brain imaging, motion artifacts pose a significant challenge for analyzing electroencephalography (EEG) data collected during whole-body movement. Artifact Subspace Reconstruction (ASR) is a widely adopted algorithm for addressing this issue, with its performance heavily dependent on the selection of the threshold parameter k. This guide provides an objective comparison between ASR and iCanClean, another prominent artifact removal method, focusing on their efficacy in recovering neural signals during dynamic tasks like running. We summarize experimental data, detail methodologies, and provide visual resources to aid researchers in selecting and optimizing these tools for their specific applications.

Technical Comparison: ASR vs. iCanClean

The table below summarizes the core characteristics and performance metrics of ASR and iCanClean based on recent comparative studies.

Feature Artifact Subspace Reconstruction (ASR) iCanClean
Core Algorithm Principal component analysis (PCA) based on clean calibration data [1]. Canonical correlation analysis (CCA) using reference or pseudo-reference noise signals [1] [7].
Primary Use Case Preprocessing for mobile EEG to remove high-amplitude artifacts prior to ICA [1]. An all-in-one cleaning solution for multiple artifact types (motion, muscle, eye, line-noise) [7].
Key Parameter Threshold k (standard deviation cutoff). Lower k is more aggressive [1]. Correlation criterion R². Higher R² is less aggressive [1].
Optimal Parameter (from studies) k = 20-30 is recommended generally [1]; k = 10 may be better for running [1]. R² = 0.65 with a 4-s sliding window for locomotion [1].
Data Quality Score (Phantom Head, All Artifacts) Improved from 15.7% to 27.6% [7]. Improved from 15.7% to 55.9% (target for pure brain: 57.2%) [7].
ICA Component Dipolarity Improved dipolarity, facilitating better brain source separation [1] [8]. Produced more dipolar brain components than ASR [1].
Power Reduction at Gait Frequency Significant reduction [1] [8]. Significant reduction [1] [8].
P300 ERP Congruency Effect Recovery Produced ERP components similar to the standing task [1]. Recovered the expected P300 amplitude difference, similar to the standing task [1].

Experimental Protocols and Methodologies

Protocol: Comparative Evaluation During Overground Running

This protocol was designed to evaluate motion artifact removal during a dynamic Flanker task [1] [8].

  • Objective: To compare ASR and iCanClean based on ICA decomposition quality, reduction of gait-related spectral power, and recovery of stimulus-locked event-related potentials (ERPs).
  • Task: Participants performed a Flanker task while jogging on a treadmill and during static standing.
  • EEG Recording: Mobile EEG was recorded from young adult athletes.
  • Data Analysis:
    • ICA Dipolarity: The number of brain-independent components with a dipolar topography was used to measure ICA decomposition quality.
    • Spectral Power: Power at the fundamental gait frequency and its harmonics was analyzed before and after cleaning.
    • ERP Analysis: The P300 component from the Flanker task was compared between jogging and standing conditions to assess the recovery of neural signals.

Protocol: Phantom Head Validation with Ground Truth

This protocol used a phantom head with known brain sources to quantitatively assess cleaning performance [7].

  • Objective: To validate iCanClean's ability to remove multiple artifacts while preserving brain activity and compare it against ASR and other methods.
  • Apparatus: An electrically conductive phantom head with 10 embedded brain source antennae and 10 contaminating artifact sources.
  • Conditions: Data were collected under six scenarios: Brain only, and Brain combined with various artifacts (eyes, neck muscles, facial muscles, walking motion, and all artifacts combined).
  • Performance Metric: A Data Quality Score (0-100%) was calculated as the average correlation between the known ground-truth brain sources and the cleaned EEG channels.

Experimental Workflow and Signaling Pathways

The following diagram illustrates the logical workflow and key decision points for optimizing the ASR parameter k and comparing it with the iCanClean approach, based on the experimental protocols.

G Start Start: Raw Mobile EEG Data Subgraph1 Approach Selection Start->Subgraph1 Goal Goal: Clean EEG for Analysis Node_ASR Artifact Subspace Reconstruction (ASR) Subgraph1->Node_ASR Node_iCanClean iCanClean Subgraph1->Node_iCanClean Subgraph2 Core Parameter Tuning Node_ASR->Subgraph2 Node_R2 Select Correlation R² Node_iCanClean->Node_R2 Node_k Select Threshold k Subgraph2->Node_k Node_ASR_Param Node_ASR_Param Node_k->Node_ASR_Param Low k (e.g., 10) More Aggressive Node_ASR_Param2 Node_ASR_Param2 Node_k->Node_ASR_Param2 High k (e.g., 20-30) Less Aggressive Node_iCC_Param Node_iCC_Param Node_R2->Node_iCC_Param High R² (e.g., 0.65) Less Aggressive Subgraph3 Evaluation & Validation Node_Eval1 Assess ICA Component Dipolarity Subgraph3->Node_Eval1 Node_Eval2 Measure Power at Gait Frequency Subgraph3->Node_Eval2 Node_Eval3 Validate ERP Recovery (e.g., P300) Subgraph3->Node_Eval3 Node_Eval4 Calculate Data Quality Score (Phantom Head) Subgraph3->Node_Eval4 Node_Eval1->Goal Node_Eval2->Goal Node_Eval3->Goal Node_Eval4->Goal Node_ASR_Param->Subgraph3 Node_ASR_Param2->Subgraph3 Node_iCC_Param->Subgraph3

The Scientist's Toolkit: Essential Research Reagents and Materials

This table details key hardware, software, and methodological components essential for conducting research in mobile EEG motion artifact removal.

Item Function & Description
High-Density Mobile EEG System Enables the recording of brain activity during whole-body movement. Systems with 100+ channels are often recommended for optimal ICA decomposition [7].
Dual-Layer or Noise Sensors Specialized electrodes mechanically coupled to scalp electrodes but not in contact with the scalp. They capture only motion-related noise, providing ideal reference signals for algorithms like iCanClean [1] [7].
Electrical Phantom Head A benchtop apparatus with embedded simulated brain sources and contaminating artifact sources. It provides known ground-truth signals for quantitative validation and comparison of artifact removal algorithms [7].
Artifact Subspace Reconstruction (ASR) A PCA-based algorithm implemented in software (e.g., EEGLAB/BCILAB) for removing high-amplitude artifacts from continuous EEG. Its aggressiveness is controlled by the k parameter [1] [7].
iCanClean Algorithm A software-based cleaning solution that uses CCA to identify and subtract noise subspaces from EEG data. It can use signals from dual-layer sensors or create pseudo-reference signals from the EEG itself [1] [7].
Independent Component Analysis (ICA) A blind source separation method used after preprocessing to isolate brain and non-brain sources. The quality of its decomposition is a key metric for evaluating preprocessing methods [1].
ICLabel A machine learning-based EEGLAB plugin used to automatically classify ICA components as brain or various artifact types (e.g., eye, muscle, heart) [1].

Mobile brain imaging with electroencephalography (EEG) enables unprecedented opportunities for studying brain dynamics during natural movement, but motion artifacts significantly compromise data quality and subsequent independent component analysis (ICA). Among artifact removal algorithms, iCanClean has emerged as a powerful approach that uses canonical correlation analysis (CCA) and reference noise signals to remove artifact subspaces from contaminated EEG data. Unlike methods that require clean calibration data, iCanClean operates by identifying and subtracting noise components based on their correlation with reference signals, making it particularly suitable for mobile brain-body imaging (MoBI) applications. However, the algorithm's performance critically depends on two key parameters: the R² threshold (cleaning aggressiveness) and window length (temporal scope for correlation analysis). This guide provides evidence-based recommendations for optimizing these parameters through direct comparison with Artifact Subspace Reconstruction (ASR), supported by experimental data from recent validation studies.

Understanding iCanClean's Core Parameters

R² Threshold: Balancing Aggressiveness and Signal Preservation

The R² threshold determines which correlated components are removed during processing. This parameter represents the squared canonical correlation value between EEG channels and reference noise signals, with lower values resulting in more aggressive cleaning (removing components with lower correlation to noise) and higher values being more conservative [24]. Proper calibration is essential—overly aggressive cleaning (R² too low) risks removing neural signals, while overly conservative cleaning (R² too high) preserves more brain activity but may leave significant motion artifacts.

Window Length: Temporal Dynamics of Artifact Correlation

The window length defines the duration of data segments used for correlation analysis between EEG and noise signals [24]. Shorter windows (e.g., 1-2 seconds) can capture rapidly changing artifacts but may be susceptible to overfitting random correlations, while longer windows (e.g., 4+ seconds) provide more stable correlation estimates but might miss transient artifacts. The "infinite" window option uses the entire dataset for a single correlation calculation, potentially useful for stationary artifacts but less adaptable to dynamic motion patterns during mobility.

Experimental Evidence for Optimal Parameter Selection

Comprehensive Parameter Sweep Findings

A systematic parameter sweep study evaluated iCanClean performance across multiple populations (young adults, high-functioning older adults, and low-functioning older adults) during walking tasks with varying terrain difficulty [24]. Researchers conducted a full factorial investigation of R² thresholds (0.05 to 1.0 in 0.05 increments) and window lengths (1s, 2s, 4s, and infinite) using high-density dual-layer EEG (120 scalp electrodes + 120 noise electrodes).

Table 1: Optimal iCanClean Parameters from Experimental Evidence

Study Optimal R² Optimal Window Length Performance Improvement Experimental Context
Gonsisko et al. (2023) [24] 0.65 4 seconds 57% increase in good ICA components (8.4 to 13.2) Walking on varied terrain, dual-layer EEG
Downey & Ferris (2023) [7] 0.3-0.7 1-4 seconds Data Quality Score: 15.7% → 55.9% Phantom head with simulated artifacts

The parameter sweep revealed that optimal performance was achieved with an R² threshold of 0.65 and a 4-second window length [24]. At these settings, iCanClean improved the average number of "good" independent components (well-localized dipoles with high brain probability) from 8.4 to 13.2, representing a 57% increase in usable brain sources for subsequent analysis.

Impact on ICA Decomposition Quality

The primary metric for evaluating parameter optimization was the quality of subsequent ICA decomposition, quantified by the number of components meeting strict criteria for brain sources (residual variance < 15% in dipole fitting and >50% brain probability from ICLabel) [24]. The 4-second window significantly outperformed shorter windows (1s and 2s) across most R² values, while the infinite window showed inconsistent performance across participants. The R²=0.65 value represented the optimal balance between removing artifacts and preserving neural signals across all participant groups and terrain conditions.

Table 2: Performance Comparison Across Different Parameter Combinations

Parameter Combination Good ICA Components Artifact Removal Neural Signal Preservation Computational Efficiency
R²=0.65, 4s window 13.2 (optimal) Optimal Optimal Good
R²=0.3, 4s window 10.1 Overly aggressive Reduced Good
R²=0.8, 4s window 9.8 Insufficient Excellent Good
R²=0.65, 1s window 9.5 Variable Reduced Excellent
R²=0.65, infinite window 11.7 (inconsistent) Moderate Good Poor

Comparative Performance: iCanClean vs. ASR

Direct Method Comparison in Mobile EEG

Recent comparative studies have evaluated iCanClean against ASR under various movement conditions. During overground running with an adapted Flanker task, both methods improved ICA decomposition quality, but iCanClean demonstrated superior performance in several key metrics [1] [8]. Specifically, iCanClean with pseudo-reference noise signals produced more dipolar brain components and better preserved the expected P300 event-related potential congruency effects compared to ASR.

Phantom Head Validation

A comprehensive phantom head study with known ground-truth brain signals provided objective evidence of iCanClean's superiority under controlled conditions [7]. When confronted with multiple simultaneous artifacts (motion, muscle, eye, line-noise), iCanClean improved the Data Quality Score from 15.7% to 55.9%, substantially outperforming ASR (27.6%), Auto-CCA (27.2%), and Adaptive Filtering (32.9%). For context, the clean "Brain only" condition scored 57.2%, indicating iCanClean's ability to approximate artifact-free conditions.

