BCG Artifact in Simultaneous EEG-fMRI: A Comprehensive Guide from Mechanisms to Advanced Removal Strategies

Abstract

Simultaneous EEG-fMRI is a powerful multimodal neuroimaging technique that combines high temporal resolution with high spatial resolution. However, its utility is significantly challenged by the ballistocardiogram (BCG) artifact, a complex, cardiac-induced signal that contaminates EEG recordings inside the MRI scanner. This article provides a comprehensive resource for researchers and drug development professionals, addressing the fundamental mechanisms, methodological landscape, optimization challenges, and empirical validation of BCG artifact removal. We explore the physics behind BCG generation, critically compare prevalent software and hardware correction methods—including Average Artifact Subtraction (AAS), Optimal Basis Set (OBS), Independent Component Analysis (ICA), and emerging deep learning and harmonic regression techniques—and provide a framework for troubleshooting and performance evaluation. By synthesizing current literature and validation studies, this review aims to guide the selection of optimal artifact removal strategies to ensure data integrity for both event-related potentials and functional connectivity analyses in clinical and research settings.

Unraveling the BCG Artifact: Origins, Characteristics, and Impact on Neural Data

The Promise and Challenge of Multimodal Integration

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging approach that integrates the complementary strengths of each technique. EEG provides direct measurement of neural electrical activity with millisecond temporal resolution, while fMRI measures hemodynamic changes linked to neural activity with millimeter spatial precision [1] [2]. This integration enables investigators to explore brain dynamics across spatiotemporal scales that neither method could achieve alone, offering unprecedented insights into brain function during cognitive tasks, rest, and in various neurological and psychiatric conditions [3] [4].

However, the acquisition of high-quality EEG data inside the MRI scanner presents significant technical challenges. The EEG signals are contaminated by severe artifacts induced by the MRI environment, which can obscure the much weaker neuronal signals of interest [4] [2]. The most problematic of these is the ballistocardiogram (BCG) artifact, a complex artifact related to cardiac activity that remains difficult to remove completely without distorting neural signals [3] [5]. Effective BCG artifact reduction is thus a critical prerequisite for reliable EEG-fMRI studies.

Understanding the BCG Artifact: Origins and Characteristics

The BCG artifact arises from multiple physiological and physical phenomena associated with cardiac activity within the strong static magnetic field of the MRI scanner. Several concurrent mechanisms contribute to its generation:

• Cardiac-induced head movement: The heartbeat causes subtle pulsatile movements of the head and electrodes within the strong magnetic field, inducing electrical currents explained by Faraday's law of electromagnetic induction [3] [2]
• Scalp pulsation: Electrodes located near blood vessels experience mechanical motion from arterial pulsation, generating electrical potential changes [3]
• Hall effect: As an electrically conductive fluid, pulsatile blood flow generates potential differences across blood vessels in the presence of the magnetic field [3] [2]

The BCG artifact is particularly challenging because it is time-locked to cardiac activity yet exhibits considerable variability in shape and topography across individuals, electrodes, and time [5]. With typical amplitudes exceeding 50 μV in 3T scanners [2], the BCG artifact often obscures neural signals of interest, with most of its power concentrated in the frequency range below 25 Hz [2]—directly overlapping with key neural oscillatory bands including delta, theta, alpha, and beta rhythms.

Figure 1: Multiple physiological mechanisms contribute to BCG artifact generation during simultaneous EEG-fMRI recordings. Cardiac activity drives head movement, scalp pulsation, and Hall effects in blood flow, all of which contaminate the EEG signal within the MRI's strong magnetic field.

Methodologies for BCG Artifact Reduction

Various computational approaches have been developed to mitigate the BCG artifact, each with distinct strengths and limitations. These methods can be broadly categorized into template-based, blind source separation, and hybrid approaches.

Template-Based Methods

Template-based methods operate on the principle of identifying the characteristic BCG artifact waveform and subtracting it from the contaminated EEG signal.

• Average Artifact Subtraction (AAS): This early approach generates an artifact template by averaging EEG segments time-locked to cardiac events (typically detected from ECG or pulse oximetry), then subtracts this average template from the EEG signal [1] [4]. While computationally straightforward, AAS assumes temporal stationarity of the BCG artifact, which often does not hold true in practice, leading to residual artifacts [5].

• Optimal Basis Set (OBS): An extension of AAS, OBS applies Principal Component Analysis (PCA) to the epoched BCG artifacts to capture the dominant temporal variations [1] [5]. The first few principal components form a basis set that is regressed out of the EEG data, providing better handling of artifact variability than simple averaging [5].

• Adaptive Optimal Basis Set (aOBS): This enhanced version addresses two key limitations of standard OBS: beat-to-beat estimation of the delay between cardiac activity and BCG occurrence, and automatic selection of which PCA components to remove based on explained variance criteria [5]. Studies show aOBS achieves significantly lower BCG residuals (5.53%) compared to AAS (12.51%) and standard OBS (9.20%) [5].

Blind Source Separation Approaches

• Independent Component Analysis (ICA): This statistical technique decomposes multichannel EEG data into independent components (ICs), which can be manually or automatically classified as neural or artifactual [1] [3]. Artifactual components are then removed before signal reconstruction. The effectiveness of ICA depends on correct component identification and the fundamental assumption of source independence, which may be violated for BCG artifacts due to their complex spatiotemporal properties [5].

Hybrid and Advanced Methods

• Combined Approaches (OBS+AAS, OBS+ICA): Hybrid methods leverage the complementary strengths of multiple techniques. For instance, applying OBS followed by ICA can remove residual BCG artifacts that survive initial template subtraction [1] [6]. One study found that OBS+ICA produced the lowest p-values across frequency band pairs in dynamic connectivity analysis [1].

• Hardware-Based Solutions (Carbon Wire Loops): This approach uses carbon wire loops placed around the head to record reference signals containing primarily MR-induced artifacts [3] [2]. These signals are used to regress out artifacts from the EEG data. Studies show CWL systems can outperform computational methods in recovering spectral contrast in alpha and beta bands and visual evoked responses [2].

• Real-Time Processing Tools: Recent developments like NeuXus implement Long Short-Term Memory (LSTM) networks for real-time R-peak detection combined with artifact average subtraction, enabling real-time artifact correction with execution times under 250 ms [7]. Similarly, the APPEAR pipeline provides fully automated, standardized processing combining OBS/AAS with ICA [6].

Figure 2: BCG artifact reduction methods can be categorized into template-based, blind source separation, and hybrid/hardware approaches. Each method follows a pathway from contaminated to cleaned EEG, with varying complexity and effectiveness.

Comparative Performance of BCG Artifact Removal Methods

The selection of an appropriate artifact removal strategy requires careful consideration of methodological performance across multiple metrics. Different methods excel in different domains—some preserve signal fidelity better, while others optimize for connectivity analysis or specific frequency bands.

