Clearing the Signal: Advanced Strategies for Handling Motion Artifacts in Overground Running EEG

Abstract

Electroencephalography (EEG) is the sole brain imaging method with the temporal precision and portability to assess electrocortical dynamics during human locomotion. However, its application during dynamic activities like overground running is severely challenged by motion artifacts that contaminate the neural signal. This article provides a comprehensive guide for researchers and drug development professionals on the latest methodologies for motion artifact handling. We cover the foundational sources of artifacts, evaluate advanced removal techniques including iCanClean, Artifact Subspace Reconstruction (ASR), and deep learning models, and provide a framework for troubleshooting and validation. By synthesizing recent comparative studies and validation protocols, this review aims to empower robust mobile brain imaging in ecologically valid settings, thereby accelerating research in neurophysiology and clinical assessment.

Understanding the Enemy: The Genesis and Impact of Motion Artifacts in Mobile EEG

FAQs: Understanding and Troubleshooting Motion Artifacts

What makes motion artifacts during running particularly challenging for EEG?

Motion artifacts during running are especially problematic due to their large amplitude and broadband spectral characteristics. When you run, the resulting artifacts are typically at least ten times greater in amplitude than the actual brain signals you are trying to measure [1]. Furthermore, the artifact produced is not a simple, single-frequency noise. Instead, overground running produces broadband spectral power, particularly at the step frequency and its harmonics, which can spread across the same frequency bands as neural signals of interest, making them difficult to separate and filter out without damaging the underlying brain signal [2].

What are the most effective methods for removing motion artifacts from running EEG data?

Current research indicates that the most effective preprocessing methods leverage advanced algorithms. A 2025 comparative study highlighted that iCanClean and Artifact Subspace Reconstruction (ASR) are particularly effective for data collected during overground running [2] [3].

• iCanClean: This method uses canonical correlation analysis (CCA) to identify and subtract noise subspaces from the scalp EEG. It can work with dedicated "dual-layer" noise sensors or create "pseudo-reference" noise signals from the raw EEG itself. It has been shown to produce high-quality Independent Components (ICs) and help recover expected event-related potential components like the P300 [2].
• Artifact Subspace Reconstruction (ASR): This is a popular method that uses a sliding-window principal component analysis (PCA) to identify and remove high-variance, artifact-laden components from the continuous EEG data. Its performance depends on a calibration period and a user-defined threshold (often a "k" value between 20-30). An overly aggressive (low) k-value can "overclean" the data and remove genuine brain activity [2].
• Deep Learning: Newer approaches using 1D Convolutional Neural Networks (CNNs), such as MLMRS-Net and Motion-Net, show significant promise. These are signal reconstruction networks trained to map motion-corrupted EEG to clean signals, achieving high marks in artifact reduction and signal-to-noise ratio improvement [4] [5].

How can my experimental setup help minimize motion artifacts?

Your hardware and setup choices are the first line of defense against motion artifacts.

• Dual-Layer Electrode Systems: A highly effective hardware solution involves using dual-layer electrodes. In this setup, standard scalp electrodes are mechanically coupled to "noise" electrodes that are electrically isolated from the scalp. The scalp electrode records a mix of brain signal and motion artifact, while the noise electrode records only the motion artifact. This pure noise reference can then be used to clean the contaminated signal [6].
• Proper Electrode Placement and Impedance: There is no substitute for clean data [7]. Meticulous application of the EEG cap is crucial. Ensure proper placement according to the 10-20 system and keep electrode impedances as low as possible, as high impedances make the recording more susceptible to motion artifacts [7] [8]. Quality control should include photos and ratings of capping quality [8].

How do I evaluate the success of my motion artifact removal pipeline?

You should evaluate your processing pipeline using multiple, complementary metrics:

• ICA Component Dipolarity: A successful cleanup should lead to an ICA decomposition that yields a higher number of dipolar brain independent components. This indicates that the algorithm has effectively separated brain sources from non-brain artifacts [2].
• Power at Gait Frequency: Check the power spectrum of your cleaned data. Effective processing should show a significant reduction in spectral power at the fundamental frequency of your gait (step rate) and its harmonics [2].
• Recovery of Expected Neural Signatures: Finally, validate your data by checking if you can recover well-established neural responses. For example, in a Flanker task, a successful pipeline should reveal the expected P300 event-related potential with its characteristic congruency effect (greater amplitude for incongruent trials) [2].

The Scientist's Toolkit: Research Reagents & Essential Materials

Table 1: Key hardware, software, and algorithmic "reagents" for motion artifact research.

Item Name	Type	Primary Function	Key Consideration / Parameter
Dual-Layer EEG Electrodes [6]	Hardware	Provides a mechanically coupled "noise" reference that records only motion artifacts, enabling highly effective signal cleaning.	Requires specialized hardware and a compatible amplifier system.
iCanClean [2]	Algorithm	Uses Canonical Correlation Analysis (CCA) to identify and subtract motion artifact subspaces highly correlated with a noise reference.	Can use pseudo-reference signals if dedicated noise sensors are unavailable. An R² threshold of ~0.65 is a suggested starting point [2].
Artifact Subspace Reconstruction (ASR) [2]	Algorithm	Removes high-variance, artifact-laden segments in continuous EEG using a sliding-window PCA approach.	Performance is sensitive to the "k" threshold. A value that is too low can overclean data; values of 20-30 are often recommended [2].
Independent Component Analysis (ICA)	Algorithm	A blind source separation technique that decomposes multi-channel EEG into maximally independent components, which can then be manually or automatically classified and rejected.	Quality of decomposition is reduced by large motion artifacts, making pre-cleaning with methods like ASR or iCanClean crucial [2].
1D CNN Networks (e.g., MLMRS-Net) [4]	Algorithm (Deep Learning)	A signal reconstruction network trained to map motion-corrupted EEG signals to their clean versions, offering a state-of-the-art, data-driven approach.	Requires a substantial dataset of clean and corrupted EEG signals for training, which can be a barrier to entry.

Experimental Protocols & Performance Data

Comparative Performance of Key Methods

The following table summarizes quantitative findings from recent studies that have directly compared artifact removal techniques in dynamic contexts.

Table 2: Quantitative comparison of motion artifact removal approaches based on recent studies.

Method / Algorithm	Key Performance Metrics (Reported Averages)	Best For / Key Advantage	Study Context
iCanClean (with pseudo-reference)	- Improved ICA dipolarity [2]- Reduced power at gait frequency [2][2]<="" congruency="" effect="" identified="" p300="" td="">	Recovering stimulus-locked ERPs in demanding conditions like running.	Overground running with Flanker task [2]
Artifact Subspace Reconstruction (ASR)	- Improved ICA dipolarity [2]- Reduced power at gait frequency [2]- Produced ERP components similar to standing task [2]	A robust, widely available method that improves data quality for subsequent ICA.	Overground running with Flanker task [2]
MLMRS-Net (1D CNN)	- ΔSNR: 26.64 dB [4]- Artifact Reduction (η): 90.52% [4]- MAE: 0.056 [4]	High-performance, automated denoising for single-channel EEG.	Benchmark dataset from PhysioNet (Leave-one-out cross-validation) [4]
Motion-Net (1D CNN)	- Artifact Reduction (η): 86% ±4.13 [5]- ΔSNR: 20 ±4.47 dB [5]- MAE: 0.20 ±0.16 [5]	Subject-specific training; effective with smaller datasets by using Visibility Graph features.	Real-world motion artifacts, subject-specific framework [5]

Workflow for a Motion-Robust EEG Experiment

The diagram below outlines a recommended experimental and processing workflow, integrating hardware and software solutions to tackle motion artifacts.

Visualizing the Dual-Layer Electrode Principle

The core principle behind one of the most effective hardware solutions is illustrated below.

In mobile electroencephalography (EEG) research, particularly during dynamic activities like overground running, mechanical sources introduce significant artifacts that can compromise data quality. These artifacts originate from multiple locations within the recording chain: the skin-electrode interface, connecting cables, and the electrode-amplifier system itself [9]. Understanding these mechanical sources is fundamental to developing effective artifact mitigation strategies for obtaining clean neural signals during whole-body movement.

Motion artifacts are particularly problematic because their amplitude can be orders of magnitude greater than the underlying neural signals of interest, which are typically in the microvolt range [9] [10]. Unlike some physiological artifacts, motion-related artifacts are often time-locked to the gait cycle and highly variable in shape and spectral content, making them difficult to remove with standard post-processing techniques alone [2] [9]. For researchers investigating electrocortical dynamics during running, addressing these mechanical sources is a critical prerequisite for valid data interpretation.

Troubleshooting Guides & FAQs

Frequently Asked Questions

Q1: What are the primary mechanical sources of motion artifacts in mobile EEG? The three primary mechanical sources are:

• Electrode-Skin Interface Artifacts: Caused by relative motion between the electrode and skin, leading to changes in the ionic distribution and impedance at the interface [9].
• Cable Movement Artifacts: Result from triboelectric effects (friction within the cable components) and the motion of conductors within magnetic fields as cables sway during movement [9] [11] [10].
• Electrode-Amplifier System Artifacts: Include sudden changes in electrode-skin impedance that modulate power line interference (PLI), appearing as spike-like artifacts or baseline shifts [9].

Q2: Why are standard artifact removal techniques often insufficient for motion artifacts? Motion artifacts are challenging because they are often non-stationary and non-repetitive in shape [9]. Their spectral components frequently overlap with the EEG bandwidth of interest (0.1–100 Hz), making it difficult to filter them out without also removing neural signals [9]. Techniques like wavelet transforms or blind source separation excel at removing physiological artifacts like eye blinks but are less effective for complex motion artifacts [9].

Q3: How does cable sway specifically affect my EEG signal? Cable sway generates artifacts through triboelectric noise, where the friction and deformation of the cable insulator during movement create an additive voltage potential that is amplified along with the neural signal [9]. These artifacts often have a spike-like morphology and their spectral content is broad, overlapping with the typical EEG bandwidth [9]. Experimental studies using phantom heads have confirmed that cable sway is a major contributor to signal degradation during motion [11].

Q4: What is the link between electrode-skin impedance and motion artifacts? A stable, low-impedance connection is crucial for high-fidelity EEG recording [12]. Motion can cause sudden changes in this impedance, leading to baseline shifts and transient spikes (often called "electrode pops") in the signal [9] [10]. Higher impedance interfaces are also more susceptible to motion artifacts and exhibit increased thermal noise, which degrades the overall signal-to-noise ratio [12] [13].

Troubleshooting Guide: Common Problems and Solutions

Problem	Possible Mechanical Cause	Recommended Solution
Slow baseline drifts synchronized with gait	Electrode movement relative to the skin at the electrode-skin interface [9].	Ensure secure electrode attachment; use adequate electrolyte gel; apply appropriate skin preparation (e.g., light abrasion) to stabilize impedance [12].
High-frequency noise or spike-like artifacts	Cable sway causing triboelectric effects [9] [11].	Use wireless systems; if wired, secure cables to the subject's clothing or body to minimize movement; use specialized low-noise cables [11].
Sudden, large-amplitude transients ("pops") on individual channels	Sudden impedance changes from poor electrode contact or drying electrolyte [10].	Check electrode contact and re-apply if necessary; ensure sufficient electrolyte gel is used; avoid pulling on electrode wires [10].
50/60 Hz power line interference modulated by movement	Motion-induced changes in electrode-skin imbalance, modulating residual input-referred PLI [9].	Improve electrode-skin contact at both recording and reference sites to stabilize impedance; ensure proper grounding of the system [9].
Broadband spectral power, particularly at step frequency and harmonics	Overall head motion during whole-body movements like running [2].	Employ advanced preprocessing algorithms like iCanClean or Artifact Subspace Reconstruction (ASR) before ICA to reduce these motion-related artifacts [2].

Experimental Protocols & Methodologies

Protocol: Assessing the Impact of Cable Sway

Objective: To isolate and quantify the contribution of cable motion to EEG artifacts.

Materials:

• EEG system with active electrodes
• Head phantom with embedded signal source [11]
• Motion platform (e.g., hexapod)
• Methods to secure cables (e.g., tape, wraps)

Methodology:

• Place the phantom head on the motion platform and attach EEG electrodes.
• Generate a clean, known signal (e.g., 7 Hz sine wave) from the phantom's internal antenna [11].
• Record baseline EEG with the platform stationary and cables secured.
• Experiment 1 (Electrode Motion): Subject the head to sinusoidal vertical displacements (e.g., 4 cm amplitude at 1-2 Hz) to simulate head motion during running, with cables firmly secured to minimize their movement [11].
• Experiment 2 (Cable Sway): With the head stationary, detach cables and use the platform to create horizontal displacements (e.g., 4 cm amplitude) to simulate cable sway [11].
• Record EEG during both experimental conditions.
• Calculate the Signal-to-Noise Ratio (SNR) for each condition by comparing the recorded signal to the ground-truth baseline [11].

Expected Outcome: A significant decrease in SNR during the cable sway condition will demonstrate the substantial impact of cable motion on signal quality, independent of head movement [11].

Protocol: Evaluating Skin-Preparation Techniques for Impedance Stability

Objective: To determine the most effective skin treatment for maintaining a low and stable electrode-skin impedance during long recordings.

Materials:

• EEG/EMG recording system
• Ag/AgCl electrodes
• Impedance measurement device
• Different skin preparation materials (e.g., abrasive tape, alcohol pads, skin prepping gel).

Methodology:

• Select multiple equivalent sites on a subject's skin.
• Apply different skin treatments to each site (e.g., no treatment, alcohol rub, light abrasion with specialized tape) [12].
• Apply electrodes to each treated site.
• Measure the electrode-skin impedance at each site immediately after application (time zero).
• Continue to measure impedance at regular intervals (e.g., every 15-30 minutes) over several hours while the subject may be at rest or performing mild activities [12] [13].
• Monitor the impedance value and its stability over time.

Expected Outcome: While abrasive treatments typically provide the lowest initial impedance, the impedance across different treatments may equilibrate over longer periods (e.g., 24 hours) due to natural skin processes like sweating [12]. This protocol helps identify the best balance between initial impedance reduction and long-term stability for a specific study design.

Data Presentation: Quantitative Findings

Impact of Electrode Properties and Cable Sway on Signal Quality

Table 1: Quantitative Effects of Hardware Factors on Motion Artifacts (from Phantom Studies)

Factor	Experimental Manipulation	Key Quantitative Finding	Impact on Signal-to-Noise Ratio (SNR)
Cable Sway	Horizontal displacement (4 cm amplitude) of cables [11].	A major contributor to motion artifacts, producing significant high-amplitude noise [11].	Substantial decrease in SNR, identified as one of the most impactful factors [11].
Electrode Surface Area	Comparison of electrodes with different surface areas under motion [11].	Larger electrode surface area can improve signal quality during motion [11].	Small but significant improvement in SNR with larger surface area [11].
Electrode Mass	Comparison of electrodes with different masses under motion [11].	Increased mass can cause greater displacement relative to the scalp [11].	Can lead to a decrease in SNR due to increased motion artifacts [11].

Performance of Signal Processing Algorithms for Motion Artifact Removal

Table 2: Performance Comparison of Motion Artifact Removal Algorithms (from Human Running Studies)

Algorithm	Key Principle	Input Parameters	Performance Metrics
iCanClean [2]	Uses canonical correlation analysis (CCA) to detect and subtract noise subspaces correlated with pseudo-reference or dedicated noise signals [2].	R² correlation threshold (e.g., 0.65); sliding window length (e.g., 4 s) [2].	Effective at reducing power at gait frequency; improves ICA component dipolarity; can recover expected ERP components (e.g., P300) [2].
Artifact Subspace Reconstruction (ASR) [2]	Uses sliding-window PCA to identify and remove high-variance components exceeding a threshold relative to a clean baseline period [2].	Standard deviation threshold ("k"; e.g., 10-30) [2].	Reduces power at gait frequency; improves ICA dipolarity; aggressive cleaning (low k) may risk over-cleaning and signal distortion [2].
Independent Component Analysis (ICA)	Blind source separation to isolate and remove artifactual components [2].	Requires high-quality data for decomposition; often works best after preprocessing with methods like iCanClean or ASR [2].	Decomposition quality is reduced if large motion artifacts are present; effectiveness improves when combined with other preprocessing methods [2].

Signaling Pathways and Workflows

Motion Artifact Generation Pathways

The following diagram illustrates the primary mechanical pathways through which motion introduces artifacts into the EEG signal.

Mechanical Sources of EEG Motion Artifacts. This diagram maps the causal pathways from physical motion to specific types of artifacts in the recorded EEG signal, stemming from three key mechanical sources: interface motion, cable effects, and impedance instability [9].

Motion Artifact Mitigation Workflow

A combined hardware and software approach is recommended for effective motion artifact management. The following workflow outlines a comprehensive mitigation strategy.

Motion Artifact Mitigation Workflow. This diagram outlines a two-stage strategy for handling motion artifacts, combining preventive hardware setup with a multi-step signal processing pipeline to yield analyzable neural data [2] [9] [11].

The Scientist's Toolkit: Essential Materials

Research Reagent Solutions

Table 3: Key Materials and Equipment for Mitigating Mechanical Motion Artifacts

Item	Function & Rationale
Abrasive Skin Prep Gel/Tape	Reduces initial skin impedance by removing dead skin cells and oils from the stratum corneum, promoting a more stable interface and lower baseline impedance [12].
Electrolyte Gel	Creates a stable ionic connection between the electrode and the skin. Using sufficient gel helps buffer against motion-induced impedance changes and prevents "pops" [11] [10].
Active Electrode Systems	Incorporate a pre-amplifier directly at the electrode site. This minimizes the distance the tiny EEG signal travels before amplification, reducing susceptibility to noise induced by cable sway [11] [1].
Low-Noise, Shielded Cables	Engineered to minimize triboelectric effects. Securing these cables to the subject's body is a critical step to reduce motion-related artifacts [11].
Wireless EEG Systems	Eliminate cable sway artifacts entirely by removing the physical connection between the subject and the recording unit, ideal for highly dynamic tasks like running [11].
Conductive Adhesive Rings/Secure Caps	Provide physical stabilization of electrodes, minimizing relative motion between the electrode and the scalp during movement [11].
Phantom Head with Signal Source	Provides a ground-truth signal for controlled testing and validation of artifact removal methods without biological variability [11].

