Validating EEG Artifact Removal: A Comprehensive Guide to Simulated Data and Ground-Truth Methodologies

Abstract

This article provides a comprehensive framework for researchers and drug development professionals on validating electroencephalography (EEG) artifact removal techniques using simulated data. It covers the fundamental necessity of ground-truth data for reliable validation, explores established and emerging methodologies for generating and utilizing simulated EEG, addresses common troubleshooting and optimization challenges, and establishes rigorous protocols for the comparative evaluation of artifact removal performance. By synthesizing current best practices and validation strategies, this guide aims to enhance the rigor and reproducibility of EEG preprocessing in both clinical and research settings, ultimately leading to more robust neural biomarkers and neurophysiological insights.

The Critical Role of Ground-Truth Data in EEG Artifact Removal

Why Simulated Data is Indispensable for Validation

In electroencephalography (EEG) research, the ability to accurately remove artifacts—unwanted noise from muscle movement, eye blinks, or cardiac activity—is fundamental to interpreting brain signals. However, validating the algorithms designed to remove these artifacts presents a unique challenge: in real-world recordings, the underlying true brain signal is never known. Simulated data provides an indispensable solution to this problem by creating scenarios where the "ground truth" is explicitly defined, enabling rigorous and quantitative validation of artifact removal techniques.

The Fundamental Challenge of Unknown Truth

Electroencephalography is a non-invasive technique used to detect and record human brain electrical activity by placing electrodes on the scalp. Due to its high temporal resolution, portability, and non-invasiveness, EEG is widely utilized across various fields from monitoring sleep quality and recognizing emotions to detecting Alzheimer's disease and epilepsy [1].

However, EEG signals are notoriously susceptible to contamination from multiple sources:

• Physiological artifacts: Including Electromyography (EMG) from muscle activity, Electrooculography (EOG) from eye movements, and Electrocardiography (ECG) from heart activity [1]
• Motion artifacts: Particularly problematic in mobile EEG (mo-EEG) systems used in real-world settings [2]
• Technical artifacts: Stemming from environmental noise and recording system limitations [2]

The central problem for validation is straightforward yet profound: when analyzing real EEG data, researchers can never be certain what constitutes genuine brain activity versus artifact because both are mixed in the recorded signal. This makes it impossible to quantitatively assess how well artifact removal algorithms work, as there's no benchmark for comparison [1] [2].

Simulated Data as a Validation Solution

Simulated data addresses this fundamental limitation by creating scenarios where the ground truth brain signal is known. The core approach involves:

• Starting with clean EEG recordings obtained under ideal, restricted-movement conditions
• Artificially introducing controlled artifacts into these clean signals
• Applying artifact removal algorithms to the contaminated signals
• Comparing the results against the original clean recordings [1]

This methodology enables researchers to calculate precise performance metrics because both the input (clean EEG) and output (processed EEG) can be compared against a known standard.

Performance Metrics Enabled by Simulated Data

The use of simulated data allows quantification of artifact removal performance through standardized metrics:

Table 1: Key Performance Metrics for EEG Artifact Removal Validation

Metric	Description	Interpretation
Signal-to-Noise Ratio (SNR)	Ratio of signal power to noise power	Higher values indicate better artifact removal
Correlation Coefficient (CC)	Linear relationship between processed and clean EEG	Values closer to 1.0 indicate better preservation of original signal
Relative Root Mean Square Error (RRMSE)	Magnitude of differences between signals	Lower values indicate more accurate reconstruction
Artifact Reduction Percentage (η)	Percentage reduction in artifact components	Higher values indicate more effective artifact removal

Experimental Evidence from Recent Advances

Recent research demonstrates how simulated data has driven innovations in EEG artifact removal by enabling quantitative validation:

CLEnet: Dual-Scale CNN and LSTM Architecture

The CLEnet model integrates dual-scale convolutional neural networks (CNN) with Long Short-Term Memory (LSTM) networks and an improved EMA-1D attention mechanism. When validated on simulated datasets, CLEnet demonstrated significant improvements:

• In removing mixed artifacts (EMG + EOG), CLEnet achieved an SNR of 11.498 dB and CC of 0.925 [1]
• For ECG artifact removal, CLEnet showed a 5.13% increase in SNR and 8.08% decrease in temporal RRMSE compared to previous models [1]
• The model exhibited 2.45% and 2.65% improvements in SNR and CC respectively for multi-channel EEG containing unknown artifacts [1]

These quantitative results, made possible by simulated data validation, demonstrated CLEnet's capability to handle various artifact types more effectively than previous approaches.

Motion-Net: Subject-Specific Motion Artifact Removal

Motion-Net represents a specialized approach for removing motion artifacts in mobile EEG applications. Using simulated data with known ground truth references, researchers demonstrated:

• Motion artifact reduction percentage (η) of 86% ± 4.13 [2]
• SNR improvement of 20 ± 4.47 dB [2]
• Mean Absolute Error (MAE) of 0.20 ± 0.16 [2]

The study incorporated visibility graph features to enhance model accuracy with smaller datasets, with performance rigorously quantified through simulated data validation [2].

A²DM: Artifact-Aware Denoising Model

The A²DM framework introduced an innovative approach by fusing artifact representation into the time-frequency domain. When evaluated on the benchmark EEGdenoiseNet dataset:

• A²DM demonstrated a 12% improvement in correlation coefficient metrics compared to previous state-of-the-art methods [3]
• The model effectively removed both ocular and muscle artifacts in a unified framework [3]

This performance validation relied critically on simulated data where ground truth signals were available for comparison.

Experimental Protocols for Simulation-Based Validation

The validation of EEG artifact removal methods follows structured experimental protocols when using simulated data:

Protocol 1: Semi-Synthetic Data Generation

This approach combines clean EEG segments with recorded artifact signals:

Diagram 1: Semi-Synthetic Data Validation Workflow

This protocol involves:

• Source Data Collection: Clean EEG is typically recorded under restricted movement conditions, while artifacts (EOG, EMG, ECG) are separately recorded [1]
• Mixing Process: Artifacts are artificially introduced into clean EEG at controlled signal-to-noise ratios [1]
• Algorithm Testing: The contaminated signals are processed through artifact removal algorithms
• Performance Quantification: Processed signals are compared to original clean EEG using standardized metrics [1]

Protocol 2: Real Data with Ground Truth Reference

This approach uses specialized experimental setups to capture both contaminated and reference signals simultaneously:

Diagram 2: Real Data with Reference Validation Workflow

This method includes:

• Specialized Recording: Using systems specifically designed to capture both clean and contaminated signals [2]
• Motion Tasks: Subjects perform specific movements known to generate artifacts [2]
• Reference Signals: Obtained through electrode placements minimally affected by motion [2]
• Cross-Validation: Comparing algorithm output against reference signals [2]

Table 2: Key Research Resources for EEG Artifact Removal Validation

Resource	Function	Application in Validation
EEGdenoiseNet [1] [3]	Benchmark dataset with semi-synthetic EEG	Provides standardized data for algorithm comparison
Mockaroo [4]	Data generation tool	Creates synthetic datasets with controlled parameters
Independent Component Analysis (ICA) [5]	Blind source separation technique	Baseline method for performance comparison
Visibility Graph Features [2]	Signal transformation method	Enhances artifact detection in deep learning models
Monte Carlo Simulations [6] [7]	Statistical sampling method	Assesses algorithm robustness across variations

Advancing the Field Through Simulation

The indispensable role of simulated data in EEG artifact removal validation extends beyond initial algorithm development. It enables:

• Objective benchmarking of different approaches [1] [3]
• Identification of algorithm strengths and weaknesses across different artifact types [3]
• Robustness testing under varying signal conditions [7]
• Reproducible research through standardized evaluation protocols [1]

As the field progresses toward more sophisticated applications—including mobile EEG, brain-computer interfaces, and clinical monitoring systems—the role of simulated data in validating artifact removal techniques becomes increasingly critical. Only through rigorous simulation-based validation can researchers ensure that the brain signals they analyze truly represent neurological activity rather than methodological artifacts.

Common EEG Artifacts and Their Impact on Data Integrity

Electroencephalography (EEG) is a vital tool for investigating brain function, but the signals it records are notoriously susceptible to contamination from unwanted sources, known as artifacts. These artifacts, which can originate from the patient's own body or the external environment, pose a significant threat to data integrity, potentially leading to the misinterpretation of neural signals and incorrect conclusions in both research and clinical settings [8] [9]. This guide provides an objective comparison of common EEG artifacts and the methodologies used to mitigate them, with a specific focus on experimental data and protocols relevant to validating artifact removal techniques using simulated data.

EEG artifacts are undesired signals that introduce changes in measurements and can obscure the neural signal of interest. Given that the amplitude of cerebral EEG is typically in the microvolt range, it is highly sensitive to contamination from sources with much larger amplitudes [8] [10]. Artifacts are broadly categorized as physiological (originating from the subject's body) or non-physiological (technical, from external sources) [9] [11].

The table below summarizes the characteristics and impacts of the most prevalent physiological artifacts.

Table 1: Characteristics and Impact of Common Physiological EEG Artifacts

Artifact Type	Origin	Typical Waveform/Morphology	Spectral Characteristics	Primary Electrodes Affected	Potential Impact on Data Integrity
Ocular (Eye Blinks/Movements)	Corneo-retinal dipole (eye as dipole: cornea +, retina -) [8] [11]	Slow, high-amplitude deflections; blinks cause positive waveform in frontal electrodes [8]	Dominant in low frequencies (Delta, Theta bands) [9]	(Pre)frontal (Fp1, Fp2); lateral movements affect F7, F8 [8] [9]	Mimics slow-wave activity; can be misinterpreted as interictal discharges in epilepsy [8]
Muscle (EMG)	Contraction of head, face, or neck muscles [8]	High-frequency, sharp, irregular activity [9]	Broadband, but primarily >30 Hz (Beta/Gamma bands) [8] [9]	Widespread, but often localized to temporal regions [8]	Obscures genuine brain rhythms; reduces SNR in high-frequency bands critical for cognitive/motor studies [9]
Cardiac (ECG/Pulse)	Electrical activity of the heart (ECG) or pulsation of blood vessels under electrodes (pulse) [8] [11]	Rhythmic, stereotyped waveforms synchronized with heartbeat [11]	Overlaps multiple EEG bands; pulse artifact ~1.2 Hz [10]	Left-side brain electrodes (due to heart's position); electrodes over pulsating vessels [8] [10]	Rhythmic sharp waves may be confused with cerebral abnormal activity like sharp waves [11]

Non-physiological artifacts include electrode pops (sudden, high-amplitude transients from impedance changes), cable movement, 60 Hz AC line noise, and incorrect reference placement [9] [11]. These are often addressed through proper experimental setup and hardware solutions, but can require specific post-processing if introduced [10].

# Methodologies for Artifact Detection and Removal

A wide array of techniques has been developed to manage EEG artifacts. The choice of method often involves a trade-off between the amount of data preserved and the risk of introducing new distortions or failing to remove the artifact.

Table 2: Comparison of Common EEG Artifact Removal Methodologies

Method	Underlying Principle	Common Applications	Key Experimental Findings	Advantages	Limitations
Regression (Time/Frequency Domain)	Estimates artifact contribution via reference channels (e.g., EOG) and subtracts it from EEG signal [10]	Primarily for ocular artifacts [10]	Traditional method; can be affected by bidirectional interference (EEG contaminating EOG reference) [10]	Conceptually simple; requires separate reference channel [10]	Requires exogenous reference channels; may remove neural signals along with artifact [10]
Independent Component Analysis (ICA)	Blind source separation (BSS); decomposes EEG into statistically independent components, which are classified and removed [8] [10]	Ocular, muscle, and cardiac artifacts; BCG artifacts in EEG-fMRI [8] [12]	A 2025 study found ICA did not significantly improve decoding performance in most cases but was essential to avoid artifact-related confounds [13]. For BCG removal, ICA showed sensitivity to frequency-specific patterns in dynamic connectivity graphs [12].	Does not require reference channels; can isolate and remove multiple artifact types [8]	Requires manual component inspection (expertise-dependent); computationally intensive; performance degrades with low-channel counts [8] [14]
Artifact Subspace Reconstruction (ASR)	Identifies and removes high-variance components in multi-channel EEG using a sliding window and statistical thresholding [15]	Non-stereotyped artifacts, movement artifacts; adaptable for mobile EEG and infant data [15]	The NEAR pipeline, combining ASR with bad channel detection, successfully reconstructed established EEG responses from noisy newborn datasets [15].	Effective for non-stereotypical artifacts; can be used in real-time; suitable for low-channel setups [15]	Performance depends on calibration data and parameter tuning; may remove neural data if too aggressive [15]
Wavelet Transform	Decomposes signal into time-frequency space, allowing for targeted removal of artifact coefficients [10] [14]	Ocular and muscular artifacts, particularly in wearable EEG [14]	Emerging as a frequently used technique for managing ocular and muscular artifacts in wearable EEG pipelines, often using thresholding [14].	Good for non-stationary signals; does not require multi-channel data [10]	Choice of mother wavelet and threshold rules can impact performance [10]

→ Experimental Protocol: Impact of Artifact Correction on Decoding Performance

A 2025 study systematically evaluated how artifact correction and rejection affect the performance of Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA)-based decoding of EEG signals [13].

• Objective: To determine whether eliminating artifacts during preprocessing enhances the performance of multivariate pattern analysis (MVPA), given that artifact rejection reduces the number of trials available for training the decoder [13].
• Methodology: The researchers used Independent Component Analysis (ICA) to correct for ocular artifacts. Artifact rejection was employed to discard trials with large voltage deflections from other sources, such as muscle artifacts. They assessed decoding performance across seven common event-related potential (ERP) paradigms and more challenging multi-way decoding tasks [13].
• Key Result: The combination of artifact correction and rejection did not significantly improve decoding performance in the vast majority of cases. However, the study strongly recommends using artifact correction prior to decoding analyses to reduce artifact-related confounds that might artificially inflate decoding accuracy [13].

# Validation and Emerging Frontiers in Artifact Management

Validation of artifact removal techniques is increasingly relying on sophisticated computational approaches and multi-modal data integration. A key development is the use of data assimilation (DA) to estimate neurophysiological parameters like cortical excitation-inhibition (E/I) balance from EEG data. A 2025 study validated a DA-based computational approach by comparing its E/I estimates with concurrent Transcranial Magnetic Stimulation and EEG (TMS-EEG) measures, finding significant correlations [16]. This demonstrates that computational methods can provide neurophysiologically valid estimations, offering a framework for benchmarking artifact removal pipelines.

Simultaneous EEG-fMRI presents a unique artifact removal challenge in the form of the ballistocardiogram (BCG) artifact. A 2025 systematic evaluation of BCG removal methods—Average Artifact Subtraction (AAS), Optimal Basis Set (OBS), and ICA—found that the choice of method significantly impacts subsequent functional connectivity analysis [12]. AAS achieved the best signal fidelity (MSE = 0.0038, PSNR = 26.34 dB), while OBS best preserved structural similarity (SSIM = 0.72). ICA, though weaker on signal metrics, was sensitive to frequency-specific patterns in dynamic network graphs [12]. This highlights that the optimal method depends on the downstream analysis goal.

The workflow for validating artifact removal pipelines often involves a combination of simulated and real data, as illustrated below.

For researchers designing experiments involving EEG artifact validation, the following tools and data are essential.

Table 3: Essential Research Reagents and Resources for EEG Artifact Research

Tool/Resource	Function/Role in Research	Application Notes
Independent Component Analysis (ICA)	A blind source separation algorithm to decompose multi-channel EEG into independent components for artifact identification and removal [8] [10].	Implemented in toolboxes like EEGLAB; requires expertise for manual component selection. Performance is best with high-density EEG systems [8] [14].
Simulated Data Generation	Creates EEG signals with known ground truth and controlled artifact properties to quantitatively evaluate removal algorithms [16].	Critical for initial validation. Can be generated using neural mass models or by adding real artifacts to clean baseline data [16].
Artifact Subspace Reconstruction (ASR)	An adaptive, automated method for cleaning continuous EEG data by identifying and removing high-variance segments [15].	Particularly useful for non-stereotypical artifacts in mobile EEG and motion-heavy recordings (e.g., infant studies) [15].
Auxiliary Reference Sensors	Sensors (EOG, ECG, IMU) that provide direct measurements of physiological artifacts (eye, heart, movement) [10] [14].	Used for regression-based removal and for validating the performance of other methods like ICA. Still underutilized in wearable EEG [10] [14].
Public EEG Datasets with Artifacts	Benchmarks for comparing the performance of different artifact removal pipelines across laboratories [14].	Should include data from varied populations (adults, infants) and recording setups (lab, mobile) to ensure generalizability [14] [15].
Multi-Modal Neuroimaging (TMS-EEG)	Provides neurophysiological benchmarks (e.g., E/I balance indices) for validating the neurophysiological integrity of data after artifact removal [16].	Serves as a "gold standard" for validating that artifact cleaning preserves genuine brain signals, not just removes noise [16].

The integrity of EEG data is fundamentally linked to the effective management of artifacts. While techniques like ICA, ASR, and wavelet transforms offer powerful solutions, the choice of method is not one-size-fits-all. Experimental evidence indicates that the optimal pipeline depends on the artifact type, the recording context (lab-based vs. wearable), and the ultimate goal of the analysis, whether it is ERP decoding, functional connectivity, or estimating neurophysiological parameters. A robust validation strategy, combining simulated data with known ground truth and multi-modal benchmarks like TMS-EEG, is essential for developing and selecting artifact removal methods that ensure the reliability of neuroscientific and clinical findings.

Spectral and Temporal Characteristics of Key Artifacts

Electroencephalography (EEG) is a vital tool in clinical and cognitive neuroscience, prized for its high temporal resolution and non-invasive nature [1]. However, its diagnostic accuracy is consistently challenged by the presence of artifacts—extraneous signals that obscure underlying brain activity. These artifacts originate from diverse sources, including physiological processes like eye movements and muscle activity, as well as non-physiological sources such as environmental interference and electrode motion [2] [17]. In wearable EEG systems, which employ dry electrodes and are used in mobile, real-world settings, these challenges are exacerbated due to reduced electrode-skin stability and the uncontrolled nature of the recording environment [18] [17]. The reliable removal of these artifacts is therefore a critical step in EEG analysis, forming an essential foundation for downstream applications in brain-computer interfaces, neurological diagnosis, and cognitive monitoring. This guide provides a comparative analysis of the spectral and temporal characteristics of key EEG artifacts and the performance of modern methods designed for their removal, with a specific focus on validation using simulated data.

Spectral and Temporal Profiles of Major Artifact Types

EEG artifacts exhibit distinct signatures in both the spectral (frequency) and temporal (time) domains. Understanding these characteristics is the first step in developing effective artifact removal strategies. The table below summarizes the defining features of the most common artifact types.

Table 1: Spectral and Temporal Characteristics of Key EEG Artifacts

Artifact Type	Spectral Characteristics	Temporal Characteristics	Primary Sources
Ocular (EOG)	Low-frequency content (< 4 Hz); overlaps with delta rhythm [17].	Slow, large-amplitude deflections; correlated with eye blinks and movements [17].	Eye movements, blinks [2].
Muscular (EMG)	Broad-spectrum, high-frequency activity (20-200 Hz); often overlaps with beta and gamma rhythms [17].	Rapid, irregular, high-frequency spikes [17].	Muscle contractions in head, neck, jaw [2].
Motion	Can contaminate a broad frequency range [2].	Sharp transients, baseline shifts, and periodic oscillations; patterns can be arrhythmic [2].	Head movements, gait cycles, electrode displacement [2].
Cardiac (ECG)	Periodic component around 1-1.7 Hz [1].	Stereotypical, periodic spike patterns synchronized with heartbeat [18].	Heartbeat [2].
Non-Physiological/Technical	Specific to noise source (e.g., 50/60 Hz line noise) [1].	Sudden, large-amplitude shifts (e.g., electrode pops); continuous interference (e.g., static) [19].	Electrode pops, static, line noise, instrumental interference [2] [17].

Comparative Performance of Artifact Removal Algorithms

A variety of algorithms, from classical signal processing to modern deep learning, have been developed to tackle artifact removal. Their performance varies significantly based on the artifact type and the experimental setup. The following table synthesizes quantitative results from recent studies, providing a basis for comparison.

Table 2: Performance Comparison of Representative Artifact Removal Algorithms

Algorithm	Artifact Type	Key Performance Metrics	Reported Experimental Setup
CLEnet [1]	Mixed (EMG + EOG)	SNR: 11.50 dB; CC: 0.925; RRMSEt: 0.300 [1].	End-to-end removal from multi-channel EEG; tested on semi-synthetic and real 32-channel data [1].
Motion-Net [2]	Motion	Artifact Reduction (η): 86% ±4.13; SNR Improvement: 20 ±4.47 dB; MAE: 0.20 ±0.16 [2].	Subject-specific CNN; trained/tested on real EEG with motion artifacts and ground truth [2].
Fingerprint + ARCI + Improved SPHARA [18]	Multiple (Dry EEG)	SD: 6.15 μV; SNR: 5.56 dB [18].	Combination of ICA-based methods and spatial filtering; tested on 64-channel dry EEG during motor tasks [18].
ICA-based Framework [20]	TMS-Evoked Muscle	High reproducibility of TEPs with 35+ training trials [20].	Real-time, two-step ICA; validated on pre-published TMS-EEG datasets [20].
Deep Lightweight CNN [19]	Eye, Muscle, Non-Physiological	Eye (ROC AUC): 0.975; Muscle (Accuracy): 93.2%; Non-Physio (F1-score): 77.4% [19].	Artifact-specific CNN models; trained/tested on Temple University Hospital EEG Corpus [19].

Experimental Protocols for Key Studies

The performance data presented above is derived from rigorous experimental protocols. The following workflow generalizes the common steps involved in validating artifact removal methods using simulated or semi-synthetic data, a cornerstone of research in this field.

Diagram 1: Workflow for Validating Artifact Removal Methods.

The methodology for validating artifact removal algorithms often follows a structured pipeline [1] [2]:

• Data Acquisition: The process begins with the collection of clean EEG data. Simultaneously, artifacts are either recorded from real subjects (e.g., during specific tasks to induce eye movements or muscle activity) or are mathematically simulated [1].
• Semi-Synthetic Data Generation: To establish a reliable ground truth for quantitative evaluation, clean EEG signals are artificially contaminated with the recorded or simulated artifacts. This creates a semi-synthetic dataset where the true, artifact-free signal is known [1].
• Algorithm Training: The removal algorithm is trained. For deep learning models, this is typically done in a supervised manner, where the model learns to map the contaminated input to the clean output. Other methods, like Independent Component Analysis (ICA), operate in an unsupervised fashion, learning to separate sources without labeled examples [1] [20].
• Execution and Evaluation: The trained or configured algorithm is applied to the contaminated data. The output is then quantitatively compared against the known ground truth using metrics like Signal-to-Noise Ratio (SNR), Correlation Coefficient (CC), and Root Mean Square Error (RMSE) [1] [2].

To facilitate replication and further research, the table below details key computational tools and data resources used in the featured studies.