Table 1: Performance Comparison of Major BCG Artifact Removal Methods

Method	Best Performance Metrics	Key Limitations	Impact on Network Topology
AAS	Best signal fidelity (MSE = 0.0038, PSNR = 26.34 dB) [1]	Assumes artifact stationarity; leaves residuals with non-stationary artifacts [5]	Moderate effect on functional connectivity patterns [1]
OBS	Highest structural similarity (SSIM = 0.72) [1]	Fixed component selection; sensitive to ECG-BCG misalignment [5]	Significant effects on network structure, especially in dynamic analyses [1]
aOBS	Lowest BCG residuals (5.53%); best cross-correlation reduction (0.028) [5]	Complex implementation; computationally intensive [5]	Not specifically reported in evaluated studies
ICA	Sensitivity to frequency-specific patterns in dynamic graphs [1]	Difficult component selection; potential neural signal loss [3] [5]	Greater sensitivity in dynamic graph metrics [1]
OBS+ICA	Lowest p-values across frequency band pairs [1]	Potential error propagation; complex pipeline [1] [6]	Enhanced differentiation in beta and gamma bands [1]
CWL	Superior spectral contrast in alpha/beta bands; best VEP recovery [2]	Requires specialized hardware; cannot be applied retrospectively [3] [2]	Not specifically reported in evaluated studies

Table 2: Quantitative Performance Metrics Across Artifact Removal Methods

Method	BCG Residual (%)	Cross-Correlation with ECG	SNR Improvement	Classification Accuracy
AAS	12.51 [5]	0.051 [5]	Not reported	Not applicable
OBS	9.20 [5]	0.042 [5]	Not reported	Not applicable
ICA	20.63 [5]	0.067 [5]	Not reported	Not applicable
aOBS	5.53 [5]	0.028 [5]	Not reported	Not applicable
Hybrid Model (BiGRU-FCN)	Not applicable	Not applicable	Not reported	98.61% [8] [9]

Experimental Protocols for Method Validation

Rigorous validation of BCG artifact removal methods requires standardized experimental paradigms and evaluation metrics. Below are detailed protocols employed in recent comprehensive studies:

Resting-State and Task-Based Validation

• Data Acquisition: EEG data is typically collected using MRI-compatible systems (e.g., 64-channel SynAmps2) inside 3T MR scanners during simultaneous fMRI acquisition (common parameters: TR = 3s, TE = 30ms, 47 slices) [3]. Parallel recordings outside the scanner serve as benchmark data [2].

• Cardiac Monitoring: Heartbeat detection employs General Electric MR-compatible physiological pulse oximetry (50 Hz sampling) or ECG recordings via dedicated electrodes [6]. The synchronized cardiac signal is essential for template-based methods.

• Experimental Paradigms: Multiple paradigms are used for comprehensive validation:

  • Resting-state: Subjects observe a black screen while remaining calm for 6 minutes (120 volumes) [3]
  • Visual stimulation: Presentation of checkerboard patterns or similar stimuli to elicit robust visual evoked potentials (VEPs) [5] [2]
  • Motor tasks: Finger tapping protocols to assess beta-band event-related desynchronization [2]
  • Auditory evoked potentials: Simulated or actual auditory stimuli to evaluate time-domain performance [3]

Evaluation Metrics and Analysis

Comprehensive method validation employs multiple quantitative and qualitative metrics:

• Signal Quality Metrics: Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Signal-to-Noise Ratio (SNR), Structural Similarity Index (SSIM), Dynamic Time Warping (DTW), and Peak-to-Peak Ratio (PPR) [1]

• Residual Artifact Measurement: BCG residual intensity is quantified as the percentage of residual artifact power after correction [5]. Cross-correlation between EEG and ECG signals should approach zero after effective correction [5]

• Spectral Analysis: Power Spectral Density (PSD) analysis examines preservation of frequency components and oscillatory activity across bands (delta, theta, alpha, beta, gamma) [1] [2]

• Event-Related Potential Assessment: For task-based data, Signal-to-Noise Ratio (SNR) and inter-trial variability of ERPs quantify neural signal preservation [5]

• Connectivity and Graph Metrics: Connection Strength (CS), Clustering Coefficient (CC), and Global Efficiency (GE) analyze effects on functional connectivity patterns in static and dynamic contexts [1]

Table 3: Key Research Reagents and Tools for BCG Artifact Research

Tool/Resource	Function/Purpose	Example Implementations
Carbon Wire Loops (CWL)	Hardware reference system to capture MR-induced artifacts for regression [2]	Custom-built loops following van der Meer et al. (2016) specifications [2]
MR-Compatible EEG Systems	Safe EEG acquisition in high magnetic field environments	SynAmps2 (Neuroscan), BrainAmp MR plus (Brain Products)
Pulse Oximetry/ECG	Cardiac monitoring for artifact template synchronization	General Electric MR-compatible physiological monitoring system [6]
EEGLAB with FMRIB Plugin	Offline artifact removal using AAS, OBS, and ICA	FMRIB's Fast Template Regression algorithm (FASTR) [6]
APPEAR Toolbox	Automated pipeline for comprehensive EEG-fMRI artifact reduction	Combines OBS/AAS with ICA for fully automatic processing [6]
NeuXus	Real-time artifact reduction toolbox	Implements LSTM networks for R-peak detection and artifact subtraction [7]
BESA Research	Commercial software for surrogate artifact reduction methods	PCA-S and ICA-S approaches for spatial filtering [3]

Current Research Directions and Future Outlook

The field of BCG artifact reduction continues to evolve with several promising research directions:

• Real-Time Processing: Tools like NeuXus and EEG-LLAMAS now enable low-latency artifact removal (under 250 ms), opening possibilities for closed-loop EEG-fMRI neurofeedback paradigms [7] [1]. These implementations use Long Short-Term Memory (LSTM) networks for improved R-peak detection and adaptive subtraction methods [7].

• Machine and Deep Learning: Hybrid models integrating deep learning with traditional signal processing show exceptional accuracy in artifact detection. The BiGRU-FCN model achieves 98.61% classification accuracy for motion artifacts in BCG signals [8] [9]. These approaches leverage temporal Bidirectional Gated Recurrent Units combined with Fully Convolutional Networks to handle the complexity and variability of artifacts [9].

• Dynamic Connectivity Preservation: Recent research emphasizes evaluating artifact removal methods based on their impact on functional network topology rather than just signal-level metrics [1]. Studies reveal that different methods significantly affect network structure interpretation, with dynamic analyses showing more pronounced frequency-specific effects, particularly in beta and gamma bands [1].

• Standardization and Automation: Tools like APPEAR provide fully automated processing pipelines that minimize researcher bias and enable reproducible processing of large EEG-fMRI cohorts [6]. This addresses the critical need for standardized methodologies in the field.

• Multimodal Data Fusion: Advanced fusion techniques like Generalized Coupled Matrix Tensor Factorization (GCMTF) utilize normalized mutual information to capture both linear and nonlinear dependencies between EEG and fMRI modalities, potentially providing additional constraints for artifact separation [1].

The continued development of BCG artifact reduction methods remains essential for advancing simultaneous EEG-fMRI research. Future work will likely focus on optimizing the trade-off between artifact removal and neural signal preservation, validating methods across diverse populations and states, and developing increasingly sophisticated real-time processing frameworks for dynamic brain imaging applications.

In simultaneous electroencephalogram-functional magnetic resonance imaging (EEG-fMRI), the ballistocardiogram (BCG) artifact represents a significant challenge for data quality. This artifact manifests as a complex signal distortion contaminating EEG recordings, originating from the interplay between the human body and the high-strength magnetic field of the MRI scanner. The BCG artifact is fundamentally tied to cardiac activity, occurring rhythmically with each heartbeat, and exhibits non-stationary characteristics that make its removal particularly difficult compared to other MRI-related artifacts [10]. Unlike the gradient artifact (GA) caused by switching magnetic field gradients, the BCG artifact arises from physiological processes and persists even when no fMRI acquisition is performed, primarily obscuring the clinically relevant EEG frequency bands below 20 Hz, including delta, theta, alpha, and beta oscillations [10] [11]. Understanding the physics underlying BCG generation is essential for developing effective artifact removal strategies and ensuring accurate interpretation of neural signals in simultaneous EEG-fMRI studies.

The historical context of BCG traces back to Starr's pioneering work, which demonstrated that BCG captures signals generated by repetitive body motions due to sudden blood ejection into the great vessels with each heartbeat [12]. In contemporary neuroscience research, the resurgence of BCG research is driven by the expanding applications of simultaneous EEG-fMRI in mapping neural oscillations, localizing epileptic events, and investigating sleep physiology [13] [10]. The artifact's complex spatio-temporal dynamics, variability across subjects and recording channels, and changing characteristics over time necessitate physics-based approaches for effective mitigation [14]. This technical guide examines the fundamental physical principles governing BCG generation, with particular emphasis on cardiac-induced motion and the Hall effect, to establish a theoretical foundation for artifact removal methodologies within EEG-fMRI research.