FAQs: Understanding and Identifying Gait-Related Artifacts

Q1: What is a "spectral footprint" in the context of mobile EEG? A spectral footprint refers to the characteristic pattern of contaminating signals left in the EEG power spectrum by motion artifacts. During overground running, this is typically dominated by a pronounced peak at the fundamental gait frequency (the step rate) and its harmonic frequencies (integer multiples of the step rate) [2]. These artifacts are caused by head motion, electrode displacement, and cable sway that are time-locked to the gait cycle [3] [2].

Q2: Why do gait artifacts appear as a fundamental frequency with harmonics? The repetitive, quasi-periodic nature of the running motion produces rhythmic mechanical forces on the EEG system. The primary oscillation occurs at the step frequency, but the movement is not a perfect sine wave; it contains more complex shapes. These non-sinusoidal, periodic movements are decomposed in the frequency domain into a fundamental frequency (the step rate) and a series of harmonics, creating a distinctive comb-like pattern in the power spectral density [2].

Q3: How can I distinguish a gait artifact from actual brain activity? Gait artifacts are identified by their precise temporal coupling to the gait cycle. Key indicators include:

• A sharp spectral peak precisely at the calculated step frequency.
• Identical, smaller peaks at integer multiples (harmonics) of this frequency.
• The artifact is often broadband, affecting a wide range of frequencies, and can be most prominent in specific channels due to head motion geometry [2]. In contrast, cognitively-relevant brain oscillations (like sensorimotor rhythms) are typically confined to a specific band (e.g., alpha 8-13 Hz, beta 13-30 Hz) and are not perfectly locked to the harmonic series of the step frequency [14].

Q4: What is the impact of these artifacts on Independent Component Analysis (ICA)? Large motion artifacts significantly reduce the quality of ICA decomposition. ICA is a blind source separation method that identifies maximally independent components in the data. Excessive motion artifact can overwhelm the algorithm, reducing its ability to isolate clean brain sources and resulting in fewer components with dipolar scalp maps that are characteristic of brain activity [3] [2]. Effective artifact removal before ICA is often necessary for mobile EEG studies [2].

Troubleshooting Guides

Guide 1: Protocol for Visualizing the Spectral Footprint

• Objective: To confirm the presence and severity of gait-frequency contamination in your raw or preprocessed EEG data.
• Required Tools: EEG data acquisition system, motion capture system or foot switches for gait event detection, signal processing software (e.g., EEGLAB, MATLAB, Python).
• Procedure:
  • Record Synchronized Data: Simultaneously collect EEG and gait event data (e.g., heel strikes from a motion capture system or force plates).
  • Calculate Gait Frequency: For a given data segment, calculate the primary step frequency (e.g., Steps per second / Hz).
  • Generate Power Spectral Density (PSD): Compute the PSD for your EEG channels, particularly focusing on frontal, central, and parietal sites which are often affected.
  • Visual Inspection: Plot the PSD and visually identify prominent peaks. Check if the largest peaks align with the calculated step frequency (F0) and its harmonics (2F0, 3F0, etc.).
  • Topographic Mapping: Create scalp topography maps focused on the power at the fundamental gait frequency and its first harmonic. This often shows a characteristic global or anterior-posterior pattern distinct from brain activity [15].

Guide 2: Mitigating Gait Artifacts Using Advanced Preprocessing

This guide outlines a comparative approach using two modern methods: Artifact Subspace Reconstruction (ASR) and iCanClean.

• Objective: To significantly reduce motion artifact power prior to ICA, thereby improving the yield of brain-related independent components.
• Method Comparison Table:

Feature	Artifact Subspace Reconstruction (ASR)	iCanClean
Principle	Uses sliding-window Principal Component Analysis (PCA) to identify and remove high-variance components exceeding a threshold ("k") compared to a clean baseline period [2].	Leverages Canonical Correlation Analysis (CCA) to identify and subtract noise subspaces that are highly correlated with a reference noise signal [3] [2].
Noise Reference	Learned from a clean segment of the data itself [2].	Uses mechanically coupled "dual-layer" noise sensors or creates a pseudo-reference from the raw EEG (e.g., by notch-filtering) [3] [2].
Key Parameter	k: Standard deviation threshold (k=20-30 is common; lower values are more aggressive). A k of 10 may be needed for running [2].	R²: Correlation threshold for noise subspace removal (R²=0.65 was effective for walking data [2]).
Primary Effect	Removes high-amplitude, transient artifacts [2].	Targets structured noise correlated with the reference [3].
Performance in Running	Effectively reduces power at the gait frequency and harmonics; improves ICA dipolarity [2].	In analysis, somewhat more effective than ASR at improving ICA dipolarity and recovering expected ERPs like the P300 [3] [2].

• Validation Steps:
  • Spectral Check: Recompute the PSD after processing. Successful mitigation will show a clear reduction in the amplitude of peaks at the gait frequency and its harmonics.
  • ICA Quality: Run ICA and check the number of brain-like components using ICLabel. An increase indicates better decomposition.
  • ERP Validation: If using an event-related paradigm (e.g., a Flanker task), check if known components (e.g., P300) are recoverable and show expected effects [3].

Experimental Protocols from Key Studies

• Objective: To compare the efficacy of ASR and iCanClean in recovering stimulus-locked ERPs during overground running.
• Participants: Young adult athletes.
• Task Design:
  • Dynamic Condition: Participants perform a Flanker task (responding to congruent/incongruent arrows) while jogging on a treadmill or overground.
  • Static Control Condition: The same task is performed while standing still.
• Data Acquisition:
  • EEG: Recorded using a wireless mobile EEG system (e.g., 64-channel LiveAmp).
  • Gait Events: Recorded via motion capture or integrated force plates.
• Processing Pipeline:
  • Preprocessing: Apply a 1 Hz high-pass filter.
  • Artifact Removal: Process data with either ASR (k=10) or iCanClean (pseudo-reference, R²=0.65).
  • ICA: Perform ICA decomposition (e.g., Adaptive Mixture ICA).
  • Component Rejection: Classify and remove non-brain components using ICLabel.
  • ERP Analysis: Epoch data relative to Flanker stimulus onset and average to compute ERPs.
• Outcome Measures:
  • Component Dipolarity: The number of ICs with a dipolar scalp map (residual variance < 15%).
  • Spectral Power: Reduction in power at the gait frequency (F0) and its harmonics.
  • ERP Congruency Effect: Successful identification of the P300 component and its expected larger amplitude for incongruent vs. congruent trials in the dynamic condition.

• Objective: To investigate brain activation patterns during a fatiguing motor task, requiring a stable EEG signal free from movement artifact.
• Task: Intermittent submaximal elbow flexion contractions (e.g., 5s contraction, 2s rest) at 40% Maximum Voluntary Contraction until exhaustion.
• Data Collected:
  • EEG: High-density EEG.
  • EMG: From the elbow flexor muscles.
  • Force: From a force transducer.
• Analysis:
  • Time-Frequency Analysis: Compute Event-Related Spectral Perturbation (ERSP) to see changes in power relative to baseline.
  • Fatigue Levels: Compare EEG spectral power from trials at the beginning (mild fatigue), middle (moderate), and end (severe fatigue) of the session.
  • Key Finding: Theta, alpha, and beta power spectral densities in cortical motor areas vary significantly with fatigue level, highlighting the importance of clean data for detecting these subtle spectral changes [16].

Signaling Pathway & Workflow Visualizations

Artifact Mitigation and Validation Pathway

Gait Artifact Spectral Profile Identification

The Scientist's Toolkit: Essential Research Reagents & Materials

This table details key hardware, software, and methodological "reagents" essential for conducting overground running EEG research.

Item	Type	Function / Explanation
High-Density Mobile EEG System (e.g., 64-channel LiveAmp [17])	Hardware	Allows for full scalp coverage and source localization while permitting free movement. Essential for capturing spatial information needed for effective ICA [2].
Motion Capture System (e.g., VICON [17]) or Force Plates	Hardware	Provides gold-standard timing for gait events (heel strike, toe-off). Critical for calculating the fundamental gait frequency (F0) and time-locking analysis [17].
Artifact Subspace Reconstruction (ASR)	Software/Algorithm	A robust, automated method for removing large-amplitude motion artifacts from continuous EEG data prior to ICA, improving subsequent decomposition [2].
iCanClean Algorithm	Software/Algorithm	An effective method for removing motion artifacts using reference noise signals (real or pseudo), shown to be particularly effective for recovering ERPs during running [3] [2].
Independent Component Analysis (ICA)	Software/Algorithm	A blind source separation technique fundamental for isolating brain, eye, muscle, and residual noise components after initial preprocessing [2] [14].
ICLabel	Software/Classifier	An EEGLAB plugin that automatically classifies ICA components by type (brain, muscle, eye, heart, line noise, channel noise, other). Speeds up the component rejection process [2].
Pseudo-Reference Noise Signal	Method	A signal derived from the raw EEG itself (e.g., by applying a temporary notch filter) to serve as a noise reference for iCanClean when dedicated noise sensors are not available [3] [2].
Flanker Task Paradigm	Method	A cognitive task used to elicit a well-known event-related potential (the P300). Its use during running provides a functional validation check for the artifact removal pipeline [3] [2].

Frequently Asked Questions

How do motion artifacts specifically degrade the quality of an ICA decomposition? Motion artifacts introduce large-amplitude, non-brain signals that violate ICA's core assumption of statistical independence between sources. These artifacts dominate the input data, causing ICA to "waste" components on representing motion instead of neural activity. This results in fewer components for brain signals and a less stable decomposition, ultimately reducing the number of identifiable brain-related components [18] [3].

Can artifact rejection alone solve the problem of motion artifacts in ERP studies? While rejecting trials with extreme artifacts can reduce noise, it is often insufficient for motion-laden data like that from overground running. Aggressive rejection leads to a critical loss of trials, which severely compromises the signal-to-noise ratio (SNR) and statistical power of the averaged ERPs. A combined approach of preprocessing (e.g., with ASR or iCanClean) followed by selective rejection is recommended [19] [3].

What are the consequences of not adequately addressing motion artifacts before ICA? Failure to reduce motion artifacts prior to ICA leads to two primary consequences:

• Confounds: Artifacts that occur more often in one experimental condition than another (e.g., more head movement during a difficult task) can create false, artifact-driven differences in ERP waveforms, leading to incorrect scientific conclusions [19].
• Reduced Power: The uncontrolled variance from random motion artifacts increases noise across all conditions, which can mask true underlying neural effects and prevent them from reaching statistical significance [19].

Which preprocessing method is more effective for running data: ASR or iCanClean? A 2025 comparative study on overground running found that both Artifact Subspace Reconstruction (ASR) and iCanClean were effective, but iCanClean demonstrated a slight advantage. It was more effective at recovering dipolar brain components from ICA and was the only method in that study to successfully capture the expected P300 ERP congruency effect during running [3].

---

Troubleshooting Guides

Problem: Poor ICA Decomposition with Non-Brain Components

Description After running ICA, the component scalp maps appear noisy, non-dipolar, or dominated by components with frontal (eye), temporal (muscle), or other non-neural characteristics. There is a lack of clear, dipolar brain components.

Diagnosis Checklist

• Data Quantity: Ensure sufficient data length for a stable decomposition. For high-channel counts, considerable data is required [20].
• Data Quality: Check for excessive high-amplitude motion artifacts before ICA.
• Channel Locations: Confirm all channels have proper locations specified, as this is critical for source separation [20].
• Filtering: Apply a high-pass filter (e.g., 1 Hz) to remove slow drifts that can impede ICA [21].

Resolution

• Apply Preprocessing: Use a dedicated motion artifact reduction tool before performing ICA.
• Choose iCanClean or ASR: For running data, implement iCanClean (preferably with dual-layer sensors, or with pseudo-reference signals) or ASR with an appropriate threshold (e.g., k=20-30 to avoid over-cleaning) [18] [3].
• Re-run ICA: Perform ICA on the preprocessed data. The decomposition should yield a higher proportion of components with dipolar scalp topographies, indicative of brain sources [18].

Problem: Absent or Unreliable ERP Components

Description The expected event-related potential (ERP) components, such as the P300, are absent, distorted, or do not show the expected differences between experimental conditions after data collection involving movement.

Diagnosis Checklist

• Residual Artifacts: Inspect the epoched data for remaining high-frequency muscle noise or low-frequency drift time-locked to the event.
• Trial Count: Ensure an adequate number of clean trials remain after artifact rejection for a robust average.
• Spectral Contamination: Plot the power spectrum of your ERP baseline; a peak at the gait frequency (e.g., ~2 Hz for running) indicates significant residual motion artifact [3].

Resolution

• Aggressive Preprocessing: Focus on preprocessing methods proven to handle motion. The table below compares two effective methods.
• Targeted Artifact Removal: Use a method like iCanClean, which was specifically shown to recover the P300 effect during running [3].
• Validate with Metrics: Assess success by a reduction in power at the gait frequency and the emergence of the expected ERP effect (e.g., larger P300 for incongruent vs. congruent stimuli) [3].

---

Quantitative Data on Artifact Impact and Method Efficacy

Table 1: Consequences of Motion Artifacts on ICA and ERP Metrics

Metric	Impact of Motion Artifacts	Quantitative Measure
ICA Component Quality	Reduced number of brain-like components [18]	Lower number of dipolar components identified by ICLabel [18].
Spectral Power	Introduces spurious power at motion-related frequencies [3]	Significant power increase at gait frequency (e.g., ~2 Hz) and its harmonics [3].
ERP Fidelity	Obscures true neural components and introduces confounds [19]	Failure to detect expected ERP effects (e.g., P300 congruency effect) [3].

Table 2: Comparison of Motion Artifact Removal Methods for Running EEG

Method	Mechanism	Key Parameters	Effectiveness for Running
iCanClean [18] [3]	Uses canonical correlation analysis (CCA) to subtract noise subspaces correlated with reference noise signals.	R² threshold (e.g., 0.65); sliding window (e.g., 4 s) [18].	Superior; produced more dipolar ICs and recovered the P300 effect [3].
Artifact Subspace Reconstruction (ASR) [18] [3]	Uses sliding-window PCA to identify and remove high-variance components exceeding a threshold.	Standard deviation threshold k (e.g., 20-30) [18].	Effective; improved IC dipolarity and reduced gait frequency power, but did not recover P300 effect in one study [3].
ICA alone [18] [3]	Blind source separation without prior cleaning.	N/A	Inadequate; motion artifacts reduce decomposition quality and obscure brain sources [18] [3].

---

Experimental Protocols

Protocol 1: Preprocessing for ICA on Mobile EEG Data

This protocol is adapted from methods used to successfully analyze EEG data during overground running [18] [3].

• Data Acquisition: Record EEG using a mobile system. If possible, use a dual-layer electrode setup where one layer contacts the scalp and a second, mechanically coupled layer does not, to provide a pure noise reference for iCanClean [18].
• Initial Filtering: Bandpass filter the raw data (e.g., 1-70 Hz) using a zero-phase FIR filter [22].
• Motion Artifact Reduction: Apply a dedicated motion cleaning algorithm.
  • For iCanClean: If dual-layer sensors are unavailable, create pseudo-reference noise signals from the raw EEG. Process the data with iCanClean using an R² criterion of 0.65 and a 4-second sliding window [18].
  • For ASR: Calibrate ASR on a clean segment of data (e.g., during quiet standing) and process the data with a k parameter of 20-30 [18].
• Bad Channel Removal & Interpolation: Identify and interpolate channels that are consistently noisy or flat.
• ICA Decomposition: Run ICA (e.g., using the Infomax algorithm in EEGLAB) on the cleaned, continuous data [20].
• Component Classification: Use ICLabel to automatically classify components as brain, muscle, eye, heart, or noise [18].

Protocol 2: Recovering ERPs in a Dynamic Flanker Task

This protocol details the specific approach used to identify stimulus-locked ERPs during running [3].

• Experimental Design: Adapt a cognitive task like the Flanker task for dynamic conditions. Include matched static (standing) and dynamic (jogging) conditions.
• Data Preprocessing: Follow Protocol 1 to preprocess the data and obtain cleaned, ICA-weighted data.
• Epoching: Segment the data into epochs time-locked to the stimulus presentation (e.g., -200 ms to 800 ms).
• Baseline Correction: Apply baseline correction using the pre-stimulus interval.
• Final Artifact Rejection: Reject any remaining epochs containing extreme voltage values (e.g., exceeding ±100 µV) that were not corrected by the previous steps [19].
• Averaging & Analysis: Average the accepted epochs separately for each condition (e.g., Congruent/Incongruent, Standing/Running). Quantify the ERP component of interest (e.g., P300 mean amplitude) and test for the expected experimental effects.

---

Logical Workflow: From Artifacts to Analysis

The following diagram illustrates the logical pathway through which motion artifacts degrade data quality and the critical decision points for effective remediation.