Table 3: Key Research Reagents and Computational Resources

Tool/Resource	Type	Function in Research	Example Use Case
EEGdenoiseNet [1]	Benchmark Dataset	Provides semi-synthetic single-channel EEG data with clean and artifact components for standardized algorithm testing [1].	Training and benchmarking of artifact removal models like CLEnet [1].
Temple University Hospital (TUH) EEG Corpus [19]	Clinical EEG Dataset	Offers a large corpus of real, clinical EEG recordings with expert-annotated artifact labels, enabling validation in real-world conditions [19].	Training and testing artifact-specific deep learning models [19].
Independent Component Analysis (ICA) [18] [20]	Algorithm	A blind source separation method that decomposes multi-channel EEG into statistically independent components, allowing for manual or automatic identification and removal of artifact components [18] [20].	Removal of ocular and muscular artifacts; suppression of TMS-evoked artifacts in real-time [20].
Convolutional Neural Network (CNN) [1] [2] [19]	Deep Learning Architecture	Excels at extracting spatial and morphological features from data; can be designed for 1D signals or 2D topographic maps. Used for end-to-end artifact removal or detection [1] [19].	Motion-Net for motion artifact removal; lightweight CNNs for detecting specific artifact classes [2] [19].
Long Short-Term Memory (LSTM) [1]	Deep Learning Architecture	A type of recurrent neural network designed to learn long-range dependencies and temporal features in sequential data like EEG signals [1].	Integrated in CLEnet to capture the temporal dynamics of EEG for better separation from artifacts [1].

The effective removal of artifacts from EEG signals hinges on a deep understanding of their unique spectral and temporal fingerprints. As demonstrated, ocular artifacts dominate the low-frequency range, while muscular and motion artifacts present complex, broad-spectrum challenges. The contemporary landscape of artifact removal is increasingly dominated by deep learning approaches, such as CLEnet and Motion-Net, which show superior performance in handling complex and mixed artifacts in multi-channel and mobile settings. However, traditional methods like ICA remain highly relevant, especially in scenarios with sufficient channels and for specific artifact types like those evoked by TMS. Validation through semi-synthetic data with known ground truth remains a critical and standard practice for the objective quantification of algorithm performance. As the field progresses, the fusion of spatial, spectral, and temporal processing techniques, coupled with the availability of robust public datasets, will continue to enhance the reliability of EEG analysis across clinical and research applications.

The Critical Role of Ground-Truth in Validating EEG Artifact Removal

Electroencephalography (EEG) functional connectivity (FC) research provides invaluable insights into brain network dynamics in both healthy and clinical populations. However, the accurate interpretation of EEG FC patterns is critically dependent on successfully removing artifacts from the signal. Artifacts from physiological sources (e.g., eye blinks, muscle activity, cardiac rhythms) and non-physiological sources (e.g., environmental noise, motion) can significantly distort connectivity metrics, leading to false conclusions about brain network organization [1] [21]. Consequently, establishing reliable ground-truth connectivity patterns is fundamental for validating the performance of artifact removal algorithms.

Research demonstrates that methodological choices in EEG processing pipelines significantly impact the estimation of functional connectivity, creating considerable variability across studies [22] [23]. Simulated data has emerged as an essential validation tool because researchers know the precise "ground truth" of the underlying neural connections, enabling objective evaluation of how different artifact removal techniques affect connectivity estimates [23]. Without this ground-truth benchmark, it is impossible to determine whether an artifact removal method preserves genuine neural signals while effectively eliminating non-neural contaminants.

Comparative Performance of Artifact Removal Methodologies

Traditional Signal Processing Approaches

Traditional approaches to EEG artifact removal include regression-based methods, blind source separation (BSS) techniques like Independent Component Analysis (ICA), wavelet transforms, and hybrid methods [1]. Among these, ICA remains one of the most widely used methods in both research and clinical applications [24] [25]. ICA operates on the principle of separating statistically independent components from multidimensional data, effectively isolating neural signals from artifactual sources [25]. However, ICA's performance is contingent on meeting specific statistical assumptions and is influenced by measurement uncertainty [24].

Table 1: Performance Characteristics of Traditional Artifact Removal Methods

Method	Key Mechanism	Advantages	Limitations	Impact on FC Metrics
ICA (FastICA, Infomax)	Blind source separation of statistically independent components	Effective for ocular, muscle, and line noise artifacts; Widely available in toolboxes	Requires manual component inspection; Performance degrades with measurement uncertainty (SNR <15dB) [24]	Can preserve connectivity with proper component rejection [22]
Wavelet-Enhanced ICA (wICA)	Combines wavelet decomposition with ICA	Improved artifact separation; Better preservation of neural signal morphology	Increased computational complexity; Parameter sensitivity	Provides high test-retest reliability for alpha band FC [22]
Artifact Subspace Reconstruction (ASR)	Statistical rejection of high-variance components	Suitable for online processing; Effective for large-amplitude motion artifacts	May remove neural signals with high amplitude; Threshold selection critical	Limited evidence on specific FC impact

Modern Deep Learning Architectures

Recent advances in deep learning have transformed EEG artifact removal by enabling automated, end-to-end processing without manual intervention. These approaches leverage convolutional neural networks (CNNs), long short-term memory (LSTM) networks, transformers, and generative adversarial networks (GANs) to learn complex patterns in artifact-contaminated EEG data [1] [21].

Table 2: Comparative Performance of Deep Learning Artifact Removal Models

Model	Architecture	Artifact Types Addressed	Performance Metrics	Experimental Validation
CLEnet [1]	Dual-scale CNN + LSTM with EMA-1D attention	EMG, EOG, ECG, and unknown artifacts	SNR: 11.498dB; CC: 0.925; RRMSEt: 0.300; RRMSEf: 0.319 (mixed artifacts)	Semi-synthetic datasets with known ground truth
AnEEG [21]	LSTM-based GAN	Ocular, muscle, powerline interference	Lower NMSE and RMSE; Higher CC vs. wavelet methods	Multiple public EEG datasets
IMU-Enhanced LaBraM [26]	Fine-tuned transformer with IMU fusion	Motion artifacts during physical activities	Improved robustness under diverse motion scenarios vs. ASR-ICA	Mobile BCI dataset with standing/walking/running conditions
Unsupervised Encoder-Decoder [27]	Deep encoder-decoder with outlier detection	Task-specific artifacts without pre-labeling	10% relative improvement in downstream classification	Clinical EEG data for coma prognostication

The comparative data reveals that deep learning approaches generally outperform traditional methods, particularly for complex artifact types and in real-world conditions with motion [1] [26]. CLEnet demonstrates particular strength in handling mixed artifacts and unknown noise sources, while IMU-enhanced approaches show promise for mobile brain-computer interface applications where motion artifacts are prevalent.

Experimental Protocols for Ground-Truth Validation

Simulated EEG Functional Connectivity Data

Rigorous validation of artifact removal methods requires experimental protocols that incorporate known ground-truth connectivity patterns. One established approach involves simulating EEG data with predefined functional connectivity networks, enabling precise quantification of how artifact removal affects connectivity estimation [23].

Simulation Methodology:

• Generate synthetic EEG signals with predetermined connectivity patterns between brain regions
• Introduce controlled artifacts with varying properties (amplitude, frequency, spatial distribution)
• Apply artifact removal techniques to the contaminated signals
• Compare estimated connectivity with the original ground-truth patterns

Key Design Considerations:

• Incorporate realistic artifact properties based on empirical measurements [24]
• Systematically vary signal-to-noise ratios to test robustness [24]
• Include both physiological (EOG, EMG, ECG) and non-physiological artifacts
• Test across multiple connectivity metrics (phase-based, amplitude-based, directed)

This simulated approach enables researchers to determine optimal processing parameters for accurate FC estimation. Studies using this methodology have revealed that combining specific preprocessing steps significantly enhances connectivity measurement accuracy [23].

Diagram Title: Simulated Data Validation Workflow

Semi-Synthetic Benchmark Datasets

An alternative validation approach employs semi-synthetic datasets created by adding real artifacts to clean EEG recordings or combining artifact-free EEG segments with recorded artifactual signals [1]. This method preserves the statistical properties of genuine artifacts while maintaining ground-truth knowledge of the underlying neural signals.

Protocol Implementation:

• Data Collection: Obtain clean EEG segments during resting state or specific tasks
• Artifact Acquisition: Record pure artifact signals (EOG, EMG, ECG) separately
• Mixing Procedure: Linearly combine clean EEG and artifacts with controlled weighting
• Algorithm Validation: Apply artifact removal methods and quantify signal preservation

This approach has been utilized in benchmark studies such as those employing the EEGdenoiseNet dataset [1], enabling standardized comparison across multiple artifact removal algorithms. The semi-synthetic paradigm offers a compelling balance between experimental control and physiological realism.

Optimal Processing Pipelines for Connectivity Research

Evidence from ground-truth validation studies provides specific guidance for constructing optimal EEG processing pipelines for functional connectivity research. The most reliable approaches combine multiple techniques in a sequential manner to address different artifact types.

Integrated Processing Pipeline

The most effective pipelines incorporate artifact reduction techniques, appropriate re-referencing methods, and carefully selected functional connectivity metrics [22]. Research indicates that the combination of wavelet-enhanced ICA (wICA) artifact cleaning, current source density (CSD) re-referencing, and real magnitude squared coherence (rMSC) as a FC metric provides particularly high accuracy and test-retest reliability for detecting age-related differences in alpha band functional connectivity [22].

Optimal Parameters for FC Estimation:

• Epoch Length: 6 seconds or longer [23]
• Number of Epochs: 40 or more [23]
• Re-referencing: Reference Electrode Standardization Technique (REST) or common average reference (CAR) for phase-based metrics; CSD for magnitude-based metrics [23]
• FC Metrics: Imaginary coherence (iCOH) and weighted phase lag index (wPLI) for phase-based connectivity; rMSC for amplitude-based connectivity [22] [23]

Diagram Title: Optimal EEG-FC Processing Pipeline

Method Selection Guidelines

Choosing the appropriate artifact removal method depends on multiple factors, including the research question, artifact types, available computational resources, and the specific functional connectivity metrics of interest. The following guidelines emerge from ground-truth validation studies:

• For clinical applications with limited technical expertise: Deep learning approaches (e.g., CLEnet, AnEEG) offer automated processing with minimal manual intervention [1] [21]
• For motion-rich environments: IMU-enhanced methods provide superior artifact removal during physical activities [26]
• For phase-based connectivity metrics: ICA with wavelet enhancement combined with REST re-referencing delivers optimal performance [22] [23]
• For amplitude-based connectivity metrics: Current source density re-referencing with rMSC metrics is preferable [22]
• When measurement uncertainty is high (SNR <15dB): ICA performance degrades significantly, necessitating alternative approaches or quality checks [24]

Essential Research Reagents and Computational Tools

Table 3: Essential Research Resources for EEG Artifact Removal Validation

Resource Category	Specific Tools & Datasets	Primary Function	Access Information
Public EEG Datasets	BrainClinics Repository [22]; EEGdenoiseNet [1]; Mobile BCI Dataset [26]	Benchmarking artifact removal algorithms; Ground-truth validation	Publicly available through respective repositories
Processing Toolboxes	EEGLAB [24]; MNE-Python [24]	Implementation of ICA and other artifact removal methods	Open-source platforms with extensive documentation
Deep Learning Frameworks	TensorFlow; PyTorch	Developing and training custom artifact removal models	Open-source with active community support
Simulation Platforms	MATLAB; Python (MNE, NumPy, SciPy)	Generating ground-truth connectivity patterns; Method validation	Commercial and open-source options available
Performance Metrics	SNR; CC; RRMSEt; RRMSEf [1]	Quantitative evaluation of artifact removal efficacy	Standardized calculation methods

Future Directions and Emerging Approaches

The field of EEG artifact removal continues to evolve with several promising research directions. Multi-modal approaches that combine EEG with complementary physiological recordings (e.g., IMU, EOG, EMG) show significant potential for improved artifact identification and removal [26]. Additionally, self-supervised and unsupervised learning methods address the challenge of obtaining labeled training data by leveraging the inherent statistical properties of clean versus artifact-contaminated EEG segments [27].

Future validation efforts should focus on developing more sophisticated simulation frameworks that better capture the complex spatial, temporal, and spectral properties of both neural signals and artifacts. Furthermore, standardized benchmarking protocols using shared ground-truth datasets will enable more direct comparison between existing and emerging artifact removal methodologies, ultimately advancing the reliability of EEG functional connectivity research.

Limitations of Real Data for Methodological Benchmarking

The validation of electroencephalogram (EEG) artifact removal methods represents a critical challenge in computational neuroscience and biomedical engineering. While real EEG data provides the ultimate test environment, significant methodological limitations complicate its use for standardized benchmarking of new algorithms. This analysis examines these constraints within the broader context of validation research, where simulated data offers complementary advantages for controlled, reproducible evaluation of algorithmic performance.

Limitations of Real EEG Data for Benchmarking

Using real EEG data for benchmarking artifact removal methods introduces several fundamental challenges that can compromise validation integrity. The table below summarizes the primary limitations identified in current research.

Table 1: Key Limitations of Real EEG Data for Methodological Benchmarking

Limitation Category	Specific Challenge	Impact on Benchmarking
Unknown Ground Truth	Inability to precisely separate true neural activity from artifacts [28]	Prevents accurate calculation of performance metrics and recovery fidelity
Artifact Variability	Uncontrolled, subject-specific artifact composition and intensity [28]	Introduces uncontrolled variables that complicate performance comparisons
Channel Correlations	Poor performance on multi-channel data due to overlooked inter-channel relationships [28]	Limits generalizability of single-channel focused algorithms
Data Scarcity	Difficulty obtaining sufficient real data representing all artifact types [21]	Restricts training data for deep learning models and comprehensive testing
Subjective Annotation	Manual component rejection requiring expert intervention and prior knowledge [28]	Introduces human bias and limits reproducibility across studies
Resource Intensity	Requirement for reference signals and manual inspection in traditional methods [28]	Increases operational complexity and cost of data collection

The absence of known ground truth presents the most fundamental constraint. Without precise knowledge of the underlying neural signal, researchers cannot accurately quantify how effectively an algorithm removes artifacts while preserving genuine brain activity [28]. This problem is compounded by the presence of unknown artifacts in real recordings, whose proportion relative to the original signal remains unquantified [28].

Artifact variability further complicates benchmarking. Real biological artifacts (EOG, EMG, ECG) exhibit substantial inter-subject variability in characteristics and intensity, creating uncontrolled variables that hinder fair algorithm comparison [28]. This variability is particularly problematic for deep learning approaches that require extensive, diverse datasets for training [21].

Simulated Data as a Complementary Validation Tool

To address these limitations, researchers have developed sophisticated simulation approaches that enable controlled benchmarking. The following workflow illustrates a standard methodology for creating semi-synthetic EEG data, which combines clean EEG segments with recorded artifacts.

Semi-synthetic datasets created through this process provide researchers with contaminated signals alongside pristine ground truth, enabling precise quantification of artifact removal performance using standardized metrics [28] [21].

Quantitative Performance Metrics for Algorithm Benchmarking

The validation of EEG artifact removal methods relies on specific quantitative metrics that enable objective comparison between different algorithms. The following table outlines key performance indicators derived from experimental protocols in recent literature.

Table 2: Quantitative Metrics for EEG Artifact Removal Benchmarking

Metric	Description	Interpretation	Experimental Results from Recent Studies
Signal-to-Noise Ratio (SNR)	Ratio of signal power to noise power [28]	Higher values indicate better artifact removal	CLEnet: 11.498 dB for mixed artifacts [28]
Correlation Coefficient (CC)	Linear correlation between processed and clean EEG [28]	Values closer to 1.0 indicate better signal preservation	CLEnet: 0.925 for mixed artifacts [28]
Relative Root Mean Square Error (RRMSE)	Temporal (t) and frequency (f) domain reconstruction error [28]	Lower values indicate higher fidelity	CLEnet: RRMSEt 0.300, RRMSEf 0.319 [28]
Normalized Mean Square Error (NMSE)	Normalized reconstruction error [21]	Lower values indicate better agreement with original signal	AnEEG demonstrated lower NMSE vs. wavelet techniques [21]
Signal-to-Artifact Ratio (SAR)	Ratio of signal power to residual artifact power [21]	Higher values indicate more complete artifact removal	AnEEG showed improvements in SAR values [21]

These metrics provide complementary insights into algorithm performance. For instance, CLEnet demonstrated significant improvements across multiple metrics when evaluated on semi-synthetic datasets, achieving a 2.45% increase in SNR and 2.65% increase in CC compared to other models, while reducing temporal and frequency domain errors by 6.94% and 3.30% respectively [28].

Experimental Protocols for Benchmarking Studies

Standardized experimental protocols enable fair comparison between different artifact removal approaches. The following diagram illustrates a comprehensive benchmarking workflow that incorporates both simulated and real validation stages.

This protocol emphasizes initial validation on semi-synthetic datasets with known ground truth, followed by confirmation on real EEG recordings. For example, recent studies have utilized EEGdenoiseNet as a standardized semi-synthetic dataset, combining clean EEG with recorded EOG and EMG artifacts at controlled signal-to-noise ratios [28]. Additional datasets incorporate ECG artifacts from the MIT-BIH Arrhythmia Database to evaluate algorithm performance on cardiac artifacts [28] [21].

Table 3: Research Reagent Solutions for EEG Artifact Removal Studies

Resource	Type	Function	Example Implementation
EEGdenoiseNet	Benchmark Dataset	Provides standardized semi-synthetic data with ground truth [28]	Mixed EEG, EOG, and EMG artifacts with known mixing ratios [28]
MIT-BIH Database	Artifact Source	Supplies clean ECG signals for cardiac artifact simulation [28] [21]	Combined with EEGdenoiseNet for ECG artifact evaluation [28]
CLEnet	Deep Learning Architecture	Dual-scale CNN with LSTM for multi-channel artifact removal [28]	Incorporates EMA-1D attention mechanism for temporal feature enhancement [28]
AnEEG	Deep Learning Framework	LSTM-based GAN for artifact removal [21]	Generator produces clean EEG, discriminator evaluates quality [21]
GCTNet	Hybrid Architecture	GAN-guided parallel CNN with transformer network [21]	Captures both global and temporal dependencies in EEG [21]
1D-ResCNN	Baseline Algorithm	One-dimensional residual convolutional neural network [28]	Uses multiple convolutional kernels for multi-scale feature extraction [28]

These resources enable standardized implementation and comparison of artifact removal methods. For instance, CLEnet's architecture specifically addresses the limitation of previous models that performed poorly on multi-channel data by effectively capturing inter-channel correlations [28]. Similarly, AnEEG's adversarial training approach enables the generation of artifact-free EEG signals while maintaining original neural activity patterns [21].

The limitations of real EEG data for methodological benchmarking underscore the critical importance of simulated and semi-synthetic datasets in validation research. While real data remains essential for final performance confirmation, controlled simulations enable rigorous, reproducible evaluation of artifact removal algorithms using quantitative metrics with known ground truth. Future benchmarking efforts should leverage both approaches, utilizing standardized datasets and evaluation protocols to enable meaningful comparison across the rapidly evolving landscape of EEG artifact removal methodologies.

Methodologies for Generating and Utilizing Simulated EEG Data

Electroencephalography (EEG) is a fundamental tool for investigating brain function in clinical, neuroscience, and cognitive research. A significant challenge in developing EEG analysis techniques, particularly for artifact removal, is the absence of a known ground truth in real neural data. Without this reference, validating the accuracy and efficacy of new algorithms becomes problematic. Simulated EEG data with precisely known properties provides an essential solution, creating a controlled test bench for method validation [29] [30].

This guide compares three prominent toolboxes for simulated EEG generation: SEED-G, EEGSourceSim, and SEREEGA. We focus on their application within a research thesis dedicated to validating EEG artifact removal methods, providing objective performance data and experimental protocols to inform researchers' selections.

Comparative Analysis of Simulated EEG Toolboxes

The table below summarizes the core characteristics and capabilities of the three featured toolboxes.

Toolbox Name	Primary Simulation Approach	Key Features & Strengths	Best Suited for Validating	Accessibility
SEED-G [29] [31] [32]	Multivariate Autoregressive (MVAR) Models	Designed for testing connectivity estimators Imposes known ground-truth connectivity patterns Controls network parameters (density, nodes) Models non-stationary and inter-trial variable connectivity [29]	Connectivity-based artifact removal, Dynamic network analysis	MATLAB; Publicly available on GitHub [32]
EEGSourceSim [33]	MRI-based Forward Models & Realistic Noise Embedding	High anatomical realism with individual head models Embeds signal in realistic biological noise Suitable for source estimation & connectivity [33]	Source localization methods, Spatially-focused artifact removal	MATLAB; Open-source toolbox and dataset [33]
SEREEGA [30]	Lead Field Projection & Configurable Signal Mixing	Modular and general-purpose design Simulates event-related potentials (ERPs) and oscillations Configurable head models and signal types [30]	ERP analysis methods, Temporal artifact filtering	MATLAB; Free and open-source [30]

Quantitative Performance and Experimental Data

SEED-G Performance Metrics

SEED-G is optimized for computational efficiency and realistic spectral properties. Performance testing demonstrates that datasets with up to 60 time series can be generated in less than 5 seconds [29] [31]. The toolbox successfully produces signals with spectral features similar to real EEG data, a critical factor for meaningful validation [29].

To illustrate the impact of data length on connectivity estimation accuracy—a key consideration for artifact removal validation—SEED-G documentation provides the following experimental results [32]:

Number of Samples	False Positive Rate (FPR)	False Negative Rate (FNR)
500 samples	2%	50%
1500 samples	1%	11%
2500 samples	0%	6%

This data underscores that longer simulated epochs significantly improve the reliability of the ground truth, which is crucial for robustly testing artifact removal algorithms.

EEGSourceSim emphasizes realism through its use of a large set of 23 individual MRI-based head models and surface-based regions of interest brought into registration for each subject [33]. This approach allows for simulation studies that account for individual-subject variability in structure and function, providing a more rigorous test for artifact removal methods that may be sensitive to anatomical differences.

Experimental Protocols for Artifact Removal Validation

Here, we outline a general experimental workflow for validating an EEG artifact removal method using simulated data, adaptable to any of the toolboxes above.

Core Experimental Workflow

The following diagram visualizes the multi-stage process of creating a benchmark and validating a method against it.

Protocol 1: Validating a Deep Learning Denoiser for tES Artifacts

This protocol is inspired by a study that benchmarked deep learning models, including Complex CNN and State Space Models (SSMs), for removing transcranial Electrical Stimulation (tES) artifacts [34].

• Data Generation: Use a toolbox like SEREEGA to generate clean, synthetic EEG signals. Alternatively, use real, artifact-free EEG recordings as your baseline.
• Artifact Introduction: Create a semi-synthetic dataset by adding mathematically generated tES artifacts (e.g., for tDCS, tACS, tRNS) to the clean EEG. This provides a known ground truth for the underlying brain signal [34].
• Method Application: Apply the deep learning-based denoising method (e.g., Complex CNN for tDCS, multi-modular SSM for tACS/tRNS) to the contaminated dataset [34].
• Performance Quantification: Compare the denoised output to the original clean signal using metrics like:
  • Root Relative Mean Squared Error (RRMSE) in temporal and spectral domains.
  • Correlation Coefficient (CC) between the cleaned and ground-truth signals [34].

Protocol 2: Testing a Connectivity-Based Method with Non-Stationary Data

This protocol leverages SEED-G's unique capabilities to test methods in dynamic scenarios.

• Ground-Truth Design: Use SEED-G to generate a pseudo-EEG dataset with a known, time-varying connectivity pattern. For example, impose a connectivity link that changes in intensity at a specific time point [29] [32].
• Artifact Simulation: Introduce a controlled, non-brain artifact (e.g., an ocular blink simulated by a spike or slow wave) that overlaps with the period of connectivity change.
• Validation:
  • Apply the artifact removal algorithm to the contaminated dataset.
  • Run a dynamic connectivity estimator (e.g., Partial Directed Coherence) on both the clean ground-truth data and the artifact-corrected data.
  • Assess whether the artifact removal process preserved or distorted the underlying, known dynamic change in connectivity.