Fundamental Physical Principles of BCG Generation

Cardiac-Induced Body Mechanics

The primary mechanism underlying BCG artifact generation involves cardiac-induced mechanical motions transmitted throughout the body. With each cardiac cycle, the sudden ejection of blood from the heart into the major vessels generates reactive forces that cause subtle but measurable movements of the entire body [12]. In the context of EEG-fMRI recordings, these mechanical oscillations affect the subject's head position and the relationship between EEG electrodes and the scalp, inducing artifactual potentials in the EEG recordings.

The mathematical foundation for describing these mechanics was established by Starr and Noordergraaf, who expressed the displacement of the body's center of mass along the head-to-toe direction at any time (t) as:

[ Y(t) = \frac{\rhob}{M} \sum{i=1}^{N} Vi(t) yi + c ]

where (\rhob) represents blood density, (M) is the total body mass, (N) is the number of vascular compartments, (Vi(t)) denotes the blood volume in compartment (i) at time (t), (yi) is the fixed coordinate of compartment (i), and (c) is a constant term representing the body frame [12]. The corresponding BCG signals associated with velocity ((BCG{vel}(t))) and acceleration ((BCG_{acc}(t))) can be derived through time differentiation of this fundamental equation [12].

In simultaneous EEG-fMRI, these mechanical oscillations become particularly problematic as the head moves within the strong static magnetic field, inducing electrical currents according to Faraday's law of electromagnetic induction [3]. Even microscopic head movements in the range of micrometers can generate significant artifacts that overwhelm the microvolt-scale neural signals measured by EEG [3] [10]. Additional contributions come from pulse-driven expansion of the scalp and mechanical vibrations of EEG electrodes themselves, which further complicate the artifact morphology [13] [11].

Table 1: Physical Mechanisms Contributing to BCG Artifacts in EEG-fMRI

Physical Mechanism	Description	Key Characteristics
Cardiac-Induced Head Motion	Head movement in static magnetic field due to cardiac pulsation	Follows heartbeat rhythm; dominant in anterior-posterior direction [15]
Scalp Expansion	Pulsatile expansion of scalp arteries	Correlated with cardiac cycle; affects electrode-scalp interface [13]
Hall Effect	Electrical potential from blood flow in magnetic field	Intrinsic to conductive blood flow; independent of motion [13] [10]
Electrode Movement	Mechanical vibration of EEG electrodes	Causes impedance changes; contact-dependent variability [10]

The Hall Effect in Biological Context

The Hall effect represents a distinct electromagnetic mechanism contributing to BCG artifacts, separate from mechanical motion. When an electrically conductive fluid such as blood flows through a magnetic field, a voltage difference develops perpendicular to both the flow direction and the magnetic field direction [13]. This phenomenon arises from the Lorentz force acting on moving charges in the blood, creating a potential difference that can be measured as electrical signals on the scalp surface.

In the specific context of EEG-fMRI, the Hall effect generates artifacts through several pathways. As pulsatile blood flow travels through cerebral vessels within the strong static magnetic field of the MRI scanner (typically 3 Tesla or higher), the resulting electrical potentials directly contaminate EEG recordings [13] [10]. This effect is particularly significant when electrodes are positioned near major scalp veins, where the artifact amplitude increases substantially [3]. The Hall effect contribution differs fundamentally from motion-based artifacts because it originates from the conductive properties of blood itself rather than mechanical displacement, making it particularly challenging to remove through motion-correction approaches alone [13].

The complexity of BCG artifacts stems from the superposition of multiple physical phenomena occurring simultaneously and exhibiting varying spatial distributions across the scalp. Different mechanisms may dominate at different electrode locations, with cardiac-induced motions generally producing more global artifact patterns while Hall effect contributions and scalp pulsations show more localized effects [3] [11]. This multifaceted origin explains why BCG artifacts display complex spatio-temporal dynamics with varying morphology across channels, subjects, and recording sessions, presenting a persistent challenge for artifact removal algorithms [14] [11].

Quantitative Models and Experimental Data

Cardiovascular Modeling of BCG Signals

The development of closed-loop mathematical models of the cardiovascular system has advanced the quantitative understanding of BCG signal generation. These models simulate the fundamental mechanisms producing BCG signals by representing blood circulation using analogies to electrical systems, where fluid pressure corresponds to electric potential, blood volume to electric charge, and flow rates to electric current [12]. Such models typically incorporate resistors representing vascular resistance, capacitors representing vessel wall compliance, and inductors accounting for blood inertia, arranged into interconnected compartments representing the heart, systemic circulation, pulmonary circulation, and cerebral circulation [12].

Simulations using these cardiovascular models successfully reproduce the characteristic I, J, K, L, M, and N peaks observed in experimental BCG measurements [12]. Furthermore, they predict specific changes in BCG morphology under pathological conditions, including reduced ventricular contractility and increased arterial stiffness, demonstrating the method's potential for clinical interpretation of BCG signals [12]. These models provide a virtual laboratory for investigating how hemodynamic alterations manifest in BCG measurements, establishing a foundation for using BCG not merely as an artifact to be removed but as a source of valuable cardiovascular information.

Table 2: Characteristic BCG Signal Components and Their Cardiovascular Correlates

BCG Peak	Timing in Cardiac Cycle	Physiological Correlate	Amplitude Range
I Wave	Early systole	Atrial contraction	0.5-1.2 mV
J Wave	Peak systole	Maximum ventricular ejection	1.0-2.5 mV
K Wave	Late systole	Closure of aortic valve	0.8-1.8 mV
L Wave	Early diastole	Rapid ventricular filling	0.3-0.9 mV
M Wave	Mid-diastole	Slow ventricular filling	0.2-0.7 mV
N Wave	Late diastole	Atrial contraction	0.1-0.5 mV

Frequency Domain Characteristics

BCG artifacts exhibit distinctive spectral properties that inform removal strategies. The artifact predominantly contaminates low-frequency EEG bands below 20 Hz, directly overlapping with clinically important neural oscillations including delta (1-4 Hz), theta (4-8 Hz), and alpha (8-13 Hz) rhythms [10] [11]. This spectral overlap complicates simple frequency-based filtering approaches, as such methods would inevitably remove neural signals of interest along with the artifact. The table below summarizes the key spectral characteristics of BCG artifacts and their overlap with neural oscillations of interest.

Table 3: Spectral Characteristics of BCG Artifacts and Neural Oscillations

Frequency Band	Frequency Range (Hz)	BCG Artifact Presence	Neural Significance
Delta	1-4	Strong contamination	Deep sleep, pathological states
Theta	4-8	Moderate to strong contamination	Drowsiness, meditation
Alpha	8-13	Moderate contamination	Relaxed wakefulness
Beta	13-30	Mild contamination	Active thinking, focus
Gamma	>30	Minimal contamination	Information processing

Experimental Methodologies for BCG Investigation

Suspended Bed Accelerometry

The traditional approach for BCG measurement utilizes a suspended bed system that captures body movements resulting from cardiac activity. This method employs a lightweight bed suspended by long cables, allowing the bed to swing freely in response to forces generated by blood ejection [12]. Accelerometers mounted on the bed measure these subtle movements, producing a clean BCG signal uncontaminated by other physiological processes. Recent research has replicated this historical approach to validate theoretical models, with studies building replicas of Starr's original suspended bed to compare experimental measurements with simulated BCG waveforms derived from cardiovascular models [12]. This validation strategy ensures that models accurately represent the physical processes generating BCG signals, providing a foundation for predicting how specific cardiovascular pathologies alter BCG morphology.

Imaging Ballistocardiography (iBCG)

Advances in computer vision have enabled video-based BCG measurement through imaging ballistocardiography (iBCG). This non-contact technique quantifies subtle rhythmic head movements caused by cardiac activity through video analysis of facial regions [15]. Unlike traditional approaches, iBCG requires no physical contact with subjects, making it particularly suitable for monitoring applications. Recent methodological improvements have focused on motion artifact reduction in iBCG through anterior-posterior (Z-axis) signal reconstruction based on the law of perspective, which is more sensitive to cardiac-induced head motions in seated subjects compared to conventional vertical (Y-axis) measurements [15]. The iBCG approach demonstrates that BCG signals can be captured through multiple modalities, each with distinct advantages for specific research or clinical applications.