---

The Scientist's Toolkit: Key Research Reagents & Solutions

Table 3: Essential Tools for Motion-Robust Mobile EEG Research

Tool / Solution	Function	Example & Notes
iCanClean Plugin [18] [3]	Motion artifact reduction using canonical correlation analysis (CCA).	An EEGLAB plugin. Most effective with dual-layer electrodes, but can use pseudo-reference signals created from the raw EEG [18].
Artifact Subspace Reconstruction (ASR) [18] [3]	Removes high-amplitude, non-stereotyped artifacts in continuous data.	Available in EEGLAB's CleanRAW plugin. The k parameter is critical; a value of 20-30 is recommended to avoid over-cleaning [18].
ICLabel Classifier [18]	Automates the classification of ICA components into categories (Brain, Muscle, Eye, etc.).	An EEGLAB plugin. Essential for objectively identifying which components to reject after decomposition [18].
Mobile EEG System	Enables data collection outside the lab during whole-body movement.	Systems from companies like Brain Vision, Wearable Sensing, or ENOBIO. Must be lightweight and cable-free to minimize motion artifacts.
Dual-Layer Electrodes [18]	Provides a dedicated noise reference for optimal artifact separation.	The top layer is disconnected from the scalp and records only motion-induced noise, used as a reference for iCanClean [18].

The Modern Toolkit: Signal Processing and Deep Learning for Artifact Removal

Frequently Asked Questions

Q1: My research involves overground running, and I am not using a custom dual-layer EEG system. Can I still use iCanClean? Yes, you can. iCanClean can be effectively implemented using pseudo-reference noise signals derived from your standard EEG data, making it suitable for standard systems without dedicated noise sensors [2]. This approach involves applying a temporary notch filter (e.g., below 3 Hz) to the raw EEG to identify noise-dominated subspaces, which then serve as the reference for the canonical correlation analysis (CCA) cleaning process [2]. Studies on overground running have successfully used this method to improve independent component analysis (ICA) decompositions and recover event-related potentials (ERPs) [2] [3].

Q2: I am getting inconsistent cleaning results when using iCanClean on data from different participants. What are the key parameters I should optimize? The two most critical parameters to optimize for consistent results are the R² cleaning aggressiveness threshold and the sliding window length [23].

• R² Threshold: A lower R² value (e.g., 0.05) results in more aggressive cleaning, while a value closer to 1.0 is more conservative. For mobile EEG data during activities like walking or running, an R² value of 0.65 is often a robust starting point [2] [23].
• Window Length: This defines the temporal segment of data used for the CCA. A 4-second window has been identified as optimal for data with motion artifacts from locomotion [2] [23]. We recommend performing a small parameter sweep on a subset of your data to find the ideal values for your specific experimental setup.

Q3: After cleaning with iCanClean, how can I objectively validate that brain signals were preserved and not accidentally removed? A robust method for validation is to evaluate the quality of your Independent Component Analysis (ICA) decomposition post-cleaning [23]. You should look for an increase in the number of "good" independent components, which are defined as those that are:

• Well-localized as dipoles: Residual variance (RV) from a single dipole model should typically be less than 15% [23].
• Identified as brain activity: Using an automated classifier like ICLabel, the component should have a high probability (e.g., >50%) of being a brain component [23]. An increase in the count of components meeting these criteria indicates successful noise removal without degrading the neural signal of interest [23].

Q4: For a dual-layer setup, how many reference noise channels are necessary for effective cleaning? Research shows that good performance can be maintained even with a reduced set of noise channels [23]. While a full set of noise channels is ideal, studies have found that using 64, 32, or even 16 noise channels still provided a significant improvement in the number of good brain components recovered after ICA compared to no cleaning [23]. This is valuable information for designing more practical and cost-effective dual-layer EEG systems.

Performance Data and Optimal Parameters

The following tables summarize key quantitative findings from recent studies to guide your experimental setup.

Table 1: iCanClean Performance on Data Quality Score (Phantom Head Study) [24]

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

Note: The Data Quality Score is the average correlation between known ground-truth brain sources and the recorded EEG channels. The "Brain" condition serves as a reasonable target for cleaning performance [24].

Table 2: Recommended iCanClean Parameters for Human Locomotion EEG [2] [23]

Parameter	Recommended Value	Application Context
R² Threshold	0.65	Balanced aggressiveness for motion and muscle artifacts during walking/running.
Sliding Window Length	4 seconds	Optimal for capturing motion artifact subspaces in locomotion data.
Noise Channels (Dual-Layer)	16 - 64	Provides significant cleaning improvement; more channels yield marginally better results.

Experimental Protocols

Protocol 1: Implementing iCanClean with a Dual-Layer EEG Setup This protocol is based on studies using high-density EEG during walking and table tennis [23] [25].

• Equipment Setup: Use a custom dual-layer EEG cap where each scalp electrode is mechanically coupled to an inverted, electrically isolated noise electrode via a 3D-printed coupler. The wires from the coupled electrodes should be bundled together [25].
• Data Collection: Record data from both the scalp layer (contains brain signal + noise) and the noise layer (contains only noise). Ensure separate referencing for each layer (e.g., average reference for scalp channels, separate average reference for noise channels) [23].
• Basic Preprocessing: High-pass filter the data (e.g., 1 Hz cutoff). Perform an initial channel rejection to remove channels with abnormally high amplitude (e.g., standard deviation >3x the median across channels) [23].
• Apply iCanClean: Use the preprocessed scalp and noise data as inputs. Set parameters according to Table 2 (R²=0.65, 4-second window). The algorithm will use CCA to identify and subtract noise subspaces from the scalp data [24] [23].
• Validation: Perform ICA on the cleaned data. Calculate the number of "good" brain components (Residual Variance <15%, ICLabel brain probability >50%) and compare it to the number obtained from data cleaned with other methods or uncleaned data [23].

Protocol 2: Implementing iCanClean with Pseudo-Reference Signals This protocol is adapted from research on overground running where dual-layer hardware was not available [2].

• Data Collection: Record EEG data using your standard system during the overground running task.
• Generate Pseudo-References: From your raw, contaminated EEG data, create pseudo-reference noise signals by applying a temporary notch filter. A common approach is to use a high-pass filter with a cutoff below 3 Hz to isolate low-frequency noise dominated by motion artifacts [2].
• Apply iCanClean: Input the raw EEG data and the newly created pseudo-reference signals into the iCanClean algorithm. Use the parameters recommended in Table 2.
• Validation: As with the dual-layer protocol, validate the cleaning efficacy by examining the ICA decomposition. Additionally, for ERP studies, check if the expected components (like the P300 in a flanker task) are recovered and show the anticipated effects (e.g., greater amplitude for incongruent stimuli) [2] [3].

iCanClean Signal Processing Workflow

The diagram below illustrates the core signal processing workflow of the iCanClean algorithm, showing how both dual-layer and pseudo-reference setups integrate into the cleaning pipeline.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Materials and Solutions for iCanClean Experiments

Item	Function / Purpose
Dual-Layer EEG Cap	A custom cap with paired scalp and noise electrodes. The mechanical coupling ensures both experience similar motion, making the noise electrode an ideal reference [23] [25].
3D-Printed Electrode Couplers	Used to physically and rigidly connect a scalp electrode to its corresponding noise electrode, which is crucial for the dual-layer approach [25].
Conductive Fabric	Acts as an artificial skin circuit for the noise electrode layer, bridging the electrodes to capture environmental and motion-based artifacts effectively [25].
iCanClean Algorithm	The core processing algorithm that uses CCA to find and remove subspaces in the EEG data that are highly correlated with the reference noise signals [24] [23].
Pseudo-Reference Signal	A software-generated noise reference, created by filtering the raw EEG (e.g., with a sub-3 Hz notch filter), enabling iCanClean use on standard EEG systems [2].

Core Principles of ASR

What is the fundamental principle behind Artifact Subspace Reconstruction?

Artifact Subspace Reconstruction (ASR) is an adaptive method for cleaning continuous EEG data in real-time or offline. Its core principle is to learn the statistical properties of clean, "calibration" EEG data from a user and then use this reference to identify and reconstruct data segments in subsequent recordings where artifacts (like motion) create high-amplitude, anomalous signals. It operates on the assumption that non-brain artifacts introduce a variance that is large and statistically deviant compared to the baseline brain activity [26].

How does ASR differentiate between brain signals and artifacts?

ASR uses a sliding window to process the EEG data. For each short data segment (e.g., 500 ms), it performs a Principal Component Analysis (PCA). The principal components of the current segment are compared to the statistical distribution of components from the clean calibration data. Any component in the current data with a variance (root mean square) that exceeds a user-defined standard deviation threshold (often referred to as "k") is identified as an artifact. These artifactual components are then removed, and the data segment is reconstructed using the remaining "clean" components and the calibration mixing matrix [27] [26].

Troubleshooting Common ASR Implementation Issues

Why does my ASR-calibrated data fail to clean motion artifacts during running, and how can I fix this?

A primary reason for failure is the inability of the original ASR (ASR~original~) algorithm to identify sufficient high-quality calibration data from recordings involving movement. The standard method for selecting calibration data can be too conservative, rejecting large portions of usable data and resulting in a poor statistical model of clean EEG [28] [29].

Solution: Implement newer variants of ASR designed for high-motion scenarios:

• ASR~DBSCAN~: Uses a non-parametric clustering method (Density-Based Spatial Clustering) to identify clean calibration data on a point-by-point basis. Empirical data shows it finds, on average, 42% of data usable for calibration [28] [29].
• ASR~GEV~: Employs a parametric approach based on the Generalized Extreme Value distribution to define calibration data, finding on average 24% usable data [28] [29].

Both methods significantly outperform ASR~original~, which typically identifies only 9% of data as usable, and subsequently produce Independent Components (ICs) that account for more variance in the original data [29].

How do I choose the right threshold parameter ('k') to avoid overcleaning or undercleaning?

The k parameter is a critical threshold that determines the sensitivity of artifact detection. A lower k value makes the algorithm more aggressive, potentially removing weaker brain signals ("overcleaning"), while a higher k makes it more conservative, possibly leaving artifacts in the data ("undercleaning") [18].

The table below summarizes recommendations based on different research contexts.

Research Context	Recommended k value	Rationale and Expected Outcome
Standard Lab Studies (Non-locomotion)	20–30	A higher threshold is conservative, effectively handling standard artifacts (e.g., eye blinks) while minimizing the risk of removing brain activity [18].
Human Locomotion (e.g., walking)	≥10	A threshold below 10 is not recommended for walking, as it can overclean the data. A value of 10 or higher helps preserve dipolar brain sources during ICA [18].
High-Motion Tasks (e.g., juggling, running)	Use ASR~DBSCAN~ or ASR~GEV~	For intense motor tasks, the improved calibration of these new methods is more critical than fine-tuning k in the original ASR. They better handle non-stationary noise [28].

My Independent Component Analysis (ICA) results are poor after ASR. What is happening?

The presence of large, residual motion artifacts after preprocessing can corrupt the ICA decomposition, reducing its ability to identify maximally independent brain sources [18]. The quality of ICA is often measured by the dipolarity of its components, as true brain sources are typically dipolar.

Solution: Preprocessing with a well-configured ASR or iCanClean has been shown to improve subsequent ICA. Studies on running data show that both ASR and iCanClean lead to the recovery of more dipolar brain independent components compared to no cleaning or other methods [18] [3]. If using ASR does not yield good ICA results, consider trying iCanClean, which in some direct comparisons was found to be "somewhat more effective than ASR" in producing dipolar components and recovering expected ERP effects like the P300 [18] [2].

ASR Workflow and Experimental Protocol

The following diagram illustrates the two-stage workflow of the standard ASR algorithm, from calibration to processing.

Detailed Methodology for an Overground Running Experiment

The protocol below is adapted from recent research comparing motion artifact removal techniques [18] [2].

• Participants: Recruit young adult athletes (e.g., 18-30 years old).
• EEG System: Use a high-density, wireless mobile EEG system (e.g., 205-channel or 32-channel setup).
• Accelerometer: Place a 3-axis accelerometer on the participant's forehead to monitor head motion and determine stepping frequency [30].
• Experimental Task: Employ a Flanker task adapted for dynamic conditions. Participants respond to congruent or incongruent arrows while both jogging on a treadmill/overground and standing.
• Data Acquisition:
  • Record a baseline of clean EEG during standing or sitting rest (1-2 minutes) for ASR calibration.
  • Record EEG during the standing Flanker task (artifact-free control condition).
  • Record EEG during the dynamic jogging Flanker task.
• Data Preprocessing:
  • Apply a band-pass filter (e.g., 1-50 Hz) to remove slow drifts and high-frequency noise.
  • Apply ASR: Use the resting baseline data for calibration. Process the continuous data from both standing and running conditions. Parameters such as k=20 can be a starting point, adjusted based on the troubleshooting guide above.
• Downstream Analysis & Evaluation:
  • Perform Independent Component Analysis (ICA) on the ASR-cleaned data.
  • Evaluate ICA quality by calculating the number and proportion of dipolar brain components.
  • Evaluate artifact removal by analyzing power spectral density at the gait frequency and its harmonics; successful cleaning shows significant power reduction [18].
  • Evaluate neural signal preservation by comparing the recovered Event-Related Potentials (ERPs), like the P300 from the Flanker task, between the standing and running conditions.

Comparative Performance Data

The following table quantifies the performance of different ASR approaches and other leading methods in handling motion artifacts, based on recent studies.

Table 1: Comparative Performance of Artifact Removal Methods in Mobile EEG Studies

Method	Key Principle	Reported Performance in Motion Artifact Removal
ASR~original~	PCA-based reconstruction using a clean calibration period.	Found only 9% of data usable for calibration during juggling. Produced brain ICs explaining 26% of data variance [29].
ASR~DBSCAN~	Uses clustering to improve calibration data selection.	Found 42% of data usable for calibration. Produced brain ICs explaining 30% of data variance [29].
ASR~GEV~	Uses extreme value statistics for calibration.	Found 24% of data usable for calibration. Produced brain ICs explaining 29% of data variance [29].
iCanClean	Uses canonical correlation analysis (CCA) with noise references.	Somewhat more effective than ASR in producing dipolar ICs during running. Enabled identification of the expected P300 congruency effect [18] [3].
rASR	Uses Riemannian geometry for covariance matrix processing.	Outperformed original ASR in reducing eye-blink artifacts and improving Visual-Evoked Potential (VEP) signal-to-noise ratio, with favorable computation time [26].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Materials and Software for ASR-based EEG Research

Item	Specification / Example	Function in Experiment
High-Density Mobile EEG System	32+ channel wireless system (e.g., ANT Neuro eego sports)	Records brain electrical activity with minimal movement constraints [30].
Passive Wet Electrodes	Ag/AgCl electrodes	Ensure stable signal quality. Impedance should be kept low (e.g., < 20 kΩ) [30].
3-Axis Accelerometer	Lightweight sensor (often part of mobile EEG systems)	Placed on the forehead to monitor head motion and identify gait frequency for artifact analysis [30].
Signal Processing Software	EEGLAB with clean_rawdata plugin	Provides the standard implementation of the ASR algorithm for offline analysis [26].
Calibration Data	1-2 minutes of resting-state EEG	Serves as the clean reference for building the ASR statistical model [27] [26].
iCanClean Algorithm	Alternative to ASR; requires pseudo-reference or dual-layer sensor noise signals	Provides a high-performance alternative for motion artifact removal, especially effective with dedicated noise sensors [18].

Technical Support Center

Troubleshooting Guides

Issue 1: Model Performance is Poor on Small Datasets

• Problem: The deep learning model does not converge or achieves low artifact removal accuracy when training data is limited.
• Solution: Integrate Visibility Graph (VG) features into your model input. VG features convert time-series EEG into graph structures, providing supplemental structural information that enhances learning stability and model accuracy, making it particularly effective for smaller datasets [5].

Issue 2: Inconsistent Signal Integrity After Cleaning

• Problem: The cleaning process successfully removes artifacts but also distorts or removes genuine neural signals.
• Solution: For subject-specific frameworks, ensure data is processed separately per individual. Implement a loss function during training that includes a correlation metric to enforce signal morphology preservation. Using separate encoding pathways for VG features and raw EEG can also improve consistency [5].

Issue 3: High Computational Load During Model Training

• Problem: Training a 1D CNN model like Motion-Net is computationally intensive and slow on a local machine.
• Solution: Optimize training by using a batch size of one subject. Ensure your hardware meets the recommended specification of a single GPU with at least 20 GB of memory to handle the computational requirements efficiently [31].

Frequently Asked Questions (FAQs)

Q: What is the key advantage of a subject-specific model like Motion-Net over generalized approaches? A: Subject-specific models are trained and tested on data from individual participants. This accounts for the high variability in both EEG signals and motion artifact characteristics across different people, leading to more robust and accurate artifact removal compared to a one-size-fits-all model [5].

Q: Can I use these models with the data from my wireless EEG system recorded during running? A: Yes, models like Motion-Net are designed for mobile EEG (mo-EEG) applications. For optimal results, ensure your dataset includes a proper ground-truth reference. Preprocessing techniques such as iCanClean or Artifact Subspace Reconstruction (ASR) have also been validated specifically for running data and can be considered [5] [2] [3].

Q: Are pre-trained models available, or do I need to train from scratch? A: You are welcome to use existing foundation models. However, for a subject-specific approach, you will likely need to fine-tune any pre-trained model on your specific dataset. You must clearly document any pre-trained models used in your work [31].

Q: My research involves analyzing Event-Related Potentials (ERPs). Can deep learning cleaning preserve these? A: Yes, when properly configured. For instance, preprocessing mobile EEG data with methods like iCanClean has been shown to successfully recover expected ERP components, such as the P300 congruency effect, during running tasks [2] [3].

Experimental Protocols & Quantitative Performance

Motion-Net Experimental Workflow

The following diagram illustrates the core experimental workflow for developing and validating a subject-specific deep learning model for motion artifact removal.