The Scientist's Toolkit: Key Research Reagents

The table below lists essential "reagents" or components for designing realistic EEG simulation experiments.

Research Reagent	Function in the Simulation Experiment
Head Model (Forward Model)	Prescribes how electrical activity from brain sources is projected to scalp electrodes. Realistic boundary element method (BEM) models are crucial for simulating volume conduction effects [33] [30].
Multivariate Autoregressive (MVAR) Model	Acts as a generator filter to produce synthetic time series with specific, user-imposed statistical connectivity patterns between signals, creating a ground-truth network [29].
Synthetic Artifact Model	A mathematical or data-driven model of a specific artifact (e.g., ocular blink, muscle activity, tES stimulation) that can be added to clean EEG with controlled amplitude and timing [34].
Realistic Biological Noise	A model of ongoing, background brain activity, often derived from fitting components to measured resting-state EEG, which provides a plausible noise floor for the simulated signal [33].

Selecting the ideal simulated EEG toolbox depends directly on the validation goals of the artifact removal research. For studies focusing on the integrity of functional connectivity networks before and after cleaning, SEED-G is the superior choice due to its dedicated feature set for imposing and testing ground-truth connectivity. When the research question involves the spatial accuracy of source reconstruction following artifact removal, EEGSourceSim offers unparalleled anatomical realism. For more general-purpose validation, particularly of methods targeting event-related potentials or oscillatory activity, SEREEGA provides the necessary flexibility and modularity. By leveraging the experimental protocols and performance data outlined in this guide, researchers can make an informed decision and build a robust validation framework for their specific thesis on EEG artifact removal.

Leveraging Multivariate Autoregressive (MVAR) Models

Multivariate Autoregressive (MVAR) modeling is a powerful parametric approach for estimating dynamic, directed interactions from physiological signals like electroencephalography (EEG). In neuroscience, it is particularly valued for its ability to quantify directed functional connectivity with high temporal resolution, helping researchers understand how different brain areas interact over time scales as brief as tens of milliseconds [35]. The core principle of an MVAR model is that the current value of a multivariate time series can be predicted by a linear combination of its own past values. For a d-dimensional time series, the general form of a time-varying MVAR process of order p at each time step n is represented as:

Y(n)=∑r=1pAr(n)Y(n−r)+E(n)

Where Ar(n) is the matrix of time-varying MVAR coefficients at time n, and E(n) is a zero-mean, uncorrelated white noise vector process [35]. The model order p determines the number of past observations included in the model. A key advantage of MVAR models in the context of artifact removal validation is that, when fitted to clean neural data, they can generate synthetic EEG signals with known ground-truth properties, free from artifacts. This makes them an indispensable tool for creating realistic benchmark datasets where the true underlying brain activity is known, thereby allowing for objective evaluation of artifact removal algorithms [36] [37].

Comparative Analysis of MVAR Algorithm Performance

Several recursive algorithms exist for estimating time-varying MVAR (tvMVAR) models from non-stationary neural data. The choice of algorithm and its parameter settings significantly impacts the accuracy and reliability of the resulting connectivity estimates and synthetic data generation. The following table provides a structured comparison of four prominent tvMVAR algorithms.

Table 1: Comparison of Time-Varying MVAR (tvMVAR) Estimation Algorithms

Algorithm Name	Core Methodology	Key Adaptation Parameters	Strengths	Weaknesses & Sensitivity
Recursive Least Squares (RLS) [35]	Extends Yule-Walker equations using a forgetting factor (λ) to weight errors over time.	Forgetting factor (λ); Model order (p)	Lower computational complexity; Suitable for single-trial modeling followed by averaging.	Performance degrades with signal downsampling; Sensitive to choice of λ.
General Linear Kalman Filter (GLKF) [35]	Models state process as a random walk using observation and state equations.	Two adaptation constants (c1, c2); Model order (p)	Allows for both single-trial and multi-trial modeling; Effective with well-tuned constants.	High c1/c2 values increase estimate variance; Low values slow adaptation.
Multivariate Adaptive Autoregressive (MVAAR) [35]	Kalman filter variant updating measurement noise from prior prediction error.	Adaptation coefficient (c); Model order (p)	Effective for single-trial analysis.	Limited to single-trial modeling; Performance varies with model order and sampling.
Dual Extended Kalman Filter (DEKF) [35]	Simultaneously estimates states and parameters of the dynamical system.	Adaptation coefficient; Model order (p)	Efficient for nonlinear dynamical systems.	Limited to single-trial modeling; Sensitive to parameter initialization.

Performance and Sensitivity Insights

Experimental comparisons using both simulated data and benchmark EEG recordings have revealed critical performance insights. Across a broad range of model orders, all algorithms can correctly reproduce interaction patterns, demonstrating a degree of robustness to this parameter [35]. However, signal downsampling often degrades connectivity estimation accuracy for most algorithms, though in some cases it can reduce estimate variability by lowering the number of model parameters [35]. Furthermore, the strategy for handling multiple trials significantly impacts outcomes. Single-trial modeling followed by averaging can achieve optimal performance with larger adaptation coefficients than previously suggested, but it exhibits slower adaptation speeds compared to multi-trial modeling, where one tvMVAR model is fitted simultaneously across all trials [35].

Experimental Protocols for Artifact Removal Validation

A rigorous protocol for validating EEG artifact removal methods using MVAR models involves two main stages: 1) generating a simulated, ground-truth EEG dataset, and 2) applying and evaluating the artifact removal techniques on this controlled data.

Protocol 1: Generating Synthetic EEG with Ground Truth

This protocol creates realistic, artifact-free EEG signals with known connectivity properties.

• Step 1: Define Head Model and Source Locations: Begin by using a realistic head model, such as the New-York head model, to define the locations of neural sources and their projections to scalp electrodes [36].
• Step 2: Fit MVAR Model to Clean Data: Fit a high-dimensional MVAR model to long segments of clean, high-quality intracranial or scalp EEG data. This step captures realistic neural interaction dynamics. Techniques like the group Least Absolute Shrinkage and Selection Operator (gLASSO) can be effectively used to fit these models, even with a high number of recording sites (~100-200) [37].
• Step 3: Generate Synthetic EEG Signals: Use the fitted MVAR model to generate synthetic multichannel EEG time series. The dynamic directed connectivity in this synthetic data is known by design, providing a perfect ground truth [36] [38].
• Step 4: Simulate Artifacts: Introduce controlled artifacts into the clean, synthetic EEG. For example, to simulate motion artifacts, one can model the effects at the skin-electrode interface, connecting cables, and the electrode-amplifier system [39]. This results in a final dataset with known brain signals and known artifacts.

The following diagram illustrates this workflow for creating validated synthetic EEG data.

Protocol 2: Benchmarking Artifact Removal Methods

This protocol tests the efficacy of different artifact removal algorithms on the simulated data.

• Step 1: Apply Artifact Removal: Process the artifact-laden synthetic dataset from Protocol 1 with various artifact removal techniques. Prominent methods include:
  • Independent Component Analysis (ICA): A blind source separation method that separates mixed signals into statistically independent components [40] [41].
  • ERASE: A modified ICA approach that uses additional EMG reference channels to force more artifact power into identifiable components [41].
  • RMD-SVD: A method combining Regenerative Multi-Dimensional Singular Value Decomposition with ICA, effective for single-channel artifact removal [40].
• Step 2: Compare to Ground Truth: Compare the "cleaned" output from each method against the original, known ground-truth EEG from Protocol 1.
• Step 3: Quantify Performance: Calculate performance metrics to objectively compare methods. Key metrics include:
  • Signal-to-Noise Ratio (SNR): Measures the level of desired signal relative to noise/artifacts [40].
  • Mean Squared Error (MSE): Quantifies the average squared difference between the cleaned and ground-truth signals [40].
  • Sensitivity & False Positive Rates: Assess the ability to correctly identify and remove artifacts without discarding neural signals [41].

Table 2: Key Performance Metrics for Artifact Removal Validation

Metric	Definition	Interpretation in Validation
Signal-to-Noise Ratio (SNR)	Ratio of signal power to noise power.	A higher SNR indicates more effective artifact suppression and better preservation of the neural signal.
Mean Squared Error (MSE)	Average of the squares of the errors between cleaned and true signal.	A lower MSE indicates the cleaned signal is closer to the true, artifact-free neural signal.
Sensitivity	Proportion of true artifacts correctly identified and removed.	Measures the method's ability to detect artifacts. High sensitivity means fewer artifacts remain.
False Positive Rate	Proportion of neural signal incorrectly identified as artifact.	A low false positive rate indicates the method preserves brain activity well, minimizing data loss.

The following diagram outlines this benchmarking process.

The Scientist's Toolkit: Essential Research Reagents

Successfully implementing the aforementioned experimental protocols requires a combination of specific computational tools, software, and methodological components.

Table 3: Essential Reagents for MVAR-based Validation Research

Research Reagent	Function & Role in Validation	Exemplars & Notes
Computational Head Model	Provides a biophysically realistic volume conductor to simulate scalp potentials from neural sources.	New-York Head Model [36]; Allows for accurate forward modeling of EEG signals.
MVAR Model Fitting Toolbox	Software package for estimating MVAR parameters from time series data.	Custom MATLAB scripts [36]; gLASSO for high-dimensional data [37]; SEED-G toolbox for inter-brain simulation [38].
Artifact Simulation Module	Introduces controlled, realistic artifacts into clean EEG signals for validation.	Models for motion artifacts at skin-electrode interface and cables [39]; Models for EMG artifacts [41].
Blind Source Separation (BSS) Algorithm	Core computational engine for separating neural signals from artifacts in mixed recordings.	Independent Component Analysis (ICA) [40] [41]; Canonical Correlation Analysis (CCA) [41].
Performance Metric Calculator	Quantitatively assesses the fidelity of artifact-cleaned signals against ground truth.	Scripts to calculate SNR, MSE, PSNR [40]; Sensitivity and False Positive rate calculators [41].

MVAR models provide a rigorous mathematical framework for generating synthetic EEG with known ground-truth properties, making them a cornerstone for the objective validation of artifact removal algorithms. Systematic comparison of tvMVAR algorithms reveals that while all can recover basic interaction patterns, their performance is sensitive to parameters like adaptation coefficients and sampling rate. The experimental protocols outlined—involving synthetic data generation followed by systematic benchmarking—offer a robust pathway for evaluating the next generation of EEG artifact removal techniques. This approach is critical for advancing the reliability of EEG analysis in both basic neuroscience and applied settings such as clinical drug development.

Incorporating Realistic Non-Idealities and Non-Stationarities

Validating electroencephalography (EEG) artifact removal algorithms requires frameworks that incorporate realistic non-idealities and non-stationarities inherent in real-world data acquisition. As research into neural dynamics during natural movement and real-world tasks accelerates, the limitations of traditional artifact removal methods become increasingly apparent. These approaches often struggle with the complex, time-varying artifacts encountered in mobile EEG scenarios, where motion-induced signals and physiological interferences exhibit non-stationary characteristics that overlap with neural signals of interest both temporally and spectrally. This comparison guide objectively evaluates contemporary artifact removal methodologies, focusing on their performance validation using simulated data and controlled experimental setups that incorporate these challenging real-world conditions.

The transition from laboratory-based EEG systems to mobile brain imaging approaches has created an urgent need for validation frameworks that can accurately replicate the non-ideal conditions encountered during movement. These frameworks employ sophisticated simulation techniques, including head phantoms with electrical dipoles, semi-synthetic datasets combining clean EEG with recorded artifacts, and experimental protocols that systematically introduce non-stationarities. By testing algorithms against known ground truth signals in controlled yet realistic environments, researchers can establish meaningful performance benchmarks and identify the most suitable approaches for specific application contexts, from clinical monitoring to athletic performance optimization.

Performance Comparison of Contemporary Artifact Removal Methods

Table 1: Quantitative Performance Metrics of Deep Learning-Based Methods

Method	Architecture	Artifact Types Addressed	SNR Improvement (dB)	Correlation Coefficient	Artifact Reduction (%)	RMSE
Motion-Net	1D CNN with Visibility Graph features	Motion artifacts	20 ±4.47 [2]	N/R	86 ±4.13 [2]	0.20 ±0.16 [2]
CLEnet	Dual-scale CNN + LSTM with EMA-1D	EMG, EOG, ECG, Mixed artifacts	11.50 [1]	0.925 [1]	N/R	0.300 (temporal) [1]
A²DM	Artifact-aware CNN with frequency enhancement	Ocular (EOG), Muscle (EMG)	N/R	12% improvement over NovelCNN [3]	N/R	N/R
LSTEEG	LSTM-based Autoencoder	Multiple artifact types	N/R	Superior to convolutional autoencoders [42]	N/R	N/R
Complex CNN	Convolutional Neural Network	tDCS artifacts	Best performance for tDCS [34]	N/R	N/R	N/R
M4 Network	State Space Models (SSMs)	tACS, tRNS artifacts	Best performance for tACS/tRNS [34]	N/R	N/R	N/R

SNR: Signal-to-Noise Ratio; RMSE: Root Mean Square Error; N/R: Not Reported

Table 2: Performance Comparison of Non-Deep Learning Methods for Motion Artifacts

Method	Core Approach	Applicable Scenarios	Key Performance Advantages	Computational Considerations
iCanClean	Canonical Correlation Analysis with reference signals	Walking, running	Better P300 congruency effect recovery [43], Improved dipolarity [43]	Effective with pseudo-reference signals [43]
Artifact Subspace Reconstruction (ASR)	Sliding-window PCA with thresholding	Walking, running	Reduced power at gait frequency [43], Similar ERP latency to standing task [43]	Less effective than iCanClean for P300 [43]
onEEGwaveLAD	Wavelet Transform + Isolation Forest	Real-time single-channel applications	Fully automated, no reference signals required [44]	Configurable window length tradeoffs [44]
Tripolar Concentric Ring Electrodes	Hardware-based Laplacian filtering	High-density mobile EEG	Improved spatial selectivity [45], Better localization accuracy at high artifact amplitudes [45]	Hardware solution requiring specialized equipment [45]

The quantitative comparison reveals distinctive performance profiles across methodological categories. Deep learning approaches generally excel at handling complex, non-stationary artifacts when sufficient training data is available, with architectures like CLEnet demonstrating versatility across multiple artifact types [1]. Motion-Net shows exceptional performance specifically for motion artifacts, achieving approximately 86% artifact reduction while maintaining signal integrity through its subject-specific training approach [2]. The incorporation of visibility graph features provides structural information that enhances performance with smaller datasets, addressing a critical limitation of many deep learning methods [2].

Non-deep learning methods offer compelling advantages in scenarios requiring real-time processing or where training data is limited. iCanClean demonstrates particular effectiveness for motion artifact removal during dynamic tasks like running, successfully recovering expected P300 congruency effects that other methods miss [43]. Hardware-based solutions like tripolar concentric ring electrodes provide unique value in high-artifact environments, maintaining performance even when software approaches struggle with extreme amplitude artifacts [45]. The choice between these methodological approaches ultimately depends on specific application requirements, including computational constraints, available channels, and the nature of expected artifacts.

Experimental Protocols and Validation Methodologies

Simulation-Based Validation with Head Phantoms

Advanced head phantom systems provide the most physiologically realistic validation environments for evaluating artifact removal performance under controlled conditions. These systems incorporate electrical dipoles at anatomically relevant positions to simulate neural sources with precise spatial and temporal characteristics. In a comprehensive validation study, researchers constructed a head phantom using ballistics gelatin with fourteen dipolar sources: ten simulating neural generators in regions including occipital lobes, sensorimotor cortices, cerebellum, frontal and parietal lobes, premotor cortex, and anterior cingulate gyrus; and four simulating myoelectric sources in neck muscles (sternocleidomastoids and semispinalis capitis) [45]. This configuration enabled direct comparison of conventional disk electrodes versus tripolar concentric ring electrodes in recovering known neural signals amid contaminating muscle artifacts.

The experimental protocol broadcast simulated neural signals as random, time-varying, single-frequency sinusoidal bursts within standard EEG spectral bands (5-37 Hz), using prime number frequencies to avoid harmonic resonance in recorded signals [45]. Simultaneously, actual recorded human neck muscle activity during walking was broadcast at scaled amplitudes ranging from 0× to 2× typical surface recording levels. This approach systematically tested the robustness of each method across varying artifact intensities while maintaining ground truth knowledge of neural signals. Performance was evaluated through spectral power peak detection, scalp map spatial entropy, and localization accuracy metrics, providing comprehensive assessment of both signal preservation and spatial fidelity [45].

Semi-Synthetic Dataset Construction

Semi-synthetic approaches provide a flexible framework for evaluating artifact removal performance when specific artifact types are of interest. The CLEnet validation employed three distinct datasets: Dataset I combined single-channel EEG with EMG and EOG artifacts; Dataset II incorporated ECG artifacts from the MIT-BIH Arrhythmia Database; while Dataset III utilized real 32-channel EEG collected during a 2-back task containing unknown physiological artifacts [1]. This progressive validation approach tested generalizability from controlled semi-synthetic conditions to realistic unknown artifacts.

The creation of semi-synthetic data follows a standardized methodology: clean EEG segments are first identified from recordings during resting states or tasks with minimal artifact contamination. Artifact signals are then recorded separately under conditions that elicit specific artifacts (e.g., muscle tension for EMG, eye movements for EOG, walking for motion artifacts). These artifact signals are scaled to appropriate amplitudes and added to the clean EEG, typically with varying signal-to-noise ratios to test robustness across contamination levels [1]. This approach maintains the physiological characteristics of real artifacts while preserving ground truth knowledge of the underlying neural signals.

Experimental Protocols for Real-World Motion Artifacts

Validating motion artifact removal during naturalistic movement requires specialized protocols that capture the non-stationarities of whole-body motion. One approach employed a dynamic Flanker task during jogging compared to a static standing condition, enabling direct comparison of event-related potential components with and without motion artifacts [43]. This protocol specifically evaluated the preservation of neural signals of interest (P300 ERP components) alongside artifact reduction metrics.

Performance assessment in motion artifact studies typically employs multiple complementary metrics: (1) ICA component dipolarity, measuring how well independent components reflect physiologically plausible brain sources; (2) spectral power reduction at gait frequency and harmonics, quantifying removal of movement-related periodic artifacts; and (3) recovery of expected ERP components compared to stationary conditions [43]. This multi-metric approach balances artifact removal effectiveness against neural signal preservation, addressing the critical challenge of avoiding over-cleaning that removes neural signals along with artifacts.

Signaling Pathways and Experimental Workflows

Simulation-Based Validation Workflow

The following diagram illustrates the comprehensive workflow for validating EEG artifact removal methods using simulated data and head phantom systems:

Simulation-Based Validation Workflow

This workflow illustrates the systematic approach for incorporating non-idealities and non-stationarities into validation frameworks. The process begins with simultaneous simulation of neural sources and recording of actual artifact signals, preserving the non-stationary characteristics of real-world artifacts [45]. The combined signals are broadcast through an electrically realistic head phantom at varying amplitude levels, testing algorithm robustness across different artifact intensities [45]. Multi-electrode recordings capture the contaminated signals using both conventional and specialized electrodes, enabling direct comparison of hardware and software approaches. Finally, comprehensive validation against known ground truth signals provides quantitative performance assessment using multiple complementary metrics [45].

Method Comparison Framework

The following diagram outlines the conceptual framework for comparing different artifact removal approaches:

Method Comparison Framework

This framework categorizes artifact removal approaches based on their fundamental methodology and highlights their suitability for different application contexts. Hardware-based solutions like dual-layer EEG and tripolar concentric ring electrodes address artifacts at the acquisition stage, providing inherent noise rejection through specialized electrode geometries and signal processing [45]. Software-based algorithms encompass traditional methods like ICA and ASR that leverage statistical properties of the signals, alongside modern deep learning approaches that learn complex artifact patterns from data [2] [1] [3]. Real-time adaptive methods offer specialized solutions for mobile brain-computer interface applications where immediate processing is essential [44] [46]. The optimal approach depends heavily on the specific application requirements, with clinical settings often prioritizing accuracy, research focusing on method development, and mobile applications emphasizing real-time capability.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Materials for EEG Artifact Removal Validation

Category	Item	Function and Application	Key Characteristics
Validation Platforms	Electrical Head Phantom	Provides ground truth testing with simulated neural sources and artifacts [45]	Ballistics gelatin medium with embedded dipoles; Anatomically positioned sources
	Semi-Synthetic Datasets	Controlled performance evaluation with known artifact types [1]	Combines clean EEG with recorded artifacts; Scalable artifact amplitudes
Reference Algorithms	iCanClean Pipeline	Reference method for motion artifact comparison [43]	Uses CCA with reference signals; Effective for gait-related artifacts
	Artifact Subspace Reconstruction (ASR)	Benchmark for real-time artifact removal [43]	PCA-based sliding window approach; Adjustable threshold parameter (k)
	ICA-based Approaches (ICLabel)	Standard for component-based artifact removal [43]	Blind source separation; Requires multiple channels
Specialized Electrodes	Tripolar Concentric Ring Electrodes	Hardware-based artifact reduction [45]	Surface Laplacian calculation; Enhanced spatial selectivity
	Dual-layer EEG Electrodes	Motion artifact rejection through noise reference [45]	Mechanically coupled electrodes; Primary and secondary sensing elements
Software Tools	EEGDENoiseNet	Benchmark dataset and evaluation framework [1]	Contains multiple artifact types; Standardized performance metrics
	onEEGwaveLAD Framework	Real-time single-channel artifact removal [44] [46]	Wavelet-based decomposition; Isolation Forest anomaly detection

This toolkit represents the essential resources for rigorous validation of EEG artifact removal methods under conditions incorporating non-idealities and non-stationarities. Electrical head phantoms stand as particularly valuable tools, enabling controlled introduction of realistic artifacts while maintaining precise knowledge of ground truth neural signals [45]. The combination of simulated neural sources and actual recorded artifacts preserves the non-stationary characteristics critical for meaningful algorithm evaluation. Specialized electrodes like tripolar concentric ring configurations provide hardware-based alternatives that can complement software approaches, particularly in high-artifact environments [45].

Reference algorithms establish performance baselines across different methodological categories, from traditional statistical approaches like ASR and ICA to more recent developments like iCanClean [43]. Standardized datasets such as EEGDENoiseNet enable direct comparison across studies and methods, providing consistent evaluation frameworks [1]. For real-time applications, frameworks like onEEGwaveLAD offer fully automated processing without requirement for reference signals or multi-channel setups, addressing critical constraints in mobile BCI applications [44] [46]. Together, these tools support comprehensive validation workflows that stress-test artifact removal methods under realistically challenging conditions.

Neural Mass Models (NMMs) and State-Space Frameworks (SSFs) are two powerful computational approaches for modeling brain dynamics, each with distinct strengths and applications. NMMs provide a biophysically grounded representation of the mean-field activity of neural populations, making them ideal for simulating the source signals that generate neuroimaging data like EEG. In contrast, SSFs offer a robust mathematical structure for separating true neural signals from noise and artifacts, and for tracking the dynamic evolution of latent brain states in real-time. While their primary functions differ, their combination is increasingly vital for creating validated, clinically relevant neurotechnologies, particularly in the critical task of EEG artifact removal. The following comparison delineates their performance, supported by experimental data, to guide researchers and drug development professionals in selecting and integrating these tools.

Understanding the core principles of each framework is essential for appreciating their comparative performance.

Neural Mass Models (NMMs) are biophysical models that describe the average electrical activity—such as mean membrane potentials and firing rates—of a population of neurons. They operate on the principle of mean-field approximation, where the complex interactions of thousands of individual neurons are summarized into a few key state variables. NMMs are typically formulated as systems of coupled differential equations that can generate realistic brain rhythms (e.g., alpha, gamma) and are used to simulate the cortical source activity that gives rise to macroscale signals like EEG [47]. Their strength lies in their biological interpretability, as their parameters often correspond to physiological quantities like synaptic gains and time constants.