G iCanClean vs. ASR Performance Comparison in Motion Artifact Removal cluster_inputs Input Conditions cluster_methods Processing Methods cluster_outputs Performance Metrics Contaminated_EEG Contaminated EEG (Brain + Motion Artifacts) iCanClean iCanClean (R²=0.65, 4s window) Contaminated_EEG->iCanClean ASR Artifact Subspace Reconstruction (ASR) Contaminated_EEG->ASR Reference_Signals Reference Noise Signals Reference_Signals->iCanClean Calibration_Data Clean Calibration Data Calibration_Data->ASR Good_Components Good ICA Components (Residual Variance <15%) iCanClean->Good_Components 13.2 components (+57%) Data_Quality Data Quality Score (0-100%) iCanClean->Data_Quality 55.9% ERP_Recovery ERP Component Recovery (P300 Congruency Effect) iCanClean->ERP_Recovery Preserved ASR->Good_Components ~9-11 components ASR->Data_Quality 27.6% ASR->ERP_Recovery Reduced/absent

Detailed Experimental Protocols

Parameter Optimization Study Design

The foundational parameter sweep study employed the following rigorous methodology [24]:

Participants and Data Collection:

  • 45 participants across three groups: young adults (YA), high-functioning older adults (HFOA), and low-functioning older adults (LFOA)
  • High-density dual-layer EEG recording with 120 scalp electrodes and 120 outward-facing noise electrodes
  • Walking tasks on a custom treadmill with varying terrain difficulty (Flat, Low, Medium, High) and walking speeds (0.25, 0.50, 0.75, 1.00 m/s)
  • Approximately 48 minutes of recording per participant

Data Processing Pipeline:

  • Basic preprocessing: 1 Hz high-pass filtering and average re-referencing
  • Robust channel rejection based on amplitude thresholds
  • iCanClean processing with parameter sweeps:
    • R² threshold: 0.05 to 1.0 in increments of 0.05
    • Window length: 1s, 2s, 4s, and infinite
  • ICA decomposition using Adaptive Mixtures ICA (AMICA)
  • Component evaluation:
    • Dipole localization quality (residual variance < 15%)
    • Brain source probability (ICLabel > 50%)

Performance Metrics:

  • Primary: Number of "good" independent components meeting both dipole and brain probability criteria
  • Secondary: Dipole residual variance distribution, spatial localization patterns

Comparative Validation Protocol

The comparative study between iCanClean and ASR employed this experimental design [1] [8]:

Participants and Task:

  • 21 young adult athletes performing adapted Flanker tasks
  • Two conditions: dynamic jogging and static standing
  • Wireless mobile EEG recording during overground running

Processing and Analysis:

  • Parallel processing pipelines for iCanClean and ASR
  • iCanClean implementation with pseudo-reference noise signals
  • ASR processing with recommended parameters (k=20-30)
  • Evaluation metrics:
    • ICA component dipolarity
    • Power spectral density changes at gait frequency and harmonics
    • P300 event-related potential recovery and congruency effects
    • Comparison to ground-truth standing condition

Table 3: Essential Research Materials for iCanClean Implementation and Validation

Resource Category Specific Tools/Solutions Function/Purpose Implementation Notes
EEG Hardware Dual-layer EEG systems (120+120 electrodes) Provides mechanical coupling between scalp and noise electrodes Enables direct noise reference recording [24]
Reference Signals Pseudo-reference noise signals Artificial noise references created via notch filtering Alternative when physical noise sensors unavailable [1]
Software Platforms MATLAB with EEGLAB toolbox Core processing environment for iCanClean Custom scripts + EEGLAB integration [24]
ICA Algorithms Adaptive Mixtures ICA (AMICA) High-quality source separation after cleaning Computationally intensive but superior to Infomax [24]
Component Validation ICLabel classifier Automated component categorization (brain vs. artifacts) Based on crowd-labeled training dataset [24]
Dipole Localization DIPFIT plugin for EEGLAB Quantitative component localization quality Residual variance <15% indicates good dipole fit [24]
Performance Benchmarking Phantom head apparatus Ground-truth validation with known brain sources Objective performance quantification [7]

Implementation Guidelines and Practical Considerations

Parameter Adaptation for Specific Research Contexts

While R²=0.65 with a 4-second window provides optimal performance for general mobile EEG applications, certain research contexts may benefit from parameter adjustments:

High-Motion Scenarios: For activities with intense, whole-body movement (e.g., running, sports), consider slightly more aggressive cleaning (R²=0.6-0.65) to address more pronounced motion artifacts [1].

Low-Motion Cognitive Tasks: For primarily stationary tasks with minor head movements, a more conservative approach (R²=0.7-0.75) may better preserve subtle neural signals associated with cognitive processing.

Limited Noise Channels: When using reduced noise channel configurations (64, 32, or 16 channels instead of 120), maintain the R²=0.65 threshold but consider slightly longer window lengths (4-5 seconds) to stabilize correlation estimates with fewer spatial samples [24].

Integration with Existing Processing Pipelines

Successful iCanClean implementation requires proper integration with standard EEG processing workflows:

  • Preprocessing Sequence: Apply iCanClean after basic filtering and channel rejection but before ICA decomposition [24]
  • Reference Signal Quality: Verify noise channel integrity using standard deviation thresholds and visual inspection
  • Validation Steps: Always include component quality metrics (dipolarity, ICLabel probabilities) when evaluating new parameter combinations
  • Comparative Baseline: Process a subset of data with ASR (k=20-30) as a performance reference point [1]

Optimal iCanClean parameter selection significantly enhances mobile EEG data quality and subsequent brain source identification. The evidence-based recommendations of R²=0.65 with a 4-second window length provide researchers with a validated starting point for motion artifact removal across diverse movement conditions. iCanClean consistently outperforms ASR in recovering usable brain components and preserving event-related neural dynamics, particularly when properly calibrated. These parameter guidelines enable researchers to maximize data quality in mobile brain imaging studies, advancing investigation of neural processes during natural human movement.

In mobile brain imaging and automatic speech recognition (ASR), the primary goal of data cleaning is to remove artifacts without distorting the underlying neural or linguistic signals. Overcleaning, the excessive removal of data, can strip away meaningful biological information, leading to inaccurate conclusions. This guide objectively compares the performance of Artifact Subspace Reconstruction (ASR) and iCanClean in balancing effective artifact removal with crucial data preservation.

Experimental Protocols and Performance Metrics

Key Comparative Experiments in Mobile EEG

Research directly comparing iCanClean and ASR has employed several rigorous experimental paradigms to evaluate their performance in removing motion artifacts, particularly during dynamic activities like walking and running.

  • Overground Running Flanker Task: A 2025 study recorded EEG from young adults during a Flanker task performed while jogging overground and while standing still. The approaches were evaluated on three key outcomes: the quality of the subsequent Independent Component Analysis (ICA) decomposition (measured by component dipolarity), the reduction in spectral power at the gait frequency and its harmonics, and the ability to recover expected event-related potential (ERP) components, such as the P300 congruency effect [8] [19].
  • Parameter Sweep on Human Walking Data: An earlier study aimed to find the optimal settings for iCanClean for data collected during human walking. Using a dual-layer EEG cap (120 scalp electrodes + 120 noise electrodes), researchers processed data from 45 participants across different age and functional groups. They performed a parameter sweep, varying the iCanClean window length (1s, 2s, 4s, infinite) and R² cleaning aggressiveness threshold (0.05 to 1.0). Performance was measured by the number of "good" independent components post-ICA—those that were well-localized as dipoles (residual variance < 15%) and had a high brain probability (>50% per ICLabel) [4].
  • Phantom Head Validation: A 2023 study used an electrically conductive phantom head with embedded, known brain sources to provide absolute ground truth. The phantom was subjected to six conditions: Brain alone, and Brain combined with various artifacts (Eyes, Neck Muscles, Facial Muscles, Walking Motion, and All Artifacts combined). The performance of cleaning algorithms was quantified using a Data Quality Score (0-100%), calculated as the average correlation between the known brain sources and the cleaned EEG channels [7].

Performance Data and Comparison

The following tables summarize key quantitative findings from these experiments, providing a direct comparison of the effectiveness of iCanClean and ASR.

Table 1: Performance in Preserving Brain Sources during Human Locomotion

Metric iCanClean Performance ASR Performance Experimental Context
ICA "Good" Components Increased from 8.4 to 13.2 (+57%) [4] Not directly compared in this study Walking on a treadmill; optimal settings (4s window, R²=0.65) [4]
ICA Component Dipolarity "Somewhat more effective" than ASR; produced more dipolar brain components [8] [19] Led to recovery of more dipolar brain components [8] [19] Overground running Flanker task [8] [19]
ERP Component Recovery Identified the expected P300 amplitude difference between congruent/incongruent stimuli [8] [19] Produced ERP components similar in latency to the static task [8] [19] Overground running Flanker task [8] [19]
Power at Gait Frequency Significantly reduced [8] [19] Significantly reduced [8] [19] Overground running [8] [19]

Table 2: Performance in Controlled Phantom Head Study

Condition iCanClean Data Quality ASR Data Quality Auto-CCA Data Quality Adaptive Filtering Data Quality
Brain (Target) 57.2% (uncleaned) [7] 57.2% (uncleaned) [7] 57.2% (uncleaned) [7] 57.2% (uncleaned) [7]
Brain + All Artifacts 55.9% [7] 27.6% [7] 27.2% [7] 32.9% [7]

Signaling Pathways and Workflows

The following diagrams illustrate the core logical workflows for the iCanClean and ASR algorithms, highlighting the points where parameter choices influence the risk of overcleaning.

iCanClean Workflow

G Start Start with Raw EEG NoiseSig Obtain Reference Noise Signals Start->NoiseSig PseudoNoise (Alternative) Create Pseudo-Reference via Notch Filter Start->PseudoNoise CCA Perform Canonical Correlation Analysis (CCA) NoiseSig->CCA PseudoNoise->CCA Identify Identify EEG Subspaces Correlated with Noise CCA->Identify UserParam User Defines Parameters: • Window Length (e.g., 4s) • R² Threshold (e.g., 0.65) UserParam->CCA Subtract Subtract Correlated Noise Components Identify->Subtract Output Output Cleaned EEG Subtract->Output

Artifact Subspace Reconstruction (ASR) Workflow

G Start Start with Raw EEG Calibrate Calibration: Identify 'Clean' Reference Data Start->Calibrate PCA Sliding-Window Principal Component Analysis (PCA) Calibrate->PCA UserParam User Defines Parameter: • Standard Deviation Threshold (k) Detect Detect Artifactual Components (Exceed k * SD of Reference) UserParam->Detect PCA->Detect Reconstruct Reconstruct Data using Reference PCA Subspace Detect->Reconstruct Output Output Cleaned EEG Reconstruct->Output

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for Mobile EEG Artifact Removal Research

Item / Solution Function & Description Relevance to Avoiding Overcleaning
Dual-Layer EEG Cap A specialized cap with paired electrodes: scalp electrodes (brain signal + noise) and mechanically coupled, outward-facing noise electrodes (noise only). Provides ideal reference noise signals for algorithms like iCanClean. [7] [4] Provides a true physical ground truth for noise, enabling highly precise subtraction and minimizing guesswork that can lead to brain signal removal.
Electrical Phantom Head A bench-top apparatus with embedded, known electrical "brain" sources. Allows for controlled introduction of motion, muscle, and other artifacts with a known ground truth signal. [7] The gold standard for validating cleaning algorithms without the risk of damaging precious human subject data. Enables absolute quantification of data preservation.
iCanClean Algorithm A novel cleaning algorithm that uses Canonical Correlation Analysis (CCA) and reference noise signals to detect and subtract noise subspaces from the EEG signal. [8] [7] [4] Its parameter-based approach (R², window length) allows for calibrated aggressiveness. Its use of noise references makes its cleaning more targeted.
Artifact Subspace Reconstruction (ASR) An algorithm that uses a calibration period of clean data to identify and remove high-variance, artifactual components in a sliding-window PCA approach. [8] [19] The k parameter controls aggressiveness; too low a value (e.g., <10) risks "overcleaning" and distorting the brain signal. [19]
ICLabel A standardized, automated component classification tool built into EEGLAB. Uses a neural network to label Independent Components (ICs) as brain, muscle, eye, etc. [4] Provides an objective metric for evaluating cleaning success. An increase in brain-classified ICs post-cleaning indicates successful data preservation.