BCG Generation and Investigation Workflow

Simultaneous EEG-fMRI Protocols

The primary experimental context for BCG artifact investigation involves simultaneous EEG-fMRI recording protocols. These protocols typically include visual stimulation paradigms, resting-state measurements, and event-related potential (ERP) tasks conducted inside MRI scanners [13] [10]. The standard data acquisition involves MRI-compatible EEG systems with 64 or more channels synchronized with the fMRI scanner clock, recording at high sampling rates (typically 5-10 kHz) to adequately capture both neural signals and artifacts [3] [13]. Critical to these protocols is the simultaneous recording of electrocardiogram (ECG) or pulse oximetry signals to precisely identify cardiac events for subsequent artifact removal procedures [13] [10]. These comprehensive datasets enable researchers to analyze BCG characteristics across different brain states and validate artifact removal methods using known neural responses to controlled stimuli.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 4: Essential Equipment and Materials for BCG Research

Equipment/Material	Technical Specifications	Research Function
MRI-Compatible EEG System	64+ channels, 5-10 kHz sampling rate, synchronized clock	Records neural activity simultaneous with fMRI acquisition [3] [10]
Accelerometer Array	High-sensitivity (0.1-0.5 mV/g), low-frequency response (0.1-50 Hz)	Measures mechanical body vibrations from cardiac activity [12]
ECG Recording System	MRI-compatible electrodes, filtering for gradient artifacts	Provides precise cardiac timing reference for BCG events [13] [10]
Video Recording System	High-speed (60+ fps), high-resolution facial imaging	Enables imaging BCG through head motion analysis [15]
Carbon Fiber Loops/Slings	Conductive reference layers placed near head	Measures BCG artifact independently for reference subtraction [3] [10]
Suspended Bed Platform	Low-friction suspension system, precision accelerometers	Captures reference BCG signals without fMRI interference [12]

Methodologies for BCG Artifact Removal

The complex physics of BCG generation necessitates sophisticated removal strategies in simultaneous EEG-fMRI research. Current approaches can be broadly categorized into template-based methods, which include Average Artifact Subtraction (AAS) and Optimal Basis Set (OBS), and blind source separation techniques, primarily Independent Component Analysis (ICA) [3] [13]. Template-based methods construct an average artifact waveform time-locked to cardiac events and subtract it from the EEG signal, while blind source separation approaches decompose the EEG into components that are statistically independent, allowing identification and removal of artifact-related components [3] [13].

Recent advances include hybrid methods that combine multiple approaches to address the limitations of individual techniques. The EMD-PCA method employs empirical mode decomposition followed by principal component analysis to separate BCG artifacts from neural signals without requiring reference cardiac measurements [11]. Similarly, the adaptive OBS (aOBS) method improves upon standard OBS by estimating the variable delay between cardiac activity and BCG occurrence on a beat-to-beat basis, enabling more accurate artifact template creation [13]. Machine learning approaches, particularly Generative Adversarial Networks (GANs), represent the cutting edge in BCG removal, learning to transform BCG-contaminated EEG into clean EEG through unpaired signal translation without requiring reference signals [14] [16].

BCG Artifact Removal Method Classification

The physics of BCG generation encompasses multiple interrelated phenomena, with cardiac-induced motion and the Hall effect representing fundamental mechanisms that contaminate EEG recordings during simultaneous fMRI. The complex nature of these artifacts, stemming from the interplay between cardiovascular physiology and electromagnetic principles, necessitates sophisticated removal approaches that account for both temporal and spatial characteristics. Current research directions focus on developing increasingly adaptive methods that minimize neural signal distortion while effectively suppressing BCG components, with machine learning approaches showing particular promise for handling the non-stationary characteristics of these artifacts. A comprehensive understanding of the physical principles underlying BCG generation provides researchers with the foundational knowledge required to select appropriate artifact mitigation strategies and interpret cleaned EEG signals accurately within the context of simultaneous EEG-fMRI studies.

The ballistocardiogram (BCG) artifact represents a significant challenge in simultaneous EEG-fMRI, fundamentally limiting the utility of this multimodal approach. This technical guide details the core characteristics of BCG artifacts, with particular focus on their complex spatiotemporal dynamics and substantial amplitude overlap with neurophysiologic signals. We synthesize current research findings and present quantitative data demonstrating how these artifacts obscure neural activity across both temporal and spatial domains. The paper further provides detailed methodologies for key experimental protocols used in BCG investigation and removal, visual workflows of artifact dynamics and processing pipelines, and essential research tools for scientists working in this field. Understanding these characteristics is paramount for developing effective artifact removal strategies and advancing EEG-fMRI applications in both cognitive neuroscience and clinical drug development.

Simultaneous electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) recording provides a powerful neuroimaging tool that combines high temporal resolution from EEG with high spatial resolution from fMRI [17] [18]. This integration enables researchers to capture identical brain activity through complementary measures, offering unprecedented insights into brain function [18]. However, the full potential of EEG-fMRI remains constrained by significant artifacts induced in the EEG data by the MRI environment, with the ballistocardiogram (BCG) artifact being particularly problematic [17] [19].

The BCG artifact arises from multiple physical phenomena associated with cardiac pulsation. When recorded inside the MRI scanner, EEG signals become contaminated by artifacts generated through cardiac-related head movement within the static magnetic field, local scalp movements caused by expansion and contraction of scalp arteries, and the Hall effect produced by pulsatile blood flow [20]. These artifacts manifest as repetitive, high-amplitude distortions that are time-locked to the cardiac cycle but exhibit complex variability in both timing and morphology [17] [21].

The prevalence and impact of BCG artifacts have made them a central focus in EEG-fMRI methodology research. As static magnetic field strengths increase in modern MRI systems, BCG artifacts pose even greater challenges due to their amplitude scaling with field strength [17] [19]. This technical guide examines the fundamental characteristics of BCG artifacts, with particular emphasis on their spatiotemporal dynamics and amplitude characteristics relative to neural signals, providing researchers with a comprehensive framework for understanding and addressing these artifacts in experimental settings.

Fundamental BCG Artifact Characteristics

Spatiotemporal Dynamics

The BCG artifact exhibits complex spatiotemporal properties that complicate its removal from EEG signals. Unlike gradient artifacts, which are highly reproducible and can be effectively removed using template subtraction methods, BCG artifacts demonstrate substantial variability both temporally and spatially [20] [21].

Temporally, BCG artifacts are characterized by their harmonic structure and variability across cardiac cycles. Spectrograms of EEG recorded inside MRI scanners reveal distinctive comb-like harmonic streaks that obscure underlying neural rhythms [17] [19]. These artifacts maintain a fundamental relationship with the heart rate but exhibit variations in timing, shape, and amplitude from beat to beat due to natural fluctuations in heart rate, blood pressure, and pulsatile head motion [17]. This temporal variability means that a unique template describing all cardiac events cannot be derived, making simple subtraction approaches insufficient [20].

Spatially, BCG artifacts demonstrate inhomogeneous distributions across the scalp, with varying shapes and intensities at different electrode locations [17] [20]. This spatial heterogeneity arises from the different underlying physical mechanisms: global head movements affect all electrodes, while local scalp pulsations predominantly influence electrodes near blood vessels, and the Hall effect introduces additional spatially variable components [20] [22]. The complex spatial topography of BCG artifacts further complicates their separation from neural signals using standard spatial filtering techniques.