State-Space Frameworks (SSFs) are a broad class of statistical models used to describe systems that evolve over time. They consist of two primary equations: a state equation that models the dynamics of an underlying, unobserved process (the "state"), and an observation equation that describes how these hidden states are measured in the presence of noise. In neuroscience, SSFs are exceptionally powerful for denoising and temporal tracking. For instance, they can model the analytic signal of a brain rhythm as a latent state, allowing for real-time phase estimation without the need for bandpass filtering, which often couples signal and noise [48]. The Kalman filter is a classic algorithm used to optimally estimate the state in these models.

Performance Comparison and Experimental Data

The table below summarizes the objective performance of NMMs and SSFs across key metrics relevant to neuroimaging research, particularly in the context of signal generation and artifact removal.

Table 1: Quantitative Performance Comparison of Neural Mass Models and State-Space Frameworks

Performance Metric	Neural Mass Models (NMMs)	State-Space Frameworks (SSFs)
Primary Function	Biophysically realistic signal & connectivity generation [47]	Noise suppression, latent state tracking, and real-time estimation [48]
Temporal Tracking Accuracy	N/A (Used for simulation)	Outperforms bandpass-filtering methods in real-time phase estimation under broadband rhythms, phase resets, and low SNR [48]
Connectivity Estimation	Can generate ground-truth interconnected signals for validation [49]	Infers effective connectivity from observed data via multivariate autoregressive models in the state equation [50]
Computational Efficiency	Efficient for generating complex, multi-frequency signals [49]	Linear-time efficiency and high parallelizability for long-sequence modeling; lower cost than Transformers [51]
Noise & Artifact Handling	Limited intrinsic handling; signals are often simulated without artifact [49]	Excellently separates state dynamics from observation noise; directly models and mitigates artifacts [48]
Validation Outcome (Example)	Recovered original antenna signals with cross-correlations >0.8 after ICA [49]	Provides credible intervals for phase estimates; improves accuracy of brain-behavior relationship detection [48]

Detailed Experimental Protocols

To contextualize the data in Table 1, here are the detailed methodologies from key experiments that benchmarked these frameworks.

Protocol 1: Validating EEG Connectivity with a Phantom Head and NMMs

This experiment was designed to quantify the accuracy of connectivity measures in the presence of real-world volume conduction and head motion, using NMMs to provide a ground truth [49].

• Signal Generation: Six separate neural mass models were used to generate source signals with peak frequencies in standard EEG bands (delta, theta, alpha, beta, low gamma, high gamma). These sources were summed with different weightings to create three primary antenna signals (low: 6.5 Hz, mid: 10 Hz, high: 41 Hz) and three distractor signals.
• Physical Setup: The predefined signals were fed into antennae embedded within a mannequin head. The head was filled with a dental plaster mixture to simulate tissue conductance and mounted on a motion platform that mimicked human head motion at various walking speeds.
• Data Acquisition & Processing: High-density EEG was recorded from the moving phantom head. Independent Component Analysis (ICA) was applied to the scalp channel data to recover the original antenna signals.
• Validation Metrics: The performance was assessed using:
  • Signal Recovery: Cross-correlation between recovered independent components and original antenna signals.
  • Connectivity Estimation: Multiple connectivity measures (e.g., ffDTF, dDTF, gPDC, WPLI) were computed on the recovered signals to see if they could identify the true interconnections implanted by the NMMs.

Protocol 2: Real-Time Neural Phase Estimation with a State-Space Model

This study introduced a State-Space Phase Estimator (SSPE) to overcome limitations of filter-based real-time phase estimation methods [48].

• Model Formulation: The analytic signal (a complex value representing the instantaneous amplitude and phase of a rhythm) was defined as the latent state to be tracked. The state equation modeled this latent signal as a harmonic oscillator rotating at a fixed frequency, with added state noise to allow for dynamic variations. The observation equation summed the real parts of multiple such oscillators to predict the recorded neural data.
• Parameter Fitting: The model parameters (e.g., oscillator frequencies, noise covariances) were first fit acausally to a segment of data assuming spectral stationarity.
• Real-Time Estimation: The Kalman filter was applied for causal, real-time estimation. Upon receiving each new data sample, the filter would:
  • Predict the next state (analytic signal).
  • Compare the prediction to the actual observation.
  • Update the state estimate to optimally balance the model's prediction and the new data.
• Benchmarking: The SSPE was tested against established real-time methods in simulations featuring narrowband rhythms, broadband rhythms, phase resets, and multiple concurrent rhythms at varying signal-to-noise ratios (SNR). Performance was measured by the accuracy of the phase estimate compared to the known ground truth.

Signaling Pathways and Workflows

The following diagrams illustrate the core logical workflows for employing NMMs in a validation pipeline and for implementing a State-Space Framework.

Neural Mass Model Validation Pipeline

Diagram 1: NMM Validation Workflow. This flowchart outlines the experimental procedure for using Neural Mass Models to validate EEG processing techniques. The process begins by defining a ground truth and using NMMs to generate known, interconnected neural signals. These signals are then transmitted through a physical phantom head, which introduces real-world volume conduction. EEG is recorded while the head is in motion, adding motion artifacts. The contaminated EEG is then processed (e.g., using Independent Component Analysis) to recover the underlying sources. Finally, connectivity metrics are calculated from the recovered signals and compared against the original ground truth to quantify accuracy [49].

State-Space Model Framework for Denoising

Diagram 2: SSF Denoising Loop. This diagram illustrates the recursive operation of a State-Space Framework for tracking a neural signal. The State Model represents the hidden, true dynamics of the brain process (like a rhythmic oscillation). The Observation Model defines how this clean state is manifested in the recorded, noisy EEG data. The Kalman Filter Update is the core algorithm that continuously integrates the model's prediction with incoming, real-world measurements to produce an optimal, denoised estimate of the underlying neural state, such as the instantaneous phase of a rhythm [48].

The Scientist's Toolkit

This table details key computational and experimental "reagents" essential for working with these frameworks.

Table 2: Essential Research Reagents and Tools

Item Name	Function/Brief Explanation	Relevance to Framework
Phantom Head	A physical mannequin head with embedded antennae that simulates the electrical properties of real tissue and volume conduction.	Critical for validating NMM-generated signals and EEG processing methods in a realistic but controlled environment [49].
Independent Component Analysis (ICA)	A blind source separation algorithm used to decompose multichannel EEG into statistically independent components, often separating neural signals from artifacts.	A standard preprocessing step used to recover simulated NMM signals from mixed scalp recordings [49].
Kalman Filter	An optimal recursive estimation algorithm that updates the state of a system based on a series of measurements observed over time, containing statistical noise.	The core computational engine for many State-Space Frameworks, enabling real-time tracking and denoising [48].
Directed Transfer Function (DTF)	A multivariate spectral measure used to quantify the directional flow of information (effective connectivity) between EEG sources.	A connectivity metric validated using ground-truth connections generated by NMMs [49].
Open Ephys (SSPE Plugin)	An open-source software platform for electrophysiology data acquisition. The SSPE plugin integrates the state-space phase estimation method.	Provides a ready-to-use implementation of the State-Space Phase Estimator for real-time experiments [48].

The validation of Electroencephalogram (EEG) artifact removal algorithms requires rigorous benchmarking against ground truth data. Since clean EEG is often unobtainable in real-world scenarios due to inherent biological and environmental noise, semi-synthetic datasets have become a cornerstone of methodological development [1]. These datasets are constructed by deliberately adding well-characterized artifacts to clean EEG recordings, enabling precise quantification of an algorithm's performance in separating neural signal from noise [52]. This case study examines the creation and application of such datasets, focusing on their critical role in objectively comparing the performance of state-of-the-art artifact removal techniques within the broader context of validation research for electrophysiological signal processing.

Experimental Protocols & Methodologies

Semi-Synthetic Dataset Generation

The foundational protocol for creating a semi-synthetic dataset involves the controlled addition of artifactual signals to clean EEG epochs.

• Core Principle: Artifact-free segments of multi-channel EEG data are combined with recorded or simulated artifact signals (e.g., Electrooculography (EOG) for eye movements, Electromyography (EMG) for muscle activity) in a specific manner [52] [1]. This process creates a contaminated signal where the original clean EEG is known, providing a target for reconstruction algorithms.
• Standardized Benchmarks: Datasets like EEGDenoiseNet [52] [1] provide a benchmark for this approach. They offer a structured collection of semi-synthetic data, formed by combining single-channel EEG with EMG and EOG artifacts, which facilitates consistent training and evaluation of deep learning models [52].

The following workflow outlines the standard procedure for generating and utilizing a semi-synthetic EEG dataset for artifact removal research.

Featured Artifact Removal Algorithms

This section details the experimental methodologies of several advanced artifact removal algorithms, which are trained and evaluated on semi-synthetic datasets.

CLEnet: A Dual-Scale Hybrid Model

CLEnet integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks with an improved attention mechanism for end-to-end artifact removal [1].

• Workflow:
  • Morphological Feature Extraction: Two convolutional kernels of different scales extract spatial features from the input EEG. An embedded EMA-1D (One-Dimensional Efficient Multi-Scale Attention) module enhances temporal feature preservation during this stage [1].
  • Temporal Feature Extraction: The extracted features are dimensionality-reduced via fully connected layers. An LSTM network then processes these features to capture long-term temporal dependencies in the genuine EEG [1].
  • EEG Reconstruction: The fused features are flattened and passed through fully connected layers to reconstruct the artifact-free EEG signal. The model is trained in a supervised manner using Mean Squared Error (MSE) as the loss function [1].

LSTEEG: An LSTM-based Autoencoder

LSTEEG is a novel deep learning approach based on LSTM layers within an autoencoder architecture, designed for both artifact detection and correction [52].

• Workflow:
  • Unsupervised Training for Detection: The autoencoder is trained exclusively on clean EEG data to minimize the reconstruction error (MSE). Epochs containing artifacts, which are dissimilar to the training data, yield a high reconstruction error. This error metric is leveraged as an anomaly score for automated artifact detection [52].
  • Artifact Correction: For the correction task, the network is trained to map artifact-contaminated input signals to their corresponding clean versions, learning the complex, non-linear relationships within the sequential EEG data [52].

Wavelet-ICA: A Traditional Blind Source Separation Approach

This method combines Discrete Wavelet Transform (DWT) with Independent Component Analysis (ICA) for automatic artifact removal [53].

• Workflow:
  • Wavelet Decomposition: An L-level decomposition tree is applied to multi-channel signals, producing approximate and detail coefficients [53].
  • Blind Source Separation: The combined wavelet coefficients from all channels are used as input to the FastICA algorithm, which estimates independent wavelet-domain components (WICs) [53].
  • Automatic Artifact Identification: A priori artifact information (e.g., individually recorded eye blink or teeth clenching epochs) is used. The power spectrum density (PSD) features of reconstructed WICs are correlated with artifact labels to automatically identify artifactual components [53].
  • Signal Reconstruction: Identified artifact components are removed by setting their corresponding values in the entity matrix to zero before projecting the components back to the sensor space and applying the inverse wavelet transform [53].

Performance Comparison of Artifact Removal Algorithms

The following tables summarize the quantitative performance of various algorithms on standardized tasks, as reported in the literature. These metrics are crucial for objective comparison.

Table 1: Performance on Single-Channel Semi-Synthetic Datasets (EMG + EOG Artifact Removal)

Algorithm	Architecture	SNR (dB)	CC	RRMSEt	RRMSEf
CLEnet [1]	CNN + LSTM + EMA-1D	11.498	0.925	0.300	0.319
DuoCL [1]	CNN + LSTM	Information Not Provided	Information Not Provided	Information Not Provided	Information Not Provided
NovelCNN [1]	Convolutional Neural Network	Information Not Provided	Information Not Provided	Information Not Provided	Information Not Provided
1D-ResCNN [1]	1D Residual CNN	Information Not Provided	Information Not Provided	Information Not Provided	Information Not Provided

Table 2: Performance on Multi-Channel EEG Data with Unknown Artifacts

Algorithm	Architecture	SNR (dB)	CC	RRMSEt	RRMSEf
CLEnet [1]	CNN + LSTM + EMA-1D	(Baseline) +2.45%	(Baseline) +2.65%	(Baseline) -6.94%	(Baseline) -3.30%
DuoCL [1]	CNN + LSTM	Baseline	Baseline	Baseline	Baseline

Table 3: Algorithm Performance and Characteristics Overview

Algorithm	Artifact Types Addressed	Key Strengths	Computational Demand
CLEnet [1]	EMG, EOG, ECG, Unknown	Superior performance on multi-channel data & unknown artifacts; integrates spatial and temporal features.	Presumably High
LSTEEG [52]	General Artifacts	Accurate artifact detection via anomaly detection; meaningful latent space.	Presumably High
Wavelet-ICA [53]	EOG, EMG	Does not require massive offline training samples; automatic component identification.	Lower than DL models
Fast BSS Algorithm [54]	Ocular, Cardiac, Muscle, Powerline	Fast computation; suitable for online/ongoing correction.	Low

The Scientist's Toolkit: Research Reagents & Essential Materials

This section lists key software and data resources essential for conducting research in EEG artifact removal.

Table 4: Essential Research Tools for EEG Artifact Removal

Tool Name	Type	Primary Function	Application in This Context
EEGDenoiseNet [52] [1]	Benchmark Dataset	Provides semi-synthetic data pairs (clean + artifactual)	Training and benchmarking deep learning models for artifact removal.
EEGLAB [55]	Software Toolbox	Processing EEG and MEG data; includes ICA.	Implementing and comparing traditional BSS methods like ICA.
MNE-Python [55]	Software Framework	A complete toolkit for EEG/MEG data analysis.	Building full analysis pipelines, including preprocessing, source localization, and statistical testing.
FieldTrip [55]	Software Toolbox	A wide range of functions for MEG/EEG analysis.	Creating highly customized analysis scripts for specific research questions.
BioSig [55]	Software Library	Biomedical signal processing for EEG and other biosignals.	Handling various data formats and providing standardized processing tools.

Optimizing Artifact Removal Pipelines and Overcoming Pitfalls

Balancing Computational Cost with Estimation Accuracy

Electroencephalography (EEG) artifact removal represents a critical preprocessing step in neuroscience research, clinical diagnosis, and brain-computer interface applications. The fundamental challenge in this domain centers on the inherent trade-off between computational efficiency and estimation accuracy—two often competing priorities that researchers must carefully balance when selecting artifact removal methodologies. With the advent of deep learning approaches, this balance has become increasingly complex, as sophisticated models promise superior artifact suppression but often at the expense of significant computational resources [56]. This comparison guide objectively evaluates current EEG artifact removal techniques through the lens of this critical trade-off, providing researchers with experimental data and methodological insights to inform their selection process within simulated data validation frameworks.

Performance Comparison of EEG Artifact Removal Methods

Quantitative Performance Metrics

Table 1: Performance Metrics Across Deep Learning Architectures

Model	Architecture	SNR (dB)	CC	RRMSEt	RRMSEf	Key Artifacts Targeted
CLEnet [1]	Dual-scale CNN + LSTM + EMA-1D	11.498	0.925	0.300	0.319	EMG, EOG, Mixed, Unknown
AnEEG [21]	LSTM-based GAN	Improved	Higher	Lower	-	Ocular, Muscle, Environmental
Multi-modular SSM [34]	State Space Models	-	Higher	Lower (spectral)	-	tACS, tRNS artifacts
Complex CNN [34]	Convolutional Neural Network	-	-	Lower (temporal)	-	tDCS artifacts
1D-ResCNN [1]	Residual CNN	Lower	Lower	Higher	Higher	EMG, EOG
NovelCNN [1]	CNN-based	Lower	Lower	Higher	Higher	EMG-specific
DuoCL [1]	CNN + LSTM	Lower	Lower	Higher	Higher	General artifacts

Computational Requirements Analysis

Table 2: Computational Complexity and Resource Requirements

Model	Training Complexity	Inference Speed	Hardware Demands	Scalability	Suitable Applications
CLEnet [1]	Medium-High	Medium	GPU recommended	High (multi-channel)	Research, clinical analysis
AnEEG [21]	High (GAN training)	Medium	GPU required	Medium	Offline processing
Multi-modular SSM [34]	High	Medium	GPU recommended	Medium	tES-EEG applications
Lightweight CNN [19]	Low	High	CPU feasible	High	Real-time monitoring, embedded systems
ICA-based Methods [56]	Low-Medium	Medium	CPU sufficient	Limited	General purpose, research
iCanClean [43]	Low	High	CPU feasible	High	Mobile EEG, locomotion studies

Experimental Protocols and Methodologies

CLEnet Dual-Branch Architecture Protocol

The CLEnet framework employs a sophisticated dual-branch approach that systematically extracts both morphological and temporal features from EEG signals. The experimental protocol consists of three critical stages [1]:

• Morphological Feature Extraction with Temporal Enhancement: The model utilizes two convolutional kernels of different scales to identify and extract morphological features at multiple resolutions. The primary architecture consists of stacked CNN layers with an embedded EMA-1D (One-Dimensional Efficient Multi-Scale Attention Mechanism) module to maximize genuine EEG morphological feature extraction while preserving temporal characteristics.

• Temporal Feature Extraction: Features from the initial stage undergo dimensional reduction through fully connected layers to eliminate redundant information. Subsequently, Long Short-Term Memory (LSTM) networks process these refined features to capture temporal dependencies inherent in genuine EEG signals.

• EEG Reconstruction: The fused and enhanced features are flattened, and fully connected layers reconstruct them into artifact-free EEG signals. The entire model is trained in a supervised manner using mean squared error (MSE) as the loss function.

The validation protocol employed three distinct datasets: Dataset I (semi-synthetic EEG with EMG/EOG), Dataset II (semi-synthetic EEG with ECG), and Dataset III (real 32-channel EEG from healthy university students performing a 2-back task). This comprehensive validation approach demonstrates CLEnet's effectiveness across both controlled and real-world scenarios [1].

GAN-Based Artifact Removal Protocol

Generative Adversarial Networks have emerged as powerful tools for EEG artifact removal, with AnEEG representing an LSTM-based GAN implementation [21]. The experimental methodology involves:

• Adversarial Training Framework: The generator network processes artifact-contaminated EEG and attempts to produce clean EEG signals, while the discriminator network evaluates the authenticity of these generated signals compared to ground-truth clean EEG.

• LSTM Integration: The incorporation of Long Short-Term Memory layers enables the model to capture temporal dependencies and contextual information critical for EEG data processing.

• Comprehensive Validation: Quantitative metrics including NMSE (Normalized Mean Square Error), RMSE (Root Mean Square Error), CC (Correlation Coefficient), SNR (Signal-to-Noise Ratio), and SAR (Signal-to-Artifact Ratio) are calculated to verify model effectiveness. Performance benchmarks are established against traditional methods like wavelet decomposition techniques.

This approach demonstrates that GAN-based models can achieve lower NMSE and RMSE values, indicating better agreement with original signals, while obtaining higher CC values reflecting stronger linear agreement with ground truth signals [21].

State Space Models for tES Artifact Removal

For Transcranial Electrical Stimulation artifacts, which present unique challenges due to their characteristics, a multi-modular State Space Model (SSM) architecture has demonstrated particular efficacy [34]. The experimental protocol includes:

• Synthetic Dataset Creation: Clean EEG data is combined with synthetic tES artifacts representing tDCS, tACS, and tRNS stimulation types to create controlled evaluation datasets with known ground truth.

• Multi-modular Architecture: The SSM-based approach utilizes state space representations to model the dynamic nature of tES artifacts, effectively separating them from neural signals.

• Cross-Stimulation Validation: The model is evaluated across different stimulation sources, demonstrating superior performance for complex tACS and tRNS artifacts compared to other approaches.

Evaluation metrics include RRMSE (Root Relative Mean Squared Error) in both temporal and spectral domains, and Correlation Coefficient (CC), providing comprehensive assessment across multiple signal dimensions [34].

CLEnet Architecture Workflow

Signaling Pathways and Methodological Relationships

The evolution of EEG artifact removal methodologies reveals a clear progression from simple linear approaches to increasingly sophisticated architectures that better model the complex, non-linear nature of neural signals and artifacts [56]. This progression directly reflects the field's ongoing effort to balance computational efficiency with estimation accuracy.

Methodological Evolution in EEG Denoising

Table 3: Critical Research Reagents and Computational Resources

Resource Category	Specific Tools/Solutions	Function/Purpose	Implementation Considerations
Benchmark Datasets	EEGdenoiseNet [1], Temple University Hospital EEG Corpus [19], Semi-simulated EEG/EOG dataset [57]	Provides standardized data for training and evaluation	Ensure proper licensing; Check annotation quality
Validation Metrics	SNR, CC, RRMSE(t/f), SAR, NMSE [21] [1]	Quantifies artifact removal performance	Use multiple metrics for comprehensive evaluation
Computational Frameworks	TensorFlow, PyTorch, EEGLAB [58]	Implements deep learning architectures	Consider GPU compatibility for large models
Traditional Baselines	ICA, Wavelet Transform, ASR [43]	Establishes performance benchmarks	Useful for comparative analysis
Specialized Architectures	CLEnet, AnEEG, Multi-modular SSM [21] [1] [34]	Addresses specific artifact types	Match architecture to artifact characteristics
Real-time Processing	Lightweight CNN [19], iCanClean [43]	Enables online artifact removal	Optimize for latency and resource constraints

The balance between computational cost and estimation accuracy in EEG artifact removal remains a dynamic research frontier with no universal solution. Current evidence suggests that researchers must carefully consider their specific application requirements when selecting methodologies. For offline analysis where accuracy is paramount, sophisticated architectures like CLEnet and GAN-based approaches deliver superior performance despite higher computational demands [21] [1]. In contrast, real-time applications benefit from lightweight CNNs and specialized tools like iCanClean, which provide satisfactory artifact suppression with significantly lower computational overhead [43] [19]. The emerging trend toward hybrid architectures that combine multiple computational principles represents the most promising direction for optimizing this critical trade-off, potentially enabling new capabilities in both clinical and research settings while maintaining practical computational requirements.

Electroencephalogram (EEG) is a fundamental tool for analyzing brain activity in research and clinical applications, but its signals are frequently contaminated by artifacts originating from ocular movements (EOG), muscle activity (EMG), cardiac activity (ECG), and motion [10] [3] [17]. These artifacts can severely bias the interpretation of neural data, making their removal a critical preprocessing step. The expansion of wearable EEG systems into real-world, non-clinical domains—such as wellness tracking, neurofeedback, and brain-computer interfaces (BCIs)—has intensified the need for robust, automated artifact handling pipelines [17].

Within this context, ensemble methods and threshold-based techniques have emerged as powerful approaches for artifact management. Ensemble learning, which combines multiple models to improve overall predictive performance and stability, presents a key tuning parameter: the number of base learners (ensemble size) [59]. Similarly, many artifact detection algorithms rely on setting thresholds to differentiate between neural signals and contaminants [60] [17]. The optimal configuration of these parameters—ensemble sizes and thresholds—is not universal; it depends on the artifact type, data complexity, and available computational resources. This guide provides a comparative analysis of these key parameters, framing the discussion within the methodology of validating EEG artifact removal using simulated data.

Comparative Analysis of Ensemble Methods and Thresholding Techniques

Performance and Computational Trade-offs in Ensemble Learning

Ensemble complexity, defined as the number of base learners, directly influences algorithm performance, time cost, and computational resource consumption [59]. A 2025 comparative analysis quantified these trade-offs for Bagging and Boosting, two core ensemble algorithms.