The comparative data indicates that iCanClean often holds an advantage in preserving data integrity, especially in high-motion scenarios, as evidenced by its superior performance in phantom studies and its ability to recover more brain components during human locomotion. However, ASR remains a robust and effective method, particularly when configured with appropriate thresholds (e.g., k=20-30). The choice between them should be guided by the specific research context, availability of noise references, and a careful consideration of their respective parameters to navigate the critical pitfall of overcleaning.

Strategies for Handling High-Density EEG Configurations

Electroencephalography (EEG) is a fundamental tool in neuroscience research, offering non-invasive measurement of brain activity with millisecond temporal resolution. The advent of high-density EEG systems (typically with 64+ electrodes) has improved spatial resolution, enabling more precise source localization and detailed topographical analysis [25]. However, this advancement coincides with growing interest in mobile brain-body imaging (MoBI), where EEG is recorded during natural movement, making signals highly susceptible to motion artifacts [1] [2]. These artifacts originate from multiple sources, including cable sway, electrode-skin interface changes, and muscle activation, presenting as high-amplitude, non-brain signals that can obscure neural activity of interest [2] [7]. Effectively removing these artifacts is crucial for advancing research in areas such as gait analysis, athletic performance, and neurorehabilitation. Among various solutions, Artifact Subspace Reconstruction (ASR) and iCanClean have emerged as prominent preprocessing strategies. This guide provides a detailed comparison of their performance, supported by experimental data and methodological protocols.

Methodological Approaches

Artifact Subspace Reconstruction (ASR)

ASR is an automated, data-driven approach integrated into popular toolboxes like EEGLAB and BCILAB. It functions by identifying and removing high-variance components in EEG data that exceed a statistical threshold relative to a clean baseline calibration period [1] [7]. The method employs a sliding-window principal component analysis (PCA) to decompose the data. Components whose amplitude (root mean square) exceeds a user-defined standard deviation threshold ("k") are identified as artifactual. These components are then reconstructed using the calibration data's covariance matrix [1]. The critical parameter in ASR is the "k" value, which controls sensitivity. A lower k-value (e.g., k=10) results in more aggressive cleaning but risks "overcleaning" and potential removal of brain signals, whereas a higher k-value (e.g., k=20-30) is more conservative but may leave significant artifacts [1]. ASR operates without requiring dedicated reference noise sensors, instead leveraging clean data segments from the recording itself for calibration.

iCanClean

iCanClean is a more recent framework that leverages canonical correlation analysis (CCA) to identify and subtract noise subspaces from the EEG signal [1] [7]. Its core innovation lies in using reference noise signals to guide the cleaning process. Ideally, these are recorded from dedicated "dual-layer" sensors mechanically coupled to the EEG electrodes but not in contact with the scalp, capturing only motion-related noise. When such hardware is unavailable, iCanClean can generate pseudo-reference noise signals from the raw EEG itself, for instance, by applying a temporary notch filter to isolate activity below 3 Hz, which is likely non-neural [1]. CCA identifies components in the scalp EEG that are highly correlated with these noise references. The user selects a correlation threshold (R²), and components exceeding this threshold are projected back to the channel space and subtracted [1] [7]. This makes iCanClean particularly effective at targeting motion artifacts while preserving brain activity, even in real-time applications.

Emerging and Comparative Methods

Research in motion artifact removal is rapidly evolving, with new methods providing valuable performance benchmarks.

  • IMU-Enhanced Deep Learning: A recent approach fine-tunes a large brain model (LaBraM) using data from Inertial Measurement Units (IMUs) attached to the head. The IMU data, which directly quantifies motion, is used to guide an attention mechanism that identifies and removes motion-related artifacts from the EEG [5]. This represents a shift towards multi-modal, deep learning-based cleaning.
  • Motion-Net: This is a subject-specific, CNN-based deep learning model designed for motion artifact removal. It incorporates visibility graph (VG) features to enhance performance on smaller datasets and has been shown to achieve an average artifact reduction of 86% and a 20 dB improvement in signal-to-noise ratio (SNR) [2].
  • Conventional Filters and ICA: Standard high-pass and low-pass filters are often ineffective as motion artifact frequencies overlap with neural signals [2]. Independent Component Analysis (ICA) is a powerful blind source separation method but is computationally intensive, requires long data recordings for a good decomposition, and its performance is degraded by the presence of large motion artifacts [1] [7].

Table 1: Overview of Motion Artifact Removal Methods

Method Core Principle Key Parameters Reference Signals Required? Computational Load
ASR PCA-based reconstruction of high-variance segments "k" threshold (e.g., 10-30) No (uses clean calibration data) Low to Moderate [1] [7]
iCanClean CCA-based subtraction of noise subspaces R² correlation threshold (e.g., 0.65) Yes (physical or pseudo-reference) Moderate [1] [7]
IMU-LaBraM Deep learning with IMU-guided attention Model architecture, training data Yes (IMU sensors) High (requires training) [5]
Motion-Net Subject-specific 1D CNN Network layers, VG features No High (per-subject training) [2]
ICA Blind source separation Algorithm choice (e.g., Infomax, AMICA) No Very High (hours for decomposition) [7]

G cluster_ASR ASR Pathway cluster_iCanClean iCanClean Pathway Start Raw HD-EEG Data with Motion Artifacts Preproc Basic Filtering (High-pass, Notch) Start->Preproc ASR1 Calibration: Identify Clean Segments Preproc->ASR1 ICC1 Create Noise Reference (Dual-layer or Pseudo) Preproc->ICC1 ASR2 Sliding Window PCA ASR1->ASR2 ASR3 Reject Components > k ASR2->ASR3 ASR4 Reconstruct Data ASR3->ASR4 ASR_Out ASR-Cleaned EEG ASR4->ASR_Out Eval Performance Evaluation: Dipolarity, Power Spectrum, ERPs ASR_Out->Eval ICC2 Canonical Correlation Analysis (CCA) ICC1->ICC2 ICC3 Subtract Components > R² ICC2->ICC3 ICC_Out iCanClean-Cleaned EEG ICC3->ICC_Out ICC_Out->Eval

Diagram 1: Experimental workflow for comparing ASR and iCanClean cleaning pathways for high-density EEG data.

Experimental Performance Comparison

Phantom Head and Simulation Studies

Controlled studies using phantom heads with known ground-truth brain signals provide the most direct evidence of cleaning efficacy. A 2023 study used a conductive phantom head with 10 simulated brain sources and 10 contaminating sources (eyes, neck muscles, facial muscles, walking motion) [7]. The results were quantified using a Data Quality Score (DQS), which measures the average correlation between the known brain sources and the cleaned EEG channels.

Table 2: Performance on Phantom Head Data (Brain + All Artifacts Condition)

Method Initial DQS (%) Post-Cleaning DQS (%) Improvement (Percentage Points)
iCanClean 15.7 55.9 +40.2
ASR 15.7 27.6 +11.9
Auto-CCA 15.7 27.2 +11.5
Adaptive Filtering 15.7 32.9 +17.2

The data show that iCanClean dramatically outperformed other methods, more than tripling the DQS improvement of ASR. The study concluded that iCanClean offers superior ability to remove multiple artifact sources while faithfully preserving the underlying brain signal [7].

Human Studies During Locomotion

Validation in human participants during actual movement is critical. A 2025 study compared ASR and iCanClean for EEG data recorded during overground running while participants performed a Flanker task (a cognitive task that elicits a P300 event-related potential) [1] [26]. The evaluation was based on three key metrics:

  • ICA Component Dipolarity: Measures the quality of source separation; more dipolar components indicate better isolation of neural sources from artifacts [1].
  • Power at Gait Frequency: Effective cleaning should reduce power at the step frequency and its harmonics [1].
  • P300 ERP Congruency Effect: The ability to recover the expected neural response (higher P300 amplitude for incongruent Flanker stimuli) [1].

The study found that both iCanClean and ASR improved ICA decomposition, leading to a greater number of dipolar brain components compared to uncorrected data. However, iCanClean was "somewhat more effective than ASR" on this metric [1]. Both methods also significantly reduced power at the gait frequency. Crucially, only preprocessing with iCanClean successfully recovered the expected P300 congruency effect during running, demonstrating its superior ability to preserve task-relevant brain signals under high-motion conditions [1].

Table 3: Performance During Human Running (Flanker Task)

Evaluation Metric Uncleaned EEG After ASR After iCanClean
Dipolar Brain Components Low Improved Most Improved [1]
Power at Gait Frequency High Reduced Reduced [1]
P300 Congruency Effect Not Detected Not Detected Detected [1]

The Researcher's Toolkit

Implementing these strategies requires a suite of specific hardware and software solutions. Below is a list of essential "research reagents" for high-density EEG motion artifact research.

Table 4: Essential Research Reagents for HD-EEG Motion Artifact Research

Item Function & Importance Examples / Notes
High-Density EEG System Base system for recording neural data with high spatial resolution. 64+ channel systems; Dry MXene-based electrodes show promise for easier setup [27].
Active Electrodes Amplify signal at the source to reduce cable sway artifact. Critical for mobile EEG [7].
Dual-Layer Noise Sensors For optimal iCanClean performance; provide pure motion noise reference. Electrodes coupled to cap but not contacting scalp [1] [7].
Inertial Measurement Units (IMUs) Record head kinematics as a reference for motion artifact removal. Used in adaptive filtering and modern deep learning approaches [5].
Electrical Phantom Head Provides ground-truth signals for controlled algorithm validation. Contains embedded "brain" sources and contaminating sources [7].
EEGLAB Open-source MATLAB toolbox for EEG processing; includes ASR. Core platform for many analysis pipelines [1].
iCanClean Algorithm Software for CCA-based artifact removal. Available as a standalone implementation [1] [7].
BCILAB MATLAB toolbox for brain-computer interface research. Includes real-time capable implementation of ASR [7].

The empirical evidence from both phantom and human studies indicates that iCanClean generally outperforms ASR in removing motion artifacts from high-density EEG, particularly in challenging scenarios like running. Its use of noise references allows for more targeted cleaning, which better preserves the integrity of brain signals, including cognitive ERPs. However, ASR remains a powerful and widely accessible option, especially when external noise sensors are not available.

Future developments are leaning towards multi-modal sensing and deep learning. Integrating IMU data directly into cleaning algorithms, as seen with the IMU-enhanced LaBraM framework, shows significant promise [5]. Furthermore, subject-specific deep learning models like Motion-Net offer high accuracy, though they often require substantial computational resources and training data [2]. For researchers today, the choice between ASR and iCanClean depends on experimental setup, available hardware, and the level of cleaning precision required. iCanClean is recommended for studies involving vigorous motion where preserving fine neural details is paramount, while ASR provides a robust and efficient solution for less extreme scenarios.