Detailed Methodology for Motion-Net

The protocol for the Motion-Net framework, as described in the search results, involves several key stages [5]:

• Data Acquisition and Preprocessing:

  • Collect real EEG recordings with motion artifacts alongside a ground-truth (artifact-free) reference for each subject separately.
  • Synchronize data using experiment triggers. Preprocessing may include resampling and baseline correction (e.g., using a polynomial fit) to improve signal alignment.

• Feature Engineering:

  • Extract Visibility Graph (VG) features from the preprocessed EEG signals. This step converts the 1D time-series signal into a graph structure, capturing non-linear properties and dynamics that enhance the model's learning capability, especially with smaller datasets.

• Model Training (Subject-Specific):

  • Design a 1D Convolutional Neural Network (CNN) based on a U-Net architecture, named Motion-Net. This model is trained to map artifact-contaminated EEG signals to their clean counterparts.
  • The training is performed separately for each subject, ensuring the model adapts to individual-specific artifact patterns and brain signals.

• Model Validation and Testing:

  • Evaluate the model on held-out data from the same subject using quantitative metrics such as Artifact Reduction Percentage (η), Signal-to-Noise Ratio (SNR) improvement, and Mean Absolute Error (MAE).

Performance Metrics Table

The following table summarizes the quantitative performance of the Motion-Net model as reported in its source study, providing benchmarks for expected outcomes [5].

Metric	Reported Performance	Interpretation
Artifact Reduction (η)	86% ± 4.13	High percentage of motion artifact successfully removed from the signal.
SNR Improvement	20 ± 4.47 dB	Significant enhancement in the signal-to-noise ratio after processing.
Mean Absolute Error (MAE)	0.20 ± 0.16	Low error between the cleaned signal and the ground-truth reference.

Comparative Methods Table

The table below compares other contemporary artifact removal methods validated for use during locomotion, such as running [2] [3].

Method	Principle	Key Parameters	Use Case in Locomotion
Motion-Net	1D CNN (U-Net) with VG features	Subject-specific training	Subject-specific motion artifact removal for mobile EEG.
iCanClean	Canonical Correlation Analysis (CCA) with noise references	R² threshold (e.g., 0.65), sliding window (e.g., 4s)	Effective for preprocessing running EEG; improves ICA dipolarity and recovers P300 ERPs.
Artifact Subspace Reconstruction (ASR)	Principal Component Analysis (PCA) & calibration data	Standard deviation threshold k (e.g., 10-30)	Effective for preprocessing running EEG; reduces power at gait frequency.

The Scientist's Toolkit

Essential Research Reagents & Materials

Item Name	Function / Explanation
Mobile EEG System with Accelerometer	A wireless EEG system is fundamental for recording during overground running. An integrated accelerometer provides motion data that can be used for artifact analysis and synchronization [5] [32].
Ground-Truth Reference EEG	Clean, artifact-free EEG recordings from the same subject are crucial for training and validating supervised deep learning models like Motion-Net [5].
Visibility Graph (VG) Algorithm	A tool to convert EEG time-series into graph structures, providing additional features that improve deep learning model performance on smaller datasets [5].
Dual-Layer or Pseudo-Reference Electrodes	Dedicated sensors that capture only noise (motion artifacts). These are mechanically coupled to scalp electrodes and are used by algorithms like iCanClean to identify and subtract noise subspaces [2] [1].
High-Performance Computing (HPC) GPU	A single GPU with at least 20 GB of memory is recommended to handle the computational demands of training deep learning models like CNNs efficiently [31].

Signaling Pathway of Motion Artifact Generation

Understanding the sources of motion artifacts is key to addressing them. The diagram below maps the primary pathways through which motion artifacts corrupt the EEG signal.

## FAQs: Core Concepts and Method Selection

What is the core advantage of an integrated preprocessing and ICA pipeline? An integrated pipeline standardizes the initial "clean-up" of EEG data (e.g., handling line noise and bad channels) to create a better-quality signal for subsequent ICA. This is crucial because the presence of large motion artifacts can contaminate ICA's ability to effectively separate brain from non-brain sources. A robust preprocessing stage improves the quality of the ICA decomposition, leading to more dipolar and physiologically plausible independent components [33] [18].

Should I use Artifact Subspace Reconstruction (ASR) or iCanClean for motion artifacts during running? The choice depends on your equipment and specific goals. Recent comparative studies indicate that both are effective, but with some differences:

• iCanClean is somewhat more effective at improving ICA component dipolarity and can help recover expected ERP components like the P300 during running. It is ideal if you have access to dual-layer electrodes that provide pure noise references [18] [2] [3].
• ASR also significantly reduces motion artifacts and improves ICA. It is a powerful option when using standard EEG systems without dedicated noise sensors. Its performance is highly dependent on the chosen threshold parameter (k); a value that is too low can "overclean" the data [18] [2].

My ICA results are poor during locomotion tasks. What should I check? Poor ICA decomposition during locomotion is often due to incomplete removal of large-amplitude motion artifacts prior to running ICA. Consider the following:

• Aggressive Preprocessing: Implement a robust preprocessing pipeline like PREP to handle line noise and bad channels before ICA [33].
• Pre-ICA Cleaning: Use methods like ASR or iCanClean to reduce the motion artifact burden in the continuous data, which leads to the recovery of more dipolar brain components [18].
• Parameter Tuning: Ensure you are using parameters validated for locomotion. For iCanClean with pseudo-references, an R² of 0.65 with a 4-second sliding window has been shown to be effective. For ASR, avoid overly aggressive k values below 10 [18].

How can I validate that my artifact removal worked without a ground-truth signal? You can use multiple, indirect performance-based metrics to build confidence in your results:

• ICA Dipolarity: Calculate the dipolarity of your independent components. Cleaner data should yield a higher number of components with a dipolar scalp topography, which is a hallmark of a cerebral source [18] [2].
• Spectral Power: Examine the power spectrum at the gait frequency and its harmonics. Successful motion artifact reduction should show significantly reduced power at these frequencies [18].
• ERP Recovery: If your paradigm includes event-related potentials, check if you can recover known ERP components (e.g., the P300 in a Flanker task) with the expected latency and experimental effects after processing [18] [2].

## Troubleshooting Guides

### Problem: Persistent Motion Artifact After Standard Preprocessing

Symptoms: High-amplitude, rhythmic noise in the EEG signal time-locked to the gait cycle; unsuccessful ICA decomposition; muscle artifacts obscuring brain signals.

Solution A: Implement iCanClean with Pseudo-References

Step	Action	Key Parameters & Tips
1.	Create Pseudo-Noise Signals	Generate reference noise signals from your raw EEG data. Apply a temporary high-pass or notch filter (e.g., below 3 Hz) to isolate low-frequency motion artifacts [18].
2.	Run iCanClean	Use Canonical Correlation Analysis (CCA) to identify and subtract noise subspaces from the scalp EEG.	Set the correlation criterion (R²) to 0.65 and use a 4-second sliding window, as validated for locomotion data [18].
3.	Proceed to ICA	Perform ICA on the cleaned data.	The resulting independent components should show improved dipolarity, facilitating better classification of brain and artifact components [18].

Solution B: Configure Artifact Subspace Reconstruction (ASR)

Step	Action	Key Parameters & Tips
1.	Select Calibration Data	Identify a clean segment of your recording to use as a reference. ASR can automatically select periods with RMS z-scores within -3.5 to 5.0 for at least 92.5% of electrodes [18] [2].
2.	Set the Threshold (k)	Choose a k parameter that balances artifact removal and signal preservation.	For locomotion, start with a k between 10 and 20. A lower k is more aggressive; values below 10 may overclean the data [18].
3.	Reconstruct Data	ASR will use a sliding-window PCA to identify and remove artifact components that exceed the threshold, reconstructing the data based on the clean calibration period [18].

### Problem: Inability to Recover Expected ERP Components

Symptoms: The P300 (or other target ERP) is absent, delayed, or shows an unexpected amplitude in the processed data from a dynamic task.

Solutions:

• Benchmark Against a Static Condition: Always run the same experimental task (e.g., Flanker task) under static standing conditions. Use the ERP from the static condition as a benchmark to compare against the dynamic (running) condition [18] [2].
• Compare Preprocessing Pipelines: Process your dynamic data with different methods (e.g., standard preprocessing vs. preprocessing with iCanClean/ASR). The pipeline that produces an ERP in the dynamic condition most similar to your static benchmark is likely the most effective. Research shows iCanClean may be particularly effective for recovering the P300 congruency effect during running [18].
• Verify Task Performance: Ensure that participants' behavioral accuracy during the running task is high. A poorly defined ERP could be due to a lack of task engagement or performance degradation caused by the dual-task nature of running and cognitive testing.

## Experimental Protocols for Method Validation

### Protocol: Validating Motion Artifact Removal During Overground Running

Objective: To compare the efficacy of different preprocessing pipelines in reducing motion artifacts and preserving neural signals during running.

1. Experimental Design:

• Participants: Young adult athletes recommended for consistency in locomotion [18].
• Task: An adapted Eriksen Flanker task performed under two conditions:
  • Dynamic Condition: While jogging overground.
  • Static Control Condition: While standing still.
  • This design allows for direct comparison of ERPs (e.g., P300) between movement states [18] [2].
• Data Recording: Use a mobile EEG system with a sufficient number of channels to support ICA.

2. Data Processing and Comparison Pipelines: Process the data from the dynamic condition through the following pipelines for comparison:

• Pipeline 1: Standard preprocessing (filtering, bad channel removal) + ICA.
• Pipeline 2: Standard preprocessing + ASR + ICA.
• Pipeline 3: Standard preprocessing + iCanClean (with pseudo-references) + ICA.

3. Performance-Based Metrics for Comparison: Use the following quantitative measures to evaluate each pipeline:

Table: Key Metrics for Pipeline Validation

Metric	How to Calculate/Measure	Interpretation of a Good Result
Component Dipolarity	Use tools like ICLabel or measure the dipole fit of independent components.	A higher number of dipolar brain components indicates a better decomposition [18] [2].
Power at Gait Frequency	Compute the power spectral density and extract power at the fundamental step frequency and its harmonics.	Significant reduction in power at these frequencies indicates effective motion artifact suppression [18].
ERP Quality	Calculate the average ERP for congruent vs. incongruent trials in the Flanker task.	The recovery of a P300 with greater amplitude for incongruent trials, similar to the static condition, indicates preserved neural information [18].

## The Scientist's Toolkit: Research Reagent Solutions

Table: Essential Tools for Mobile EEG and Motion Artifact Correction

Tool / Method	Function	Application Context
PREP Pipeline	Standardized early-stage preprocessing for large-scale EEG. Handles line noise removal, robust average referencing, and bad channel detection.	Provides a consistent, automated baseline preprocessing step before applying more specialized artifact removal methods [33].
ICA (EEGLAB)	Blind source separation to decompose EEG data into maximally independent components (brain and non-brain).	Core method for isolating and removing artifacts like eye blinks and muscle activity, but works best on data pre-cleaned of severe motion artifacts [34] [18].
ICLabel	Automated classifier for ICA components. Labels components as brain, muscle, eye, heart, line noise, or channel noise.	Speeds up the component selection process after ICA, though it may be less reliable for motion artifacts not in its training set [18].
Artifact Subspace Reconstruction (ASR)	Identifies and removes high-variance, high-amplitude artifacts from continuous data using a sliding-window PCA approach.	Effective for real-time or offline cleaning of motion artifacts in mobile EEG. Performance depends on calibration data and the k parameter [18] [2].
iCanClean	Uses canonical correlation analysis (CCA) and reference noise signals to subtract motion artifact subspaces from the EEG.	Highly effective for motion artifact removal, especially with dual-layer electrodes. Can also use pseudo-references derived from the EEG itself [18] [2].
Mobile EEG System	A portable, amplifier-integrated EEG system that allows for unrestricted movement.	Essential for recording brain activity during whole-body movement like overground running [35].

## Integrated Workflow and Decision Diagrams

### EEG Preprocessing and ICA Integration Pipeline

### Motion Artifact Removal Method Selection

Fine-Tuning for Success: Parameter Selection and Hardware Considerations

This technical support guide provides researchers with detailed, evidence-based protocols for parameter selection in two prominent EEG artifact removal algorithms, iCanClean and Artifact Subspace Reconstruction (ASR), with a specific focus on handling motion artifacts during overground running.

Troubleshooting Guide: Algorithm Selection and Parameter Optimization

1. How do I choose between iCanClean and ASR for my running study? The choice depends on your experimental setup and the nature of your analysis. iCanClean generally outperforms ASR in removing motion artifacts and preserving brain signals, making it particularly suitable for recovering event-related potentials (ERPs) during running [18]. However, ASR remains a powerful and widely used method, especially when clean calibration data is available. For studies aiming to detect specific ERP components like the P300 during running, iCanClean has shown superior efficacy [18]. ASR can be an excellent choice for general-purpose cleaning and when working with standard EEG systems without dedicated noise sensors [36].

2. My data is still noisy after using default parameters. Should I adjust the R² or k? Yes, default parameters are a starting point and may require optimization for specific tasks like running. The key is to adjust parameters aggressively while avoiding "over-cleaning" that removes brain signals [18].

• For iCanClean, if noise persists, consider using a more aggressive (lower) R² threshold (e.g., 0.65) [18]. Ensure you are using an appropriate sliding window (e.g., 4 seconds) [18].
• For ASR, if artifacts remain, use a more aggressive (lower) k value (e.g., 10-20) [18] [36]. Be cautious, as values below 10 may remove neural signals and are not generally recommended [36].

3. What is the risk of setting an overly aggressive parameter? Overly aggressive cleaning can remove genuine brain activity, distorting your neural signals.

• In iCanClean, an R² threshold that is too low might remove high-amplitude, low-frequency brain signals that are correlated with the noise reference [37].
• In ASR, a k value that is too low leads to a higher percentage of data being modified. For example, k=10 can modify over 60% of data, while k=100 modifies only about 3% [36]. Over-cleaning reduces the validity of your findings.

Parameter Performance Tables

Table 1: iCanClean R² Threshold Performance Guide

R² Threshold	Cleaning Aggressiveness	Recommended Use Case	Empirical Support
~0.65	More Aggressive	Optimal for motion artifact removal during human running; produces more dipolar ICA components and recovers P300 effects [18].	Young adults during overground running; Flanker task [18].
Varies by data	Adaptive	General use with pseudo-reference signals; performance depends on accurate noise subspace identification [18] [37].	Phantom head testing with simulated walking artifacts [37].

Table 2: ASR 'k' Value Performance Guide

k Value	Cleaning Aggressiveness	Recommended Use Case	Empirical Support & Notes
10	Very Aggressive	Not generally advised; may remove brain activity. Maximum ICA quality index in some motor tasks [36].	Use with extreme caution; no clear benefit over k=20 [36].
10 - 20	Aggressive	Motor tasks with high-amplitude motion artifacts (e.g., running, juggling) [18] [28].	Improves ICA decomposition quality during running [18].
20 - 30	Moderate	Recommended standard range; good balance between artifact removal and brain signal preservation [36].	Effective for ocular and muscle artifacts; optimal for subsequent ICA [36].
Up to 100	Conservative	Removing only extreme, high-amplitude artifacts [36].	Modifies a very small portion of data (~3%) [36].

Detailed Experimental Protocols

Protocol 1: Validating iCanClean with Pseudo-References for Running This protocol is adapted from a study that successfully recovered ERP components during an overground running task [18].

• Objective: To remove motion artifacts and recover stimulus-locked ERPs during running.
• EEG Recording: Record EEG from participants performing a dynamic task (e.g., adapted Flanker task) during both running and static standing.
• iCanClean Preprocessing:
  • Create Pseudo-Reference Noise Signals: Temporarily apply a notch filter to the raw EEG (e.g., below 3 Hz) to isolate noise subspaces, as dedicated noise sensors may not be available [18].
  • Apply iCanClean: Use a sliding window of 4 seconds and an R² threshold of 0.65 [18].
  • Execute Canonical Correlation Analysis (CCA): iCanClean uses CCA to identify and subtract noise subspaces from the scalp EEG that are correlated with the pseudo-reference [18].
• Validation Metrics:
  • ICA Dipolarity: Assess the quality of subsequent Independent Component Analysis (ICA) by counting the number of brain-independent components with dipolar properties [18].
  • Spectral Power: Check for a significant reduction in power at the gait frequency and its harmonics [18].
  • ERP Analysis: Confirm that expected ERP components (e.g., P300) with correct latencies and congruency effects (e.g., higher amplitude for incongruent stimuli) are recovered [18].

Protocol 2: Tuning ASR for High-Motion Scenarios This protocol is suitable for cleaning EEG data from whole-body movements like running or juggling [18] [28].

• Objective: To aggressively remove high-amplitude motion artifacts without over-cleaning.
• EEG Recording: Record high-density EEG (e.g., 64+ channels) during the motor task. A short baseline recording (e.g., 1-minute resting state) is needed for calibration.
• ASR Preprocessing:
  • Calibration: Use the clean baseline data to calibrate the ASR statistical model. Newer methods like ASR_DBSCAN or ASR_GEV can automatically identify higher-quality calibration data from non-stationary recordings [28].
  • Set Parameter: Apply ASR with a k value of 10-20 [18]. Start with 20 and move to 10 if artifacts persist.
• Validation Metrics:
  • Data Modification: Be aware that a k=10 may modify over 60% of the data [36].
  • ICA Output: Evaluate the number of brain-related independent components and the variance they account for. Data preprocessed with optimized ASR should produce brain ICs that explain more variance [28].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials for Motion Artifact Research

Item / Solution	Function in Research
Dual-Layer EEG System	The primary tool for optimal iCanClean performance. The upper "noise" layer, mechanically coupled but not in contact with the scalp, provides a pure motion artifact reference [18] [37].
High-Density EEG Cap (64+ channels)	Enables better spatial filtering and source separation using ICA, which complements both iCanClean and ASR cleaning [36].
Electrical Phantom Head	Provides known ground-truth brain and artifact signals for quantitative validation and benchmarking of cleaning algorithms without neural variability [37].
Robust Motion Platform	Simulates human gait cycles in a controlled manner for method development and testing against a reliable ground truth [38].