Table 1: Performance and Cost of Bagging vs. Boosting at Different Ensemble Complexities (MNIST Dataset)

Ensemble Complexity (Number of Base Learners)	Bagging Performance (Accuracy)	Boosting Performance (Accuracy)	Relative Computational Time (Boosting vs. Bagging)
20	0.932	0.930	~14x
200	0.933 (plateaus)	0.961	~14x

The data reveals a critical divergence: as ensemble complexity increases, Boosting's performance improves significantly but at a substantially higher computational cost—approximately 14 times longer than Bagging for a similar number of base learners [59]. Bagging exhibits diminishing returns, with performance quickly plateauing, making it a more cost-effective choice for applications where computational resources or time are constrained. For researchers prioritizing maximum accuracy and who have sufficient computational resources, Boosting with a larger ensemble size is more beneficial, though it risks overfitting at very high complexities [59].

Thresholding in Signal Processing-Based Artifact Removal

Thresholding is a fundamental decision rule in many artifact detection pipelines. Its practical application is illustrated in a study on motion artifact removal using Singular Spectrum Analysis (SSA). The researchers identified artifact components by setting a threshold of 0.1 for the local mobility of eigenvectors, a signal complexity measure where lower values correspond to higher artifact probability [60]. This threshold was used to group artifact-related components, which were then subtracted from the contaminated signal. The study reported an improvement of 0.92 dB in Signal-to-Noise Ratio (SNR) and an 11.39% improvement in the percentage reduction in artifact, validating the effectiveness of the chosen threshold [60].

Another approach involves multi-metric thresholding. A single-channel EOG removal method used a combination of kurtosis (KS), dispersion entropy (DisEn), and power spectral density (PSD) metrics to automatically identify artifact-laden components after signal decomposition [61]. This multi-faceted thresholding strategy reduces reliance on manual intervention and increases the robustness of artifact detection.

Experimental Protocols for Parameter Validation

Protocol 1: Validating Ensemble Size for EEG Artifact Classification

A rigorous protocol for evaluating ensemble size involves using publicly available benchmark datasets and a structured experimental workflow [59] [3].

Table 2: Key Research Reagents and Computational Tools

Item Name	Function in Experimental Protocol
EEGdenoiseNet [3]	A benchmark dataset providing clean EEG segments and artificially contaminated EEG with known artifacts (EOG, EMG).
Synthetic EEG Data [61]	Simulated EEG signals with precisely controlled artifact injections, enabling ground-truth validation of removal algorithms.
Visibility Graph (VG) Features [2]	A method to transform EEG time series into graph structures, providing features that improve model learning on smaller datasets.
SMOTE (Synthetic Minority Oversampling) [62]	A class balancing technique used to address imbalanced datasets, mitigating bias against minority groups.

Workflow Description:

• Data Preparation: Select an appropriate dataset (e.g., EEGdenoiseNet) or generate simulated data. Partition the data into training, validation, and test sets.
• Model Training: Train multiple ensemble models (e.g., Bagging and Boosting), varying the number of base learners (e.g., from 20 to 200).
• Performance Evaluation: On the validation set, record performance metrics (e.g., accuracy, Correlation Coefficient) and computational costs (time and memory) for each ensemble size.
• Threshold Determination: Identify the "knee-of-the-curve"—the ensemble size beyond which performance gains become marginal relative to the increasing cost.
• Final Validation: Apply the optimally-sized ensemble model to the held-out test set to obtain an unbiased performance estimate.

The diagram below illustrates this structured workflow.

Protocol 2: Optimizing Thresholds for Artifact Component Identification

Threshold tuning is critical for methods that separate artifacts in a decomposed signal space, such as SSA or Empirical Wavelet Transform (EWT).

Workflow Description:

• Signal Decomposition: Apply a decomposition algorithm (e.g., SSA, EWT, EMD) to the contaminated EEG signal to break it into constituent components (e.g., eigenvectors or intrinsic mode functions).
• Feature Extraction: Calculate relevant features (e.g., local mobility [60], kurtosis, dispersion entropy [61]) for each decomposed component.
• Threshold Application & Component Classification: Apply a predetermined threshold to the calculated features to classify each component as either an "artifact" or "neural signal."
• Signal Reconstruction: Reconstruct the cleaned EEG signal using only the components classified as "neural signal."
• Performance Quantification: Compare the reconstructed signal against a ground-truth clean signal using metrics like Signal-to-Noise Ratio (SNR), Correlation Coefficient (CC), and Mean Absolute Error (MAE) [61] [60] [2].

The iterative process of tuning the threshold parameter to maximize these performance metrics is key to the protocol.

Discussion and Implementation Guidelines

The choice between complex ensemble models and simpler threshold-based methods, and the tuning of their respective parameters, hinges on the specific research context.

• For High-Accuracy Demands with Ample Resources: Boosting ensembles with a larger number of base learners (e.g., 100-200) are recommended, provided computational costs are acceptable and measures are taken to prevent overfitting [59].
• For Resource-Constrained or Real-Time Applications: Bagging with smaller ensemble sizes or efficient threshold-based methods like modified SSA offer a better balance of performance and speed. The modified SSA technique, for instance, achieved a significant artifact reduction with approximately six times lower computational complexity than a comparable EEMD-CCA method [60].
• For Single-Channel or Wearable EEG: Thresholding techniques integrated with signal decomposition are particularly suitable due to their lower computational footprint and adaptability to low-channel-count configurations [61] [60] [17]. The emergence of deep learning models like A²DM [3] and Motion-Net [2], which can learn complex, non-linear thresholds from data, presents a powerful alternative. These models can automatically identify artifact types and apply tailored removal strategies, sometimes using hard attention mechanisms as an adaptive threshold [3].

In conclusion, there is no one-size-fits-all parameter for EEG artifact removal. Validation using simulated data is crucial, as it provides the ground truth necessary to objectively compare the performance of different ensemble sizes and threshold values. Researchers must carefully consider the trade-offs between performance, computational cost, and the specific constraints of their application domain when tuning these key parameters.

Addressing the Challenge of Over-cleaning and Signal Loss

In electroencephalography (EEG) research, the removal of artifacts—unwanted noise from sources like eye movements, muscle activity, or motion—is a critical preprocessing step. However, an overly aggressive approach to artifact removal can lead to over-cleaning, a significant problem where genuine neural signals are inadvertently subtracted or distorted alongside the noise. This signal loss compromises data integrity, potentially excising neurophysiologically relevant information and biasing experimental results. The central challenge lies in optimizing artifact removal techniques to maximize noise reduction while minimizing the impact on the underlying brain signals of interest. This guide objectively compares the performance of modern artifact removal methods, leveraging simulated data research to provide a controlled framework for quantifying the trade-off between cleanliness and signal preservation.

Quantitative Comparison of Artifact Removal Methods

The following tables synthesize performance data from recent studies, providing a direct comparison of various artifact removal techniques based on key metrics such as Signal-to-Noise Ratio (SNR), artifact reduction percentage, and Root Mean Square Error (RMSE). These metrics are crucial for evaluating a method's effectiveness; higher SNR and artifact reduction indicate better performance, while lower RMSE values suggest superior signal preservation.

Table 1: Performance Comparison of Deep Learning-Based Methods

Method	Architecture	Key Performance Metrics	Best For	Signal Loss Risk
Motion-Net [2]	1D CNN (U-Net)	Artifact reduction (η): 86% ±4.13; SNR improvement: 20 ±4.47 dB; MAE: 0.20 ±0.16	Subject-specific motion artifact removal	Low (Subject-specific training)
AnEEG [21]	LSTM-based GAN	Improved SNR and SAR; Lower NMSE and RMSE vs. wavelet techniques	General artifact removal (Biological & environmental)	Medium (Adversarial training aims to preserve signal)
CLEnet [1]	Dual-scale CNN + LSTM	SNR: 11.50 dB; CC: 0.925; RRMSEt: 0.300 (on mixed artifacts)	Multi-channel EEG; Unknown artifacts	Low (Dual-scale feature extraction)

Table 2: Performance of Other Prevalent Methods

Method	Category	Key Performance Metrics / Findings	Best For	Signal Loss Risk
iCanClean [43]	Reference-based (CCA)	Produced most dipolar ICs; Recovered P300 congruency effect during running	Motion artifacts in locomotion studies	Low (Leverages noise references)
Artifact Subspace Reconstruction (ASR) [43]	Statistical (PCA)	Reduced power at gait frequency; Recovered ERP components (but weaker P300 effect)	Real-time correction; High-amplitude artifact removal	Medium-High (Depends on 'k' parameter setting)
ICA + SPHARA [18]	Hybrid (Source Separation + Spatial Filter)	SNR: 5.56 dB; SD: 6.15 μV (Improved over baseline)	Dry EEG; Movement artifacts	Low (Complementary techniques)
Independent Component Analysis (ICA) [63] [64]	Blind Source Separation	More sensitive detection of small non-brain artifacts than raw scalp data	Ocular, cardiac, and noise artifacts	Medium (Requires correct component identification)

Detailed Experimental Protocols and Methodologies

To critically evaluate and reproduce the findings on artifact removal, a clear understanding of the experimental protocols is essential. This section details the methodologies behind key experiments cited in the comparison tables.

Motion-Net: A Subject-Specific Deep Learning Approach

Objective: To develop a subject-specific convolutional neural network (CNN) for removing motion artifacts from EEG signals using relatively small datasets [2].

• Data Acquisition & Preprocessing: Real EEG recordings with ground-truth references were used. Data was synchronized with accelerometer readings to identify motion artifact peaks. Preprocessing included resampling and baseline correction using a fitted polynomial, which increased the Pearson correlation between motion artifact and ground truth signals from 0.52 to 0.80 in clean signal segments [2].
• Model Architecture & Training: The model was based on a 1D U-Net CNN architecture. A key innovation was the incorporation of Visibility Graph (VG) features, which convert time-series signals into graph structures, providing enhanced structural information that improves model accuracy on smaller datasets. The model was trained and tested separately for each subject (subject-specific) [2].
• Evaluation Metrics: Performance was quantified using Artifact Reduction Percentage (η), Signal-to-Noise Ratio (SNR) improvement, and Mean Absolute Error (MAE) against a known ground truth [2].

iCanClean vs. ASR for Motion-Rich Paradigms

Objective: To compare the effectiveness of iCanClean and Artifact Subspace Reconstruction (ASR) in removing motion artifacts during overground running and recovering stimulus-locked Event-Related Potentials (ERPs) [43].

• Experimental Paradigm: Young adults performed a Flanker task (a cognitive conflict task) under two conditions: while jogging on a treadmill and during static standing. The standing task served as a motion-free benchmark for ERP comparison [43].
• Preprocessing Protocols:
  • iCanClean: This method employs Canonical Correlation Analysis (CCA) to identify and subtract noise subspaces from the scalp EEG. In the absence of dedicated noise sensors, it can create "pseudo-reference" noise signals by applying a notch filter (e.g., below 3 Hz) to the raw EEG. A user-defined R² threshold (e.g., 0.65) determines which correlated noise components are removed [43].
  • ASR: This method uses a sliding-window Principal Component Analysis (PCA). It first calibrates on a "clean" reference period from the data (automatically selected). Then, in new data, it identifies and removes components whose variance exceeds a standard deviation threshold (the "k" parameter). A higher k value (e.g., 20-30) is less aggressive, while a lower value (e.g., 10) is more aggressive and risks over-cleaning [43].
• Outcome Measures:
  • ICA Dipolarity: The number of brain-like independent components obtained after preprocessing.
  • Spectral Power: Reduction in power at the gait frequency and its harmonics.
  • ERP Analysis: Ability to recover the expected P300 ERP component and its congruency effect (larger amplitude for incongruent stimuli) during running [43].

Hybrid Spatial-Temporal Filtering for Dry EEG

Objective: To investigate if combining temporal/statistical and spatial denoising techniques improves signal quality in dry EEG, which is particularly prone to movement artifacts [18].

• Methods Combined:
  • Fingerprint + ARCI: An ICA-based pipeline for automatically identifying and removing physiological artifacts (e.g., eye blinks, muscle activity, cardiac interference).
  • SPHARA (Spatial Harmonic Analysis): A spatial filtering method that acts as a generalized Laplacian denoising technique, improving the signal-to-noise ratio across the sensor array. An improved version involved an additional step of zeroing artifactual jumps in single channels before application [18].
• Experimental Procedure: Dry 64-channel EEG was recorded from healthy volunteers during a motor execution paradigm (hand, feet, and tongue movements). The signal quality after applying each method individually and in combination was assessed using Standard Deviation (SD), Signal-to-Noise Ratio (SNR), and Root Mean Square Deviation (RMSD) [18].

Visualization of Methodologies and Workflows

The following diagrams illustrate the logical workflows and key decision points for two prominent classes of artifact removal strategies, highlighting how they aim to mitigate over-cleaning.

The Scientist's Toolkit: Research Reagent Solutions

This table details key hardware, software, and data resources essential for conducting rigorous validation experiments in EEG artifact removal, particularly those involving simulated data.

Table 3: Essential Research Tools for Artifact Removal Validation

Tool Name / Type	Function in Research	Application Context
Dual-Layer EEG Setup [43]	Provides a dedicated noise reference; electrodes in the upper layer are mechanically coupled but not in contact with the scalp, capturing only motion-induced noise.	Critical for optimal performance of reference-based methods like iCanClean in real-world motion studies.
Carbon-Wire Loops (CWL) [65]	Act as isolated reference sensors placed on the scalp to exclusively record MR-induced artifacts during simultaneous EEG-fMRI.	Superior reduction of imaging and ballistocardiogram (BCG) artifacts in EEG-fMRI research.
Accelerometers (Acc) [2]	Synchronized motion recording to objectively identify and time-lock motion artifact events within the EEG data stream.	Validating and training subject-specific models like Motion-Net; quantifying motion-artifact correlations.
eego / waveguard Touch [18]	High-density dry EEG systems with PU/Ag/AgCl electrodes enabling rapid setup and recording in ecological scenarios.	Studying artifact removal performance in dry EEG systems, which are more susceptible to motion artifacts.
EEGdenoiseNet [1]	A publicly available semi-synthetic benchmark dataset containing clean EEG, EOG, and EMG signals for controlled mixing.	Standardized training and fair comparison of deep learning models for ocular and muscular artifact removal.
ICLabel [43]	A trained classifier that automatically labels independent components from ICA decomposition as brain or various artifact types.	Automating the component rejection step in ICA, though requires caution with mobile EEG data it was not trained on.
Simulated Data Pipelines [63] [64]	Frameworks for adding simulated artifacts (e.g., eye blinks, muscle noise) to clean, ground-truth EEG data.	Essential for quantifying method performance and signal loss, as the true underlying brain signal is known.

The challenge of over-cleaning and signal loss in EEG artifact removal necessitates a careful, method-specific approach. Quantitative comparisons reveal that while modern deep learning methods like Motion-Net and CLEnet offer high performance with low signal loss, their success often depends on subject-specific training or computational resources. Hybrid approaches, such as ICA combined with SPHARA, demonstrate that combining complementary techniques can yield superior results for challenging data like dry EEG. For motion-heavy paradigms, iCanClean shows a slight edge over ASR in preserving cognitive ERPs, though ASR remains a robust, real-time capable tool when configured with an appropriate k parameter.

Ultimately, the selection of an artifact removal method should be guided by the specific artifact type, EEG hardware, and experimental paradigm. The use of simulated data and benchmark datasets is non-negotiable for rigorous, quantitative validation of any pipeline, ensuring that the pursuit of clean data does not come at the cost of losing the genuine neural signals researchers seek to understand.

Selecting Optimal Pre-processing Steps for Functional Connectivity

Functional connectivity (FC) analysis measures the statistical dependencies between neurophysiological time series to infer brain network organization. Unlike structural connectivity, FC is a statistical construct with no direct physical ground truth, making its estimation highly dependent on methodological choices [66]. Pre-processing serves as the foundational step that transforms raw, artifact-laden neuroimaging data into clean signals suitable for FC estimation. The selection of pre-processing steps directly controls the balance between preserving genuine neural information and removing confounding noise, ultimately determining the biological validity and reproducibility of connectivity findings.

Within electroencephalography (FC-EEG) research, the absence of ground truth in real data presents a fundamental validation challenge. Simulated EEG data with known source configurations and connectivity patterns provides the essential ground truth required to objectively benchmark pre-processing pipelines [33] [30]. This guide leverages evidence from simulation-based validation studies to compare the performance of alternative pre-processing strategies, providing researchers with evidence-based recommendations for optimizing FC analysis in both basic neuroscience and clinical drug development applications.

Comparative Analysis of Artifact Removal Methods

Artifact contamination represents a major threat to FC-EEG validity. Different removal strategies offer distinct trade-offs between artifact suppression and signal preservation, with performance varying significantly across artifact types.

Table 1: Performance Comparison of EEG Artifact Removal Methods

Method	Key Principle	Best For Artifact Type	Performance Advantages	Limitations & Requirements
Regression Methods [10]	Linear subtraction of artifact templates derived from reference channels (EOG, ECG)	Ocular artifacts	Simple computation, well-established	Requires reference channels; suffers from bidirectional contamination [10]
Blind Source Separation (BSS) [10]	Decomposes data into components statistically independent from neural signals	Multiple coexisting artifacts (EOG, EMG, ECG)	No reference channels needed; handles multiple artifacts	Requires manual component inspection; needs many channels (>20) [10]
Wavelet Transform [10]	Multi-resolution analysis separating signal and artifact in time-frequency domain	Transient muscle artifacts	Effective for non-stationary signals	Complex parameter tuning; can distort neural signals
Deep Learning (CLEnet) [1]	Dual-scale CNN with LSTM and attention mechanism extracts morphological/temporal features	Mixed/unknown artifacts in multi-channel data	SNR: 11.50 dB; CC: 0.925; Adaptive to unknown artifacts [1]	Requires extensive training data; complex architecture
Hybrid Methods [10] [1]	Combines multiple approaches (e.g., BSS + Wavelet)	Complex artifact combinations	Enhanced performance over single methods	Increased computational complexity; parameter optimization challenges

The performance metrics in Table 1 demonstrate that deep learning approaches like CLEnet achieve superior artifact suppression for mixed and unknown artifacts, with a signal-to-noise ratio (SNR) improvement of 11.50 dB and a correlation coefficient (CC) of 0.925 with clean EEG [1]. However, traditional methods remain valuable for targeted applications with specific artifact types and limited computational resources.

Experimental Protocols for Method Validation

Benchmarking Functional Connectivity Methods

A comprehensive benchmarking study evaluated 239 pairwise interaction statistics for FC mapping, revealing substantial variation in network organization depending on the choice of FC metric [66]. The experimental protocol assessed multiple features of resting-state FC using data from N=326 healthy young adults from the Human Connectome Project (HICP). Functional time series were processed using the pyspi package to estimate 239 FC matrices from 49 pairwise interaction measures across 6 statistic families [66].

Key benchmarked features included:

• Hub identification: Weighted degree distributions across brain regions
• Structure-function coupling: Correlation between FC and diffusion MRI-estimated structural connectivity
• Individual fingerprinting: Capacity to differentiate individuals based on FC patterns
• Brain-behavior prediction: Relationship between FC organization and behavioral measures

Results demonstrated that precision-based statistics consistently showed the strongest correspondence with structural connectivity and greatest capacity for individual differentiation [66]. This benchmarking protocol provides researchers with a template for comprehensive FC method evaluation.

Validating Pre-processing Pipeline Effects on Dynamic FC

Research on dynamic functional network connectivity (dFNC) in mild traumatic brain injury (mTBI) systematically evaluated how pre-processing sequence affects downstream analysis [67]. The study utilized a sample cohort of 50 mTBI patients and 50 matched healthy controls. All participants completed a 5-minute resting-state fMRI run using a Siemens Trio scanner with standard parameters (TR=2000ms, TE=29ms, flip angle=75°, FOV=240mm) [67].

The experimental protocol tested four different pre-processing pipelines varying the placement of motion correction steps:

• Pipeline A: Motion regression before spatial smoothing
• Pipeline B: Motion regression after spatial smoothing
• Pipeline C: Motion regression after group-independent component analysis (gICA)
• Pipeline D: Motion regression before smoothing + despiking after gICA

Classification accuracy between mTBI and healthy controls was evaluated using a linear support vector machine. Results indicated that Pipeline D (motion regression, smoothing, gICA, then despiking) produced data most suitable for differentiating patient groups, achieving the highest mean classification accuracy [67]. This protocol demonstrates the critical importance of pre-processing sequence for clinical applications.

Figure 1: Comprehensive EEG Pre-processing Workflow. This flowchart outlines the sequential steps in a standard EEG pre-processing pipeline, with the artifact removal stage offering multiple methodological pathways.

Simulation-Based Validation Framework

Synthetic Data Generation with Known Ground Truth

Simulated EEG data provides the essential ground truth required to validate pre-processing pipelines and FC estimation methods. Several sophisticated simulation tools have been developed for this purpose:

EEGSourceSim generates biologically plausible EEG data by embedding signal and noise into MRI-based forward models that incorporate individual-subject variability in structure and function [33]. The framework includes pipelines for evaluating source estimation, functional connectivity, and spatial filtering methods. EEGSourceSim provides 23 individual-participant Boundary Element Method forward matrices with corresponding cortical surface meshes, offering greater realism through individual head models and anatomically-defined regions of interest [33].

SEREEGA (Simulating Event-Related EEG Activity) is a modular, open-source MATLAB toolbox that simulates epochs of EEG data by solving the forward problem of EEG [30]. The toolbox supports multiple publicly available head models and can simulate various signal types mimicking brain activity, including noise, oscillations, event-related potentials, and connectivity patterns. The fundamental equation is: x = As + ε, where x represents the simulated scalp signal, A is the lead field matrix, s represents source activities, and ε is measurement noise [30].

Simulated EEG Data Generator from the MRC BNDU provides MATLAB functions generating simulated EEG data according to different theories of event-related potentials [68]. The implementation includes both the "classical theory" (where ERPs reflect phasic bursts of activity added to ongoing EEG) and "phase-resetting theory" (where events reset the phase of ongoing oscillations) [68].

Validation Metrics for Performance Quantification

Simulation frameworks enable quantitative evaluation using standardized metrics:

• Source Estimation Accuracy: Localization error between simulated and reconstructed sources
• Connectivity Precision: Deviation between simulated and estimated connectivity patterns
• Signal-to-Noise Ratio (SNR): Improvement in signal quality after pre-processing
• Correlation Coefficient (CC): Similarity between processed and ground-truth signals
• Spatial Precision: Accuracy in recovering simulated network topology

For example, in evaluating deep learning approaches, studies report SNR improvements up to 11.50 dB and correlation coefficients of 0.925 with clean EEG benchmarks [1].

Figure 2: Simulation-Based Validation Workflow. This diagram illustrates the methodology for validating pre-processing pipelines using simulated EEG data with known ground truth.

Table 2: Essential Tools for FC-EEG Pre-processing Research

Tool Name	Type/Format	Primary Function	Application in Research
EEGSourceSim [33]	MATLAB Toolbox	Realistic EEG simulation with individual head models	Validation of source estimation, FC measures, and spatial filtering
SEREEGA [30]	MATLAB Toolbox	General-purpose event-related EEG simulation	Benchmarking of analysis pipelines and classification methods
CLEnet [1]	Deep Learning Model	Artifact removal using CNN-LSTM architecture	Elimination of mixed/unknown artifacts in multi-channel EEG
SPM [67] [69]	Software Package	Statistical parametric mapping for neuroimaging	Pre-processing, normalization, and statistical analysis
EEGLAB [68]	MATLAB Toolbox	Interactive EEG processing environment	Implementing ICA-based artifact removal and visualization
pyspi [66]	Python Package	Calculation of 239 pairwise statistics	Benchmarking FC metrics for optimal selection
CONN Toolbox [69]	MATLAB Toolbox	Functional connectivity analysis	Network-based ROI-to-ROI and voxel-to-voxel connectivity

The evidence from simulation-based validation studies indicates that optimal pre-processing for functional connectivity requires careful method selection tailored to specific research goals:

• For standardized artifact removal, blind source separation methods like ICA provide a balance of effectiveness and interpretability, particularly when dealing with multiple coexisting artifacts [10].