Comparative Analysis of Computational Efficiency and Real-Time Suitability

Motion artifacts present a significant challenge for mobile electroencephalography (EEG) research, particularly in real-world applications such as brain-computer interfaces (BCIs), neurorehabilitation, and cognitive monitoring during physical activity. Two prominent computational approaches have emerged for addressing this problem: Artifact Subspace Reconstruction (ASR) and the iCanClean algorithm. This comparison guide provides an objective performance evaluation of these methods, focusing specifically on their computational efficiency and suitability for real-time implementation—critical considerations for researchers designing mobile brain imaging studies and clinical applications requiring immediate feedback.

Both algorithms employ distinct mathematical frameworks for artifact removal. ASR utilizes a sliding-window principal component analysis (PCA) to identify and remove high-variance components that exceed a statistically defined threshold, making it particularly effective for burst-type artifacts caused by motion [19] [1]. In contrast, iCanClean employs canonical correlation analysis (CCA) to identify and remove subspaces of EEG data that are strongly correlated with reference noise signals, enabling targeted removal of artifacts that correlate with measured noise sources [7] [14]. Understanding their fundamental operational differences is essential for selecting the appropriate tool based on specific research requirements, whether for offline analysis or real-time BCI applications.

Methodological Approaches

Core Algorithmic Principles

The computational frameworks of ASR and iCanClean dictate their performance characteristics and implementation suitability. ASR operates by first establishing a calibration period of clean EEG data, typically collected during a stationary baseline. This reference data is used to compute the covariance matrix and principal components representing normal brain activity [19] [1]. During processing, a sliding window moves through the continuous EEG data, performing PCA on each segment. Components whose variance exceeds a user-defined threshold ("k" parameter, typically 20-30 for stationary data but sometimes lowered to 10 for locomotion studies) are identified as artifactual. These components are then reconstructed using the clean calibration data, effectively removing high-amplitude artifacts while preserving neural activity that falls within the normal variance range [19]. This approach makes ASR particularly effective for removing sudden, high-amplitude artifacts but dependent on the quality of the initial calibration period.

iCanClean employs a different strategy based on canonical correlation analysis with reference noise signals. The algorithm identifies linear subspaces of the contaminated EEG data that are maximally correlated with subspaces of the reference noise recordings [7] [14]. When dedicated noise sensors are available (such as in dual-layer EEG systems where mechanically coupled electrodes record only environmental and motion noise), iCanClean can directly separate noise from brain signals. In the absence of dedicated noise sensors, the algorithm can generate "pseudo-reference" noise signals by applying a temporary notch filter to the raw EEG to isolate potential noise components, such as those below 3 Hz [19]. The core computation involves solving a CCA problem to find components that maximize correlation between EEG and noise subspaces, then subtracting those components that exceed a user-defined R² threshold (typically 0.65 for mobile EEG) [4]. This targeted approach allows iCanClean to remove multiple artifact types simultaneously while preserving neural signals that don't correlate with the noise references.

Experimental Validation Protocols

Research studies have employed rigorous methodologies to evaluate the performance of both algorithms, utilizing phantom head models and human subjects during various movement conditions. Phantom head experiments involve electrically conductive models with embedded artificial brain sources, allowing precise quantification of algorithm performance against known ground-truth signals [7]. In these controlled setups, artificial artifacts (motion, muscle, eye, line-noise) are introduced systematically, and algorithm effectiveness is measured using Data Quality Scores based on correlation coefficients between original brain sources and cleaned EEG outputs [7].

Human studies typically involve participants performing tasks while wearing mobile EEG systems, often with dual-layer electrodes for noise recording. Common protocols include adapted Flanker tasks during static standing, walking, and running conditions, enabling researchers to evaluate how well each algorithm preserves known neural signatures (e.g., P300 event-related potentials) while removing movement artifacts [19] [1]. Performance metrics typically include: dipolarity of independent components (residual variance <15%), number of brain-like components identified by ICLabel, reduction in spectral power at gait frequency harmonics, and preservation of expected ERP components [19] [4]. These comprehensive evaluations provide the empirical basis for comparing the real-world effectiveness of each approach.

Performance Comparison

Computational Efficiency Analysis

Computational efficiency is a critical factor determining the real-time applicability of artifact removal algorithms. Based on parameter sweeps and performance testing, iCanClean demonstrates superior computational characteristics for real-time implementation compared to ASR and traditional methods like ICA [7] [4]. The algorithm's efficiency stems from its use of canonical correlation analysis, which is computationally less demanding than the principal component analysis employed by ASR, particularly when processing high-density EEG systems with numerous channels.

Independent Component Analysis (ICA), while effective for artifact removal in offline processing, is computationally intensive and often requires 5+ hours to decompose high-density (100+ channel) EEG data on standard computers, making it unsuitable for real-time applications [7]. ASR improves upon this by using a sliding-window PCA approach, but its performance is highly dependent on the quality of the initial calibration data and the chosen threshold parameter [19] [1]. iCanClean achieves significantly faster processing times while maintaining high performance, with optimal results obtained using a 4-second window and R² threshold of 0.65 [4]. This balance of computational efficiency and cleaning effectiveness makes iCanClean particularly suitable for real-time BCI applications where low latency is critical.

Table 1: Computational Characteristics of Motion Artifact Removal Algorithms

Algorithm Computational Method Processing Speed Optimal Parameters Hardware Requirements
iCanClean Canonical Correlation Analysis (CCA) Suitable for real-time; faster than ASR and ICA 4-s window, R²=0.65 [4] Standard computing hardware
ASR Sliding-window Principal Component Analysis Moderate; suitable for near-real-time k=10-30 (lower for locomotion) [19] [1] Standard computing hardware
ICA Blind Source Separation 5+ hours for high-density EEG; not real-time [7] Data-dependent High-performance computing for large datasets
Adaptive Filtering Linear Regression Suitable for real-time Filter length and step-size dependent Minimal; suitable for embedded systems
Artifact Removal Effectiveness

Quantitative comparisons demonstrate distinct performance differences between ASR and iCanClean across various artifact conditions. In phantom head studies with known ground-truth signals, iCanClean consistently outperforms ASR across multiple artifact types. When all artifacts were simultaneously present (motion, muscle, eye, line-noise), iCanClean improved Data Quality Scores from 15.7% to 55.9%, while ASR only achieved improvement to 27.6% [7]. This substantial performance gap highlights iCanClean's effectiveness in challenging multi-artifact environments commonly encountered in mobile EEG research.

In human locomotion studies during running, both algorithms significantly reduce motion artifacts but with notable differences in outcome quality. iCanClean preprocessing produces more dipolar independent components than ASR, indicating superior source separation [19] [1]. Furthermore, iCanClean successfully recovers the expected P300 event-related potential congruency effects during a dynamic Flanker task adapted for running, whereas ASR shows limited capability in preserving this subtle neural signature [19]. Both algorithms effectively reduce spectral power at the gait frequency and its harmonics, but iCanClean demonstrates better preservation of neural signals while removing artifacts, as evidenced by improved ICA decomposition quality and component dipolarity [4].

Table 2: Performance Metrics for Motion Artifact Removal During Human Locomotion

Performance Metric iCanClean ASR Experimental Context
Data Quality Score Improvement 15.7% → 55.9% (all artifacts) [7] 15.7% → 27.6% (all artifacts) [7] Phantom head with simulated artifacts
Good Brain Components Recovered 13.2 (improvement from 8.4) [4] Fewer than iCanClean [19] Human walking with high-density EEG
P300 ERP Recovery Preserved congruency effect [19] Limited recovery [19] Flanker task during running
Power Reduction at Gait Frequency Significant reduction [19] [1] Significant reduction [19] [1] Overground running
Muscle Artifact Removal Effective [7] Effective [19] Phantom and human studies

Implementation Considerations

Reference Signal Requirements

A fundamental implementation difference between ASR and iCanClean lies in their requirements for reference data. ASR depends on a clean calibration period recorded during stationary baseline conditions, which is used to establish normal variance parameters for identifying artifacts during movement conditions [19] [1]. This dependency presents a potential limitation if the calibration data quality is compromised or if the statistical properties of brain activity change significantly between resting and task states. Recent research has identified limitations in ASR's algorithm for identifying appropriate reference periods, which may explain its reduced effectiveness with higher recommended k values for addressing high-amplitude motion artifacts [19].

iCanClean offers greater flexibility in reference signal options, working with both dedicated noise sensors and pseudo-reference signals derived from the EEG data itself. When dual-layer EEG systems are available with mechanically coupled noise electrodes, iCanClean achieves optimal performance by directly measuring motion artifacts and environmental noise [7] [4]. However, the algorithm maintains effective performance using pseudo-reference signals created by applying temporary notch filters to the raw EEG data, making it adaptable to standard EEG systems without specialized hardware [19]. This flexibility expands iCanClean's applicability across diverse research settings with varying equipment capabilities.

Parameter Optimization

Both algorithms require careful parameter selection to balance artifact removal effectiveness against potential signal distortion. For ASR, the critical parameter is the threshold "k," which determines the sensitivity for detecting artifactual components. While values of 20-30 are recommended for stationary data, more aggressive thresholds (as low as 10) may be necessary for locomotion studies, though these risk "overcleaning" and removing genuine neural activity [19] [1]. The algorithm's performance is highly sensitive to this parameter selection, requiring empirical optimization for specific movement paradigms.

iCanClean's primary parameters are the window length for correlation analysis and the R² threshold determining cleaning aggressiveness. Extensive parameter sweeps have identified optimal values of 4-second windows with R²=0.65 for mobile EEG data during walking [4]. The algorithm demonstrates robust performance across reduced noise channel counts, maintaining effectiveness with as few as 16 noise channels [4]. This consistent performance across parameter variations contributes to iCanClean's reliability in real-world applications where signal quality may vary substantially within and across recording sessions.

Advanced Hybrid Approaches

Recent research has explored hybrid approaches that combine multiple algorithms to leverage their complementary strengths. The ASR-ICA pipeline represents one such hybrid method, where ASR performs initial aggressive cleaning of high-amplitude motion artifacts, followed by ICA for finer separation of neural sources from residual artifacts [5]. This approach recognizes that large motion artifacts can compromise ICA's ability to identify maximally independent components, making preprocessing with ASR beneficial [19].

Emerging deep learning methods show promise for further enhancing motion artifact removal. The IMU-enhanced LaBraM framework incorporates inertial measurement unit data using attention mechanisms to identify motion-related artifacts in EEG signals [5]. By fine-tuning large brain models on multi-modal data, these approaches achieve improved robustness across diverse movement scenarios compared to traditional ASR-ICA pipelines [5]. However, these advanced methods currently require substantial computational resources and training data, potentially limiting their real-time applicability in resource-constrained settings.

Research Reagent Solutions

Table 3: Essential Research Tools for Mobile EEG Motion Artifact Research

Tool Category Specific Examples Function in Research
EEG Systems Dual-layer EEG (120+120 electrodes) [4], Conventional disk electrodes, Tripolar concentric ring electrodes [20] Signal acquisition with varying noise rejection capabilities
Reference Sensors Dedicated noise electrodes [7], Inertial Measurement Units (IMUs) [5], Electrooculogram (EOG) electrodes Providing noise reference signals for artifact removal algorithms
Validation Platforms Phantom head models with embedded sources [7], Robotic motion platforms [7] Ground-truth validation of algorithm performance
Software Tools EEGLAB [19], iCanClean plugin [19] [4], ASR implementation in BCILAB [7] Algorithm implementation and signal processing
Experimental Paradigms Adapted Flanker tasks during locomotion [19] [1], Walking at varying speeds [4], Uneven terrain walking [4] Creating controlled conditions for algorithm evaluation

The comparative analysis of computational efficiency and real-time suitability reveals a complex performance landscape for motion artifact removal algorithms. iCanClean demonstrates superior computational efficiency and effectiveness in preserving neural signals across diverse artifact conditions, making it particularly suitable for real-time applications where both processing speed and signal quality are critical [7] [4]. Its flexible reference signal requirements further enhance its applicability across various research settings.