Technical Workflow and Algorithm Principles

Diagram 1: Motion Artifact Cleaning Workflow for Running EEG

Diagram 2: Core Algorithm Principles for iCanClean and ASR

Frequently Asked Questions

Q1: Can I use iCanClean if I don't have a dual-layer EEG system with dedicated noise sensors? Yes, you can. iCanClean can generate "pseudo-reference" noise signals directly from your standard scalp EEG data. This is typically done by temporarily applying a notch filter (e.g., below 3 Hz) to isolate noise subspaces, which are then used in the canonical correlation analysis [18]. While not as ideal as dedicated noise sensors, this approach has been shown to be effective during running tasks [18].

Q2: What is a "dipolar" independent component, and why is it a validation metric? Independent sources of EEG activity in the brain are typically dipolar, meaning their electrical field can be modeled by a single equivalent dipole located within the brain [18]. Therefore, after cleaning, a higher number of ICA components with dipolar properties indicates a more successful separation of brain from non-brain signals and a higher quality decomposition [18].

Q3: Are there newer versions of ASR I should consider? Yes, researchers are actively developing improvements to the original ASR algorithm. Two notable versions are ASR_DBSCAN and ASR_GEV, which use different statistical approaches (Density-Based Spatial Clustering and Generalized Extreme Value distribution) to better identify clean calibration data from recordings with non-stationary noise, such as during intense motor tasks [28]. Another advanced version is Riemannian ASR (rASR), which uses Riemannian geometry for covariance matrix processing and has been shown to outperform the original ASR in some scenarios, including computation time [26].

Frequently Asked Questions (FAQs)

Q1: What are the primary hardware-related sources of motion artifacts in mobile EEG? Motion artifacts in mobile EEG primarily originate from three key hardware areas [39]:

• The Electrode-Skin Interface: Relative movement between the electrode and the skin alters the ion distribution at the interface, causing slow baseline voltage shifts that are often correlated with the movement frequency (e.g., the gait cycle).
• Connecting Cables: Movement of the cables connecting the electrodes to the amplifier generates spike-like artifacts due to triboelectric effects, where friction and deformation of the cable insulator create an additive voltage potential.
• The Electrode-Amplifier System: Unstable electrode-skin contact can cause sudden variations in impedance, which modulates power line interference (PLI), introducing unpredictable, non-repeatable artifacts across the EEG spectrum.

Q2: How do textile electrodes help mitigate motion artifacts, and what are their limitations? Textile electrodes, or "textrodes," offer a potential solution for long-term, unobtrusive monitoring [40]. Their soft, flexible nature helps avoid pressure points on the scalp, which is a significant advantage for sensitive skin or prolonged use. However, a key limitation is that they typically do not work as dry electrodes and require a contact medium such as standard electrode paste or saline solution to function effectively [40]. Furthermore, most current textile electrode systems are limited to hairless regions of the scalp (e.g., frontal and temporal areas), making them incompatible with comprehensive studies requiring coverage of the parietal sensorimotor cortices [39].

Q3: What are active electrodes, and how effective are they against motion artifacts? Active electrodes incorporate a pre-amplifier integrated directly into the electrode itself. This design is highly effective at rejecting power line interference caused by capacitive coupling between connecting cables and the environment [39]. However, studies have shown that concerning motion artifact reduction during dynamic recordings, the performance of active electrodes is comparable to that of passive electrodes. They also add to the overall encumbrance of the acquisition system, which can limit portability and usability in dynamic contexts [39].

Q4: What are the advanced hardware designs for motion artifact cancellation? Researchers are developing sophisticated electrode systems that use hardware to isolate and remove noise.

• Dual-Layer EEG Systems: This design uses pairs of mechanically coupled but electrically isolated electrodes at each recording site [6]. The primary electrode records a mix of brain activity and motion artifacts, while the secondary electrode, which is not in contact with the scalp, records only motion artifacts and environmental noise. This allows for the noise to be subtracted directly from the primary signal [6].
• Tripolar Concentric Ring Electrodes (TCREs): These electrodes feature a central disk surrounded by two concentric rings. This unique geometry enables the calculation of a surface Laplacian, which acts as a spatial filter. It enhances localized, high-frequency brain signals while attenuating broadly distributed, low-frequency artifacts like those from distant muscles, thereby improving spatial selectivity and artifact resistance [41].

Troubleshooting Guides

Problem: Low-frequency, rhythmic baseline wander in the EEG signal during walking or running.

• Potential Cause: Motion artifacts generated at the electrode-skin interface due to rhythmic shifts between the electrode and the skin [39].
• Solutions:
  • Improve Mechanical Coupling: Ensure the electrode cap is snug and that electrodes are firmly seated. Using a cap system with padding can help maintain stable pressure.
  • Optimize Skin Preparation: Clean the skin thoroughly to reduce baseline impedance. Use a high-quality conductive gel to ensure a stable and low-impedance connection.
  • Consider Electrode Type: Textile electrodes, when used with a conductive medium, can provide a softer, more conforming interface that may reduce these slow shifts [40].

Problem: Sharp, spike-like, non-repeatable artifacts in the signal.

• Potential Cause: Triboelectric noise from connecting cables. Cable movement generates friction, creating additive voltage potentials [39].
• Solutions:
  • Secure Cables: Use tape, velcro straps, or a specialized cable management system to bundle and secure cables to the subject's clothing or body, minimizing independent movement. A rear-exiting cable bundle is often effective [6].
  • Use Textile-Integrated Systems: Systems where conductive fibers are embroidered directly into a headband or cap can eliminate traditional dangling cables, thus removing this source of artifact [42].
  • Consider Hardware Shielding: Some advanced textile systems use layered conductive materials to actively shield the signal transmission wires, mimicking the properties of coaxial cables [42].

Problem: Unstable power line interference (PLI) that appears or worsens with movement.

• Potential Cause: Movement-induced modulation of electrode-skin impedance, leading to a fluctuating common-mode signal and unstable PLI rejection [39].
• Solutions:
  • Check Electrode Contact: Ensure all electrodes (including ground and reference) maintain good contact. A sudden change in a single electrode's impedance can modulate PLI for all channels.
  • Use Active Electrodes: Active electrodes are specifically designed to have high input impedance, which helps reject capacitively coupled PLI, making the system less sensitive to impedance changes at the electrode-skin interface [39].

Comparative Analysis of Hardware Solutions

The table below summarizes the key characteristics of different electrode technologies for motion artifact management.

Electrode Technology	Mechanism for Motion Artifact Reduction	Key Advantages	Key Limitations / Challenges
Textile Electrodes [42] [39] [40]	Soft, flexible interface reduces pressure and mechanical shifts. Can be integrated into garments to eliminate cables.	Unobtrusive, comfortable for long-term use. No hard pressure points. Enables integration into headbands/caps.	Often require a contact medium (gel/saline). Generally limited to hairless scalp regions.
Active Electrodes [39]	Integrated pre-amplifier provides high input impedance, rejecting environmental PLI.	Excellent rejection of power line interference. Can be used with a wider range of electrode-skin impedances.	Comparable to passive electrodes on motion artifact. Adds bulk and weight to the electrode assembly.
Dual-Layer Electrodes [6]	Secondary "noise" electrode records only artifacts, enabling direct noise subtraction from the primary signal.	Direct hardware-based artifact isolation. Proven effective during high-motion activities like running.	More complex and cumbersome setup. Requires double the number of recording channels.
Tripolar Concentric Ring Electrodes (TCREs) [41]	Surface Laplacian derivation enhances localized brain signals and attenuates distant, broad artifacts like muscle noise.	Real-time, hardware-based spatial filtering. Improved spatial selectivity and muscle artifact resistance.	Novel technology, less established in consumer/clinical systems. More complex manufacturing.

Detailed Experimental Protocol: Dual-Layer EEG Validation

This protocol outlines the methodology for validating a dual-layer EEG system using a head phantom, a crucial step before human trials [6].

1. Head Phantom and Signal Generation:

• Construct a human-head-shaped phantom using a material that mimics the electrical properties of the head (e.g., dental plaster or ballistics gelatin).
• Implant several electrical dipolar antennas (e.g., 6-14) at spatially distributed locations within the phantom.
• Use a digital-to-analog converter to broadcast artificial brain signals from these antennas. These signals should consist of randomly occurring, overlapping sinusoidal bursts with distinct frequency content (e.g., 7, 13, 17, 23, 29, 37 Hz) to simulate independent neural sources.

2. Motion Reproduction:

• Record real head movement trajectories from a subject walking on a treadmill using an Inertial Measurement Unit (IMU).
• Program a robotic motion platform (e.g., a hexapod) to reproduce these recorded head movements while the phantom is mounted on it.

3. Dual-Layer EEG Array Setup:

• Assemble an array of dual-layer electrodes. Each pair consists of a standard scalp-interfacing pin electrode rigidly coupled to an inverted "noise" electrode.
• Ensure the noise electrode is electrically isolated from the scalp and is mechanically coupled to experience the same motion.
• Place the electrode array on the phantom scalp, using conductive gel for the scalp electrodes.
• For the noise electrodes, fit a conductive fabric cap over the entire array and inject conductive gel between the fabric and each noise electrode to create an external artificial skin circuit.

4. Data Acquisition and Analysis:

• Record data from both the scalp and noise electrodes simultaneously on separate but synchronized systems.
• Process the data by high-pass filtering (e.g., 1 Hz) and re-referencing.
• Perform Independent Component Analysis (ICA) on the combined dataset (scalp + noise channels).
• Validation: Compare the extracted independent components to the ground-truth input signals using cross-correlation and power spectral analysis. Successful artifact removal is indicated by components that show a clear spectral peak at the broadcast frequency with negligible motion-related noise.

Hardware Solutions for Motion Artifact Mitigation

The diagram below illustrates how different hardware solutions target specific sources of motion artifacts in the EEG signal acquisition chain.

The Scientist's Toolkit: Essential Research Reagents & Materials

Item	Function in Research
Conductive Textile Headband [42]	A platform integrating textile electrodes (textrodes) for recording from hairless frontal regions, enabling assessment of new EEG systems in simple paradigms (EO/EC).
Conductive Gel / Electrode Paste [40]	A necessary contact medium for textile and many other electrodes to ensure a stable, low-impedance connection by hydrating the skin and creating a conductive pathway.
Dual-Layer Electrode Array [6]	A set of mechanically coupled but electrically isolated electrode pairs for primary (EEG + noise) and secondary (noise only) recording, enabling hardware-based motion artifact subtraction.
Tripolar Concentric Ring Electrodes (TCREs) [41]	Electrodes with a central disk and two concentric rings that allow for surface Laplacian derivation, providing inherent spatial filtering and enhanced resistance to muscle artifacts.
Head Phantom (e.g., Ballistics Gelatin) [41]	A physical model of the human head with implanted electrical antennas for broadcasting simulated brain signals, allowing for controlled validation of EEG systems against a ground truth.
Robotic Motion Platform [6]	A system to reproduce real human head movement trajectories on a head phantom, creating realistic and repeatable motion artifacts for testing hardware solutions.
Inertial Measurement Unit (IMU) [6]	A sensor placed on the forehead to capture the timing and kinematics of head movements during locomotion, which can be used for motion analysis and artifact modeling.

FAQs: Balancing Artifact Removal and Signal Integrity

What is "overcleaning" in the context of mobile EEG? Overcleaning occurs during the preprocessing of electroencephalography (EEG) data when aggressive artifact removal strategies inadvertently distort or remove the underlying neural signals of interest. This is a significant risk in mobile EEG studies, such as those involving overground running, where motion artifacts and brain signals can occupy overlapping frequency domains. Overcleaning can lead to data that appears clean but has lost critical neurophysiological information, potentially resulting in false conclusions [18].

Why is overcleaning a particular concern for EEG research during running? During running, motion artifacts are pervasive and can have large amplitudes and broad spectral content, making them challenging to separate from brain activity. Researchers may be tempted to use very aggressive filtering or artifact correction settings to achieve a clean-looking signal. However, this can remove genuine brain dynamics related to motor control, obstacle avoidance, and cognitive processing that are the primary targets of investigation in locomotor studies [18] [6].

How can I tell if my data has been overcleaned? There is no single definitive test, but several indicators can signal potential overcleaning:

• Unphysiological Signals: The cleaned EEG appears too "perfect" or lacks the characteristic complexity of neural data.
• Loss of Expected Neural Patterns: Known brain responses, such as event-related potentials (ERPs) like the P300 during a cognitive task, are absent or severely attenuated in the cleaned data, even when they are present in the raw recordings or data from static conditions [18].
• Reduced Component Dipolarity: In Independent Component Analysis (ICA), overcleaning can degrade the quality of the decomposition, resulting in independent components (ICs) that are less dipolar, meaning they are less likely to represent signals from a single neural generator [18].

Troubleshooting Guides

Guide 1: Optimizing Artifact Subspace Reconstruction (ASR) Parameters

Problem: The ASR algorithm is too aggressive, removing both motion artifacts and neural signals.

Background: ASR identifies and removes high-variance components in the EEG data based on a calibration period and a threshold parameter, often called "k". A lower k-value makes the algorithm more sensitive and aggressive, increasing the risk of overcleaning [18].

Solution Steps:

• Calibration Data: Ensure you are using a high-quality calibration period for ASR. This should be a segment of your data that is as clean as possible, ideally from a period of quiet standing or sitting before the running task [18].
• Adjust the 'k' Parameter: Instead of using the default value, systematically test higher k-values. The literature suggests that a k-value that is too low (e.g., below 10) can overclean the data. A less aggressive threshold (e.g., k=20-30) is often recommended to preserve brain sources [18].
• Validate with Known Responses: After cleaning with a chosen k-value, check for the presence of expected neural signatures. For example, in a Flanker task during running, you should be able to identify a P300 event-related potential component. If this component is missing, your parameters may be too aggressive [18].

Table 1: Effect of ASR Parameter 'k' on Data Quality

k-value	Aggressiveness	Impact on Motion Artifact	Risk to Neural Signal	Recommendation for Running EEG
Low (e.g., 5)	High	Powerful removal	High	Not recommended; high risk of signal loss.
Medium (e.g., 20)	Moderate	Good removal	Moderate	A good starting point for testing.
High (e.g., 30)	Low	Milder removal	Low	Safer for signal integrity; may leave residual artifact.

Guide 2: Implementing a Subject-Specific Deep Learning Approach

Problem: Standardized cleaning pipelines do not generalize well across all subjects, leading to inconsistent results and potential overcleaning for some individuals.

Background: Motion artifacts can vary significantly between subjects due to differences in head shape, electrode fit, and individual gait patterns. A one-size-fits-all cleaning model may be too aggressive for some and insufficient for others. The Motion-Net framework proposes a subject-specific convolutional neural network (CNN) that is trained individually for each participant, improving artifact removal consistency and preserving signal integrity [5].

Solution Steps:

• Data Collection: For each subject, collect EEG data with ground-truth references. This involves recording during motion (e.g., running) and also capturing clean, artifact-free neural signals, which can be challenging but is possible with specialized protocols [5].
• Incorporate Signal Features: Enhance the model by incorporating features like Visibility Graph (VG) that provide structural information about the EEG signal. This has been shown to improve the model's accuracy and stability, especially with smaller datasets [5].
• Train Individual Models: Train a separate Motion-Net model for each subject using their own data. This tailored approach allows the model to learn the specific characteristics of both the neural signal and the motion artifact for that individual [5].

Table 2: Performance Metrics of Subject-Specific Motion-Net Table based on results from a study that developed Motion-Net for motion artifact removal [5].

Performance Metric	Result (Mean ± Std)	Interpretation
Artifact Reduction (η)	86% ± 4.13	High level of artifact removal achieved.
Signal-to-Noise Ratio (SNR) Improvement	20 ± 4.47 dB	Substantial improvement in signal quality.
Mean Absolute Error (MAE)	0.20 ± 0.16	Low error between cleaned signal and ground truth.

Guide 3: Leveraging Hardware-Based Solutions with Dual-Layer EEG

Problem: Software-based cleaning methods are struggling to separate motion artifacts from brain signals, forcing you to choose between noisy data and overcleaned data.

Background: A hardware solution can provide a more direct way to isolate artifacts. The dual-layer EEG system uses two sets of electrodes: one layer records the standard scalp EEG (a mixture of brain signal and artifact), while a second, mechanically coupled but electrically isolated layer records only the motion artifacts. This provides a pure noise reference that can be used to clean the scalp signal without relying on aggressive statistical assumptions [6].

Solution Steps:

• Equipment Setup: Use a dual-layer EEG electrode array. The scalp-interfacing electrodes are connected to one amplifier system, while the inverted "noise" electrodes are connected to a separate system [6].
• Data Recording: During your overground running experiment, record from both the scalp and noise electrodes simultaneously.
• Signal Processing: Use algorithms like iCanClean, which leverage canonical correlation analysis (CCA) to identify and subtract the noise subspaces from the scalp EEG that are highly correlated with the reference noise signals. This method has been shown to effectively reduce motion artifacts while preserving brain signals, even during running [18] [6].