• For complex or unknown artifacts in multi-channel data, deep learning approaches like CLEnet offer superior performance, with demonstrated SNR improvements of 2.45-5.13% over traditional methods [1].

• For FC metric selection, precision-based statistics consistently show stronger structure-function coupling and better individual differentiation, though the optimal choice depends on the specific neurophysiological mechanism under investigation [66].

• For pipeline validation, simulation tools like EEGSourceSim and SEREEGA provide essential ground truth for objective performance quantification, enabling evidence-based method selection rather than reliance on convention [33] [30].

Future methodological development should focus on optimizing pre-processing sequences for dynamic FC analysis, adapting deep learning approaches for limited-data scenarios, and establishing standardized validation frameworks using simulated data with known ground truth.

Specialized electrode designs represent a critical frontier in electroencephalography (EEG) research, directly influencing data quality and the efficacy of subsequent artifact removal pipelines. Within validation studies utilizing simulated data, the choice of electrode technology—spanning traditional wet, modern dry, and concealed around-the-ear systems—establishes the foundation upon which noise and neural signals are captured. This guide provides an objective comparison of these technologies, supported by experimental data, to inform researchers and drug development professionals in selecting appropriate hardware solutions for robust EEG artifact removal research.

Electrode Technologies: A Comparative Analysis

Electrode design primarily differs in its interface with the scalp, balancing signal integrity against practical application requirements in research settings.

Table 1: Comparison of Primary EEG Electrode Technologies

Electrode Type	Scalp Interface	Key Advantages	Key Limitations	Best-Suited Research Contexts
Wet Active Electrodes	Conductive gel [70]	High signal quality; Gold standard for low-noise data [71] [70]	Lengthy setup; Requires expertise and cleaning [70]	Clinical studies; High-fidelity lab research [72]
Dry Active Electrodes	Metal pins or silicon-based contact [70]	Rapid setup; No skin preparation or gel [70]	Higher impedance with movement; Potential discomfort [70]	Field studies; Longitudinal monitoring; Populations sensitive to gel
Concealed Around-the-Ear (cEEGrid)	Flexible, gel-filled adhesive array [71]	High comfort; Minimal visibility; Suitable for long recordings [71]	Captures only a subset of neural information [71]	Real-world, mobile brain-computer interfaces (BCIs) [71]

Performance Evaluation: Experimental Data

Direct comparisons of these systems quantify performance trade-offs in signal quality, subject comfort, and application speed.

Table 2: Experimental Performance Metrics Across Electrode Systems

Performance Metric	Wet System (Biosemi ActiveTwo)	Dry System (F1)	Concealed System (cEEGrid)	Experimental Context
Resting-State Spectral Power (Theta, Alpha, Beta)	Reference standard	No significant difference from wet system [70]	N/A	Resting periods with eyes open/closed [70]
P3b Event-Related Potential (ERP)	Robust amplitude and topography	Comparable amplitude/topography; High correlation with wet (r=0.54-0.89) [70]	N/A	Visual target detection task [70]
Single-Trial Classification Accuracy	High performance	Marginally lower than wet system, but well above chance [70]	Suitable for frequency-domain features (e.g., Berger effect) [71]	Rapid Serial Visual Presentation (RSVP) task [70]
Subject Comfort & Application Speed	Lengthy setup [70]	Swift application; High comfort [70]	~5 minute application; Comfortable for hours [71]	User experience reports [71] [70]

Methodologies for Hardware Validation

Robust validation of EEG hardware relies on standardized experimental protocols to assess system performance across multiple domains.

Protocol 1: Direct System Comparison in Lab vs. Community Settings

A within-subjects design directly compares portable community-based EEG with traditional lab-based systems [72].

• Participants: A developmentally diverse group, such as children under four years, to test robustness [72].
• Procedure: Conduct two EEG sessions per participant (lab and community) within a short timeframe (e.g., 30 days). Standardize the recording procedure, aiming for five minutes of continuous, task-free EEG in both settings [72].
• Data Analysis: Compare data retention rates, noise levels, and spectral power measures (e.g., Delta, Theta, Alpha, Beta) across settings. Calculate Intraclass Correlation Coefficients (ICCs) for spectral power to assess individual-level consistency [72].

This protocol assesses an electrode system's capability to record well-established neural responses [70].

• Task Paradigm: Participants complete a Rapid Serial Visual Presentation (RSVP) target detection task designed to elicit a P3b event-related potential component [70].
• Measures: Quantify and compare the P3b amplitude and topography across systems. Perform single-trial classification analysis using machine learning to differentiate target from non-target trials [70].
• Validation: High correlations in P3b metrics (r = 0.54–0.89) and above-chance classification accuracy confirm a system's validity [70].

Experimental Workflow for EEG Hardware Validation

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Materials for EEG Hardware and Artifact Research

Item	Function/Description	Example Products/Brands
High-Density Wet EEG System	Gold-standard reference system for laboratory recordings with high channel count.	HydroCel Geodesic Sensor Net (EGI), ActiveTwo (Biosemi), BrainProducts actiCAP [72] [70]
Portable Amplifier with Active Electrodes	Mobile system with integrated amplification at the electrode for improved signal-to-noise ratio in field settings.	BrainVision LiveAmp with actiCAP slim, Smarting Mobi (MBrainTrain) [71] [72]
Low-Cost Open-Source Amplifier	Budget-friendly, customizable platform suitable for concealed EEG and proof-of-concept studies.	OpenBCI Cyton+Daisy boards [71]
Concealed EEG Electrode Array	Flexible, c-shaped electrode array for discrete, long-term recordings around the ear.	cEEGrid [71]
Artifact Removal Algorithm	Software pipeline for automatic identification and removal of ocular, muscle, and movement artifacts.	Artifact Subspace Reconstruction (ASR), NEAR pipeline for newborns, AnEEG deep learning model [15] [21]
Experimental Presentation Software	Software for designing and presenting standardized cognitive tasks (e.g., RSVP) to elicit neural responses.	Custom MATLAB/Python scripts, Presentation, PsychoPy [70]

The selection of specialized electrode designs is a fundamental determinant in the success of EEG artifact removal research. Wet electrode systems remain the benchmark for signal quality in controlled laboratory studies. In contrast, dry and concealed electrode designs offer compelling advantages for ecological validity and scalability, with experimental data confirming their capability to capture robust neural signals despite different operational constraints. The choice of hardware must therefore align with the specific research priorities—whether ultimate signal fidelity, participant comfort, or real-world applicability—to ensure the validity and impact of the subsequent artifact removal processes.

Benchmarking Performance: Metrics and Comparative Analysis

The validation of electroencephalogram (EEG) artifact removal algorithms represents a critical challenge in computational physiology and biomedical signal processing. As research in this field increasingly relies on simulated data to benchmark new methodologies, the selection of robust, informative, and standardized validation metrics has never been more important. This guide provides a comparative analysis of three key metric categories—Root Relative Squared Error (RRMSE), Correlation, and the emerging concept of Dipolarity—framed within the context of EEG artifact removal. We objectively compare their performance, synthesize experimental data from seminal studies, and provide detailed protocols to equip researchers with the tools necessary for rigorous algorithm evaluation.

EEG signals, which reflect the brain's electrical activity, are routinely contaminated by physiological artifacts such as electromyogram (EMG) from muscle activity, electrooculogram (EOG) from eye movements, and electrocardiogram (ECG) from heartbeats [73] [74]. These artifacts can severely mask neural signals, complicaining interpretation and leading to erroneous conclusions in both clinical and research settings. The development of effective artifact removal techniques is therefore a cornerstone of reliable EEG analysis.

The complexity of this task has led to the proliferation of diverse signal processing approaches, including Blind Source Separation (BSS) methods like Independent Component Analysis (ICA) and Second-Order Blind Identification (SOBI), as well as techniques combining singular spectrum analysis (SSA) with Canonical Correlation Analysis (CCA) [73] [74]. With numerous available algorithms, a fundamental question arises: how does one objectively determine which method performs best? The answer lies in the consistent application of robust validation metrics. In studies using simulated data, where the "ground truth" clean EEG is known, these metrics quantitatively assess how well an algorithm can separate artifact from neural signal. This guide focuses on three such metrics, evaluating their theoretical basis, practical application, and comparative performance to establish a framework for validation in EEG artifact removal research.

A Comparative Analysis of Key Validation Metrics

The evaluation of regression models and signal processing algorithms, including those for EEG artifact removal, relies on metrics that quantify the difference between predicted and actual values. The table below summarizes the core metrics relevant to this field.

Table 1: Essential Validation Metrics for Regression and Signal Processing Models

Metric	Full Name	Formula	Key Interpretation	Advantages	Disadvantages
MSE [75] [76]	Mean Squared Error	MSE = (1/n) * Σ(Ŷᵢ - Yᵢ)²	Average of squared errors. Lower values indicate better fit.	Differentiable, emphasizes larger errors.	Sensitive to outliers, unit is squared.
RMSE [75] [76]	Root Mean Squared Error	RMSE = √MSE	Square root of MSE. Lower values are better.	Interpretable in the variable's original units.	Sensitive to outliers.
MAE [75] [77]	Mean Absolute Error	MAE = (1/n) * Σ⎮Ŷᵢ - Yᵢ⎮	Average magnitude of errors. Lower values are better.	Robust to outliers, easily interpretable.	Does not penalize large errors severely.
R [77] [76]	Pearson Correlation Coefficient	R = cov(X,Y) / (σₓσᵧ)	Strength and direction of a linear relationship. -1 to +1.	Measures linear relationship strength.	Only captures linearity, not agreement.
R² [75] [76]	Coefficient of Determination	R² = 1 - (SS_res / SS_tot)	Proportion of variance explained. 0 to 1.	Intuitive interpretation of explained variance.	Misleading with multiple predictors if not adjusted.
RRMSE [78]	Root Relative Squared Error	RRMSE = √[ Σ(Ŷᵢ - Yᵢ)² / Σ(Yᵢ - Ȳ)² ]	RMSE relative to the data's variance. Lower values are better.	Unitless, allows cross-dataset comparison.	Less intuitive than RMSE or MAE.

Deep Dive on RRMSE, Correlation, and Dipolarity

• RRMSE (Root Relative Squared Error): This metric is a normalized version of RMSE, which makes it particularly valuable for comparing model performance across different datasets or signals with varying scales [78]. A lower RRMSE indicates a smaller error relative to the simple variance of the actual data. In the context of EEG artifact removal, an RRMSE value of 0.1 would mean that the error from the reconstruction is 10% of the variability inherent in the original clean signal. This normalization is crucial when validating algorithms on multiple subjects or simulated datasets with different signal amplitudes.

• Correlation (R and R²): The Pearson Correlation Coefficient (R) measures the strength and direction of a linear relationship between the clean ground-truth signal and the artifact-cleaned signal [77] [76]. Its value ranges from -1 (perfect negative correlation) to +1 (perfect positive correlation). A value of 0 suggests no linear relationship. In EEG validation, a high positive R is desired, indicating that the temporal dynamics of the cleaned signal closely follow those of the true neural data. The Coefficient of Determination (R²) goes a step further, representing the proportion of variance in the clean signal that is explained by the cleaned signal [75] [76]. For example, an R² of 0.9 means the artifact removal algorithm explains 90% of the variance in the ground-truth EEG. It is critical to remember that a high correlation does not necessarily mean low error; a model can be systematically biased (consistently over- or under-predicting) and still have a high R value [77].

• Dipolarity: While not a traditional numerical metric like RRMSE, dipolarity is a concept used in source separation techniques like ICA for validating neural-derived components. A core assumption of ICA is that brain sources have a dipolar field pattern, meaning their electrical fields propagate volumetrically from a single origin in the brain to the scalp in a predictable pattern. When an ICA component is reconstructed as a scalp map, researchers can assess whether its topography matches this physical expectation of a dipolar field. Components that do not exhibit a clear dipolar pattern are more likely to be non-neural artifacts (e.g., from muscle, eye, or line noise). Thus, dipolarity serves as a physically-grounded, qualitative validation metric to determine if a separated component is neurologically plausible.

Experimental Protocols for Metric Validation in EEG Studies

Validating artifact removal algorithms requires a structured approach, often employing simulated or semi-simulated data where the true, artifact-free signal is known. The following workflow and subsequent case studies illustrate how these metrics are applied in practice.

Diagram 1: EEG Artifact Removal Validation Workflow

Case Study 1: SSA-CCA for EMG Artifact Removal

A 2019 study introduced a novel method combining Singular Spectrum Analysis (SSA) with Canonical Correlation Analysis (CCA) to remove muscle (EMG) artifacts from multichannel EEG [73].

• Experimental Protocol: The researchers used a semi-simulated data approach. First, they collected real EEG data during rest to obtain clean neural signals. Then, they artificially added EMG artifacts recorded from facial and neck muscles to this clean EEG at varying signal-to-noise ratios (SNRs). This created a dataset where the ground-truth, clean EEG was known, enabling precise validation. The SSA-CCA method was applied, which involves using SSA to decompose each channel and then applying CCA for further noise reduction. The performance was benchmarked against established techniques like standard CCA and EEMD-CCA.

• Supporting Data and Metric Performance: The study used correlation-based metrics to demonstrate superiority. The results showed that SSA-CCA outperformed EEMD-CCA and classic CCA under multichannel circumstances. While the specific RRMSE values were not reported, the framework of using semi-simulated data and quantitative metrics like correlation is the standard for the field.

Case Study 2: A Multi-Algorithm Approach with SOBI and SWT

Another robust study published in 2022 presented a method integrating Second-Order Blind Identification (SOBI) with Stationary Wavelet Transform (SWT) and machine learning classifiers [74].

• Experimental Protocol: This research also leveraged simulated, semi-simulated, and real EEG data for comprehensive validation. The protocol was as follows:

  • Source Separation: The SOBI algorithm was applied to the multichannel EEG data to decompose it into independent source components.
  • Automatic Classification: A machine learning classifier (e.g., MLP, KNN) was trained on features derived from the phase space of these components to automatically identify which ones were artifactual.
  • Artifact Removal & Reconstruction: Instead of simply discarding artifact components, the SWT was used to suppress the artifact within them, preserving underlying neural information. The components were then reconstructed into clean EEG channels.

• Supporting Data and Metric Performance: The study employed error-based metrics for validation. On the reconstructed EEG signals, the method achieved a very low Mean Square Error (MSE) of about 2% when compared to the known ground truth [74]. This low MSE (and by extension, a low RMSE and RRMSE) quantitatively confirms the high accuracy of the reconstruction, demonstrating the effectiveness of the combined SOBI-SWT approach.

The Scientist's Toolkit: Key Reagents & Computational Solutions

The following table details essential computational tools and conceptual "reagents" used in the development and validation of EEG artifact removal algorithms.

Table 2: Research Reagent Solutions for EEG Artifact Removal Studies

Research Reagent / Solution	Function / Role in Validation	Relevance to Metrics
Semi-Simulated Datasets [73]	A hybrid of real clean EEG and artificially added artifacts. Provides a known ground truth for rigorous benchmarking.	The foundation for calculating all error (RRMSE, MSE) and correlation (R, R²) metrics.
Blind Source Separation (BSS) [74]	A family of algorithms (e.g., SOBI, ICA) that separate mixed signals into underlying sources without prior information.	Used to generate the components that will be assessed by metrics like dipolarity before being used in reconstruction.
Machine Learning Classifiers [74]	Algorithms (e.g., SVM, KNN, MLP) that automate the identification of artifactual components from neural ones.	Improves objectivity; their accuracy can be validated with metrics (e.g., % accuracy) before being used in the main workflow.
Canonical Correlation Analysis (CCA) [73]	A BSS method that finds relationships between two sets of data. Used to separate sources based on temporal structure.	Often a core part of the algorithm being validated. Its output is judged by the final RRMSE and R of the reconstructed signal.
Wavelet Transform (SWT) [74]	A signal processing technique for multi-resolution analysis. Used to suppress artifacts in specific components without full removal.	Helps minimize final reconstruction error (RRMSE, MAE) by preserving neural data within artifact-labeled components.
Phase-Space Analysis [74]	A nonlinear dynamics method to study the behavior of a system over time. Used to extract features for component classification.	Provides features that help build better classifiers, leading to more accurate artifact removal and improved final validation metrics.

The validation of EEG artifact removal algorithms demands a multi-faceted approach. No single metric provides a complete picture. Based on the experimental data and protocols reviewed, the following guidance emerges:

• Use RRMSE for Overall Accuracy: When you need a normalized, unitless measure of overall reconstruction error that is comparable across different datasets or studies, RRMSE is the most robust choice among error metrics [78].
• Use Correlation for Signal Dynamics: When the primary concern is whether the temporal structure and morphology of the neural signal have been preserved after artifact removal, R and R² are the appropriate metrics. However, always supplement them with an error metric to check for bias [77] [76].
• Use Dipolarity for Source Validation: When using source separation methods like ICA, dipolarity provides a crucial, physiology-based check to confirm the neural origin of a component before it is retained in the signal [74].

In conclusion, a robust validation framework for EEG artifact removal research should concurrently report normalized error metrics like RRMSE, correlation coefficients, and—where applicable—qualitative assessments like dipolarity. This multi-metric strategy, applied to rigorously constructed simulated data, provides the comprehensive evidence needed to advance the field and develop more reliable tools for uncovering the brain's true electrical signature.

Comparative Framework for Traditional and Machine Learning Methods

The analysis of electroencephalography (EEG) signals is fundamental to neuroscience research and clinical diagnostics, but these signals are consistently contaminated by unwanted artifacts originating from ocular, muscular, cardiac, and environmental sources [10]. Effective artifact removal is therefore a critical preprocessing step to ensure the validity of subsequent neural analysis. This guide provides a comparative framework for the two predominant methodological approaches for this task: traditional signal processing techniques and modern machine learning (ML)/deep learning (DL) algorithms. The comparison is contextualized within validation protocols that utilize simulated and semi-simulated EEG data, which provide known ground-truth signals essential for objective performance assessment [79].

The core challenge in evaluating artifact removal techniques on real EEG data is the absence of a known, artifact-free "ground truth" signal for comparison [79]. This limitation is overcome by using semi-simulated datasets, where clean EEG recordings are artificially contaminated with well-characterized artifacts, allowing for precise, quantitative measurement of an algorithm's ability to recover the original neural signal [79] [21]. This framework leverages such experimental paradigms to deliver an objective comparison.

Traditional Artifact Removal Methods

Traditional methods typically rely on well-established statistical and signal processing theories to separate artifacts from neural signals based on their physiological or statistical properties [10].

• Regression Methods: These are among the most traditional techniques, which use reference channels (e.g., EOG for ocular artifacts) to estimate and subtract the artifact contribution from the EEG signal. Performance is limited by the need for clean reference signals and the assumption of a linear relationship between the reference and the contamination [10].
• Blind Source Separation (BSS): Techniques like Independent Component Analysis (ICA) are highly prevalent in the literature [10]. They work by decomposing multi-channel EEG data into statistically independent components, which can be manually or automatically classified as neural or artifactual before reconstructing the signal without the artifact-laden components.
• Filtering and Wavelet Transform: Standard band-pass filters are used to remove noise outside the typical EEG frequency band. For more sophisticated removal, Wavelet Transform decomposes the signal into time-frequency components, allowing for the selective filtering of coefficients associated with artifacts [10] [21].

Machine and Deep Learning Approaches

ML and DL methods learn to model the complex, non-linear relationships between artifact-contaminated and clean EEG signals directly from data, often eliminating the need for manual intervention [21] [1].

• General Workflow: These are typically supervised learning models. They are trained on semi-simulated datasets where pairs of contaminated and clean EEG signals are presented. The model learns a mapping function to output a cleaned signal, and its performance is quantified by how closely it matches the known ground truth [1] [34].
• Key Architectures:
  • Convolutional Neural Networks (CNNs): Excel at extracting spatial and morphological features from EEG signals, making them effective for certain artifact types [1] [2].
  • Long Short-Term Memory (LSTM) Networks: Are adept at modeling the temporal dependencies in EEG time-series data [21] [1].
  • Hybrid Models (e.g., CNN-LSTM): Combine strengths of multiple architectures to simultaneously capture spatial and temporal features, as seen in models like CLEnet [1].
  • Generative Adversarial Networks (GANs): Employ a generator network to produce cleaned EEG and a discriminator to critique its quality, driving the generator to produce increasingly realistic, artifact-free signals [21].

Experimental Protocols for Validation

The quantitative comparisons in this guide are predicated on rigorous experimental protocols that use semi-simulated data for fair and objective benchmarking.

Semi-Simulated Data Generation

A common and robust protocol involves the linear mixing of artifact-free EEG recordings with recorded artifact signals [79] [1]. The standard contamination model is represented by the equation [79]: Contaminated_EEG_i,j = Pure_EEG_i,j + a_j * VEOG + b_j * HEOG where Pure_EEG is the ground-truth signal, VEOG and HEOG are vertical and horizontal EOG artifacts, and a_j, b_j are contamination coefficients calculated for each subject and electrode. This provides a known ground truth for validation.

Performance Metrics

The following metrics, calculated by comparing the algorithm's output to the known Pure_EEG, are standard for evaluating performance [21] [1] [34]:

• Signal-to-Noise Ratio (SNR): Measures the level of desired signal relative to noise (artifact). A higher SNR is better.
• Correlation Coefficient (CC): Quantifies the linear relationship between the cleaned signal and the ground truth. A value closer to 1.0 is better.
• Root Mean Square Error (RMSE) / Relative RMSE (RRMSE): Measures the magnitude of the error between the cleaned and ground-truth signals. Lower values are better.

Experimental Workflow

The following diagram illustrates the standard workflow for training and validating artifact removal models using semi-simulated data.

Quantitative Performance Comparison

The tables below synthesize quantitative results from recent studies to compare the performance of various traditional and deep-learning methods across different artifact types.

Table 1: Performance Comparison on Ocular (EOG) and Muscular (EMG) Artifacts. Results from EEGdenoiseNet benchmark [1].

Method Category	Method Name	Artifact Type	SNR (dB)	Correlation Coefficient (CC)	Temporal RRMSE
Traditional	ICA	EOG	8.452	0.837	0.421
Traditional	Wavelet	EMG	6.443	0.703	0.592
Deep Learning	SimpleCNN	EOG	10.912	0.909	0.319
Deep Learning	NovelCNN	EMG	9.323	0.781	0.492
Deep Learning	CLEnet (Hybrid)	EOG+EMG	11.498	0.925	0.300

Table 2: Performance on Motion Artifacts and Specific Deep Learning Models. Results from Motion-Net and related studies [2] [34].

Method Name	Model Architecture	Artifact Type	Key Metric	Performance Value
Motion-Net	CNN with Visibility Graph	Motion	Artifact Reduction (η)	86% ± 4.13
Motion-Net	CNN with Visibility Graph	Motion	SNR Improvement	20 ± 4.47 dB
AnEEG	LSTM-based GAN	Mixed	Model NMSE	Lower than Wavelet
M4 Network	State Space Model (SSM)	tACS/tRNS	Spectral RRMSE	Best Performance [34]

The Scientist's Toolkit: Research Reagents & Materials

For researchers seeking to implement these methods, the following table details essential "research reagents" – key datasets, algorithms, and computational resources required for experimentation in this field.