ASR provides a robust alternative, particularly for scenarios with clear baseline periods and predominantly high-amplitude, burst-type artifacts [19] [1]. Its simpler statistical framework may be advantageous in implementation-constrained environments. Emerging hybrid approaches and deep learning methods show promising directions for future development, potentially combining the strengths of multiple algorithms while addressing their individual limitations [5].

Selection between these approaches should be guided by specific research requirements, including real-time processing needs, available sensor systems, movement characteristics, and computational resources. As mobile EEG research continues to expand into more ecologically valid and dynamic settings, ongoing algorithm development will further enhance our capability to isolate clean neural signals during natural human movement.

G Motion Artifact Removal Algorithm Decision Framework start Start: Mobile EEG with Motion Artifacts decision1 Real-time Processing Required? start->decision1 decision2 Dedicated Noise Sensors Available? decision1->decision2 No iCanClean_real iCanClean (Real-time Optimized) decision1->iCanClean_real Yes decision3 Artifact Type decision2->decision3 No iCanClean_offline iCanClean (High Performance) decision2->iCanClean_offline Yes decision4 Computational Resources decision3->decision4 Mixed Artifacts ASR_near_real ASR (Near Real-time) decision3->ASR_near_real Burst Artifacts decision4->iCanClean_offline Adequate ASR_offline ASR (Calibration Dependent) decision4->ASR_offline Limited hybrid Hybrid ASR-iCanClean or Deep Learning ICA ICA (Offline Only)

Algorithm Selection Framework

G Computational Pipeline Comparison cluster_ASR ASR Processing Pipeline cluster_iCanClean iCanClean Processing Pipeline ASR1 1. Clean Calibration Data Collection ASR2 2. Calculate Reference Covariance Matrix ASR1->ASR2 ASR3 3. Sliding Window PCA on Continuous EEG ASR2->ASR3 ASR4 4. Identify High-Variance Components (k threshold) ASR3->ASR4 ASR5 5. Reconstruct Using Reference Data ASR4->ASR5 ASR_perf Processing: Moderate Dependency: Calibration Quality Strength: Burst Artifacts ICC1 1. Obtain Reference Noise (Sensors or Pseudo-Reference) ICC2 2. Canonical Correlation Analysis (CCA) ICC1->ICC2 ICC3 3. Identify Correlated Subspaces (R² threshold) ICC2->ICC3 ICC4 4. Subtract Noise Components Using Least-Squares ICC3->ICC4 ICC5 5. Output Cleaned EEG ICC4->ICC5 ICC_perf Processing: Efficient Dependency: Noise Reference Strength: Mixed Artifacts

Computational Pipeline Comparison

Head-to-Head Validation: A Comparative Analysis of ASR and iCanClean Efficacy

The advancement of mobile brain–body imaging (MoBI) relies on the ability to extract high-fidelity brain signals from electroencephalography (EEG) data contaminated by motion artifacts. Among the various artifact correction methods available, Artifact Subspace Reconstruction (ASR) and iCanClean have emerged as two prominent, real-time-capable preprocessing techniques [7]. A critical challenge in the field is the objective evaluation of their performance using robust, quantitative validation metrics. This guide provides a comparative analysis of ASR and iCanClean, framed within the broader thesis of performance evaluation for motion artifact removal. We focus on three principal validation metrics—ICA component dipolarity, ICLabel probabilities, and spectral power reduction at the gait frequency—summarizing experimental data and detailing the protocols used to acquire it, providing researchers with a clear framework for assessment.

The following tables synthesize quantitative results from key studies that directly or indirectly compared ASR and iCanClean across the critical validation metrics.

Table 1: Comparative Performance on ICA Decomposition Quality

Metric iCanClean Performance ASR Performance Experimental Context & Notes
Number of "Good" Brain Components (Residual Variance < 15%, ICLabel > 50%) 13.2 components on average [24] Information Missing During walking; Baseline before cleaning: 8.4 components [24].
Improvement in Good Components +57% from baseline [24] Information Missing Compared to basic preprocessing [24].
Comparative Dipolarity Recovery More dipolar brain components recovered [1] More dipolar brain components recovered [1] During running; iCanClean was "somewhat more effective than ASR" [1].

Table 2: Efficacy in Artifact Power Reduction and ERP Recovery

Metric iCanClean Performance ASR Performance Experimental Context
Spectral Power at Gait Frequency Significant reduction [1] Significant reduction [1] During running [1].
Recovery of ERP Components Produced ERP components similar to static task [1] Produced ERP components similar to static task [1] P300 latency during running Flanker task [1].
Recovery of P300 Congruency Effect Successful identification [1] Not successfully identified [1] During running Flanker task [1].

Table 3: Performance Against Diverse Artifacts (Phantom Head Study)

Artifact Condition iCanClean Data Quality Score ASR Data Quality Score Notes
Brain + All Artifacts (Motion, Muscle, Eye, Line-Noise) 55.9% [7] 27.6% [7] Baseline "Brain" condition scored 57.2% without cleaning [7].
Starting Quality (Before Cleaning) 15.7% [7] 15.7% [7]

Detailed Experimental Protocols

The quantitative data presented above were derived from rigorous experimental protocols designed to stress-test the artifact removal capabilities of each algorithm.

Overground Running and Flanker Task Protocol

This protocol was designed to evaluate motion artifact removal during a dynamic, ecologically valid task [1].

  • Participants & Task: Young adults performed an adapted Eriksen Flanker task under two conditions: while jogging on an overground treadmill and during static standing [1].
  • EEG Recording: EEG data was recorded using a wireless mobile system during both task conditions.
  • Artifact Processing: The recorded data was preprocessed using both iCanClean (with pseudo-reference noise signals) and ASR.
  • Validation & Metrics:
    • ICA & Dipolarity: Data were decomposed via ICA. The quality of the decomposition was assessed by the number of independent components (ICs) that were well-localized by a dipole model (low residual variance) [1].
    • Spectral Power: Power spectral density was examined to quantify the reduction in power at the fundamental frequency of gait and its harmonics [1].
    • ERP Analysis: The ability to recover the stimulus-locked P300 event-related potential (ERP) and its expected "congruency effect" (higher amplitude for incongruent vs. congruent Flanker stimuli) was evaluated and compared to the static standing condition [1].

Treadmill Walking and Component Quality Protocol

This study focused on optimizing iCanClean parameters and establishing a benchmark for high-quality brain component yield [24] [4].

  • Participants & Task: Young adults, high-functioning older adults, and low-functioning older adults walked on a treadmill at various speeds and over terrains of varying difficulty [24].
  • EEG Recording: High-density (120-channel) EEG was recorded alongside 120 mechanically coupled but electrically isolated noise electrodes (dual-layer EEG) [24].
  • Artifact Processing: A parameter sweep was conducted for iCanClean, varying its two key parameters: the cleaning aggressiveness (R² threshold from 0.05 to 1) and the window length (1, 2, 4, and infinite seconds) [24].
  • Validation & Metrics:
    • ICA & ICLabel: Data were decomposed using ICA. A "good" component was defined as one that was both well-localized by a single dipole (residual variance < 15%) and had a high probability of being a brain source as classified by the ICLabel algorithm (>50% probability) [24].
    • Optimal Parameters: The number of "good" components was used to determine the optimal parameter set for iCanClean (4-s window, R²=0.65) [24].

Phantom Head Validation Protocol

This study employed a ground-truth paradigm to evaluate the absolute performance of cleaning algorithms in a controlled setting [7].

  • Apparatus: An electrically conductive phantom head was embedded with 10 simulated brain signal sources and 10 contaminating artifact sources [7].
  • Tested Conditions: Data were collected under six conditions: Brain only, and Brain combined with various artifacts (Eyes, Neck Muscles, Facial Muscles, Walking Motion, and All Artifacts combined) [7].
  • Artifact Processing: The corrupted EEG data was processed using iCanClean, ASR, Auto-CCA, and Adaptive Filtering.
  • Validation & Metrics: A Data Quality Score (0-100%) was calculated as the average correlation between the known ground-truth brain sources and the cleaned EEG channels. This provides a direct measure of how well each algorithm preserves the signal of interest while removing contaminants [7].

Technical Implementation and Workflows

Algorithm Processing Pipelines

The following diagram illustrates the core signal processing workflows for both the ASR and iCanClean algorithms, highlighting their conceptual differences.

G cluster_ASR Artifact Subspace Reconstruction (ASR) cluster_iCanClean iCanClean start Raw Contaminated EEG asr1 1. Identify clean calibration data from the recording start->asr1 ic1 1. Obtain reference noise signals (Dual-layer or Pseudo-reference) start->ic1 asr2 2. Sliding-window PCA on continuous data asr1->asr2 asr3 3. Detect & remove components variance > threshold (k) asr2->asr3 asr4 4. Reconstruct data using remaining components asr3->asr4 post Downstream Analysis (ICA, Time-Frequency, ERP) asr4->post ic2 2. Canonical Correlation Analysis (CCA) between EEG and noise signals ic1->ic2 ic3 3. Identify & remove EEG subspaces correlated with noise (R² threshold) ic2->ic3 ic4 4. Reconstruct cleaned EEG using remaining subspaces ic3->ic4 ic4->post

Diagram 1: ASR and iCanClean Processing Workflows

Experimental Validation Paradigm

This diagram outlines the structure of a comprehensive validation experiment, such as the overground running study, which integrates behavioral tasks, EEG recording, and multiple validation metrics.

G cluster_exp Experimental Paradigm cluster_metrics Validation Metrics task Dual-Task Design cond1 Condition 1: Dynamic Task (e.g., Running) task->cond1 cond2 Condition 2: Static Control (e.g., Standing) task->cond2 eeg Wireless Mobile EEG Recording cond1->eeg cond2->eeg proc Parallel Preprocessing (ASR vs. iCanClean) eeg->proc metric1 ICA Decomposition (Dipolarity, ICLabel) proc->metric1 metric2 Spectral Analysis (Power at Gait Frequency) proc->metric2 metric3 Event-Related Potentials (P300 Amplitude/Latency) proc->metric3

Diagram 2: Experimental Validation Paradigm

The Scientist's Toolkit: Key Research Reagents and Solutions

Table 4: Essential Materials and Tools for Mobile EEG Artifact Research

Item Function in Research Example Use Case
Dual-Layer EEG System Scalp electrodes record brain signal + noise; mechanically coupled noise electrodes record only environmental and motion noise. Provides optimal reference for noise-cancellation algorithms like iCanClean. [24] [28] Provides ground-truth noise reference for iCanClean during walking and table tennis studies. [24] [28]
Wireless EEG Amplifiers Enables data collection during full-body movement without cable-induced artifacts. Essential for ecologically valid MoBI studies. [29] Recording EEG during running, skateboarding, or table tennis. [1] [22] [28]
Inertial Measurement Units (IMUs) Measures head and body kinematics (acceleration, rotation). Can be used as a motion reference or for timing event onset. [28] Placed on forehead, body, or equipment to synchronize motion with EEG data and identify motion-related artifacts. [29] [28]
Phantom Head Apparatus Provides a ground-truth system with known brain signals and controllable artifacts. Allows for rigorous, quantitative benchmarking of algorithms without biological variability. [7] Comparing the absolute performance of ASR, iCanClean, and other methods against a known signal. [7]
Standardized Behavioral Tasks Tasks like the Flanker task elicit well-known neural responses (e.g., P300 ERP). Used to validate the preservation of brain signals after artifact removal. [1] Assessing whether neural correlates of cognition can be recovered after cleaning of data recorded during motion. [1] [22]

Electroencephalography (EEG) is a crucial tool for non-invasively recording brain activity with high temporal resolution, making it particularly valuable for studying neural dynamics during mobile conditions [7] [10]. However, a significant limitation of EEG is its vulnerability to various artifacts, including those from motion, muscle activity, eye movements, and line noise [7]. These artifacts pose substantial challenges for isolating true electrocortical signals, especially in real-world movement scenarios.