Table 1: Temporal and Spatial Characteristics of BCG Artifacts

Characteristic	Description	Research Implication
Temporal Variability	Beat-to-beat variations in timing, shape, and amplitude due to physiological fluctuations [17] [20]	Limits effectiveness of fixed template subtraction methods
Harmonic Structure	Comb-like harmonic streaks in spectrograms related to heart rate harmonics [17] [19]	Enables frequency-domain modeling approaches
Spatial Inhomogeneity	Varying artifact morphology across different scalp regions [17] [20]	Requires channel-specific or spatially adaptive processing
Multi-component Origin	Different physical phenomena (head movement, scalp pulsation, Hall effect) contribute differently across electrodes [20] [22]	Necessitates comprehensive artifact models that account for all components

Amplitude Overlap with Neural Signals

The most challenging aspect of BCG artifacts is their substantial amplitude overlap with neurophysiologic EEG signals, which creates significant obstacles for artifact removal without simultaneous distortion of neural activity of interest.

BCG artifacts typically range between 150-200 μV in amplitude in 3T MRI systems, dramatically exceeding the amplitude of most neurophysiologic EEG signals, which generally fall between 5-100 μV [17] [19]. This amplitude discrepancy means BCG artifacts can be tens to hundreds of times larger than the underlying neural signals they obscure [20]. The problem is particularly pronounced in the frequency range of 0-20 Hz, where both BCG artifacts and many clinically relevant neural signals (such as event-related potentials and oscillatory rhythms) predominantly reside [17] [19].

This convergence in both amplitude and frequency domains creates a challenging scenario for artifact removal techniques. The substantial overlap means that simply filtering or subtracting artifacts risks removing or distorting neural signals of interest, potentially creating false positives or eliminating genuine neural correlates [17] [22]. The situation is further complicated by the fact that BCG background activity can be mistaken for periodic brain rhythms, evoked responses, or ictal discharges in clinical applications [17] [19].

Table 2: Amplitude and Frequency Characteristics of BCG vs. Neural Signals

Signal Type	Amplitude Range	Dominant Frequency Range	Key Challenges
BCG Artifact	150-200 μV [17] [19]	0-20 Hz, with harmonic extensions [17]	Obscures neural signals; harmonic structure overlaps with brain rhythms
Evoked Potentials	5-20 μV [17]	<30 Hz [22]	Low SNR makes detection difficult; similar frequency content to BCG
Spontaneous EEG/Oscillations	10-100 μV [17]	Delta: 1-4 Hz; Theta: 4-8 Hz; Alpha: 8-12 Hz; Beta: 12-30 Hz [17]	Direct frequency overlap with BCG fundamentals and harmonics
Epileptic Spikes	50-500 μV [17]	Broadband, with significant <30 Hz components [17]	Similar morphology to BCG pulses; can be mistaken for artifact residuals

Experimental Protocols for BCG Investigation

Harmonic Regression Analysis

The harmonic regression method provides a reference-free approach to BCG artifact removal that leverages the fundamental harmonic structure of these artifacts [17] [19]. This protocol involves modeling BCG artifacts using a harmonic basis set and applying local regression techniques to estimate and remove artifacts while preserving neural signals.

Experimental Procedure:

• Data Acquisition: Acquire EEG data inside the MRI scanner with simultaneous recording of cardiac signals (ECG or pulse oximeter) for timing reference [17] [19].
• Gradient Artifact Removal: Apply established gradient artifact removal techniques (e.g., Average Artifact Subtraction) to eliminate MRI gradient-induced artifacts before BCG removal [21] [22].
• Harmonic Model Specification: Model the BCG artifact at each electrode using a harmonic basis functions tied to the fundamental cardiac frequency: y(t) = Σ[A_k cos(kωt) + B_k sin(kωt)] + ε(t), where ω is the fundamental frequency derived from the heart rate, k represents harmonic order, and ε(t) represents the neural signal plus noise [17] [19].
• Parameter Estimation: Use maximum likelihood estimation to fit the harmonic model to short, overlapping temporal windows (typically 1-2 seconds) to account for time-varying characteristics of both the artifact and neural signals [17] [19].
• Artifact Removal: Subtract the estimated BCG artifact from the original signal to obtain cleaned EEG data.
• Validation: Quantify performance using metrics such as root mean square error between cleaned and known signals (in simulations), signal-to-noise ratio improvement, and preservation of known neural responses [17] [19].

This method has demonstrated effectiveness in removing BCG artifacts while preserving both oscillatory and evoked neural responses, particularly in conditions with significant time-frequency overlap between artifacts and signals of interest [17] [19].

Adaptive Optimal Basis Set (aOBS) Method

The adaptive Optimal Basis Set method enhances the traditional OBS approach by incorporating beat-to-beat estimation of the delay between cardiac activity and BCG occurrence, along with automated selection of artifact-related components [21].

Experimental Procedure:

• Gradient Artifact Correction: Perform initial gradient artifact removal using established methods [21].
• Cardiac Event Detection: Identify R-peaks in simultaneously recorded ECG signals with high temporal precision [21].
• BCG Occurrence Identification: For each channel, identify BCG peaks in the gradient-corrected EEG data, accounting for variable delay between ECG R-peaks and BCG occurrences on a beat-to-beat basis [21].
• Data Epoching: Segment EEG data into epochs time-locked to the identified BCG peaks, typically using windows from 100-200 ms before to 500-600 ms after each BCG peak [21].
• Principal Component Analysis: Apply PCA to the epoched data for each channel separately to obtain spatial and temporal components [21].
• Component Selection: Automatically select principal components representing BCG artifacts based on explained variance criteria, typically retaining components that collectively explain >90% of variance in the epoched data [21].
• Artifact Reconstruction and Subtraction: Reconstruct BCG artifact for each epoch using linear combination of selected components, then subtract from original data [21].
• Performance Validation: Evaluate effectiveness using metrics such as BCG residual intensity, cross-correlation with ECG, and preservation of event-related potentials [21].

The aOBS method has shown superior performance compared to traditional AAS, ICA, and standard OBS approaches, particularly in reducing BCG residuals while maintaining neural signal integrity [21].

Surrogate Method with Spatial Filtering

The surrogate method utilizes spatial filtering principles to separate artifact and neural signals based on their distinct spatial distributions across the scalp [22]. This approach can be implemented using either PCA or ICA for artifact topography identification.

Experimental Procedure:

• Data Acquisition and Preprocessing: Acquire high-density EEG data (64+ channels) during simultaneous fMRI recording, followed by standard gradient artifact removal [22].
• Artifact Topography Identification:
  • PCA-S Approach: Perform PCA on averaged artifact template to identify dominant spatial components of BCG artifact [22].
  • ICA-S Approach: Use ICA to decompose data and manually select components representing BCG artifacts based on temporal and spatial characteristics [22].
• Surrogate Source Modeling: Create a surrogate source model consisting of regional dipole sources distributed throughout the brain to represent neural activity [22].
• Spatial Filter Construction: Construct spatial filters that maximize signal from neural sources while minimizing contribution from artifact topographies [22].
• Signal Separation: Apply spatial filters to separate neural activity from artifact components in the recorded data [22].
• Source Localization: Use the cleaned data for source analysis, comparing results with known source locations of simulated or evoked responses [22].
• Validation Metrics: Evaluate performance using number of events surviving artifact threshold, signal-to-noise ratio of evoked responses, error in source localization, and signal variance explained by dipolar models [22].

This method has demonstrated particular effectiveness for source localization applications, with significantly better performance compared to OBS, BSS, and OBS-ICA approaches [22].