Dual Layer EEG Noise Cancellation

The Scientist's Toolkit: Key Research Reagents & Solutions

Table 3: Essential Tools for Mobile EEG Motion Artifact Handling

Tool / Solution	Function / Description	Application in Running EEG
Artifact Subspace Reconstruction (ASR)	An algorithm that uses a sliding-window PCA to identify and remove high-variance artifact components based on a clean calibration period [18].	Effective for correcting large-amplitude motion artifacts; requires careful tuning of the 'k' parameter to prevent overcleaning.
iCanClean Algorithm	Leverages canonical correlation analysis (CCA) with reference noise signals (from dual-layer electrodes or created as pseudo-references) to detect and correct noise subspaces in the EEG [18].	Shown to be effective during running, improving ICA decomposition and helping recover event-related potentials like the P300.
Dual-Layer EEG Hardware	A specialized electrode setup where a secondary sensor records only motion artifacts, providing a clean noise reference for subtraction from the primary scalp EEG [6].	Provides a hardware-based solution to isolate motion artifact, reducing reliance on purely statistical software cleaning.
Motion-Net	A subject-specific, CNN-based deep learning model trained to map motion-corrupted EEG to clean EEG signals for individual participants [5].	Ideal for subject-specific studies; incorporates visibility graph features to enhance performance on smaller datasets.
Independent Component Analysis (ICA)	A blind source separation method that decomposes multi-channel EEG into maximally independent components, which can be manually or automatically classified and removed [18] [43].	A standard tool for isolating artifact components; its performance is improved when preceded by methods like ASR or iCanClean [18].

Frequently Asked Questions

• What are the most effective methods for removing motion artifacts from EEG data during running? Recent research indicates that iCanClean and Artifact Subspace Reconstruction (ASR) are highly effective preprocessing methods for reducing motion artifacts during running. iCanClean, which uses canonical correlation analysis with pseudo-reference noise signals, has been shown to be somewhat more effective than ASR in recovering brain-like independent components and identifying expected event-related potential components, such as the P300 congruency effect during a Flanker task [3] [2].

• Can I study brain activity during real-world overground walking and running? Yes, advances in mobile EEG technology and signal processing make this possible. Studies now successfully record EEG during overground walking and running. Key to this is the use of robust artifact removal techniques like ASR or iCanClean before analysis to handle the significant motion artifacts produced by whole-body movement [3] [44].

• How does gait adaptation relate to brain activity? Research shows that the rate at which individuals adapt their gait (e.g., on a split-belt treadmill) is linked to distinct patterns of brain activity. Fast adapters show lower alpha power in the posterior parietal and right visual cortices during early adaptation, suggesting enhanced sensory integration and attention. Slow adapters, in contrast, display greater alpha and beta power in the visual cortex during later stages [45].

• Is the brain activity measured during walking solely due to movement artifacts? No. Controlled studies using mobile EEG have identified neural oscillations that are modulated by walking and are not explained by artifacts. For example, a decrease in occipital alpha power occurs during walking compared to standing, even in complete darkness. This change is not correlated with head acceleration, confirming its neural origin [44].

• Does artifact correction improve the performance of EEG decoding algorithms? For multivariate pattern analysis (MVPA or "decoding"), a study found that combining artifact correction and rejection did not significantly improve decoding performance in most cases. However, artifact correction is still strongly recommended to minimize the risk of artifact-related signals artificially inflating decoding accuracy, which could lead to incorrect conclusions [46].

Experimental Protocols for Key Scenarios

Protocol 1: Investigating Cognition During Running

This protocol is designed to study event-related potentials (ERPs) during dynamic movement, such as running.

• Objective: To identify motion artifact removal approaches that enable the detection of stimulus-locked ERPs during overground running [3] [2].
• Task: An adapted Eriksen Flanker task is administered during both static standing and dynamic jogging.
• Key Steps:
  • Data Acquisition: Record EEG from young adults using a wireless mobile EEG system while they perform the Flanker task under both standing and jogging conditions.
  • Artifact Removal Preprocessing: Preprocess the continuous EEG data using different methods for comparison:
    • iCanClean with pseudo-reference noise signals.
    • Artifact Subspace Reconstruction (ASR).
  • Analysis & Evaluation: Compare the methods based on:
    • ICA Dipolarity: The number of brain-like independent components recovered.
    • Spectral Power: Reduction in power at the gait frequency and its harmonics.
    • ERP Validation: The ability to capture the expected P300 ERP "congruency effect" (greater amplitude for incongruent vs. congruent Flanker stimuli), comparing results to the stationary condition [3] [2].

Protocol 2: Isolating and Characterizing Gait-Related Movement Artifact

This methodology provides a way to directly measure the pure movement artifact uncontaminated by brain signals.

• Objective: To isolate, record, and characterize the movement artifact generated in EEG signals during walking [47].
• Task: Treadmill walking at a range of speeds (e.g., 0.4 to 1.6 m/s).
• Key Steps:
  • Block Electrophysiological Signals: Place a silicone swim cap (a non-conductive layer) over the participant's scalp to block all neural and other physiological signals from reaching the EEG electrodes.
  • Simulate a Scalp: Place a wig coated with conductive gel over the swim cap to create a simulated, electrically conductive scalp.
  • Record Artifact: Place the standard EEG cap over the simulated scalp and gel the electrodes. Record data while the participant walks on a treadmill.
  • Validation: Verify the setup by having the participant blink, clench their jaw, and move their head. No physiological signals should be visible, but large movement artifacts should appear.
  • Analysis: Analyze the recorded data to characterize the properties of the movement artifact across different speeds, subjects, and electrode locations [47].

Data Presentation: Comparing Motion Artifact Removal Techniques

The following table summarizes quantitative findings from a 2025 comparative study of artifact removal methods during overground running [3] [2].

Evaluation Metric	iCanClean with Pseudo-Reference	Artifact Subspace Reconstruction (ASR)
ICA Component Dipolarity	Recovery of more dipolar brain independent components; somewhat more effective than ASR [3] [2].	Recovery of more dipolar brain independent components [3] [2].
Power at Gait Frequency	Significant power reduction at the gait frequency and its harmonics [3].	Significant power reduction at the gait frequency and its harmonics [3].
P300 ERP Congruency Effect	Successfully identified the expected greater P300 amplitude to incongruent Flanker stimuli [3] [2].	Produced ERP components similar in latency to the standing task; P300 effect not specifically mentioned [3] [2].

The Scientist's Toolkit: Research Reagent Solutions

Item Name	Function / Explanation
Mobile EEG System (e.g., Smarting, BioSemi ActiveTwo)	High-density, wireless EEG systems that allow for data collection during full-body movement without cable sway artifacts [47] [44].
iCanClean Algorithm	A signal processing tool that uses canonical correlation analysis (CCA) to detect and subtract noise subspaces, ideally using dual-layer noise sensors or creating pseudo-reference signals from the EEG itself [3] [2].
Artifact Subspace Reconstruction (ASR)	A preprocessing algorithm that uses a sliding-window principal components analysis (PCA) to identify and remove high-variance, high-amplitude artifacts from continuous EEG data [3] [2].
Independent Component Analysis (ICA)	A blind source separation technique used to decompose EEG data into maximally independent components, which can then be classified and removed if they represent artifacts (e.g., eye, muscle, heart) [3] [47].
Non-Conductive Silicone Cap	Used in controlled experiments to physically block electrophysiological signals, allowing researchers to isolate and record pure movement artifact for characterization [47].

Experimental Workflow for Motion Artifact Handling

The diagram below outlines a logical workflow for designing and processing an experiment involving EEG and locomotion.

Neural Signatures of Gait Adaptation

This diagram illustrates the differential neural dynamics identified in the cortical areas of fast and slow gait adapters [45].

Benchmarking Performance: Validating and Comparing Cleaning Efficacy

FAQs on Validation Metrics for Motion Artifact Handling

What are the core validation metrics for assessing motion artifact removal in mobile EEG?

Three key validation metrics are essential for evaluating the effectiveness of motion artifact removal during dynamic EEG recordings, such as overground running:

• ICA Component Dipolarity: This metric evaluates the quality of the Independent Component Analysis (ICA) decomposition. Effective artifact cleaning should improve the "dipolarity" of brain-derived independent components. Physiologically plausible brain sources are typically dipolar, and this is quantified by assessing how well each component can be modeled by a single equivalent current dipole within the brain volume. Better artifact removal leads to a higher number of such dipolar components, indicating a successful separation of brain activity from non-brain artifacts [3] [18].
• Power Reduction at Gait Frequency and Harmonics: Motion artifacts from running are often periodic and manifest as distinct peaks in the power spectrum at the step frequency (gait frequency) and its harmonics. A successful processing pipeline should significantly reduce spectral power at these specific frequencies, confirming the attenuation of movement-related noise without overly suppressing the neural signal of interest [3] [18].
• Recovery of Event-Related Potentials (ERPs): This metric tests the pipeline's ability to preserve or recover brain signals time-locked to specific sensory, cognitive, or motor events. For example, in a Flanker task administered during running, a valid method should be able to reveal the expected P300 ERP component and its characteristic "congruency effect" (greater amplitude for incongruent versus congruent stimuli), matching the results obtained during static conditions [3] [18].

How do common artifact cleaning methods compare against these metrics?

The following table summarizes a quantitative comparison of two common preprocessing methods, Artifact Subspace Reconstruction (ASR) and iCanClean, based on a 2025 study of EEG during overground running [3] [18]:

Table 1: Quantitative Comparison of Motion Artifact Removal Methods

Validation Metric	Artifact Subspace Reconstruction (ASR)	iCanClean (with pseudo-reference signals)
ICA Component Dipolarity	Led to the recovery of more dipolar brain independent components [3] [18].	Led to the recovery of more dipolar brain independent components; was somewhat more effective than ASR [3] [18].
Power at Gait Frequency	Power was significantly reduced at the gait frequency after preprocessing [3] [18].	Power was significantly reduced at the gait frequency after preprocessing [3] [18].
ERP Component Recovery	Produced ERP components similar in latency to those identified in a static (standing) task [3] [18].	Produced ERP components similar in latency to those identified in a static task; successfully identified the expected greater P300 amplitude to incongruent flankers [3] [18].

Our ICA decomposition after preprocessing is poor. How can we improve dipolarity?

Low dipolarity after ICA often indicates that significant noise remains, overwhelming the algorithm's ability to separate brain sources effectively. To troubleshoot:

• Aggressive yet Conservative Preprocessing: Consider using a preprocessing method specifically designed for high-amplitude motion artifacts before performing ICA. Both ASR and iCanClean have been shown to improve subsequent ICA decomposition quality. When using ASR, the k parameter is critical; a value that is too low may over-clean the data, while a value that is too high may leave artifacts. For running data, a k parameter of 10 or above has been suggested to avoid "over-cleaning" while still improving decomposition [18]. For iCanClean, an R² threshold of 0.65 with a 4-second sliding window has been effective for promoting dipolarity in locomotion data [18].
• Verify Your Reference Data: ASR's performance is highly dependent on the quality of its calibration "reference" data, which should be as clean as possible (e.g., from a quiet standing period). Limitations in ASR's algorithm for identifying this reference period can sometimes explain failures to address high-amplitude motion artifacts [18].
• Consider Advanced Hardware: If possible, using a dual-layer EEG system, where a second layer of electrodes records only motion-related noise, can provide a perfect noise reference for algorithms like iCanClean, leading to superior artifact removal and ICA decomposition [18] [48].

We've cleaned our data, but strong power at the step frequency remains. What is wrong?

The persistence of power at the gait frequency suggests that the motion artifact has not been fully removed. This is a common challenge. Here are steps to address it:

• Re-evaluate Preprocessing Parameters: The methods you are using might not be aggressive enough for the amplitude of motion present in running. For ASR, try cautiously adjusting the k parameter to a lower value (e.g., 10-20) to remove more high-variance data, but be mindful of the risk of removing neural signals [18]. For iCanClean, you could adjust the R² correlation threshold.
• Inspect the Nature of the Residual Signal: Use time-frequency analysis and coherence measures to check if the residual power is strictly phase-locked to the gait cycle. If it is, it is more likely to be a residual artifact. If the spectral power shows the expected event-related desynchronization/synchronization patterns in sensorimotor areas, it may contain a neural component [48] [49].
• Check for Hardware-Generated Artifacts: Persistent, spike-like artifacts broadband in nature may not be biological. These can be caused by cable sway (triboelectric effect) or unstable electrode-skin contact, which can modulate power line interference. These are notoriously difficult to remove with signal processing alone, and the best solution is prevention through secure cable management and stable electrode impedance [9].

Can we reliably measure Event-Related Potentials (ERPs) during running?

Yes, but it requires robust artifact handling and careful experimental design. The recovery of ERPs like the P300 during running is a key benchmark for validating an artifact-removal pipeline [3] [18].

• Use a Validated Processing Pipeline: As shown in Table 1, both ASR and iCanClean can yield ERP latencies comparable to those in static conditions. iCanClean has demonstrated a particular ability to recover expected cognitive effects, such as the P300 congruency effect during a Flanker task performed while jogging [3] [18].
• Ensure Proper Trigger Synchronization: For gait-event-related potentials (gERPs), you need a system that can record gait events (e.g., heel strike from a pressure-sensitive treadmill or force plates) and send triggers to the EEG amplifier with millisecond precision [50].
• Leverage Grand Averaging: ERP waveforms are typically weak and require averaging across many trials to emerge from the background EEG noise. Ensure your task design includes a sufficient number of trials for a robust average [51] [50].

The Scientist's Toolkit

Table 2: Essential Research Reagents and Materials for Mobile EEG Research

Item	Function & Explanation
High-Density Wireless EEG System	Allows for unrestricted movement during overground running and provides sufficient channels for effective ICA. Active electrodes are often preferred for their superior motion artifact rejection capabilities [48] [52].
iCanClean Algorithm	A signal processing method that uses canonical correlation analysis (CCA) to identify and subtract noise subspaces from the EEG. It can use signals from dedicated noise sensors or create "pseudo-reference" signals from the EEG itself [3] [18].
Artifact Subspace Reconstruction (ASR)	An algorithm that uses a sliding-window PCA to identify and remove high-variance, high-amplitude artifacts from continuous EEG by comparing them to a clean baseline recording [3] [18].
Dual-Layer EEG Hardware	A specialized setup with primary scalp electrodes paired with mechanically coupled but electrically isolated "noise" electrodes. The noise electrodes record only motion artifact, providing an ideal reference for algorithms like iCanClean [18] [48].
Motion Capture & Trigger System	A system (e.g., inertial measurement units - IMUs, instrumented treadmills, force plates) to precisely track gait events (heel strike, toe-off). This is crucial for analyzing gait-related spectral changes and for generating triggers for gait-event-related potentials (gERPs) [50].

Experimental Protocols for Key Studies

This protocol is adapted from the 2025 comparative study that forms the core of this technical guide.

1. Participant Preparation:

• Recruit young adult athletes.
• Fit participants with a wireless mobile EEG system. Ensure electrode impedances are kept as low and stable as possible to minimize artifacts at the source.

2. Experimental Task & Design:

• Task: Employ an adapted Flanker task (e.g., arrows pointing left/right) where participants respond to the central target while ignoring congruent or incongruent flankers.
• Conditions: Each participant performs the task under two main conditions:
  • Dynamic Condition: While jogging/overground running.
  • Static Control Condition: While standing still.
• This within-subjects design allows for a direct comparison of ERP components (like the P300) between movement and non-movement states.

3. Data Acquisition:

• Record continuous EEG data throughout both conditions.
• Synchronize the presentation of the Flanker task stimuli with the EEG recording to create precise event markers for later ERP analysis.

4. Data Preprocessing & Comparison:

• Preprocess the continuous EEG data from the dynamic (running) condition using different methods for comparison. Key methods to test include:
  • Artifact Subspace Reconstruction (ASR): Apply with a recommended k parameter of 10 or higher.
  • iCanClean: Apply using pseudo-reference noise signals derived from the EEG data itself, with parameters such as an R² threshold of 0.65.
• After preprocessing, run an ICA on all datasets.
• Compare the outcomes based on the three key validation metrics.

The workflow for this validation framework is illustrated below:

Experimental Validation Workflow

Methodologies for Isolating Motion Artifact

To truly understand and validate artifact removal, it is useful to know how to study the artifact itself. The following diagram outlines a classic protocol for recording "pure" movement artifact, devoid of brain signals [53]:

Pure Motion Artifact Collection

The following table summarizes the key performance metrics for iCanClean, ASR, and traditional ICA when processing mobile EEG data during overground running.

Method	ICA Decomposition Quality	Gait Frequency Power Reduction	P300 ERP Recovery	Computational Demand
iCanClean	~13.2 "good" dipolar brain components on average; +57% improvement over basic preprocessing [23].	Effective reduction of power at step frequency and harmonics [2] [18].	Successfully identified the expected P300 congruency effect during running [2] [3].	Lower than ICA; suitable for real-time processing [24].
ASR	Recovery of more dipolar brain components compared to uncorrected data [2] [18].	Effective reduction of power at step frequency and harmonics [2] [18].	Produced ERP components similar in latency to the standing task, but P300 effect was not specifically reported [2].	Low; designed for online and offline use with short processing delays [26].
Traditional ICA	Quality is reduced by large motion artifacts, hindering its ability to identify maximally independent sources [2] [18].	Minimizes but does not eliminate gait-related spectral power [2].	Not evaluated as a standalone preprocessor; typically used after other methods [2].	Very high (e.g., can take 5+ hours for high-density data); not suitable for real-time use [24].

---

Detailed Methodologies & Experimental Protocols

Core Experiment: Overground Running with a Flanker Task

This protocol is designed to evaluate artifact removal methods by assessing their impact on ICA quality, spectral power, and event-related potential (ERP) recovery [2] [18].