Phantom head studies provide an essential methodology for addressing these challenges by enabling rigorous benchmarking of artifact removal algorithms with known ground-truth data [7]. Unlike human studies where true brain signals are unknown, phantom heads incorporate electrically simulated "brain" sources, allowing researchers to quantitatively evaluate how effectively different algorithms remove contaminants while preserving neural signals [7] [10]. This review focuses on comparing the performance of Artifact Subspace Reconstruction (ASR) and iCanClean using phantom head validation, providing researchers with evidence-based guidance for selecting appropriate motion artifact removal methods.

Experimental Protocols in Phantom Head Validation

Phantom Head Apparatus Design

The foundation of reliable benchmarking rests on phantom heads that realistically simulate human electrophysiology [7]. In the key study validating iCanClean, researchers created an electrically conductive phantom head with embedded components designed to replicate real recording conditions [7]. The apparatus included 10 simulated brain sources that served as ground-truth signals, alongside 10 contaminating sources representing common artifacts [7]. The design incorporated critical elements including a simulated scalp and hair to improve the ecological validity of the testing environment [7].

This comprehensive approach allowed researchers to systematically test artifact removal algorithms under controlled conditions with precisely known brain signals, enabling direct calculation of how much true neural information was preserved after cleaning [7].

Experimental Conditions and Testing Protocol

To ensure thorough evaluation, researchers tested algorithms across six distinct conditions representing common contamination scenarios [7]:

  • Brain: Only simulated brain activity without contaminants
  • Brain + Eyes: Brain signals with ocular artifacts
  • Brain + Neck Muscles: Brain signals with neck muscle artifacts
  • Brain + Facial Muscles: Brain signals with facial muscle artifacts
  • Brain + Walking Motion: Brain signals with gait-related motion artifacts
  • Brain + All Artifacts: Combined contamination from all artifact sources [7]

This systematic approach enabled researchers to evaluate how each algorithm performed against specific artifact types as well as complex, multi-source contamination.

Performance Quantification Method

The core metric for comparing algorithm performance was the Data Quality Score (ranging 0-100%), calculated as the average correlation between the known brain sources and the cleaned EEG channels [7]. This quantitative approach provided an objective measure of how effectively each method removed artifacts while preserving legitimate brain signals [7]. A score of 100% would indicate perfect preservation of brain activity with complete artifact removal, while lower scores reflected either insufficient cleaning or excessive removal of brain content [7].

Quantitative Performance Comparison

The following table summarizes the performance of four artifact removal methods across different contamination scenarios, based on phantom head validation studies:

Table 1: Performance Comparison of EEG Artifact Removal Methods Using Phantom Head Validation

Artifact Condition iCanClean ASR Auto-CCA Adaptive Filtering No Cleaning
Brain + All Artifacts 55.9% 27.6% 27.2% 32.9% 15.7%
Brain Only (target reference) - - - - 57.2%
Walking Motion Detailed results show iCanClean superiority [7] - - - -
Muscle Artifacts Detailed results show iCanClean superiority [7] - - - -
Ocular Artifacts Detailed results show iCanClean superiority [7] - - - -

The most striking performance difference emerged in the "Brain + All Artifacts" condition, where iCanClean achieved a Data Quality Score nearly double that of other methods [7]. This demonstrates iCanClean's particular strength in handling complex, real-world scenarios where multiple artifact types coexist.

Performance in Human Application Studies

Recent research has extended these phantom head findings to human applications during dynamic motor tasks. A 2025 comparative study evaluated motion artifact removal during overground running using a Flanker task paradigm [8] [1] [19]. The study found that both iCanClean (using pseudo-reference noise signals) and ASR improved subsequent independent component analysis (ICA) decomposition quality, with iCanClean showing somewhat greater effectiveness [8] [1].

Specifically, preprocessing with iCanClean enabled identification of the expected P300 event-related potential congruency effect during running, which has significance for studying cognitive processes during locomotion [8] [1]. Both methods significantly reduced power at the gait frequency and its harmonics, indicating effective motion artifact suppression [8] [1].

Methodological Approaches

iCanClean Algorithm Workflow

iCanClean (implementing Canonical correlation to Cancel Latent Electromagnetic Artifacts and Noise) employs a four-step process for artifact removal [7] [1]:

G A Input Corrupt EEG Data B Identify Candidate Noise Components via CCA with Reference Signals A->B C Select Noise Components Based on R² Correlation Threshold B->C D Reconstruct and Subtract Noise from Original EEG Signals C->D E Output Cleaned EEG Data D->E

Diagram 1: iCanClean Algorithm Workflow

The algorithm leverages canonical correlation analysis (CCA) to identify noise subspaces within the EEG data [1]. A key advantage is its flexibility in reference signal sources: it can utilize dedicated noise sensors (such as dual-layer electrodes) when available, or generate pseudo-reference noise signals from the raw EEG itself when dedicated sensors are absent [1]. For pseudo-reference operation, the method temporarily applies notch filtering to identify noise-dominated frequency components within the EEG, which then serve as reference signals for the CCA process [1].

Artifact Subspace Reconstruction (ASR) Approach

ASR employs a different methodological foundation based on principal component analysis (PCA) [1] [19]. The method requires clean calibration data, which can be either user-supplied (e.g., from resting-state recordings) or automatically extracted from contaminated data assuming clean segments exist [7]. The algorithm works by:

  • Calculating the root mean square (RMS) of sliding 1-second data windows
  • Converting RMS values to z-scores using a condensed Gaussian distribution
  • Identifying reference data segments with z-scores between -3.5 and 5.0 for at least 92.5% of electrodes
  • Performing sliding-window PCA on both reference and non-reference data
  • Identifying artifactual components where standard deviation of RMS exceeds a user-defined threshold ("k")
  • Reconstructing the time series based on calibration data [1] [19]

ASR's performance depends heavily on the k threshold parameter, with lower values (e.g., 10-20) producing more aggressive cleaning but potentially risking overcleaning [1].

Phantom Head Validation Workflow

The comprehensive approach to validating artifact removal methods using phantom heads involves multiple systematic stages:

G A Design Conductive Phantom Head with Embedded Brain Sources B Introduce Controlled Artifacts (Motion, Muscle, Ocular, Line Noise) A->B C Record Contaminated EEG Signals with Known Ground Truth B->C D Apply Artifact Removal Algorithms (Parameter Sweeps) C->D E Calculate Data Quality Score (Correlation with Ground Truth) D->E F Compare Algorithm Performance Across Multiple Conditions E->F

Diagram 2: Phantom Head Validation Workflow

The Scientist's Toolkit

Table 2: Essential Research Tools for Phantom Head EEG Studies

Tool/Resource Function & Application Key Features
Electrical Phantom Head Simulates human head conductivity with embedded brain sources; provides ground-truth signals for validation [7] Includes scalp, hair, 10+ simulated brain sources, 10+ contaminating sources
Dual-Layer EEG Sensors Dedicated noise reference electrodes; mechanically coupled to scalp electrodes but not in contact with skin [1] Provides reference noise signals uncontaminated by brain activity
Robotic Motion Platform Induces controlled motion artifacts simulating human movement patterns [11] Standardized, reproducible motion artifact generation
iCanClean Software Comprehensive artifact removal implementing canonical correlation analysis [7] Real-time capability; works with dedicated or pseudo-reference signals
Artifact Subspace Reconstruction (ASR) Alternative artifact removal based on principal component analysis [1] [19] Included in EEGLAB; requires clean calibration data
Data Quality Score Metric Quantitative performance assessment (0-100%) [7] Based on correlation between cleaned data and ground-truth sources

Discussion and Research Implications

The consistent superiority of iCanClean in phantom head validation studies demonstrates its value for mobile EEG applications where multiple artifact types coexist [7]. The method's ability to achieve a Data Quality Score of 55.9% in the challenging "all artifacts" condition—compared to 15.7% before cleaning and 57.2% for uncontaminated brain signals—suggests it preserves most neural information while effectively removing contaminants [7].

For researchers studying human locomotion, both iCanClean and ASR offer viable preprocessing options, though parameter optimization remains crucial [1]. The recommended iCanClean settings for human locomotion data include an R² threshold of 0.65 with a 4-second sliding window [1]. For ASR, a k threshold between 10-20 appears appropriate for locomotion studies, balancing aggressive artifact removal against overcleaning risks [1].

Future research directions should include continued refinement of phantom designs for increased anatomical realism, development of standardized performance metrics across laboratories, and exploration of hybrid approaches combining the strengths of multiple algorithms. Additionally, as mobile EEG applications expand beyond laboratory settings, optimizing computational efficiency for real-time processing remains an important consideration for both methods.

The pursuit of understanding cognitive processes during whole-body movement represents a frontier in mobile brain imaging. Event-Related Potentials (ERPs), particularly the P300 component, serve as crucial neural correlates of cognitive functions such as attention, decision-making, and working memory [30] [31]. However, capturing these delicate neural signals during dynamic activities like running presents significant technological challenges, primarily due to motion artifacts that contaminate the electroencephalography (EEG) signal [1] [7].

This guide provides an objective comparison of two prominent methodologies for motion artifact removal: Artifact Subspace Reconstruction (ASR) and iCanClean. The evaluation is contextualized within a broader thesis on performance evaluation of mobile brain imaging technologies, focusing specifically on their efficacy in recovering the P300 component during overground running. We present summarized quantitative data, detailed experimental protocols, and analytical visualizations to assist researchers, scientists, and drug development professionals in selecting appropriate methodologies for their mobile brain imaging studies.

Experimental Comparison: ASR vs. iCanClean

Performance Metrics and Quantitative Results

The following tables synthesize key performance metrics from controlled studies comparing ASR and iCanClean for EEG data processing during locomotion tasks.

Table 1: ICA Component Quality and Gait Artifact Reduction

Performance Metric iCanClean Performance ASR Performance Experimental Context
ICA Component Dipolarity Superior recovery of dipolar brain components [1] Led to recovery of more dipolar components, but was somewhat less effective than iCanClean [1] Evaluation during overground running with a Flanker task [1]
Power Reduction at Gait Frequency Significant reduction [1] Significant reduction [1] Evaluation during overground running [1]
P300 Congruency Effect Successfully identified the expected greater P300 amplitude to incongruent flankers [1] Produced ERP components similar in latency to a standing task, but P300 effect not specified [1] Dynamic vs. standing Flanker task [1]

Table 2: Comprehensive Data Quality Assessment Across Multiple Artifacts

Artifact Condition iCanClean Data Quality (%) ASR Data Quality (%) Adaptive Filtering Data Quality (%) Auto-CCA Data Quality (%)
Brain (Target) 57.2 (No cleaning needed) [7] 57.2 (No cleaning needed) [7] 57.2 (No cleaning needed) [7] 57.2 (No cleaning needed) [7]
Brain + All Artifacts 55.9 (from 15.7 baseline) [7] 27.6 (from 15.7 baseline) [7] 32.9 (from 15.7 baseline) [7] 27.2 (from 15.7 baseline) [7]

Detailed Experimental Protocols

Overground Running with Flanker Task

This protocol was designed to evaluate the efficacy of artifact removal methods in capturing ERPs during high-motion scenarios [1].