Visualization of BCG Artifact Dynamics and Processing

BCG Artifact Generation and Impact

BCG Artifact Origin and Consequences

BCG Artifact Removal Workflow

BCG Removal Processing Pipeline

Research Reagent Solutions and Tools

Table 3: Essential Research Materials and Tools for BCG Artifact Research

Research Tool	Specifications/Description	Primary Function in BCG Research
MRI-Compatible EEG Systems	64+ channels; SynAmps2 (Neuroscan) or equivalent; synchronized with MRI clock [22]	Acquires EEG data simultaneously with fMRI while minimizing interference
ECG/Pulse Oximetry	MRI-compatible electrodes or optical sensors; high sampling rate (≥1 kHz) [21]	Provides precise timing of cardiac events for artifact template creation
Reference Layer Hardware	Custom EEG caps with insulating layer and dedicated BCG electrodes [20]	Measures artifact signals without neural contamination for reference-based removal
Carbon Wire Loops	MRI-compatible conductive loops placed around head [22]	Alternative reference signal acquisition for motion-induced artifacts
PCA/ICA Software	EEGLAB, BESA Research, or custom implementations in MATLAB/Python [21] [22]	Decomposes signals for component-based artifact removal
High-Field MRI Systems	3T-7T scanners; increased gradient performance [17] [19]	Creates challenging high-amplitude BCG environments for method validation
Motion Tracking Systems	Optical (e.g., cameras) or piezoelectric sensors [20]	Quantifies head movement for motion-artifact correlation studies

The spatiotemporal dynamics and amplitude overlap characteristics of BCG artifacts present fundamental challenges for simultaneous EEG-fMRI research. The complex, time-varying nature of these artifacts, combined with their substantial amplitude and spectral overlap with neural signals of interest, necessitates sophisticated removal approaches that can separate artifacts from brain activity with minimal distortion. Methodologies such as harmonic regression, adaptive OBS, and surrogate spatial filtering have demonstrated promising results in addressing these challenges by leveraging the inherent structure of BCG artifacts while preserving neural signals. As EEG-fMRI continues to evolve as a tool for cognitive neuroscience and drug development, advancing our understanding of BCG artifact characteristics and refining removal methodologies remains essential for realizing the full potential of this powerful multimodal neuroimaging approach.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging approach that combines the millisecond temporal resolution of EEG with the millimeter spatial resolution of fMRI [1]. This integration enables researchers to investigate brain dynamics across complementary spatiotemporal scales, offering unprecedented insights into neural networks underlying cognitive processes, pathological states, and pharmacological interventions [3] [5]. However, the technique faces a significant technical hurdle: the ballistocardiogram (BCG) artifact, a persistent cardiac-induced contamination that profoundly compromises EEG signal quality and integrity [23] [20].

The BCG artifact arises from multiple physiological mechanisms associated with cardiac activity, including head movement caused by cardiac-induced momentum changes, local scalp pulsations from arterial expansion, and the Hall effect from pulsatile blood flow in the static magnetic field [20] [24]. In 3T MRI scanners, BCG artifacts can reach amplitudes exceeding 200 μV—tens of times larger than genuine neural signals—with their spectral power concentrated primarily below 25 Hz, directly overlapping with critical neural oscillatory rhythms and event-related potential components [25] [24]. This contamination presents a fundamental challenge for drug development professionals and neuroscientists relying on precise electrophysiological biomarkers, as residual BCG artifacts can generate spurious correlations between EEG and fMRI time-courses, potentially leading to erroneous interpretations of brain network dynamics [5].

Physiological Origins and Signal Characteristics of BCG Artifacts

The BCG artifact originates from three primary physical processes driven by the cardiac cycle, each contributing to the complex spatiotemporal contamination pattern observed in simultaneous EEG-fMRI recordings [20].

Physical Mechanisms and Temporal Dynamics

• Cardiac-Induced Head Movement: The sudden ejection of blood from the heart during systole generates reactive forces that cause subtle but significant head movements within the static magnetic field, inducing electrical currents in EEG electrodes via electromagnetic induction (Faraday's Law) [20]. This global head movement represents a primary contribution to the BCG artifact.

• Local Scalp Pulsations: Arterial pulsations beneath EEG electrodes cause mechanical displacements that generate potential differences as electrodes move relative to the scalp and magnetic field [20]. These localized effects vary across electrode positions depending on proximity to superficial arteries.

• Hall Effect in Pulsatile Blood Flow: As an electrically conductive fluid moving within the static magnetic field, blood generates potential differences across vessels—a phenomenon known as the Hall effect [20]. While contributing less amplitude than movement artifacts, this effect adds to the complex spatial distribution of BCG contamination.

The temporal relationship between cardiac events and BCG artifacts exhibits substantial variability, with BCG peaks typically occurring approximately 210 ms after the ECG R-peak, though this delay varies between subjects and across cardiac cycles [23]. This temporal jitter, combined with shape variability across heartbeats, complicates template-based removal approaches and distinguishes BCG from the more stereotyped gradient artifact [25].

Spectral and Spatial Properties

The spectral content of BCG artifacts predominantly resides below 25 Hz, directly overlapping with clinically and scientifically vital EEG rhythms including delta (1-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), and beta (13-30 Hz) bands [25] [24]. This spectral overlap presents particular challenges for investigating event-related potentials and neural oscillations, as conventional filtering cannot separate artifact from neural signals without substantial information loss.

Spatially, BCG artifacts demonstrate complex topographies that vary across electrodes and evolve over time due to factors including head position changes, blood pressure fluctuations, and electrode-scalp contact alterations [3] [5]. This spatial non-stationarity necessitates artifact removal approaches that adapt to changing contamination patterns throughout recording sessions.

Figure 1: Physiological mechanisms generating BCG artifacts in simultaneous EEG-fMRI. Cardiac activity produces multiple physical effects that collectively manifest as BCG contamination in EEG signals.

Methodological Approaches for BCG Artifact Removal

Multiple methodological approaches have been developed to address the BCG artifact challenge, each with distinct theoretical foundations, implementation considerations, and performance characteristics.

Traditional Signal Processing Approaches

Traditional BCG removal methods primarily operate on principles of template subtraction or blind source separation, leveraging the quasi-periodic nature of cardiac-related artifacts while attempting to preserve neural signals [26].

Average Artifact Subtraction (AAS) represents the earliest approach, creating artifact templates by averaging EEG segments time-locked to cardiac events [1] [25]. While effective for gradient artifacts, AAS performs suboptimally for BCG due to substantial shape variations across heartbeats, often leaving significant residuals that continue to obscure neural signals [5].

Optimal Basis Set (OBS) methods extend AAS by applying Principal Component Analysis (PCA) to artifact templates, using the first several principal components as an adaptive basis for artifact reconstruction and subtraction [5] [20]. This approach better captures inter-heartbeat variability but faces challenges in component selection, potentially removing neural information contained within lower-variance components [5].

Independent Component Analysis (ICA) employs blind source separation to decompose EEG data into statistically independent components, followed by manual or automated identification and removal of artifact-related components [1] [3]. While theoretically appealing, ICA assumes instantaneous mixing and statistical independence between neural signals and artifacts—assumptions often violated by BCG characteristics [3].

Advanced and Hybrid Methodologies

Adaptive Optimal Basis Set (aOBS) represents an OBS enhancement that dynamically estimates ECG-BCG delay on a beat-to-beat basis and automates component selection using variance-based criteria [5]. In comparative studies, aOBS demonstrated superior performance to standard OBS, reducing BCG residuals to 5.53% compared to 9.20% for OBS while better preserving neural signal integrity [5].

Clustering-Constrained ICA (ccICA) incorporates clustering algorithms to capture time-varying BCG features, constraining ICA decomposition to improve artifact component identification [24]. Validation studies demonstrated significantly improved performance over conventional ICA and OBS, particularly in preserving signal amplitude characteristics in both time and frequency domains [24].

Reference Layer Approaches utilize additional electrodes placed on conductive layers insulated from the scalp to directly measure BCG artifacts uncontaminated by neural signals [20]. These methods leverage the reference signals to regress artifacts from scalp EEG, demonstrating substantial improvements in signal recovery compared to software-based approaches, with one study reporting 101% improvement in alpha-wave contrast-to-noise ratios compared to OBS [20].

Emerging Machine Learning Approaches

BCGNet implements a deep learning architecture using Gated Recurrent Units (GRUs) to model nonlinear mappings between ECG and BCG-contaminated EEG, enabling accurate artifact prediction and subtraction [23]. This approach directly addresses the nonlinear relationship between cardiac events and BCG manifestations, outperforming OBS in power reduction at critical frequencies while improving task-relevant EEG classification accuracy [23] [27].