Table 3: Essential Research Materials and Resources for EEG Artifact Removal Research.

Item Name	Type	Function / Application	Example / Specification
Semi-Simulated Benchmark Dataset	Data	Provides ground truth for objective validation of algorithms.	EEGdenoiseNet [1], Semi-simulated EEG/EOG [79]
Pre-processed Experimental Data	Data	Validates algorithm performance on real, complex artifacts.	32-channel EEG from cognitive tasks (e.g., 2-back) [1]
Independent Component Analysis (ICA)	Algorithm	Traditional BSS method for decomposing and removing artifacts.	Implementations in EEGLAB, MNE-Python [10]
CLEnet Model	Software (DL)	Hybrid CNN-LSTM model for removing mixed/unknown artifacts.	Dual-scale CNN + LSTM with EMA-1D attention [1]
Motion-Net Model	Software (DL)	Subject-specific CNN for motion artifact removal.	1D U-Net utilizing visibility graph features [2]
GPU/TPU Cluster	Hardware	Accelerates training and inference of deep learning models.	Required for efficient processing of large EEG datasets [80] [81]

Architectural Comparison of Deep Learning Models

The performance differences between deep learning models stem from their underlying architectures. The following diagram contrasts the core structures of two prevalent types of models: a hybrid CNN-LSTM and a pure CNN-based model like a U-Net.

This comparative framework demonstrates a clear paradigm shift in EEG artifact removal. While traditional methods like ICA and regression remain interpretable and effective for specific, well-defined artifacts, deep learning models consistently achieve superior quantitative performance on complex, mixed, and unknown artifacts [1] [2]. The key advantages of DL models include their ability to automatically learn features without manual intervention, their scalability with data, and their robustness in challenging scenarios like mobile EEG [21] [2].

The choice of method, however, is not one-size-fits-all. Traditional methods offer computational efficiency and interpretability, which can be crucial for resource-constrained or highly regulated studies [81]. In contrast, deep learning methods are the tool of choice for applications demanding the highest possible accuracy and for which sufficient computational resources and training data are available. The ongoing development of hybrid models and novel architectures like State Space Models suggests that this is a rapidly advancing field, with future benchmarks likely to show even greater performance gains [1] [34].

The pursuit of robust and clinically valuable brain biomarkers necessitates rigorous validation against physiological gold standards. For non-invasive neuroimaging tools, a biomarker must be both reproducible (reliable) across experiments and laboratories, and accurately measure the underlying neural process of interest (valid) [82] [83]. The metrics of reliability and validity are critical before any biomarker can inform diagnosis or treatment decisions, as a lack of rigor can impede scientific progress and cast doubt on these measurement tools [82]. This guide objectively compares two advanced neuroimaging techniques—Transcranial Magnetic Stimulation combined with Electroencephalography (TMS-EEG) and Magnetic Resonance Spectroscopy (MRS)—in the context of their validation pathways. TMS-EEG measures direct brain responses to a controlled perturbation, known as TMS-evoked potentials (TEPs), to assess cortical excitability and connectivity [84] [85]. In contrast, MRS provides a non-invasive 'window' on biochemical processes within the body by quantifying the concentration of specific metabolites in tissue [86] [87]. This article details their respective experimental protocols, the gold standards used for their validation, and a comparative analysis of their performance as validated by physiological benchmarks.

TMS-EEG: Validating Causal Brain Connectivity

TMS-EEG is a causal neuroimaging tool that couples single pulses of TMS with scalp EEG recording to measure the brain's direct electrophysiological response, known as TMS-evoked potentials (TEPs) [82] [83]. TMS operates on the principle of electromagnetic induction; a high-intensity, rapidly changing current in the stimulation coil generates a magnetic field that penetrates the scalp and skull, inducing a secondary current in the underlying cortical tissue that can depolarize neurons [85]. When synchronized with EEG, this setup allows for the recording of TEPs, which are multiphasic responses lasting about 500 ms and reflecting the functional state of the stimulated brain area and its connected networks [82] [84]. A typical TMS-EEG experiment involves participants wearing a TMS-compatible EEG cap and receiving over 100 single TMS pulses at intervals of at least 3 seconds to avoid carry-over effects [85]. The resulting TEP morphology is characterized by a series of positive and negative peaks at specific latencies (e.g., P30, N45, P60, N100, P180), which are thought to originate from both the directly stimulated site and secondary activations in distributed networks [84].

Experimental Protocols and Validation Gold Standards

A standardized TMS-EEG protocol for assessing cortical excitability, particularly targeting the primary motor cortex (M1), involves the following key steps [84] [85]:

• Participant Preparation: The participant is fitted with a TMS-compatible EEG cap, often using ultra-thin active electrodes (3 mm height) to minimize coil-to-scalp distance and reduce TMS-induced decay artifacts. Electrode impedances are maintained below 5 kΩ for passive electrodes or up to 50 kΩ for active systems to ensure signal quality.
• Neuronavigation Setup: To ensure precise and consistent stimulation targeting, a neuronavigation system is used. This system co-registers the participant's anatomical MRI data with the real-time position of the TMS coil, enabling millimeter-level precision in stimulating the left M1 or other target regions.
• Motor Threshold Determination: The resting motor threshold (RMT) is determined prior to the main experiment. RMT is defined as the minimum TMS intensity required to elicit a motor evoked potential (MEP) of greater than 50 μV in the contralateral hand muscle in at least 5 out of 10 consecutive trials.
• TMS-EEG Recording: During the main experiment, single-pulse TMS is delivered to the left M1 at an intensity of 110-120% of the RMT. A minimum of 100 trials are collected, with an inter-trial interval of 3-5 seconds to allow for the EEG signal to return to baseline. The EEG is typically sampled at a high rate (e.g., 5 kHz) to capture early TEP components.
• Artifact Mitigation: Several procedures are employed to minimize contaminants:
  • Auditory Masking: White noise is played through earphones to mask the TMS click and prevent an auditory evoked potential.
  • Sensory Control: A sham TMS condition (e.g., using a placebo coil or tilting the coil 90 degrees) may be included to control for somatosensory and auditory confounds.
  • Post-Processing: Automated or semi-automated algorithms are applied to remove large TMS pulse artifacts, decay artifacts, eye blinks, muscle activity, and line noise.

The central challenge in TMS-EEG validation is that its proposed gold standard—the Motor Evoked Potential (MEP)—is itself an imperfect comparator. While the MEP is a highly reliable and valid measure of corticospinal excitability with a high signal-to-noise ratio (>1000 μV responses), it only probes the motor pathway [82] [83]. TEPs, which aim to assess non-motor cortical regions, therefore lack a direct, perfect gold standard for validation. Consequently, the field employs a multi-pronged validation strategy:

• Invasive Recordings: The validity of TEP components is assessed by comparing them to simultaneous intracranial recordings in patients, which serve as a ground truth for local cortical activation [83].
• Pharmacological Interventions: Administering drugs that target specific neurotransmitter systems (e.g., GABAergic drugs) helps link TEP components to known neurophysiological mechanisms [85].
• Clinical Correlation: TEP metrics are validated by demonstrating their sensitivity to pathological states (e.g., schizophrenia, major depression) and their change in response to effective treatment [82] [85].

Table 1: Key TMS-EEG Validation Metrics and Performance Data

Validation Metric	Description	Typical Experimental Readout	Performance Data
Internal Reliability	Consistency of TEPs within a laboratory and experimental setup [82].	Test-retest reliability of TEP components (e.g., N45, P60) over hours or weeks.	Concordance Correlation Coefficient (CCC) of 0.92-0.97 for early TEPs (15-80 ms) [84].
External Reliability	Generalizability of TEPs across different labs, setups, and operators [82].	Reproducibility of TEP morphology and amplitude between laboratories.	Lower than internal reliability; highly dependent on standardized protocols [82] [83].
Construct Validity	Closeness to the true underlying neural signal [82].	Correlation with invasive neural recordings or pharmacological modulation.	Early TEP components (<50 ms) show higher validity for local excitability but are confounded by noise [82].
Amplitude Stability	Convergence of TEP average with increasing trial count.	Number of trials required for a stable TEP average.	Stable waveforms achieved after 20-30 trials [84].

MRS: Validating Metabolic Biomarkers

Magnetic Resonance Spectroscopy (MRS) is a non-invasive analytical technique that exploits the magnetic properties of certain atomic nuclei to determine the chemical composition of tissue [86]. The most common nuclei investigated in clinical MRS are hydrogen-1 (¹H) and phosphorus-31 (³¹P). The fundamental principle of MRS is the chemical shift, where the resonant frequency of a nucleus is slightly influenced by its local chemical environment (the electron density of its surrounding molecular orbital) [86]. This effect allows different metabolites within a sample or tissue to be distinguished in the resulting frequency spectrum, which presents as a series of peaks. The area beneath each peak is proportional to the concentration of the metabolite. In vivo MRS is typically performed on standard whole-body MRI scanners (1.5–3.0 T), while higher field strengths (11–14 T) are used for in vitro studies on body fluids, cell extracts, and tissue samples to provide more definitive interpretation [86]. Clinically, MRS is used to assess metabolic changes in a wide range of conditions, from brain tumors to metabolic disorders, by quantifying metabolites such as choline (cell membrane turnover), creatine (energy metabolism), N-acetylaspartate (neuronal integrity), and lactate (anaerobic metabolism) [87].

Experimental Protocols and Validation Gold Standards

A standard protocol for single-voxel ¹H-MRS of the brain involves the following steps [86]:

• Anatomical Localization: A high-resolution MRI scan (e.g., T1-weighted or T2-weighted) is first acquired to identify the region of interest (ROI) for spectroscopic analysis.
• Voxel Placement: A voxel (a 3D volumetric pixel) is positioned within the ROI, carefully avoiding areas that could cause spectral contamination, such as skull, bone, fat, or fluid-filled spaces.
• Shimming: The magnetic field across the chosen voxel is optimized (a process called shimming) to ensure a homogeneous field, which is critical for obtaining well-resolved, sharp spectral peaks.
• Water Suppression: The dominant signal from water protons is suppressed using chemical shift selective saturation (CHESS) or similar techniques to allow the much smaller signals from metabolites to be detected.
• Signal Acquisition: The MR signal is acquired using a sequence such as Point-Resolved Spectroscopy (PRESS) or Stimulated Echo Acquisition Mode (STEAM). The echo time (TE) and repetition time (TR) are set based on the metabolites of interest.
• Signal Processing: The raw signal, known as the Free Induction Decay (FID), is processed to generate the final spectrum. This involves:
  • Fourier Transformation: Converting the time-domain FID signal into a frequency-domain spectrum.
  • Phase Correction: Correcting peaks to be positive and symmetrical.
  • Baseline Correction: Ensuring the spectrum baseline starts at zero.
  • Quantification: Metabolite concentrations are often reported as ratios to a stable reference metabolite like creatine, though absolute quantification is possible but more challenging.

The validation pathway for MRS is more established than for TMS-EEG, as it can be directly validated against invasive biochemical analysis. The gold standard for MRS is the in vitro high-resolution NMR spectroscopy conducted on tissue extracts, cell lines, or body fluids at high magnetic field strengths [86]. This method provides a definitive, high-quality metabolic profile against which the in vivo MRS findings can be compared and calibrated. This validation is reflected in the official recognition of MRS for specific clinical indications. For instance, Aetna's medical policy considers MRS medically necessary for distinguishing low-grade from high-grade gliomas, evaluating indeterminate brain lesions, and diagnosing specific metabolic disorders like Canavan disease and creatine deficiency [87]. This endorsement signifies that the validity of MRS for these applications is supported by a body of evidence linking the non-invasive spectral data to the underlying pathology confirmed by biopsy or biochemical tests.

Table 2: Key MRS Validation Metrics and Clinically Approved Applications

Validation Metric	Description	Gold Standard Comparison
Analytical Validity	Accuracy in detecting and quantifying specific metabolites [87].	In vitro NMR spectroscopy of tissue extracts or body fluids at high field strengths (11–14 T) [86].
Clinical Validity	Ability of a metabolic profile to distinguish a specific disease state [87].	Histopathological diagnosis (e.g., for brain tumor grading) or genetic/biochemical confirmation (e.g., for inborn errors of metabolism) [87].
Technical Reliability	Reproducibility of metabolite ratios/quantification across scans.	Test-retest reliability on phantoms and healthy controls; coefficients of variation for major metabolites.
Medically Necessary Applications	Distinguishing recurrent brain tumor from radiation necrosis [87].	-
	Assessing prognosis in hypoxic-ischemic encephalopathy [87].	-
	Diagnosis and monitoring of defined metabolic disorders (e.g., Canavan disease, creatine deficiency) [87].	-

Comparative Analysis: TMS-EEG vs. MRS

The validation journeys of TMS-EEG and MRS reveal fundamental differences in their technological maturity, the nature of their biomarkers, and their clinical translation. The table below provides a structured, point-by-point comparison based on the search results.

Table 3: Direct Comparison of TMS-EEG and MRS Validation Status and Characteristics

Aspect	TMS-EEG	MRS
Primary Measured Quantity	Causal brain connectivity and excitability (TMS-Evoked Potentials) [82] [85].	Biochemical metabolite concentrations (e.g., NAA, Choline, Creatine) [86].
Inherent Gold Standard	Imperfect (MEPs are limited to motor cortex) [82] [83].	Strong (In vitro NMR spectroscopy on tissue/fluid samples) [86].
Key Validation Challenge	Separating genuine neural signal (valid but unreliable) from reliable noise (e.g., muscle, sensory artifacts) [82] [83].	Accurate quantification in vivo; overlapping peaks; low signal-to-noise ratio compared to in vitro methods [86].
Typical Validation Method	Correlation with invasive recordings, pharmacological interventions, and clinical state [83] [85].	Direct correlation with biochemical analysis of tissue extracts or body fluids [86].
Clinical Adoption Status	Primarily a research tool; emerging clinical applications in psychiatry and neurology [85].	Established for specific, limited clinical indications (e.g., brain tumor differentiation, metabolic disorders) [87].
Regulatory/Policy Recognition	Not explicitly listed in reviewed policies.	Recognized as medically necessary for specific indications by major insurers [87].
Signal-to-Noise Challenge	Relatively weak genuine brain responses amid large off-target artifacts [82].	Low concentration of metabolites relative to water signal; requires suppression and long acquisition times [86].

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful execution and validation of TMS-EEG and MRS experiments require specialized equipment and analytical tools. The following table details key solutions for researchers in these fields.

Table 4: Essential Research Reagents and Materials for TMS-EEG and MRS

Item	Function/Application	Key Considerations
TMS-Compatible EEG System	Recording brain electrical activity time-locked to TMS pulses [84] [85].	Use of ultra-thin active electrodes (e.g., 3 mm height) to minimize coil distance and decay artifacts; sample-and-hold or DC amplifiers with high dynamic range [84].
Navigated TMS System	Ensuring precise and reproducible coil positioning over the target brain region [85].	Integrates individual MRI data with optical tracking for millimeter precision; critical for reliability [85].
Artifact Removal Algorithm	Offline processing to remove TMS-induced and physiological artifacts from EEG data [84] [88].	Can use Blind Source Separation (BSS), Independent Component Analysis (ICA), or novel deep learning models (e.g., CLEnet) [1] [88].
Phantom Metabolite Solutions	Calibrating and validating the MRS system for accurate metabolite quantification [86].	Solutions with known concentrations of key brain metabolites (e.g., NAA, Creatine, Choline) housed in a spherical phantom.
Spectral Processing Software	Converting raw FID data into a quantifiable spectrum and fitting metabolite peaks [86].	Software packages (e.g., LCModel, jMRUI) that use prior knowledge of metabolite spectra for linear combination modeling to estimate in vivo concentrations.

Experimental Workflow and Signaling Pathways

The following diagrams illustrate the core experimental and validation workflows for TMS-EEG and MRS, highlighting the logical relationships between key steps and the distinct pathways to establishing validity.

TMS-EEG Experimental & Validation Workflow

MRS Experimental & Validation Pathway

The validation of neuroimaging tools against physiological gold standards is a cornerstone of their translation into clinically useful biomarkers. TMS-EEG and MRS represent two powerful but distinct approaches, with the former probing dynamic brain connectivity and the latter quantifying static metabolic concentrations. Their validation pathways are consequently divergent. TMS-EEG faces the fundamental challenge of an imperfect gold standard (the MEP), forcing reliance on a multi-modal strategy involving invasive recordings and pharmacological probes to establish the validity of its signal amid substantial artifact contamination [82] [83] [85]. In contrast, MRS benefits from a direct and strong gold standard—in vitro NMR spectroscopy—against which its in vivo measurements can be rigorously calibrated [86]. This difference is reflected in their current stages of clinical adoption: MRS has achieved recognition for specific diagnostic tasks [87], while TMS-EEG remains a predominantly research-focused tool with emerging clinical potential. For researchers and drug development professionals, this comparison underscores that the choice of technique must align with the biological question of interest and that the interpretation of results must be critically framed within the context of each method's unique validation landscape and inherent limitations.

Assessing Performance in Challenging, Real-World Scenarios

Validating electroencephalography (EEG) artifact removal methods requires rigorous testing in environments that mimic real-world challenges. While controlled laboratory experiments are valuable, the true test of an algorithm's performance lies in its application to data contaminated by the complex, unpredictable artifacts encountered during natural movement and real-life tasks. This guide objectively compares the performance of leading artifact removal approaches by examining their experimental validation in dynamic scenarios, with a particular focus on studies that use simulated data to establish a known ground truth. The comparison centers on quantitative outcomes related to signal quality, preservation of neural signals, and computational applicability for research and drug development professionals.

Experimental Protocols for Performance Validation

A critical step in comparing artifact removal methods is the use of standardized, challenging experimental protocols that push algorithms to their limits. The following are key paradigms used to generate the validation data discussed in this guide.

The Dynamic Flanker Task during Locomotion

This protocol is designed to evaluate an algorithm's ability to preserve cognitive event-related potentials (ERPs) during intense motion. In one study, young adults performed an adapted Eriksen Flanker task while either jogging on a treadmill or standing still [43]. The task involves responding to a target arrow flanked by either congruent (e.g., >>>>>>) or incongruent (e.g., <<><<<) arrows, which elicits a characteristic P300 ERP component. The protocol tests algorithms on two fronts: their capacity to reduce the high-amplitude motion artifact from jogging, and their ability to preserve the functional P300 congruency effect, a marker of cognitive processing [43].

Realistic Driving Simulator with Multitasking Stressors

This protocol validates algorithms against stressors encountered in real-world operation. In one experiment, 20 subjects experienced a simulated driving course combined with a multitasking battery to induce cognitive stress [89]. The simultaneous recording of EEG and skin conductance level (SCL) provides a multi-modal validation dataset. The key challenge for algorithms is to remove motion artifacts from the lightweight EEG system (which used only two wet sensors) while preserving the neural correlates of stress, such as changes in prefrontal alpha and temporal-parietal beta power, for accurate real-time classification of stress levels [89].

Semi-Synthetic and Real-Artifact Datasets

To establish a ground truth for quantitative comparison, researchers often use semi-synthetic datasets. A common method involves clean EEG recordings, which are then contaminated with real artifacts—such as electromyographic (EMG), electrooculographic (EOG), or electrocardiographic (ECG) signals—recorded in isolation [2] [1]. This process creates a dataset where the clean, original EEG is known, allowing for precise calculation of signal reconstruction error and improvement in signal-to-noise ratio (SNR) after processing. Dataset III mentioned in one study goes a step further by using real 32-channel EEG data collected from subjects performing a 2-back task, containing a mix of known and unknown physiological artifacts, thus presenting a more complex and realistic challenge [1].

Performance Comparison of Artifact Removal Methods

The following tables summarize the experimental performance of various artifact removal techniques across different validation metrics and scenarios.

Table 1: Performance on Motion Artifact Removal during Locomotion

Method	Validation Scenario	Key Performance Metrics	Reported Results
iCanClean [43]	Flanker task during jogging	- Dipolarity of ICA components- Power at gait frequency- P300 congruency effect recovery	- Most dipolar brain components- Significant power reduction
Artifact Subspace Reconstruction (ASR) [43]	Flanker task during jogging	- Dipolarity of ICA components- Power at gait frequency- P300 component latency	- Good dipolarity, slightly less than iCanClean- Significant power reduction- Correct P300 latency identified
Motion-Net [2]	Real-world motion artifacts	- Artifact Reduction (η)- SNR Improvement- Mean Absolute Error (MAE)	- 86% ±4.13- 20 ±4.47 dB- 0.20 ±0.16
CLEnet [1]	Semi-synthetic EMG/EOG/ECG	- Signal-to-Noise Ratio (SNR)- Average Correlation Coefficient (CC)- Relative RMSE (Temporal)	- 11.498 dB (mixed artifacts)- 0.925 (mixed artifacts)- 0.300 (mixed artifacts)

Table 2: Performance on Functional Connectivity & Source Localization

Method Category	Validation Basis	Influence on Functional Connectivity (FC)	Recommended Use
Phase-based Metrics (e.g., imaginary coherence, wPLI) [23]	Simulated EEG-FC with known ground truth	Good FC detection with REST or common average reference	Robust EEG-FC assessment with ≥40 epochs of ≥6s length
Magnitude-Squared Coherence [23]	Simulated EEG-FC with known ground truth	Best performance with Current Source Density reference	-
Traditional Preprocessing (ICA, PCA) [90]	Source localization accuracy	Improved dipole fit and source reconstruction	Foundational step for electrical neuroimaging

The following diagram illustrates a generalized experimental workflow for validating an artifact removal algorithm, from data collection to performance assessment, as reflected in the cited studies.

Diagram 1: General workflow for validating artifact removal algorithms, encompassing key performance metrics.

Signaling Pathways and Workflows in Algorithm Design

Modern artifact removal algorithms, particularly deep learning models, employ sophisticated feature extraction pathways to separate neural activity from artifacts. The diagram below details the internal structure of a state-of-the-art model, CLEnet, which exemplifies this multi-branch approach.

Diagram 2: CLEnet's dual-pathway architecture for joint morphological and temporal feature analysis.

The Scientist's Toolkit: Key Research Reagents and Solutions

This section catalogs essential software tools and data resources that form the foundation for reproducible research in EEG artifact removal and analysis.

Table 3: Essential Software Tools and Data for EEG Artifact Research

Tool / Resource Name	Type	Primary Function	Key Application in Validation
MNE-Python [91]	Open-Source Python Library	Comprehensive EEG/MEG signal processing	Building end-to-end analysis pipelines; source localization
EEGLAB [43] [91]	MATLAB Toolbox	Interactive EEG analysis & ICA	Standardized ICA decomposition; ICLabel for component classification
Brainstorm [90] [91]	Standalone Application	User-friendly MEG/EEG analysis	No-code source localization and connectivity analysis
FieldTrip [90] [91]	MATLAB Toolbox	Advanced MEG/EEG analysis	Flexible scripting for custom analysis pipelines
Cartool [90]	Standalone Software	EEG source imaging & microstate analysis	Precise source localization with individual head models
Open EEG Dataset [92]	Data Resource	60 repeated EEG/MRI/ECG measurements	Assessing stability/repeatability of analysis methods

The comparative data indicates a performance trade-off between traditional, adaptive methods and emerging deep-learning approaches. For controlled locomotion studies, iCanClean demonstrates a slight but consistent advantage over ASR in recovering subtle cognitive neural signatures like the P300 congruency effect, making it a strong candidate for studies of cognition in action [43]. However, ASR remains a highly effective and widely accessible option.