  • Objective: To compare motion artifact removal approaches for identifying stimulus-locked P300 ERP components during running.
  • Participants: Young adult athletes.
  • Task Design: Participants completed an adapted Eriksen Flanker task under two conditions:
    • Dynamic Condition: Jogging overground on a treadmill.
    • Static Condition: Standing still.
    • The Flanker task presented congruent (e.g., >>>>>) and incongruent (e.g., <<><<) arrow sequences. Participants were required to respond to the central direction arrow, which elicits a P300 component with greater amplitude for incongruent trials [1].
  • EEG Recording: Mobile EEG was recorded during both task conditions.
  • Evaluation Criteria:
    • ICA Dipolarity: The quality of Independent Component Analysis decomposition, where brain-originating components typically exhibit a dipolar topography [1].
    • Spectral Power: Reduction of power at the gait frequency and its harmonics.
    • ERP Recovery: The ability to capture the expected P300 waveform and its characteristic "congruency effect" (higher amplitude for incongruent stimuli) in the dynamic condition, comparable to the static condition [1].
Phantom Head Validation

This study utilized a ground-truth paradigm to quantitatively compare cleaning algorithms without the ambiguity of unknown brain sources [7] [10].

  • Objective: To test iCanClean's ability to remove non-brain sources while preserving brain activity and compare it against other methods.
  • Apparatus: An electrically conductive phantom head with 10 embedded brain source antennae (ground-truth signals) and 10 contaminating artifact sources.
  • Tested Conditions:
    • Brain (reference)
    • Brain + Eyes
    • Brain + Neck Muscles
    • Brain + Facial Muscles
    • Brain + Walking Motion
    • Brain + All Artifacts
  • Performance Metric: Data Quality Score (0-100%), calculated as the average correlation between the known brain sources and the cleaned EEG channels [7].

Artifact Subspace Reconstruction (ASR)

ASR is a popular method for cleaning continuous EEG data, available in toolboxes like EEGLAB and BCILAB [1] [7].

  • Core Algorithm: Uses principal components analysis (PCA) to identify and remove high-variance artifacts based on a calibration period of clean data.
  • Workflow:
    • A baseline calibration period is used to compute the covariance matrix of "clean" data.
    • A sliding window PCA is performed on incoming data.
    • Principal components whose variance exceeds a user-defined threshold (k) are identified as artifactual.
    • These artifactual components are subtracted and the data is reconstructed using the calibration data.
  • Key Parameter: The k threshold (standard deviation cutoff) determines aggressiveness. A lower k removes more data but risks "overcleaning" [1].

iCanClean

iCanClean is a novel, generalized framework for real-time artifact removal that leverages canonical correlation analysis (CCA) [1] [7] [10].

  • Core Algorithm: Uses Canonical Correlation Analysis (CCA) to identify and subtract noise subspaces highly correlated with reference noise signals.
  • Workflow:
    • Reference Noise Signals are obtained, ideally from dual-layer sensors (electrodes mechanically coupled to scalp electrodes but not in contact with the skin, capturing only noise). If these are unavailable, pseudo-reference signals can be generated from the raw EEG itself (e.g., by applying a temporary notch filter) [1].
    • CCA identifies linear combinations of the scalp EEG signals and the reference noise signals that are maximally correlated.
    • The user sets an R² correlation threshold. Components exceeding this threshold are considered noise.
    • These noise components are projected back to the channel space and subtracted from the original signal.
  • Key Parameter: The R² threshold controls which correlated components are removed [1].

The fundamental workflows and logical relationships of these two algorithms are visualized below.

G SubGraph_Cluster_ASR Artifact Subspace Reconstruction (ASR) SubGraph_Cluster_iCanClean iCanClean Start_EEG Raw EEG Data ASR_Calibrate Calibrate on Clean EEG Segment Start_EEG->ASR_Calibrate iCC_NoiseInput Obtain Reference Noise Signals (Dual-layer or Pseudo-reference) Start_EEG->iCC_NoiseInput ASR_PCA Sliding Window PCA on New Data ASR_Calibrate->ASR_PCA ASR_Identify Identify Components Exceeding 'k' Threshold ASR_PCA->ASR_Identify ASR_Reconstruct Reconstruct Data Using Calibration Covariance ASR_Identify->ASR_Reconstruct ASR_Output Cleaned EEG Data ASR_Reconstruct->ASR_Output iCC_CCA Canonical Correlation Analysis (CCA) on EEG and Noise Signals iCC_NoiseInput->iCC_CCA iCC_Subtract Subtract Components Exceeding R² Threshold iCC_CCA->iCC_Subtract iCC_Output Cleaned EEG Data iCC_Subtract->iCC_Output

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Materials and Software for Mobile ERP Research

Item Name Function / Description Relevance to P300 Recovery
High-Density Mobile EEG System A portable EEG system with a high channel count (e.g., 64+ channels), often with active electrodes, designed for movement. Essential for capturing neural data during running. High density improves subsequent source separation and ICA [7].
Dual-Layer EEG Sensors Specialized electrodes where a second layer, not contacting the scalp, serves as a dedicated noise reference [1]. Provides optimal reference noise signals for iCanClean, enhancing its motion artifact removal capability [1].
Artifact Removal Software (EEGLAB/BCILAB) Software toolboxes containing implementations of ASR and other preprocessing algorithms [7]. Standardized platform for applying and comparing artifact removal methods like ASR.
iCanClean Algorithm A novel, real-time-capable artifact removal algorithm available for research use [1] [7]. The method under investigation, shown to effectively remove multiple artifact types while preserving brain signals like the P300.
Stimulus Presentation Software Software for displaying paradigms like the Flanker task and sending synchronization triggers to the EEG system. Required to elicit the P300 ERP and timestamp its onset for subsequent averaging and analysis [1] [32].

This comparative analysis demonstrates that both ASR and iCanClean offer significant advancements in the recovery of P300 ERPs during running. The choice between methodologies depends on the specific research priorities. iCanClean demonstrates superior performance in overall data quality improvement and reliable recovery of cognitive ERPs like the P300, making it a powerful tool for studies where neural signal fidelity is paramount. ASR provides a robust and more established alternative that effectively handles motion artifacts, though it may require more careful parameter tuning to avoid overcleaning.

For researchers investigating cognitive processes in ecologically valid, dynamic environments, iCanClean represents a promising solution to the critical challenge of motion artifact contamination, thereby opening new avenues for mobile brain imaging in both basic and applied neuroscience.

Direct Performance Comparison in Overground Running and Flanker Tasks

Electroencephalography (EEG) is the only brain imaging method that is both lightweight and possesses the temporal precision required to assess electrocortical dynamics during human locomotion [1] [19]. However, a significant barrier in mobile brain-body imaging (MoBI) is the presence of motion artifacts caused by head movement, electrode displacement, and cable sway during whole-body movements like running [1] [19]. These artifacts contaminate the EEG signal, reduce the quality of Independent Component Analysis (ICA) decomposition, and obscure neural signals of interest, such as event-related potentials (ERPs) [1] [19]. This comparison guide objectively evaluates the performance of two prominent motion artifact removal approaches—Artifact Subspace Reconstruction (ASR) and iCanClean—in the context of overground running combined with a cognitive Flanker task, providing researchers with experimental data to inform their methodological choices.

Experimental Protocols and Methodologies

The Flanker Task Paradigm

The Eriksen Flanker Task is a well-established cognitive paradigm used to measure executive function, specifically the ability to suppress irrelevant responses and manage cognitive conflict [33] [34] [35]. In a typical Flanker task, participants respond to a central target stimulus (e.g., an arrow) while ignoring surrounding distractor stimuli (flankers). The key manipulation lies in the congruency between the target and its flankers [34] [35]:

  • Congruent Trials: The flankers point in the same direction as the target (e.g., <<<<<). These trials induce less cognitive conflict.
  • Incongruent Trials: The flankers point in the opposite direction to the target (e.g., <<><<). These trials induce greater cognitive conflict, typically resulting in longer reaction times and higher error rates [33] [35].

The neural correlate of this conflict is often measured via the P300 component of the ERP, with incongruent trials typically eliciting a greater P300 amplitude than congruent trials [1] [19]. This paradigm is ideal for mobile EEG studies as it provides a clear, stimulus-locked neural signal against which the efficacy of artifact removal can be gauged.

Motion Artifact Removal Approaches
Artifact Subspace Reconstruction (ASR)

ASR is an automated, component-based method for removing high-amplitude artifacts from continuous EEG data [1] [19]. Its workflow can be summarized as follows:

  • Calibration: A segment of clean, baseline EEG data is used as a reference. This data is typically collected during a stationary period.
  • Principal Component Analysis (PCA): The calibration data is decomposed using PCA to identify the subspaces representing "clean" brain activity.
  • Artifact Identification & Reconstruction: A sliding window moves through the continuous EEG. For each window, the data is similarly decomposed. Components whose variance (root mean square) exceeds a user-defined threshold (standard deviation "k") compared to the calibration data are identified as artifactual.
  • Interpolation: Artifactual components are removed, and the data is reconstructed using the clean subspaces from the calibration data [1] [19].

A critical consideration is the k parameter, which controls the aggressiveness of the cleaning. A lower k value (e.g., 10) removes more data but risks "overcleaning" and removing brain signal, while a higher k value (e.g., 20-30) is more conservative but may leave more artifacts [1].

iCanClean

iCanClean is a more recent framework that leverages canonical correlation analysis (CCA) and reference noise signals to identify and subtract artifact subspaces [1] [7]. Its methodology involves:

  • Noise Signal Acquisition: Ideally, iCanClean uses dedicated "dual-layer" noise sensors that are mechanically coupled to the EEG electrodes but are not in contact with the scalp, capturing only motion-related noise [1] [7]. When such hardware is unavailable, it can generate "pseudo-reference" noise signals from the raw EEG itself, for instance, by applying a high-pass or notch filter to isolate low-frequency motion artifacts [1] [19].
  • Canonical Correlation Analysis (CCA): CCA is applied to identify linear combinations of the scalp EEG channels that are maximally correlated with linear combinations of the noise reference signals. These correlated subspaces are presumed to be dominated by artifacts.
  • Artifact Subtraction: The noise components that exceed a user-defined correlation threshold (R²) are projected back onto the EEG channels and subtracted using a least-squares solution [1] [7]. An R² value of 0.65 with a 4-second sliding window has been found effective for locomotion data [1].
Comparative Experimental Setup

A direct comparative study recorded EEG from young adults during both a static standing and a dynamic overground jogging version of the Flanker task [1] [19] [8]. The motion artifact removal approaches were evaluated based on three primary metrics:

  • ICA Component Dipolarity: The number of brain-like independent components with a dipolar scalp topography, which indicates a successful source separation [1].
  • Spectral Power at Gait Frequency: The reduction in power at the step frequency and its harmonics, indicating effective removal of the motion-locked artifact [1] [19].
  • ERP P300 Congruency Effect: The ability to recover the expected neurophysiological response—specifically, a larger P300 amplitude for incongruent Flanker trials compared to congruent ones [1] [19].

Performance Data and Comparative Analysis

The following table summarizes the quantitative findings from the comparative study during overground running [1] [19] [8].

Table 1: Performance Comparison of ASR and iCanClean for Mobile EEG during Running

Performance Metric Artifact Subspace Reconstruction (ASR) iCanClean (with pseudo-reference) Experimental Context
ICA Decomposition Quality Improved recovery of dipolar brain components [1] [19] Superior recovery of more dipolar brain components compared to ASR [1] [19] Overground running Flanker task [1]
Motion Artifact Reduction Significant power reduction at gait frequency & harmonics [1] [19] Significant power reduction at gait frequency & harmonics [1] [19] Overground running Flanker task [1]
P300 ERP Recovery Produced ERP components with latency similar to standing task [1] [19] Produced ERP components with latency similar to standing task; successfully identified the expected P300 congruency effect (higher amplitude for incongruent trials) [1] [19] Overground running Flanker task [1]
General Artifact Cleaning Effective for muscle and eye artifacts [7] Effectively removes motion, muscle, eye, and line-noise artifacts simultaneously [7] Phantom head study with simulated artifacts [7]
Data Quality Score Improved from 15.7% to 27.6% [7] Improved from 15.7% to 55.9% (Target for clean brain data: ~57.2%) [7] Phantom head with all artifact types present [7]
Interpretation of Comparative Results

The data indicates that both ASR and iCanClean are viable preprocessing methods for mobile EEG during running, offering significant improvements over raw data. However, iCanClean demonstrated a consistent performance advantage [1] [7] [19].