• Task: Participants perform a modified Eriksen Flanker task while jogging overground and during static standing. The task presents congruent (e.g., >>>>>) and incongruent (e.g., >><>>) arrows, requiring a button press to indicate the central arrow's direction [2].
• EEG Recording: Wireless mobile EEG is recorded during both dynamic (jogging) and static (standing) versions of the task [2] [18].
• Evaluation Metrics:
  • ICA Dipolarity: The number of independent components (ICs) that are well-localized as dipoles (residual variance < 15%) and have a high brain probability (>50% per ICLabel) [2] [23].
  • Spectral Power: Reduction in power at the fundamental gait frequency and its harmonics in the cleaned data [2].
  • ERP Analysis: The ability to recover the P300 component, specifically the expected greater amplitude for incongruent versus congruent flanker stimuli, in the dynamic running condition [2] [3].

iCanClean Protocol

iCanClean uses canonical correlation analysis (CCA) to identify and subtract noise subspaces from the EEG data [2] [24].

• Noise Reference:
  • Ideal: Uses signals from "dual-layer" noise sensors that are mechanically coupled to scalp electrodes but are not in contact with the scalp, thus recording only motion-related noise [2] [23].
  • Alternative (Pseudo-Reference): If dedicated noise sensors are unavailable, a temporary high-pass or notch filter (e.g., below 3 Hz) is applied to the raw EEG to create pseudo-reference noise signals [2] [18].
• Cleaning Process:
  • The algorithm performs CCA between the scalp EEG channels (signal + noise) and the reference noise channels (noise only).
  • It identifies subspaces of the EEG that are highly correlated with the noise reference.
  • Components exceeding a user-defined correlation threshold (R²) are projected back to channel space and subtracted from the original signal [2] [24].
• Key Parameters:
  • R² Threshold: Controls cleaning aggressiveness. A threshold of 0.65 is recommended for mobile EEG during walking/running [2] [23].
  • Window Length: The length of the sliding window for CCA. A 4-second window is recommended for optimal performance [23].

Artifact Subspace Reconstruction (ASR) Protocol

ASR uses a sliding-window principal component analysis (PCA) to identify and remove high-variance artifacts by comparing data to a clean calibration period [2] [27].

• Calibration Phase:
  • A segment of clean, artifact-free EEG data (e.g., 1-2 minutes of resting data or clean segments extracted from the recording) is used as a reference [2] [27].
  • The calibration data is used to compute a robust covariance matrix and a mixing matrix M via PCA. Statistical thresholds for "normal" data are established [27] [26].
• Processing Phase:
  • Incoming EEG data is processed in short, sliding windows (e.g., 500 ms).
  • PCA is performed on each window, and the resulting components are compared to the calibration statistics.
  • Components whose variance exceeds a threshold defined by the tuning parameter k are identified as artifacts.
  • The artifactual components are removed, and the data is reconstructed using the remaining components and the calibration mixing matrix [27] [26].
• Key Parameter:
  • k threshold: A lower k value leads to more aggressive cleaning. For human locomotion, a k value as low as 10 can be used, but values between 20-30 are common to avoid "overcleaning" [2].

Traditional ICA Protocol

ICA is a blind source separation technique that decomposes multi-channel EEG into maximally independent components [2] [24].

• Prerequisites:
  • Requires a large amount of high-density data (e.g., >30 minutes of 100+ channel EEG) for a stable decomposition [24].
  • Is computationally intensive and slow, often taking hours to run [24].
• Process:
  • The EEG data matrix is decomposed into a set of independent components (ICs), each with a time course and a scalp topography.
  • Components are classified as brain or non-brain (artifact) using tools like ICLabel.
  • Artifactual components are manually or automatically removed.
  • The "cleaned" data is reconstructed back to channel space from the remaining components [2] [24].
• Limitation in MoBI: Large motion artifacts can contaminate the ICA decomposition itself, reducing its ability to isolate true brain sources effectively. Therefore, it is often used after preprocessing with methods like iCanClean or ASR [2] [18].

---

The Scientist's Toolkit: Research Reagent Solutions

Item	Function in Experiment
Wireless Mobile EEG System	Enables EEG data collection during unrestricted whole-body movement like overground running [2].
Dual-Layer EEG Cap	The ideal setup for iCanClean. It features scalp electrodes paired with mechanically coupled but electrically isolated noise electrodes that record only motion artifacts [24] [23].
Artifact Subspace Reconstruction (ASR)	A preprocessing algorithm for removing high-amplitude, high-variance artifacts. Available as a plugin for EEGLAB [2] [27].
iCanClean Algorithm	A preprocessing algorithm that uses canonical correlation analysis (CCA) with reference noise signals to remove motion, muscle, and other artifacts [24] [23].
ICLabel	An EEGLAB plugin that uses a trained convolutional neural network to automatically classify independent components (ICs) as brain, muscle, eye, heart, line noise, or channel noise [2] [23].
Inertial Measurement Unit (IMU)	A sensor that measures motion (acceleration, rotation). Can be used as a reference signal for adaptive filtering or newer deep-learning artifact removal methods [54].

---

Frequently Asked Questions & Troubleshooting

General Performance

Q1: Which method is the most effective for recovering cognitive ERPs like the P300 during running?

• A: Based on direct comparisons during overground running, iCanClean has been shown to successfully recover the expected P300 congruency effect (greater amplitude for incongruent stimuli), while ASR produced ERPs of similar latency but did not specifically report this effect [2] [3]. iCanClean also consistently led to a higher number of high-quality, dipolar brain components from ICA [2] [23].

Q2: My lab doesn't have a dual-layer EEG system. Can I still use iCanClean?

• A: Yes. iCanClean can be implemented using pseudo-reference noise signals created from your existing raw EEG data. This is done by applying a temporary high-pass or notch filter (e.g., <3 Hz) to the data to isolate low-frequency noise, which is then used as the noise reference for the algorithm [2] [18]. While dual-layer sensors are ideal, studies confirm pseudo-references are effective.

Troubleshooting iCanClean

Q3: What are the optimal parameters for iCanClean with running data?

• A: For running and walking data, a parameter sweep found the optimal settings to be an R² threshold of 0.65 and a sliding window length of 4 seconds [23]. An R² of 0.65 provides a good balance between aggressive artifact removal and preservation of brain signal.

Q4: The cleaning seems too aggressive and might be removing brain signal. How can I validate this?

• A: You can use the following validation checks:
  • Spectral Analysis: Compare the power spectral density before and after cleaning. Effective cleaning should show a pronounced reduction in power at the step frequency and its harmonics, without creating a broadband loss of neural oscillations (e.g., in the alpha or beta band) [2].
  • ERP Topography: Check if the topography of recovered ERPs (e.g., P300) is physiologically plausible and resembles that obtained in a stationary condition [2].

Troubleshooting ASR

Q5: ASR isn't removing enough motion artifact from my running data. What can I do?

• A: This is a common issue. You can try:
  • Lower the k parameter: The k parameter defines the threshold for artifact detection. A lower value (e.g., 10-20) makes the algorithm more aggressive. Be cautious, as very low values can overclean the data [2].
  • Improve Calibration Data: The performance of ASR is highly dependent on the quality of the calibration data. Ensure your calibration segment is truly clean. Newer variants like ASRDBSCAN and ASRGEV have been developed to better identify clean calibration periods from highly contaminated data [28].

Q6: I'm concerned that ASR might be "overcleaning" my data. What are the signs?

• A: Overcleaning can occur if the k parameter is set too low. Warning signs include:
  • An overall "flatness" or loss of variance in the continuous data.
  • Attenuation of known, high-amplitude brain signals, such as the posterior alpha rhythm during eye closure.
  • A drastic reduction in the number of brain-like components found by ICA [2] [3].

Troubleshooting Traditional ICA

Q7: Why does ICA perform poorly on my mobile running data without preprocessing?

• A: Large motion artifacts create massive, non-stationary variances that dominate the signal. This violates ICA's underlying assumptions and degrades the quality of the decomposition, making it unable to cleanly separate brain sources from artifactual ones [2] [18]. The presence of motion artifacts reduces the "dipolarity" of the resulting independent components.

Q8: What is the recommended pipeline if I want to use ICA for source analysis on running data?

• A: The most effective current pipeline is to use a hybrid approach:
  • Preprocess with iCanClean or ASR to remove the bulk of the motion artifacts.
  • Run ICA on the preprocessed data to separate the remaining sources.
  • Use ICLabel to automatically classify the resulting components.
  • Remove non-brain components and reconstruct the data for further analysis [2] [23]. This sequence significantly improves the number and quality of recoverable brain components.

Frequently Asked Questions

Q1: What is the primary advantage of using a phantom head for EEG method validation? Phantom heads provide a known ground-truth signal, allowing researchers to rigorously test and validate EEG processing techniques, such as source separation and connectivity measures, in the presence of real-world volume conduction and motion artifacts. This is crucial because using human subjects makes it impossible to know the exact underlying neural signals, and computer simulations may avoid real-world non-linearities that can violate the assumptions of the measures being validated [55].

Q2: How effective is Independent Component Analysis (ICA) at recovering signals in motion-heavy scenarios? Studies using phantom heads with embedded antennae have shown that ICA can effectively recover most source signals even during motion. Evidence includes cross-correlations primarily above 0.8 between recovered and original signals. ICA also maintains a consistent signal-to-noise ratio (SNR) near 10 dB across various walking speeds, whereas raw scalp channel data can see SNR decrease to ~2 dB at fast walking speeds [55].

Q3: Can connectivity measures accurately identify true neural connections when motion is present? Yes, but efficacy varies by measure. Research using interconnected signals generated via neural mass models in a phantom head has demonstrated that many connectivity measures can identify true interconnections. However, some measures are susceptible to spurious high-frequency connections, which can induce large standard deviations of around 10 Hz in the estimated connectivity peaks [55].

Q4: What are the best-performing motion artifact removal techniques for running EEG? Recent comparative studies during overground running have found that iCanClean (using pseudo-reference noise signals) and Artifact Subspace Reconstruction (ASR) are highly effective. These methods lead to the recovery of more dipolar brain independent components, significantly reduce power at the gait frequency and its harmonics, and enable the identification of expected event-related potential (ERP) components like the P300 [3] [2]. iCanClean has been noted as somewhat more effective than ASR in some analyses [3].

Q5: Are there cost-effective alternatives to commercial phantom heads? Yes, 3D-printed conductive phantoms have emerged as a highly accessible alternative. One validated design using conductive PLA filament achieved an 85% cost reduction (£48.10 vs. £300–£500 for commercial units) and a fabrication time of 48 hours, while providing consistent electrical properties suitable for standardized EEG electrode testing [56].

Troubleshooting Guides

Issue 1: Poor Independent Component Analysis (ICA) Decomposition During Motion

Problem: ICA fails to separate brain activity from motion artifacts, resulting in low-dipolarity components and implausible source locations.

Solutions:

• Pre-clean data with advanced algorithms: Apply preprocessing techniques like iCanClean or Artifact Subspace Reconstruction (ASR) before running ICA. This reduces the high-amplitude motion artifacts that impair ICA's ability to find maximally independent sources [2].
• Use a calibrated phantom for validation: Employ a phantom head with embedded antennae playing known signals. This allows you to quantify ICA performance by calculating the cross-correlation between recovered independent components and the original ground-truth signals. Cross-correlations above 0.8 indicate good recovery [55].
• Leverage dual-layer electrode systems: If available, use systems with dedicated noise sensors. iCanClean can leverage these mechanically coupled noise references to more effectively identify and subtract motion artifact subspaces from the scalp EEG [2].

Issue 2: Inaccurate EEG Connectivity Estimates in Mobile Settings

Problem: Connectivity analysis yields spurious, non-physiological connections, particularly in high frequencies, due to motion artifacts and volume conduction.

Solutions:

• Validate measures with a ground-truth phantom: Use a phantom head setup where transiently interconnected signals are generated by neural mass models. This lets you test which connectivity measures (e.g., ffDTF, dDTF, gPDC, WPLI) most accurately identify the true, implanted interconnections without being fooled by volume conduction [55].
• Choose robust connectivity measures: Prefer measures less sensitive to volume conduction. Weighted Phase Lag Index (WPLI), for instance, is based on imaginary coherence and ignores spurious, instantaneous connections, increasing sensitivity to true connections [55].
• Perform connectivity at the source level: After ICA and source localization, compute connectivity on the source-projected time series. This can reduce the influence of volume conduction compared to channel-level analysis [55].

Issue 3: Persistent Gait-Frequency Artifact in Processed EEG

Problem: After standard preprocessing, a strong spectral peak at the step frequency and its harmonics remains, contaminating the data.

Solutions:

• Aggressive but careful preprocessing: Implement iCanClean with an R² threshold of 0.65 and a 4-second sliding window, which has been shown to be effective during locomotion. Alternatively, use ASR with an appropriate k parameter (e.g., 10-30), where a lower value is more aggressive but risks over-cleaning [2].
• Quantify the artifact: Use the phantom head or a "null" recording to establish the baseline characteristics of the motion artifact. Then, in human data, compute the power spectral density and check for residual peaks at the gait frequency. A successful processing pipeline should significantly reduce power at these frequencies compared to the raw data [3] [2].
• Explore deep learning methods: For subject-specific applications, consider novel deep learning models like Motion-Net, a CNN-based framework designed for motion artifact removal that has shown high artifact reduction percentages (86% ± 4.13) and significant SNR improvement (20 ± 4.47 dB) in studies [5].

Performance Data & Experimental Protocols

Table 1: Comparison of Motion Artifact Removal Techniques for Mobile EEG

Table based on studies involving overground running and phantom head validation. [3] [2] [5]

Technique	Core Principle	Key Performance Metrics	Advantages	Limitations
iCanClean	Uses Canonical Correlation Analysis (CCA) to subtract noise subspaces identified via pseudo-reference signals or dual-layer electrodes.	- Recovers more dipolar ICs [2]- Identifies P300 ERP effect during running [2]- Effective on broadband running artifacts [2]	Effective without dedicated hardware when using pseudo-reference; ideal for running.	Performance can depend on parameter selection (R², window length).
Artifact Subspace Reconstruction (ASR)	Identifies and removes high-variance components in real-time using a sliding-window PCA and calibration data.	- Improves ICA dipolarity [2]- Reduces power at gait frequency [2]	Fast, automated cleaning; works well with high-density EEG.	Risk of "over-cleaning" neural data with aggressive thresholds (low k).
Motion-Net	A subject-specific 1D CNN model that maps artifact-contaminated EEG to clean EEG.	- Artifact Reduction (η): 86% ± 4.13 [5]- SNR Improvement: 20 ± 4.47 dB [5]	High performance; subject-specific modeling handles artifact variability.	Requires a separate model to be trained for each subject.
Independent Component Analysis (ICA)	Blind source separation to isolate and remove artifactual components.	- Cross-correlation with ground truth: >0.8 [55]- Maintains SNR ~10 dB during motion [55]	Standard, widely available; effective if artifacts are separable.	Decomposition quality degrades with excessive motion; requires manual component rejection.

Table 2: Characteristics of Phantom Head Technologies for EEG Validation

Data synthesized from multiple phantom studies and material analyses. [55] [56] [57]

Phantom Type	Typical Cost	Fabrication Time	Key Characteristics	Best Use Case
3D-Printed Conductive	~£50 [56]	~48 hours [56]	- Resistivity: 821–1502 Ω (DC) [56]- Impedance @100Hz: 3.01–6.4 kΩ [56]	Low-cost prototyping, educational labs, standardized electrode testing.
Textile-Based	Information Missing	Information Missing	- 91.67% lighter than gelatin [58]- Long shelf-life (years) [58]- SNR better than gelatin [58]	Long-term, repeated experiments; electrode-skin interface studies.
Commercial Injection-Molded	£300–£500 (unit) [56]	3-7 days (after 4-6 wk tooling) [56]	- High consistency & durability [56]- Resistivity: 10–20 Ωcm [56]- Internal drive electrodes [56]	Regulatory testing, quality control, high-precision R&D.
Agarose/Gypsum/Saline (Multi-Compartment)	Low (material cost)	Hours to Days [57]	- Scalp (Agarose): ~0.31 S/m [57]- Skull (Gypsum): ~0.0017 S/m [57]- Brain (Saline): ~0.33 S/m [57]	Realistic volume conduction studies, source localization validation.

Essential Experimental Protocol: Phantom Head Validation for Motion Artifact Removal

This protocol outlines how to use a phantom head to validate the efficacy of motion artifact removal pipelines, as described in recent literature [55] [2].

1. Phantom and Signal Setup:

• Obtain or Fabricate a Phantom: Use a head-shaped phantom with realistic electrical conductivity. This can be a 3D-printed conductive model [56], a multi-compartment agarose/gypsum model [57], or a textile-based phantom [58].
• Embed Signal Sources: Internally, embed antennae or electrodes that can receive signals from an external input/output interface.
• Generate Ground-Truth Signals: Use a neural mass model to generate complex, physiologically relevant signals with known interconnectivity and peak frequencies (e.g., theta: 6.5 Hz, alpha: 10 Hz, gamma: 41 Hz). Feed these signals to the internal antennae [55].

2. Motion Induction and Data Collection:

• Mount the Phantom: Place the phantom on a motion platform that can mimic human head kinematics during various activities (e.g., walking at different speeds, running) [55].
• Record Data: Use your mobile EEG system to record data from the phantom's scalp while it is both stationary and in motion. Simultaneously, if available, record data from the internal antennae as the ground-truth reference.

3. Processing and Validation:

• Apply Artifact Removal: Process the motion-corrupted scalp EEG data using the techniques you wish to validate (e.g., iCanClean, ASR, Motion-Net).
• Quantify Performance:
  • Signal Fidelity: Calculate the cross-correlation and signal-to-noise ratio (SNR) between the processed EEG signals and the ground-truth signals from the internal antennae [55].
  • Component Quality: If using ICA, evaluate the dipolarity of the resulting independent components and their correlation with the true sources [55] [2].
  • Artifact Removal: Compute power spectral density to check for the reduction of power at the motion frequency (e.g., step frequency) and its harmonics [2].