In the realm of deep learning, CLEnet and Motion-Net show remarkable quantitative performance on semi-synthetic benchmarks, with high artifact reduction percentages and SNR improvements [2] [1]. Their subject-specific, end-to-end approach eliminates the need for manual component inspection, offering a path toward full automation. A critical finding from simulation studies is that the choice of artifact removal method is not isolated; it directly impacts downstream analyses like functional connectivity, where phase-based metrics (e.g., wPLI) combined with appropriate re-referencing (REST) provide the most robust results [23].

In conclusion, validation using simulated and challenging real-world data is paramount. No single algorithm universally outperforms all others in every scenario. The optimal choice depends on the specific research question, the nature of the artifacts, and the neural features of interest. For cognitive studies during movement, iCanClean is particularly promising. For achieving high, automated artifact rejection in complex, multi-channel data, deep learning models like CLEnet represent the cutting edge. Researchers are advised to ground their methodological choices in empirical, task-relevant validation studies to ensure the integrity of their findings in real-world applications.

Establishing Best Practices for Reporting and Reproducibility

Reproducible research, defined as the ability to recreate results given the same data, analytic code, and documentation, provides a minimum standard of scientific rigor, especially when replicating costly studies is infeasible [93]. In electroencephalography (EEG) artifact removal research, where methods range from traditional statistical approaches to advanced deep learning models, establishing robust reporting standards is paramount for validating findings and enabling scientific progress. The presence of motion and physiological artifacts in EEG signals presents a significant challenge for neuroscientists and drug development professionals, making the comparison and validation of different artifact removal approaches essential [43] [1]. This guide objectively compares current artifact removal technologies and provides a framework for their reproducible evaluation, particularly focusing on validation using simulated data research.

Comparative Analysis of EEG Artifact Removal Methods

Traditional and Hybrid Signal Processing Approaches

Traditional approaches for motion artifact removal in mobile EEG include artifact subspace reconstruction (ASR) and iCanClean, which have been systematically evaluated during dynamic tasks like running [43]. ASR employs a sliding-window principal component analysis (PCA) to identify high-amplitude artifacts based on a baseline calibration period and a standard deviation threshold (parameter "k") for artifact identification. A k threshold of 20-30 is typically recommended, though values as low as 10 may be necessary to avoid "overcleaning" during high-motion scenarios [43].

iCanClean leverages reference noise signals and canonical correlation analysis (CCA) to detect and correct noise-based subspaces, using a user-selected criterion (R²) of the correlation between corrupt EEG and noise signals [43]. When dual-layer noise sensors are unavailable, iCanClean can create 'pseudo-reference' noise signals from raw EEG by applying a temporary notch filter. Studies have shown that an R² of 0.65 with a 4-second sliding window produces the most dipolar brain components from independent component analysis (ICA) during human locomotion [43].

Deep Learning-Based Approaches

Recent advances in deep learning have transformed EEG artifact removal methods. CLEnet represents a novel approach that integrates dual-scale convolutional neural networks (CNN) with Long Short-Term Memory (LSTM) networks and incorporates an improved one-dimensional Efficient Multi-Scale Attention mechanism (EMA-1D) [1]. This architecture can extract both morphological features and temporal features from EEG signals, effectively separating neural activity from artifacts in an end-to-end manner. The network operates through three stages: (1) morphological feature extraction and temporal feature enhancement using dual convolutional kernels at different scales; (2) temporal feature extraction using LSTM after dimensionality reduction; and (3) EEG reconstruction through fully connected layers [1].

Other notable deep learning architectures include 1D-ResCNN, which utilizes three convolutional kernels of different scales to extract features from artifact-contaminated EEG [1], and Transformer-based EEGDNet, which focuses on both local and non-local features simultaneously for artifact removal [1]. DuoCL, based on CNN and LSTM, was specifically designed to capture temporal features of EEG but may disrupt original temporal features during morphological feature extraction [1].

Table 1: Performance Comparison of EEG Artifact Removal Methods on Standardized Tasks

Method	Type	SNR (dB)	CC	RRMSEt	RRMSEf	Key Strengths
iCanClean [43]	Signal Processing	N/A	N/A	N/A	N/A	Effective for motion artifacts during locomotion; preserves P300 ERP effects
ASR [43]	Signal Processing	N/A	N/A	N/A	N/A	Improves ICA dipolarity; reduces gait frequency power
CLEnet [1]	Deep Learning	11.498	0.925	0.300	0.319	Handles multiple artifact types; works with multi-channel data
1D-ResCNN [1]	Deep Learning	~9.0*	~0.89*	~0.33*	~0.34*	Multi-scale feature extraction
NovelCNN [1]	Deep Learning	~10.2*	~0.90*	~0.32*	~0.33*	Specialized for EMG artifacts
DuoCL [1]	Deep Learning	~10.9*	~0.90*	~0.32*	~0.33*	Combined CNN and LSTM architecture

Note: Approximate values extracted from performance descriptions in [1]; SNR = Signal-to-Noise Ratio; CC = Correlation Coefficient; RRMSEt = Relative Root Mean Square Error (temporal); RRMSEf = Relative Root Mean Square Error (frequency)

Performance Metrics and Evaluation Frameworks

The evaluation of artifact removal methods utilizes standardized metrics that quantify how effectively the algorithm removes artifacts while preserving underlying neural signals [1]. These include:

• Signal-to-Noise Ratio (SNR): Measures the power ratio between clean EEG and removed artifacts, with higher values indicating better performance [1].
• Correlation Coefficient (CC): Quantifies the similarity between processed EEG and ground-truth clean EEG, with values closer to 1.0 indicating better preservation of neural information [1].
• Relative Root Mean Square Error (RRMSE): Assesses reconstruction errors in both temporal (RRMSEt) and frequency (RRMSEf) domains, with lower values indicating higher fidelity [1].
• ICA Dipolarity: Evaluates the quality of independent component analysis decomposition by measuring how well components resemble dipolar brain sources rather than artifacts [43].
• Power Reduction at Gait Frequency: Specifically for motion artifacts, measures the algorithm's effectiveness at removing movement-related noise at the step frequency and its harmonics [43].

Experimental Protocols for Method Validation

Standardized Experimental Workflow

The validation of EEG artifact removal methods requires a structured experimental workflow that progresses from data preparation through to quantitative evaluation. The following diagram illustrates this standardized process:

Dataset Preparation and Synthesis Protocols

Robust validation of artifact removal methods requires diverse datasets that represent various artifact types and signal characteristics [1]. Standardized protocols include:

Semi-Synthetic Dataset Creation: Combining clean EEG segments with recorded artifacts (EMG, EOG, ECG) at controlled signal-to-noise ratios. EEGdenoiseNet provides a benchmark framework for creating such datasets by mixing single-channel EEG with artifact signals in specific ratios [1]. This approach enables precise ground truth comparisons since the clean EEG is known beforehand.

Real-World Dataset Collection: Gathering multi-channel EEG data during tasks that naturally induce artifacts, such as the 2-back working memory task, movement paradigms, or ambulatory recording scenarios [1]. These datasets typically contain "unknown artifacts" whose exact characteristics and proportions relative to neural signals are undefined, presenting a more challenging validation scenario.

Experimental Parameters: For locomotion studies, data should be collected during both dynamic (jogging, walking) and static (standing) versions of cognitive tasks like the Flanker task to enable comparison of event-related potential components such as the P300 [43]. This design allows researchers to determine whether artifact removal methods preserve cognitively relevant neural signals.

Implementation Protocols for Featured Methods

iCanClean Implementation: Apply either dual-layer noise electrodes or create pseudo-reference signals from raw EEG using a temporary notch filter (e.g., below 3 Hz). Use canonical correlation analysis with an R² threshold of 0.65 and a sliding window of 4 seconds. Subtract noise components exceeding the R² threshold from scalp EEG using a least-squares solution [43].

ASR Implementation: Select a baseline calibration period from clean data segments (z-scores of RMS values between -3.5 to 5.0 for at least 92.5% of electrodes). Apply sliding-window PCA to both reference and non-reference data. Identify artifactual components where the standard deviation of RMS exceeds the chosen threshold (k = 10-30). Reconstruct the time series based on calibration data [43].

CLEnet Implementation: Process data through the dual-branch architecture: (1) Extract morphological features using two convolutional kernels of different scales with embedded EMA-1D modules; (2) Reduce dimensionality with fully connected layers then extract temporal features using LSTM; (3) Reconstruct artifact-free EEG using fully connected layers. Train the model using mean squared error loss function [1].

The Researcher's Toolkit: Essential Materials and Reagents

Table 2: Essential Research Reagents and Computational Tools for EEG Artifact Removal Research

Item	Function/Application	Implementation Notes
EEGdenoiseNet Dataset [1]	Benchmark dataset with clean EEG, EMG, and EOG signals for semi-synthetic experiments	Provides standardized evaluation framework; enables controlled SNR conditions
Custom 32-Channel EEG Dataset [1]	Real-world data containing unknown artifacts for validation under realistic conditions	Collected during 2-back tasks; represents complex, real-world artifact scenarios
Artifact Subspace Reconstruction (ASR) [43]	Removal of high-amplitude motion artifacts using sliding-window PCA	Critical parameter: k threshold (10-30); lower values more aggressive
iCanClean Algorithm [43]	Motion artifact removal using CCA and reference noise signals	Can use dual-layer electrodes or pseudo-reference signals; R² threshold ~0.65
CLEnet Architecture [1]	Deep learning-based removal of multiple artifact types using CNN-LSTM hybrid	Handles multi-channel EEG; suitable for unknown artifacts; requires substantial training data
ICA Dipolarity Metrics [43]	Quality assessment of independent components after artifact removal	Higher dipolarity indicates better preservation of brain sources
SNR, CC, RRMSE Metrics [1]	Quantitative evaluation of artifact removal performance	Standardized measures for cross-study comparisons

Reporting Standards for Reproducible Research

Essential Elements of Reproducibility Packages

Computational reproducibility requires specific documentation practices that enable other researchers to recreate analytical results exactly [93]. For EEG artifact removal research, reproducibility packages must include:

Comprehensive README Documentation: A root-level README file should provide a summary overview of all materials, clear instructions for running code to produce all tables and figures, and specific notes indicating where each table and figure in the paper can be located in the output [94]. The README must also list software dependencies, including operating system, computational software versions, and all installed packages with version numbers [94].

Complete Data and Code Provision: Reproducibility packages should include data in as raw a form as possible, along with all code used to transform it [94]. For original data collection, the package must include instruments used to collect data, such as experimental paradigms, task parameters, and equipment specifications. When datasets cannot be shared due to ethical or legal constraints, authors must provide precise instructions for obtaining the data [94].

Structured Directory Organization: A logical directory structure separates code, data, and results, typically organized as: README.txt; master.R (or equivalent master script); data/ (with raw/ and analysis/ subdirectories); code/ (with sequentially numbered scripts); and results/ (containing all output tables and figures) [94].

Methodological Reporting Requirements

Complete methodological reporting enables other researchers to understand and build upon published work. Essential reporting elements include:

Algorithm Parameter Documentation: All method-specific parameters must be explicitly reported, including ASR's k value [43], iCanClean's R² threshold and window size [43], and deep learning architecture details (number of layers, kernel sizes, attention mechanisms) [1].

Computational Environment Specifications: Authors must document the computational environment used for analysis, including random seed locations for any procedures involving pseudo-randomness, and estimated runtime for long-running computations [94].

Validation Framework Description: Studies should clearly describe evaluation metrics, datasets used for validation, and comparison methods included in benchmarks. This includes specifying whether datasets are semi-synthetic or real-world, and the types of artifacts present [1].

Visualization Framework for Method Comparison

The following diagram illustrates the structural relationships between major artifact removal approaches and their applications:

The establishment of best practices for reporting and reproducibility in EEG artifact removal research requires concerted effort across multiple dimensions. Based on our comparative analysis, iCanClean and ASR provide effective preprocessing for motion artifacts during human locomotion [43], while deep learning approaches like CLEnet offer powerful solutions for diverse artifact types in multi-channel EEG [1]. Adopting standardized evaluation metrics, comprehensive reproducibility packages, and structured reporting guidelines will accelerate methodological advances in this critical area of neuroscience research. As the field evolves, validation frameworks must adapt to encompass both controlled semi-synthetic data and complex real-world scenarios, ensuring that artifact removal methods perform robustly across the diverse applications required by researchers and drug development professionals.

Conclusion

The validation of EEG artifact removal methods using simulated data with known ground-truth is a cornerstone of rigorous electrophysiological research. This synthesis of methodologies demonstrates that a multi-faceted approach—combining sophisticated simulation toolboxes like SEED-G, optimized processing pipelines, and rigorous benchmarking against physiological standards—is essential for progress. The future of this field points toward the increased integration of machine learning and state-space models for handling complex artifacts, the development of more dynamic and realistic simulation environments, and the establishment of standardized validation protocols. For biomedical and clinical research, particularly in drug development, these advances are crucial for identifying reliable neural biomarkers, understanding disease pathophysiology through metrics like cortical excitation-inhibition balance, and ultimately translating mobile EEG technologies into real-world clinical applications with confidence.

References

• 1. A novel EEG artifact removal algorithm based on an ... [https://www.nature.com/articles/s41598-025-98653-1]
• 2. Motion Artifact Removal from EEG Signals Using the ... [https://www.sciencedirect.com/science/article/pii/S1746809425013886]
• 3. $$A^{2}$$ DM: Enhancing EEG Artifact Removal by Fusing ... [https://link.springer.com/article/10.1007/s12559-025-10442-0]
• 4. Simulating surveys for graduate researcher development [https://www.frontiersin.org/journals/research-metrics-and-analytics/articles/10.3389/frma.2024.1360333/full]
• 5. Electrophysiological Insights in Exergaming ... [https://games.jmir.org/2025/1/e50992]
• 6. What Is Data Simulation? | Benefits & Modeling - Datamation [https://www.datamation.com/big-data/data-simulation/]
• 7. Using simulation studies to evaluate statistical methods - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6492164/]
• 8. Common physiological EEG artifacts - TMSi [https://www.tmsi.artinis.com/blog/common-physiological-artifacts-in-eeg-signals]
• 9. EEG Artifacts: Types, Detection, and Removal Techniques ... [https://www.bitbrain.com/blog/eeg-artifacts]
• 10. Removal of Artifacts from EEG Signals: A Review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC6427454/]
• 11. EEG Artifacts: Overview, Physiologic ... [https://emedicine.medscape.com/article/1140247-overview]
• 12. Enhancing Signal and Network Integrity: Evaluating BCG ... [https://www.mdpi.com/1424-8220/25/22/7036]
• 13. and LDA-based decoding of EEG signals [https://www.sciencedirect.com/science/article/pii/S1053811925003076]
• 14. A Systematic Review of Techniques for Artifact Detection ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12473706/]
• 15. NEAR: An artifact removal pipeline for human newborn ... [https://www.sciencedirect.com/science/article/pii/S1878929322000123]
• 16. Validation of an electroencephalography data assimilation- ... [https://www.nature.com/articles/s44172-025-00525-z]
• 17. A Systematic Review of Techniques for Artifact Detection ... [https://www.mdpi.com/1424-8220/25/18/5770]
• 18. Combination of spatial and temporal de-noising and artifact ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1576954/full]
• 19. A Deep Lightweight Convolutional Neural Network for ... [https://www.medrxiv.org/content/10.1101/2025.10.28.25338681v1.full]
• 20. An ICA-based artifact suppression method for online ... [https://pubmed.ncbi.nlm.nih.gov/40897195/]
• 21. AnEEG: leveraging deep learning for effective artifact ... [https://www.nature.com/articles/s41598-024-75091-z]
• 22. Estimating sensor-space EEG connectivity [https://www.sciencedirect.com/science/article/pii/S138824572500464X]
• 23. Estimating sensor-space EEG connectivity: Identifying best ... [https://www.sciencedirect.com/science/article/pii/S1388245725004651]
• 24. Simulated Eyeblink Artifact Removal with ICA [https://arxiv.org/html/2410.03261v1]
• 25. Measuring the accuracy of ICA-based artifact removal from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11636693/]
• 26. IMU-Enhanced EEG Motion Artifact Removal with Fine ... [https://arxiv.org/html/2509.01073v1]
• 27. Unsupervised EEG Artifact Detection and Correction [https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2020.608920/full]
• 28. A novel EEG artifact removal algorithm based on an advanced ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12134218/]
• 29. SEED-G: Simulated EEG Data Generator for Testing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8197139/]
• 30. SEREEGA: Simulating event-related EEG activity [https://www.sciencedirect.com/science/article/abs/pii/S0165027018302395]
• 31. SEED-G: Simulated EEG Data Generator for Testing ... [https://www.mdpi.com/1424-8220/21/11/3632]
• 32. aanzolin/SEED-G-toolbox: Pseudo-EEG data generation ... [https://github.com/aanzolin/SEED-G-toolbox]
• 33. EEGSourceSim: A framework for realistic simulation of EEG ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6815881/]
• 34. A deep learning approach to artifact removal in ... [https://www.sciencedirect.com/science/article/pii/S030645222500990X]
• 35. Time-varying MVAR algorithms for directed connectivity analysis [https://pmc.ncbi.nlm.nih.gov/articles/PMC5995381/]
• 36. How do the resting EEG preprocessing states affect ... [https://www.sciencedirect.com/science/article/pii/S1053811925001247]
• 37. Multivariate autoregressive model estimation for high ... [https://www.sciencedirect.com/science/article/pii/S1053811922001860]
• 38. An evaluation of inter-brain EEG coupling methods in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10881924/]
• 39. Motion Artifacts in Dynamic EEG Recordings [https://www.mdpi.com/1424-8220/24/19/6363]
• 40. Artifact Removal from EEG signals using Regenerative ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809421010491]
• 41. Electromyogram (EMG) Removal by ... - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC7873899/]
• 42. EEG Artifact Detection and Correction with Deep ... [https://arxiv.org/abs/2502.08686]
• 43. A Comparison of Approaches for Motion Artifact Removal from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12349273/]
• 44. Instantiating the onEEGwaveLAD Framework for Real- ... [https://www.mdpi.com/1424-8220/25/16/5018]
• 45. Gait-Induced Myoelectric EEG Artifact Removal Validation ... [https://www.mdpi.com/2076-3417/15/22/12103]
• 46. onEEGwaveLAD: A fully automated online EEG wavelet ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0313076]
• 47. Neural masses and fields: modeling the dynamics of brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4237139/]
• 48. A state space modeling approach to real-time phase ... [https://elifesciences.org/articles/68803]
• 49. Combined head phantom and neural mass model ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6448772/]
• 50. A latent state space model for estimating brain dynamics from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10894450/]
• 51. State-space models are accurate and efficient neural ... [https://arxiv.org/html/2409.03231v1]
• 52. EEG Artifact Detection and Correction with Deep Autoencoders [https://arxiv.org/html/2502.08686v1]
• 53. Automatic Artifact Removal from Electroencephalogram ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4562337/]
• 54. Ongoing EEG artifact correction using blind source ... [https://www.sciencedirect.com/science/article/pii/S1388245723009380]
• 55. 10 Top EEG Data Analysis Software Tools Reviewed [https://www.emotiv.com/blogs/news/eeg-data-analysis-software?srsltid=AfmBOoq7Aeb1SEgUpYI0S3mYT4YtlXsGVbFK50uQHY_g0aPKzbETXjDH]
• 56. A comprehensive review of deep learning models for ... [https://link.springer.com/article/10.1007/s42452-025-07808-2]
• 57. A semi-simulated EEG/EOG dataset for the comparison of ... [https://data.mendeley.com/datasets/wb6yvr725d/3]
• 58. EEG is better when cleaning effectively targets artifacts [https://pubmed.ncbi.nlm.nih.gov/41092567/]
• 59. Comparative analysis of algorithmic approaches in ... [https://www.nature.com/articles/s41598-025-15971-0]
• 60. Motion artifact removal from single channel ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809416300775]
• 61. Single-channel EOG artifact removal using fixed frequency ... [https://www.nature.com/articles/s41598-025-10276-8]
• 62. Evaluating ensemble models for fair and interpretable ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12339690/]
• 63. Enhanced detection of artifacts in EEG data using higher ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2895624/]
• 64. Enhanced detection of artifacts in EEG data using higher ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811906011098]
• 65. Evaluation and comparison of most prevalent artifact reduction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10406266/]
• 66. Benchmarking methods for mapping functional connectivity ... [https://www.nature.com/articles/s41592-025-02704-4]
• 67. The effect of preprocessing in dynamic functional network ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5651393/]
• 68. Simulated EEG data generator - Data sets | MRC BNDU [https://data.mrc.ox.ac.uk/data-set/simulated-eeg-data-generator]
• 69. Comparison of local activation, functional connectivity, and ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1337976/full]
• 70. Systematic comparison between a wireless EEG system ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6568010/]
• 71. A Systematic Comparison of High-End and Low-Cost EEG ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10181552/]
• 72. A direct comparison of laboratory and community EEG ... [https://www.nature.com/articles/s41598-025-03569-5]
• 73. Removal of EMG Artifacts from Multichannel EEG Signals ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6955116/]
• 74. EEG artifact removal using sub-space decomposition, ... [https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2022.910368/full]
• 75. MSE, RMSE, MAE, R², and Adjusted R² | by FARSHAD K [https://farshadabdulazeez.medium.com/essential-regression-evaluation-metrics-mse-rmse-mae-r%C2%B2-and-adjusted-r%C2%B2-0600daa1c03a]
• 76. Study on Accuracy Metrics for Evaluating the Predictions of ... [https://www.mdpi.com/2076-3417/11/5/2314]
• 77. Understanding Diagnostic Metrics for Regressors [https://my-hdr-blogs.shaveensingh.com/article/understanding-diagnostic-metrics-for-regressors/]
• 78. Performance Evaluation of Regression Models for the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8476853/]
• 79. A semi-simulated EEG/EOG dataset for the comparison of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4969208/]
• 80. Deep Learning vs. Traditional Machine Learning: Which is ... [https://www.alliancetek.com/blog/post/2025/03/11/deep-learning-vs-traditional-ml.aspx]
• 81. Deep Learning vs Machine Learning: Key Differences [https://labelyourdata.com/articles/deep-learning-vs-machine-learning]
• 82. Reliability and validity of TMS-EEG biomarkers - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10276171/]
• 83. Reliability and Validity of Transcranial Magnetic ... [https://www.sciencedirect.com/science/article/abs/pii/S2451902222003408]
• 84. Reliable and efficient transcranial magnetic stimulation ... [https://www.nature.com/articles/s44172-025-00538-8]
• 85. State-Dependent Transcranial Magnetic Stimulation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12293497/]
• 86. Magnetic Resonance Spectroscopy: Principles and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4723643/]
• 87. Magnetic Resonance Spectroscopy (MRS) - Medical ... [https://www.aetna.com/cpb/medical/data/200_299/0202.html]
• 88. Ongoing EEG artifact correction using blind source ... [https://pubmed.ncbi.nlm.nih.gov/38219404/]
• 89. Validation of a Light EEG-Based Measure for Real-Time ... [https://www.mdpi.com/2076-3425/12/3/304]
• 90. EEG Source Imaging: A Practical Review of the Analysis Steps [https://pmc.ncbi.nlm.nih.gov/articles/PMC6458265/]
• 91. Open-Source EEG Analysis Tools: Making Brain Research ... [https://www.brainaccess.ai/open-source-eeg-analysis-tools-making-brain-research-more-accessible/]
• 92. Open access EEG dataset of repeated measurements from ... [https://www.nature.com/articles/s41597-024-03241-z]
• 93. Pragmatic reproducible research: improving the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10425207/]
• 94. Guidelines for Reproducibility [https://apsanet.org/publications/journals/american-political-science-review/guidelines-for-reproducibility/]