  • Source Separation Fidelity: iCanClean's use of CCA with noise references enabled a more effective separation of artifact from brain signal, resulting in a higher yield of dipolar ICA components, which is crucial for subsequent brain source analysis [1].
  • Neural Signal Preservation: The key differentiator was the recovery of the P300 congruency effect. While both methods preserved the general latency of the ERP, only iCanClean successfully recovered the statistically significant difference in P300 amplitude between congruent and incongruent trials during running. This suggests iCanClean is better at preserving the amplitude of task-relevant neural signals while removing noise [1] [19].
  • Overall Cleaning Efficacy: The phantom head study, which provided ground-truth brain signals, unequivocally showed iCanClean's superior ability to handle multiple, concurrent artifact types, restoring data quality to near pre-contamination levels [7].

The Scientist's Toolkit: Essential Research Reagents and Materials

For researchers seeking to replicate or build upon this comparative work, the following table details the essential components of the experimental setup.

Table 2: Key Research Reagents and Materials for Mobile EEG Studies

Item Function / Description Example in Context
High-Density Mobile EEG System A wireless EEG system with a sufficient number of electrodes (e.g., 64+ channels) to support high-quality ICA and source localization. Systems used for overground running [1] [19].
Dual-Layer Electrodes / Noise Sensors Specialized electrodes where a second layer, not in contact with the scalp, records only motion-related noise. Provides an optimal reference for iCanClean [1] [7]. Ideal setup for iCanClean [1].
Flanker Task Software Software to present stimuli and record behavioral responses (reaction time, accuracy). Custom implementations or platforms like Testable [34].
Signal Processing Toolboxes Software environments containing necessary algorithms (ICA, ASR, CCA). EEGLAB (for ASR, ICA); Custom scripts for iCanClean [1] [7].
Motion Artifact Removal Algorithms The core software for preprocessing. iCanClean; ASR implementation in EEGLAB/BCILAB [1] [7] [19].

Workflow and Signaling Pathways

The diagrams below illustrate the logical flow of the two artifact removal methods and the neural signaling pathway measured by the Flanker task.

G Start Raw EEG Data (Contaminated) Calibrate Calibration with Clean Baseline Data Start->Calibrate End Cleaned EEG Data PCA1 PCA on Calibration Data Calibrate->PCA1 PCA2 Sliding-Window PCA on Continuous Data PCA1->PCA2 Identify Identify Components Exceeding 'k' Threshold PCA2->Identify Reconstruct Reconstruct Data Using Clean Subspace Identify->Reconstruct Reconstruct->End title ASR Cleaning Workflow

Diagram 1: ASR workflow. Relies on a clean calibration period and PCA to identify and remove high-variance artifacts.

G Start Raw EEG Data (Contaminated) NoiseRef Obtain Noise Reference (Dual-layer or Pseudo-ref) Start->NoiseRef End Cleaned EEG Data CCA Canonical Correlation Analysis (CCA) NoiseRef->CCA Correlate Identify Highly Correlated Subspaces (R² Threshold) CCA->Correlate Subtract Subtract Noise Components (Least-Squares Solution) Correlate->Subtract Subtract->End title iCanClean Workflow

Diagram 2: iCanClean workflow. Uses noise references and CCA to identify and subtract artifact subspaces.

G Stimulus Visual Flanker Stimulus V1 Primary Visual Cortex (V1) Stimulus->V1 Sensory Input PPC Posterior Parietal Cortex (Attention Control) V1->PPC Feature Extraction ACC Anterior Cingulate Cortex (Conflict Monitoring) PPC->ACC Conflict Detection (Higher for Incongruent) PFC Prefrontal Cortex (Cognitive Control) ACC->PFC Engage Control ERP EEG Recording (P300 ERP) ACC->ERP Manifests as P300 Amplitude MC Motor Cortex (Response Execution) PFC->MC Inhibit Incorrect Response MC->ERP Motor Preparation title Flanker Task Neural Pathway

Diagram 3: Flanker task neural pathway. The ACC plays a key role in conflict monitoring, reflected in the P300 ERP component.

Evaluating Impact on Gait-Frequency Artifact and Harmonics Attenuation

Motion artifacts present a significant challenge in mobile electroencephalography (EEG), particularly during locomotion, where they manifest as high-amplitude oscillations at the step frequency and its harmonics, effectively obscuring underlying neural signals. Among various artifact removal approaches, Artifact Subspace Reconstruction (ASR) and iCanClean have emerged as prominent preprocessing methods for mitigating these contaminants. This guide provides an objective, data-driven comparison of their performance, with a specific focus on quantifying their efficacy in attenuating gait-frequency artifacts and their harmonics—a crucial metric for evaluating signal quality in human locomotion studies. The analysis is framed within a broader research thesis on performance evaluation, synthesizing findings from recent comparative studies to inform method selection by researchers, scientists, and drug development professionals utilizing mobile brain imaging in ecologically valid settings.

Performance Comparison: ASR vs. iCanClean

Independent studies conducted on human subjects during dynamic motor tasks provide quantitative evidence for comparing the performance of ASR and iCanClean. The table below summarizes their effectiveness against key signal quality metrics.

Table 1: Quantitative Performance Comparison of ASR and iCanClean

Evaluation Metric ASR Performance iCanClean Performance Experimental Context
ICA Component Dipolarity Improved dipolarity, producing more brain-like independent components [8]. Superior improvement in dipolarity compared to ASR; more effective at recovering dipolar brain components [8] [19]. Overground running during a Flanker task [8] [1].
Power Reduction at Gait Frequency Significant reduction in spectral power at the fundamental gait frequency and its harmonics [8] [26]. Significant reduction in spectral power at the fundamental gait frequency and its harmonics [8] [26]. Overground running; power analysis at step frequency [8].
Event-Related Potential (ERP) Fidelity Produced ERP components similar in latency to those identified in a static (no motion) condition [8]. Produced ERP components similar in latency to the static condition; successfully identified the expected P300 amplitude congruency effect [8] [19]. Flanker task during jogging vs. standing; P300 component analysis [8] [1].
Overall Data Quality (Phantom Head Study) Improved data quality score from 15.7% to 27.6% in a "Brain + All Artifacts" condition [7]. Outperformed ASR, improving data quality score from 15.7% to 55.9% in the same "Brain + All Artifacts" condition [7]. Phantom head with simulated brain sources and multiple artifacts (motion, muscle, eye) [7].

Experimental Protocols and Methodologies

The comparative data presented are derived from rigorous experimental protocols designed to evaluate algorithm performance under controlled and ecologically valid conditions.

Human Subject Protocol During Overground Running

A key study providing comparative data involved recording EEG from young adults during a dynamic Flanker task, which was performed under two conditions: static standing and overground jogging [1] [19]. This design enabled a direct comparison of artifact-laden data (jogging) with a relatively clean baseline (standing).

Core Methodology:

  • Task: An adapted Eriksen Flanker task was used to elicit cognitive event-related potentials (ERPs), notably the P300 component [1].
  • EEG Recording: Wireless mobile EEG was recorded during both task conditions. The motion during jogging introduces artifacts time-locked to the gait cycle [8].
  • Evaluation:
    • ICA Dipolarity: The quality of the resulting independent components from ICA decomposition was assessed based on their dipolarity, a property of brain-like sources [8] [19].
    • Spectral Power: Power spectral density was analyzed to quantify signal power at the fundamental step frequency and its higher harmonics before and after cleaning [8].
    • ERP Analysis: The recovered P300 ERP components from the jogging condition, after preprocessing with each method, were compared to the P300 from the standing condition in terms of latency and expected amplitude difference between congruent and incongruent stimuli [8] [1].
Phantom Head Validation Protocol

To establish ground-truth performance, another critical study employed an electrically conductive phantom head [7]. This apparatus contained known simulated brain sources alongside contaminating sources (e.g., for eyes, neck muscles, and walking motion).

Core Methodology:

  • Signal Broadcasting: The phantom head broadcasted known "brain" signals mixed with various artifacts, including motion artifacts simulating walking [7].
  • Performance Metric: A Data Quality Score was calculated based on the average correlation between the original, known brain sources and the cleaned EEG channels. A higher score indicates better preservation of the true brain signal and more effective artifact removal [7].
  • Comparative Cleaning: The contaminated EEG data from the phantom was processed using iCanClean, ASR, and other methods. Their performance in improving the Data Quality Score was directly compared [7].

Signaling Pathways and Workflows

The following diagrams illustrate the logical workflows and algorithmic relationships of the ASR and iCanClean methods, based on the descriptions in the search results.

ASR Cleaning Process

G A Raw Continuous EEG B Identify Clean Calibration Data A->B D Sliding Window PCA on New Data A->D Non-Reference Data C Calculate Reference Data Covariance Matrix B->C C->D Calibration Data E Identify Artifactual Components (SD > k threshold) D->E F Reconstruct Data Using Calibration Statistics E->F G Cleaned EEG F->G

iCanClean Cleaning Process

G cluster_0 Noise Reference Options A Raw EEG Signals B Obtain Noise Reference A->B C Canonical Correlation Analysis (CCA) A->C B->C D Identify Noise Correlated Subspaces (R² threshold) C->D E Subtract Noise Components via Least-Squares D->E F Cleaned EEG E->F B1 Dual-Layer Electrodes (Ideal) B1->B B2 Pseudo-Reference (e.g., Notch Filter <3 Hz) B2->B

The Scientist's Toolkit: Key Research Reagents and Materials

The following table details essential hardware, software, and analytical tools referenced in the featured experiments for conducting similar research in mobile EEG artifact removal.

Table 2: Essential Research Tools for Mobile EEG Artifact Evaluation

Tool Name Type Primary Function in Research
Wireless Mobile EEG System Hardware Enables recording of brain electrical activity during dynamic whole-body movements like walking and running without movement constraints from cables [8] [19].
Artifact Subspace Reconstruction (ASR) Software Algorithm A PCA-based method for removing high-amplitude, non-stationary artifacts from continuous EEG by interpolating data based on clean calibration periods [8] [3].
iCanClean Software Algorithm Leverages canonical correlation analysis (CCA) with reference noise signals (real or pseudo) to detect and subtract artifact subspaces from the EEG [8] [7].
Phantom Head Testbed Experimental Apparatus An electrically conductive head model with embedded sources that simulates brain activity and artifacts, providing ground-truth data for validating cleaning algorithms [7] [20].
Independent Component Analysis (ICA) Analytical Method A blind source separation technique used to decompose EEG signals into maximally independent components, allowing for the identification and removal of artifactual sources [8] [36].
Dual-Layer or Tripolar Electrodes Hardware Specialized electrode designs that mechanically couple noise-sensing elements with scalp electrodes or use concentric rings to calculate a surface Laplacian, enhancing inherent noise cancellation [7] [20].

Conclusion

The comparative evaluation clearly establishes both ASR and iCanClean as highly effective preprocessing methods for mitigating motion artifacts in mobile EEG, with iCanClean often demonstrating a performance edge. iCanClean consistently shows superior ability to improve ICA decomposition quality, yield more dipolar brain components, and successfully recover expected ERP components like the P300 during high-motion activities such as running. For researchers, the choice between methods may depend on specific needs: iCanClean, particularly with dual-layer electrodes, offers potentially higher fidelity, while ASR remains a powerful and accessible option. Future directions should focus on developing standardized validation protocols for clinical populations, integrating these methods with real-time brain-computer interfaces for therapeutic applications, and further optimizing algorithms for the unique challenges presented by diverse patient movements in biomedical and clinical drug development research.

References