The Scientist's Toolkit: Key Research Reagents & Materials

Table 3: Essential Materials for Ground-Truth EEG Validation Studies

Compilation of key tools and their functions from the cited literature.

Item	Function in Validation	Key Details / Examples
Phantom Head	Serves as a reproducible, known-signal replacement for a human subject for method validation.	3D-printed conductive [56], textile-based [58], multi-compartment agarose/gypsum [57].
Neural Mass Model	Generates complex, synthetic EEG-like signals with controllable interconnectivity for ground-truth testing.	Used to create signals with peaks in theta, alpha, gamma bands for phantom antennae [55].
Motion Platform	Induces realistic, reproducible head motion to simulate walking or running in a lab setting.	Used to mimic recorded human head motion at various walking speeds [55].
Dual-Layer Electrodes	Provide dedicated noise references; the upper layer captures only motion, while the lower layer captures EEG + motion.	iCanClean algorithm uses signals from these to subtract motion artifacts [2].
iCanClean Software	Algorithm for motion artifact removal that uses canonical correlation analysis with noise references.	Can use dual-layer electrodes or create pseudo-reference signals from raw EEG [3] [2].

Workflow and Signaling Diagrams

Experimental Validation Workflow

iCanClean Processing Logic

Troubleshooting Guides & FAQs

Frequently Asked Questions

Q1: Why is the expected P300 congruency effect (greater amplitude for incongruent trials) absent in my data during overground running?

The most common cause is significant motion artifact obscuring the neural signal. During overground running, head motion and muscle activity produce high-amplitude noise that can overwhelm the smaller voltage fluctuations of the P300 ERP. This artifact reduces the signal-to-noise ratio, making it impossible to detect the subtle differences between congruent and incongruent trials [2] [5]. Furthermore, motion can degrade the quality of the Independent Component Analysis (ICA) decomposition, which is a critical step for isolating brain-based signals [2].

Q2: My ERP waveforms look noisy after standard preprocessing. Which motion artifact removal method is most effective for running?

Recent comparative studies indicate that iCanClean is somewhat more effective than Artifact Subspace Reconstruction (ASR) for data collected during running [2] [3]. Specifically, preprocessing with iCanClean using pseudo-reference noise signals has been shown to successfully recover the P300 congruency effect, whereas other methods may not [2]. This is because iCanClean is particularly adept at identifying and subtracting noise subspaces highly correlated with motion [2].

Q3: How can I assess the quality of my data after applying a motion correction technique?

Three key metrics are recommended [2]:

• ICA Component Dipolarity: Evaluate the quality of your ICA decomposition. Effective artifact removal should yield a higher number of dipolar brain components.
• Spectral Power at Gait Frequency: Examine the power spectrum. Successful cleaning should show a significant reduction in power at the fundamental frequency of your gait (step rate) and its harmonics.
• ERP Component Latency: Check if the latency of your recovered ERP components (like the P300) is similar to that obtained in a static (seated or standing) control condition.

Q4: Are there deep learning solutions for motion artifact removal, and are they suitable for this paradigm?

Yes, deep learning models like Motion-Net (a CNN-based model) and AnEEG (an LSTM-based GAN) have shown high performance in removing motion artifacts from EEG signals [5] [59]. However, their application in the specific context of recovering ERPs during overground running is still an emerging area of research. These models often require substantial, well-defined training data and may be more complex to implement for real-time or online processing compared to established methods like iCanClean [5] [59].

Troubleshooting Guide: Common Problems and Solutions

Problem	Potential Cause	Recommended Solution
No P300 congruency effect	Excessive motion artifact from running	Implement iCanClean with pseudo-reference signals prior to ICA [2].
Poor ICA decomposition	High-amplitude motion artifacts corrupting source separation	Use Artifact Subspace Reconstruction (ASR) with a k parameter of 10-30 during preprocessing to improve ICA dipolarity [2].
ERP latency appears delayed	Residual low-frequency drift or motion artifact	Apply additional high-pass filtering (e.g., 0.5 Hz cut-off) and validate latency against a static recording [2].
Muscle artifacts persist in frontal/temporal channels	Neck and jaw muscle tension during running	After ICA, prioritize the removal of components classified with high muscle probability by ICLabel.

Experimental Protocols & Methodologies

Protocol 1: Dynamic Flanker Task During Overground Running

This protocol is adapted from a study that successfully recovered the P300 component during jogging [2].

• Task Design: An adapted Eriksen flanker task is used. Participants respond to a central target arrow (e.g., > or <) that is flanked by either congruent (>>>>>> or <<<<<<) or incongruent (<<><<< or >><>>>) arrows.
• Equipment:
  • Mobile, wireless EEG system.
  • A safe, clear path for overground running or jogging.
  • A visual stimulus display system that can be used while running (e.g., a screen positioned at the end of the path).
• Procedure:
  • Participants complete a baseline session of the flanker task while standing still to establish a ground-truth ERP.
  • During the dynamic condition, participants jog at a self-selected pace on a treadmill or overground.
  • Flanker stimuli are presented in a random sequence. Participants are instructed to respond as quickly and accurately as possible using a handheld button response system.
  • EEG is recorded continuously throughout the running session.

Protocol 2: Motion Artifact Removal with iCanClean

This methodology details the preprocessing steps found to be effective for the dynamic flanker task [2].

• Software: EEGLAB with the iCanClean plugin.
• Steps:
  • Import Data: Load the continuous EEG data recorded during running.
  • iCanClean Preprocessing: Apply the iCanClean algorithm. Since dedicated noise sensors are often unavailable, use the pseudo-reference approach. This involves:
    • Temporarily applying a notch filter to the raw EEG to identify noise-dominated frequency bands (e.g., components below 3 Hz).
    • Using Canonical Correlation Analysis (CCA) to identify subspaces of the scalp EEG that are correlated with these noise subspaces.
    • Setting the correlation criterion (R²) to 0.65 and using a sliding window of 4 seconds, which has been shown to produce optimal results for locomotion data [2].
    • Subtracting the identified noise components from the scalp EEG signal.
  • Standard Preprocessing: Continue with standard pipeline steps: band-pass filtering (e.g., 0.1-30 Hz), epoching around stimulus events (-200 ms to 800 ms), baseline correction, and automated artifact rejection.
  • ICA: Run ICA on the iCanClean-processed data to separate neural sources from any residual artifacts.

Data Presentation: Method Comparison

The following table summarizes quantitative and qualitative findings from studies that compared motion artifact removal approaches.

Table 1: Comparison of Motion Artifact Removal Methods for Mobile EEG

Method	Underlying Principle	Key Parameters	Performance in Running Studies
iCanClean [2]	Uses canonical correlation analysis (CCA) to identify and subtract noise subspaces from EEG signals.	- R² = 0.65 (correlation threshold)- Sliding window = 4 sec- Uses pseudo-reference signals	Most effective in recovering the P300 congruency effect during running; produced the most dipolar ICA components [2].
Artifact Subspace Reconstruction (ASR) [2]	Uses principal component analysis (PCA) to identify and remove high-variance artifact components based on a clean calibration period.	- k = 10-30 (standard deviation threshold)- Lower k is more aggressive	Improved ICA decomposition and reduced gait-frequency power; less effective than iCanClean for P300 recovery in running [2].
Motion-Net [5]	A subject-specific, 1D Convolutional Neural Network (CNN) that learns to map artifact-corrupted EEG to clean signals.	- Trained per-subject- Incorporates Visibility Graph features	Achieved ~86% artifact reduction and ~20 dB SNR improvement in tests with real motion artifacts; validation in running ERP studies is ongoing [5].
AnEEG [59]	A Generative Adversarial Network (GAN) with Long Short-Term Memory (LSTM) layers to generate artifact-free EEG signals.	- Adversarial training- Captures temporal dependencies	Demonstrated superior performance on metrics like NMSE and RMSE compared to wavelet-based methods; generalizability to running data requires further testing [59].

Visualized Workflows & Signaling Pathways

Experimental Workflow for Dynamic EEG

The following diagram illustrates the end-to-end workflow for acquiring and processing EEG data during a dynamic flanker task to recover the P300 component.

Motion Artifact Removal Pipeline

This diagram details the logical sequence of the signal processing pipeline, with a focus on the critical motion artifact removal step.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Materials and Tools for Dynamic Flanker Task Research

Item	Function & Importance in Research
Wireless Mobile EEG System	Enables the recording of brain activity without restricting movement, which is fundamental for overground running studies. It eliminates cable sway artifacts [2] [5].
EEG Electrode Cap (10-20 System)	A headcap with electrodes placed according to the international 10-20 system (e.g., Fz, Cz, Pz) ensures standardized coverage of brain regions and replicable results across studies [60] [61].
iCanClean Software	A critical software tool for preprocessing. It effectively reduces motion and muscle artifacts, which is a prerequisite for obtaining clean ERPs like the P300 from dynamic movement data [2].
Artifact Subspace Reconstruction (ASR)	An alternative algorithm for cleaning continuous EEG data. It is effective as a preprocessing step to improve subsequent ICA, especially when using a k parameter between 10 and 30 [2].
Independent Component Analysis (ICA)	A blind source separation algorithm used to decompose EEG data into statistically independent components. This allows researchers to manually or automatically identify and remove components representing blinks, muscle, and heart artifacts [2] [62].

Conclusion

Effectively handling motion artifacts is no longer a barrier but a critical step in unlocking the potential of mobile EEG for studying brain dynamics during overground running. The synthesis of current evidence points to a multi-faceted approach: combining robust hardware like dual-layer electrodes with advanced algorithms such as iCanClean and carefully tuned ASR provides the most effective cleaning strategy. Validation through metrics like ICA component dipolarity and the successful recovery of expected ERP components confirms the reliability of these methods. Looking forward, the integration of subject-specific deep learning models holds great promise for further automation and accuracy. For biomedical and clinical research, these advancements pave the way for ecologically valid studies of neural control in locomotion, with direct implications for understanding neurological disorders, developing neuro-rehabilitation therapies, and creating more effective diagnostic tools.

References

• 1. Motion Artifact Removal Techniques for Wearable EEG and ... [https://www.frontiersin.org/journals/electronics/articles/10.3389/felec.2021.685513/full]
• 2. A Comparison of Approaches for Motion Artifact Removal ... [https://www.mdpi.com/1424-8220/25/15/4810]
• 3. A Comparison of Approaches for Motion Artifact Removal ... [https://pubmed.ncbi.nlm.nih.gov/40807974/]
• 4. MLMRS-Net: Electroencephalography ( EEG ) motion removal... artifacts [https://link.springer.com/article/10.1007/s00521-022-08111-6]
• 5. Motion Artifact Removal from EEG Signals Using the ... [https://www.sciencedirect.com/science/article/pii/S1746809425013886]
• 6. Human electrocortical dynamics while stepping over ... [https://www.nature.com/articles/s41598-019-41131-2]
• 7. 5 Basics of EEG 101: Data Collection, Processing & Analysis [https://imotions.com/blog/learning/best-practice/5-basics-eeg-data-processing/?srsltid=AfmBOoqVKrMIhW0r-HqN-z7A_s-_g9BPeeX5ZcM9iIOzQ7YsR7dmgBcP]
• 8. Quality Control Procedures [https://docs.hbcdstudy.org/latest/instruments/eeg/qc/]
• 9. Motion Artifacts in Dynamic EEG Recordings [https://www.mdpi.com/1424-8220/24/19/6363]
• 10. EEG Artifacts: Types, Detection, and Removal Techniques ... [https://www.bitbrain.com/blog/eeg-artifacts]
• 11. Effects of Cable Sway, Electrode Surface Area, and Electrode ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5948545/]
• 12. Time Evolution of the Skin–Electrode Interface Impedance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8348734/]
• 13. Evolution in impedance at the electrode-skin interface of ... [https://www.sciencedirect.com/science/article/abs/pii/S1050641102000974]
• 14. Removal of Artifacts from EEG Signals: A Review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6427454/]
• 15. Getting to know EEG artifacts and how to handle them in ... [https://pressrelease.brainproducts.com/eeg-artifacts-handling-in-analyzer/]
• 16. EEG-Based Spectral Analysis Showing Brainwave ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8962200/]
• 17. Effects of Matched and Mismatched Visual Flow and Gait ... [https://www.mdpi.com/2076-3425/15/5/531]
• 18. A Comparison of Approaches for Motion Artifact Removal from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12349273/]
• 19. Evaluating the effectiveness of artifact correction and rejection ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10516012/]
• 20. d. Indep. Comp. Analysis - EEGLAB Wiki [https://eeglab.org/tutorials/06_RejectArtifacts/RunICA.html]
• 21. Cleaning artifacts using ICA [https://www.fieldtriptoolbox.org/tutorial/preproc/ica_artifact_cleaning/]
• 22. Automated pipeline for EEG artifact reduction (APPEAR ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10696919/]
• 23. iCanClean Improves Independent Component Analysis of ... [https://www.mdpi.com/1424-8220/23/2/928]
• 24. iCanClean Removes Motion, Muscle, Eye, and Line-Noise ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10574843/]
• 25. Characterizing and Removing Artifacts Using Dual-Layer ... [https://www.mdpi.com/1424-8220/22/15/5867]
• 26. A Riemannian Modification of Artifact Subspace ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2019.00141/full]
• 27. Motion artifact removal using Artifact Subspace ... [https://mbraintrain.com/motion-artifact-removal-using-artifact-subspace-reconstruction/]
• 28. Juggler's ASR: Unpacking the principles of artifact ... [https://www.sciencedirect.com/science/article/pii/S0165027025001062]
• 29. Juggler's ASR: Unpacking the principles of artifact ... [https://pubmed.ncbi.nlm.nih.gov/40324599/]
• 30. Performance-based approach for movement artifact removal ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0197153]
• 31. FAQ | EEG Challenge (2025) [https://eeg2025.github.io/faq/]
• 32. Full body mobile brain-body imaging data during ... [https://www.nature.com/articles/sdata2018133]
• 33. Frontiers | The PREP pipeline : standardized preprocessing for... [https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2015.00016/full]
• 34. EEG artifact removal with Blind Source Separation [https://sccn.ucsd.edu/~jung/Site/EEG_artifact_removal.html]
• 35. Overground Walking Decreases Alpha Activity and ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2020.561755/full]
• 36. The Influence Assessment of Artifact Subspace ... [https://www.mdpi.com/2076-3417/13/3/1605]
• 37. iCanClean Removes Motion, Muscle, Eye, and Line-Noise ... [https://www.mdpi.com/1424-8220/23/19/8214]
• 38. iCanClean Removes Motion, Muscle, Eye, and Line-Noise ... [https://www.semanticscholar.org/paper/a2a60293712445cc0192297d56655d93f26c1f1c]
• 39. Motion Artifacts in Dynamic EEG Recordings [https://pmc.ncbi.nlm.nih.gov/articles/PMC11479364/]
• 40. Textile Electrodes for EEG Recording — A Pilot Study [https://www.mdpi.com/1424-8220/12/12/16907]
• 41. Gait-Induced Myoelectric EEG Artifact Removal Validation ... [https://www.mdpi.com/2076-3417/15/22/12103]
• 42. Passive Brain–Computer Interface Using Textile-Based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12526629/]
• 43. Methods for motion artifact reduction in online brain ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2023.1251690/full]
• 44. Overground Walking Decreases Alpha Activity and Entrains ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7782973/]
• 45. Exploring Electrocortical Signatures of Gait Adaptation [https://www.eneuro.org/content/11/7/ENEURO.0515-23.2024]
• 46. and LDA-based decoding of EEG signals [https://www.sciencedirect.com/science/article/pii/S1053811925003076]
• 47. Isolating gait-related movement artifacts in ... - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC4946867/]
• 48. Faster gait speeds reduce alpha and beta EEG spectral power ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7134343/]
• 49. What we have learned about the neural control of human ... [https://www.sciencedirect.com/science/article/pii/S0149763424001878]
• 50. Measuring Gait-Event-Related Brain Potentials (gERPs) ... [https://www.mdpi.com/2076-3417/10/16/5405]
• 51. analysis - Event-Related Potentials ( EEG ) ERPs [https://mne.tools/stable/auto_tutorials/evoked/30_eeg_erp.html]
• 52. (PDF) Negligible Motion in Scalp Electroencephalography... Artifacts [https://www.academia.edu/56691045/Negligible_Motion_Artifacts_in_Scalp_Electroencephalography_EEG_During_Treadmill_Walking]
• 53. Independent Component Analysis of Gait-Related ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2015.00639/full]
• 54. IMU-Enhanced EEG Motion Artifact Removal with Fine ... [https://arxiv.org/html/2509.01073v1]
• 55. Combined head phantom and neural mass model ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6448772/]
• 56. A Cost-Effective 3D-Printed Conductive Phantom for EEG ... [https://www.mdpi.com/1424-8220/25/16/4974]
• 57. Head phantoms for bioelectromagnetic applications [https://biomedical-engineering-online.biomedcentral.com/articles/10.1186/s12938-020-00830-y]
• 58. Novel textile-based anatomically realistic head phantom [https://www.ugent.be/ea/match/textiles/en/research/functional-textiles-and-applications/head-phantom.htm]
• 59. AnEEG: leveraging deep learning for effective artifact ... [https://www.nature.com/articles/s41598-024-75091-z]
• 60. 10–20 system (EEG) [https://en.wikipedia.org/wiki/10%E2%80%9320_system_(EEG)]
• 61. What Is the 10-20 System for EEG? - TMSi [https://www.tmsi.artinis.com/blog/the-10-20-system-for-eeg]
• 62. Unsupervised EEG Artifact Detection and Correction - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8521924/]