Real-Time Implementation approaches like EEG-LLAMAS provide low-latency BCG removal (under 50 ms) suitable for closed-loop EEG-fMRI paradigms, demonstrating superior power spectrum recovery compared to traditional methods [1].

Table 1: Comparative Performance of BCG Artifact Removal Methods

Method	Theoretical Basis	Key Advantages	Limitations	Performance Metrics
AAS [25]	Template averaging	Simple implementation; computationally efficient	Poor handling of BCG variability; significant residuals	BCG residuals: 12.51% [5]
OBS [5]	PCA of artifact templates	Adapts to artifact shape variations	Fixed component selection; potential neural signal removal	BCG residuals: 9.20% [5]
ICA [3]	Blind source separation	Does not require cardiac timing information	Subjective component selection; statistical assumptions	BCG residuals: 20.63% [5]
aOBS [5]	Adaptive PCA with beat-to-beat delay correction	Automated component selection; adaptive timing	Increased computational complexity	BCG residuals: 5.53%; Lowest cross-correlation with ECG [5]
ccICA [24]	ICA with clustering constraints	Captures time-varying artifact features	Complex implementation; parameter sensitivity	Lowest error in signal amplitude (Er) [24]
BRL [20]	Reference signal regression	Direct artifact measurement; minimal neural signal loss	Additional hardware required	101% improvement in alpha-wave CNR vs. OBS [20]
BCGNet [23]	Deep recurrent neural networks	Models nonlinear ECG-BCG relationships; improves task classification	Requires substantial training data	Superior power reduction at critical frequencies [23]

Experimental Protocols for BCG Method Validation

Rigorous validation of BCG artifact removal methods employs specialized experimental paradigms and analysis frameworks to quantify performance relative to ground truth neural signals.

Benchmarking with Simulated and Real Data

Hybrid Data Approach combines clean EEG recorded outside the MR environment with authentic BCG artifacts from inside-bore recordings, creating datasets with known neural signals contaminated by realistic artifacts [25]. This approach enables precise quantification of artifact removal efficacy and neural signal preservation by comparing processed outputs to original clean EEG.

Auditory Oddball Paradigms elicit well-characterized event-related potentials (particularly P300 components) during simultaneous EEG-fMRI, enabling assessment of how BCG removal impacts recovery of these neural markers [23]. Performance is quantified using signal-to-noise ratios, component amplitudes and latencies, and statistical power in single-trial analyses.

Visual Stimulation and Alpha Oscillation Protocols measure method performance in recovering steady-state visual evoked potentials and posterior alpha rhythms during eyes-open/closed conditions [5] [25]. These paradigms provide robust benchmarks for evaluating spectral power preservation in frequency bands particularly vulnerable to BCG contamination.

Quantitative Performance Metrics

Comprehensive method evaluation employs multiple quantitative metrics assessing both artifact reduction and signal preservation:

• BCG Residual Intensity: Quantifies remaining artifact power after processing, calculated as percentage of original artifact power [5]
• Cross-Correlation with ECG: Measures residual cardiac-related contamination in processed EEG [5]
• Signal-to-Noise Ratio (SNR): Computed for known evoked potentials before and after processing [5]
• Spectral Contrast Measures: Evaluates preservation of oscillatory power in frequency bands of interest (e.g., alpha-band contrast between eyes-open vs. eyes-closed conditions) [25]
• Dynamic Time Warping (DTW): Assesses temporal distortion introduced by processing algorithms [1]

Table 2: Standard Experimental Protocols for BCG Method Validation

Protocol Type	Neural Targets	Validation Metrics	Implementation Details
Hybrid Data Simulation [25]	Known clean EEG signals	Signal amplitude error; waveform correlation	Artifacts from resting-state fMRI added to outside-scanner EEG
Auditory Oddball [23]	P300 ERP components	Single-trial classification accuracy; SNR improvement	Target (20%) and standard (80%) tones; button press response
Visual Evoked Potentials [5]	Early visual ERP components (N75, P100, N145)	Component amplitude and latency; topographic accuracy	Pattern-reversal checkerboard stimulation; high-density EEG
Alpha Modulation [25]	Posterior alpha oscillations (8-13 Hz)	Spectral contrast ratio; topographic fidelity	Eyes-open vs. eyes-closed blocks; occipital ROI analysis
Resting-State Analysis [1]	Endogenous neural oscillations	Power spectral density; functional connectivity	5-10 minute resting recordings; graph theory metrics

Figure 2: Experimental workflow for validating BCG artifact removal methods. The approach combines ground truth data with comprehensive metric evaluation to quantify method performance.

Successful implementation of BCG artifact removal requires both specialized hardware components and software tools optimized for simultaneous EEG-fMRI investigations.

Table 3: Essential Research Resources for BCG Artifact Management

Resource Category	Specific Examples	Primary Function	Implementation Considerations
EEG Acquisition Systems	BrainAmp MR Plus; SynAmps2	High-quality EEG recording in MR environment	MR-compatibility; sampling rate ≥5 kHz; adequate voltage resolution [23] [3]
Artifact Monitoring Hardware	Carbon Wire Loops; Piezoelectric Sensors; Reference Layer Electrodes	Direct measurement of BCG artifacts independent of neural signals	Number and placement of reference sensors; insulation from scalp [25] [20]
Software Toolboxes	EEGLAB with FMRIB Plugin; BESA Research; MNE-Python	Implementation of standard BCG removal algorithms	Integration with data formats; batch processing capabilities; visualization tools [23] [3]
Custom Algorithm Platforms	EEG-LLAMAS; BCGNet	Advanced artifact removal using real-time processing or deep learning	Computational requirements; training data needs; latency constraints [1] [23]
Quality Assessment Tools	Power Spectral Analysis; Topographic Mapping; ERP Visualization	Quantitative evaluation of artifact removal efficacy	Comparison with ground truth; statistical analysis frameworks [5] [25]

Implications for EEG Analysis and Interpretation

The efficacy of BCG artifact removal directly impacts the validity and interpretability of EEG-derived neural measures in simultaneous fMRI studies, with particular significance for event-related potentials and neural oscillations.

Effects on Event-Related Potentials

Inadequate BCG artifact removal introduces temporally structured noise that directly obscures ERP components through amplitude reduction, latency jitter, and topographic distortion [5] [28]. The P300 component, a critical biomarker in cognitive neuroscience and pharmacological studies, proves particularly vulnerable due to its typical latency (250-500 ms post-stimulus) that often coincides with prominent BCG artifact peaks [23] [28]. Residual BCG contamination can reduce statistical power for detecting subtle cognitive effects or drug-induced changes, potentially requiring increased sample sizes up to 40% to maintain equivalent power [28].

Visual evoked potentials (VEPs) including the N75, P100, and N145 components demonstrate particular sensitivity to BCG residuals due to their maximal amplitudes over occipital regions where BCG artifacts often exhibit substantial expression [5]. Studies comparing BCG removal methods have demonstrated that inferior approaches can reduce VEP signal-to-noise ratios by over 50% compared to optimal methods, substantially impacting the detectability of these neural responses [5].

Effects on Neural Oscillations

Neural oscillations in the alpha band (8-13 Hz) represent particularly vulnerable targets for BCG contamination due to substantial spectral overlap and similar topographies [25]. Studies investigating alpha modulation during eyes-open versus eyes-closed conditions have demonstrated that conventional artifact removal methods can reduce alpha contrast-to-noise ratios by over 100% compared to reference-layer approaches, fundamentally compromising the detectability of this fundamental neural rhythm [20].

Oscillatory activity in lower frequency bands (delta, theta) faces even greater vulnerability due to higher BCG artifact power in these spectral regions [25]. This presents particular challenges for sleep research, disorder of consciousness studies, and developmental investigations where slow-wave activity provides crucial neurophysiological markers [26].

Impact on Functional Connectivity and Network Analysis

BCG artifacts introduce spurious correlations between EEG channels that profoundly distort functional connectivity measures and graph-theoretical network analyses [1]. Recent investigations have revealed method-specific differences in network topology following artifact removal, with AAS demonstrating superior signal fidelity while ICA showed greater sensitivity in dynamic graph metrics [1]. These method-dependent effects on network interpretation highlight the critical importance of appropriate artifact removal selection for studies investigating brain network dynamics.

BCG artifacts remain a significant challenge in simultaneous EEG-fMRI, directly obscuring ERPs and oscillatory rhythms through spectral and temporal overlap. The selection of artifact removal methodology profoundly impacts neural signal recovery, with advanced approaches including aOBS, reference layer methods, and deep learning demonstrating superior performance compared to conventional techniques. For researchers and drug development professionals utilizing simultaneous EEG-fMRI, method selection should be guided by specific neural targets of interest, with rigorous validation using appropriate performance metrics.

Future methodological developments will likely focus on real-time implementation for closed-loop paradigms, subject-specific adaptation to individual BCG characteristics, and multimodal integration that jointly optimizes both EEG and fMRI data quality. Through continued methodological refinement and rigorous validation, the full potential of simultaneous EEG-fMRI can be realized across basic neuroscience, clinical research, and pharmaceutical development applications.

Why BCG is a Greater Challenge than Gradient Artifacts

Simultaneous Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the millimeter spatial resolution of fMRI. This integration provides unprecedented insights into brain dynamics in both health and disease [25] [3]. However, the hostile electromagnetic environment inside the MRI scanner generates substantial artifacts that corrupt the delicate EEG signals, sometimes overwhelming neuronal activity by several orders of magnitude [26]. The two most prominent artifacts are the Gradient Artifact (GA) and the Ballistocardiogram (BCG) artifact. While both present significant technical challenges, the BCG artifact has proven notably more difficult to effectively mitigate, remaining a primary obstacle to obtaining clean, interpretable EEG data in simultaneous EEG-fMRI studies [20] [26].

This technical guide examines the fundamental reasons why BCG artifact reduction presents a greater challenge than gradient artifact correction. We explore the physiological and physical origins of both artifacts, compare their characteristics, evaluate current reduction methodologies, and synthesize experimental evidence demonstrating the superior intractability of the BCG artifact. Understanding these distinctions is crucial for researchers, scientists, and drug development professionals utilizing EEG-fMRI to investigate brain function or evaluate neurological therapeutics.

Physiological Origins and Physical Mechanisms

Gradient Artifact: A Deterministic Technical Phenomenon

The Gradient Artifact (GA) arises from the rapid switching of magnetic field gradients necessary for spatial encoding in fMRI, primarily during Echo-Planar Imaging (EPI) sequences. According to Faraday's law of induction, these time-varying magnetic fields induce electrical currents in any conductive loop, including those formed by EEG electrodes, their leads, and the patient's head [25] [29]. The induced voltage is proportional to the rate of change of the magnetic flux, leading to artifact amplitudes that can exceed 100 mV—more than 10,000 times larger than a typical evoked neural response [30].

The GA exhibits a deterministic and periodic structure dominated by harmonics of the slice repetition frequency convolved with harmonics of the volume repetition frequency [25]. Its morphology is precisely tied to the pre-programmed gradient coil switching sequence, making it highly predictable when scanner timing parameters are known [25] [31].

BCG Artifact: A Complex Physiological Phenomenon

The Ballistocardiogram (BCG) artifact originates from cardiac-induced movements within the static magnetic field (B0) of the scanner. Unlike the GA, the BCG is a physiologically-generated artifact with multiple contributing mechanisms:

• Cardiac-induced head movement: Each heartbeat causes a slight rotational movement of the head (often described as "nodding") due to momentum transfer from pulsatile blood flow [25] [32].
• Local scalp movements: Electrodes located near superficial blood vessels move with the pulsatile expansion and contraction of adjacent arteries [20] [32].
• Hall effect: As blood (an electrically conductive fluid) flows through the static magnetic field, it generates electrical potentials across blood vessels [25] [20].

The amplitude of the BCG artifact is directly proportional to the static magnetic field strength, typically exceeding 50 μV at 3T and increasing substantially at higher field strengths [25] [32]. Most of its spectral power resides below 25 Hz, creating significant overlap with the frequency range of most neuronal signals of interest [25].

The following diagram illustrates the primary mechanisms generating the BCG artifact:

Diagram 1: Primary physiological mechanisms generating the BCG artifact. The cardiac cycle interacts with the static magnetic field through multiple pathways to produce the complex BCG artifact.

Comparative Analysis: Key Distinguishing Challenges

The table below systematically compares the critical characteristics of GA and BCG artifacts that determine their relative difficulty to correct:

Table 1: Fundamental Characteristics Differentiating GA and BCG Artifacts

Characteristic	Gradient Artifact (GA)	BCG Artifact	Implication for Correction
Origin	Technical (gradient switching)	Physiological (cardiac activity)	GA predictable, BCG variable
Temporal Stability	Highly stable and periodic [25]	Variable timing and morphology [20]	Simple template effective for GA
Spectral Content	Primarily higher frequencies (>100 Hz) [30]	Overlaps neural signals (<25 Hz) [25]	Filtering removes GA but not BCG
Spatial Distribution	Consistent across repetitions [31]	Varies across cardiac cycles [32]	GA has fixed spatial pattern
Amplitude	Extremely high (up to 100 mV) [30]	Moderate (~50-200 μV) [20]	GA causes saturation issues
Primary Reduction Method	Average Artifact Subtraction (AAS) [29]	Multiple complex approaches needed [26]	GA solution is straightforward

The BCG artifact's temporal variability presents a particularly formidable challenge. While the GA repeats with nearly identical morphology across imaging cycles, the BCG artifact exhibits substantial beat-to-beat variations in shape, amplitude, and timing relative to the cardiac cycle [20] [5]. This variability stems from natural physiological fluctuations in heart rate, blood pressure, and head position, making simple averaging approaches insufficient.

Furthermore, the spatial complexity of the BCG artifact increases with magnetic field strength. At higher field strengths (3T and above), the spatial topography of the BCG becomes increasingly variable across electrodes and cardiac cycles [32]. This spatial non-stationarity violates key assumptions of many blind source separation techniques, including Independent Component Analysis (ICA), which assume stationary mixing of sources [5].

Methodological Approaches and Their Limitations

Gradient Artifact Reduction

The Average Artifact Subtraction (AAS) method, introduced by Allen et al. (2000), remains the cornerstone for GA reduction [25] [29]. This approach leverages the periodic nature of the GA by creating a template through averaging across multiple imaging cycles, then subtracting this template from the corrupted EEG signal. When combined with synchronization of EEG and MRI scanner clocks, AAS can effectively reduce the GA by over 95% [30] [31].

The workflow for this established GA reduction method is straightforward:

Diagram 2: Gradient artifact reduction using Average Artifact Subtraction (AAS). The deterministic nature of the GA enables effective reduction through template subtraction.

BCG Artifact Reduction

In contrast to the relatively straightforward GA reduction, BCG artifact correction requires more complex methodologies:

• Optimal Basis Set (OBS): An extension of AAS that applies Principal Component Analysis (PCA) to ECG-triggered EEG epochs. The first several principal components are used as adaptive templates for artifact subtraction [3] [5]. A significant limitation is determining the optimal number of components to remove without eliminating neural signals [5].

• Adaptive OBS (aOBS): An enhanced approach that estimates the delay between cardiac activity and BCG occurrence on a beat-to-beat basis and automatically identifies artifact-related components [5]. This method has demonstrated 36% lower BCG residuals compared to standard OBS [5].

• Independent Component Analysis (ICA): A blind source separation technique that identifies statistically independent components, some of which can be classified as BCG-related and removed [3] [32]. However, the spatial non-stationarity of the BCG artifact violates ICA's assumption of instantaneous mixing, limiting its effectiveness [5].

• Reference Layer Methods: Hardware-based approaches using additional electrodes placed on an electrically insulated conductive layer to record reference signals containing only BCG artifact [20]. These methods have shown 75-101% improvement in signal-to-noise ratios compared to OBS but require specialized equipment [20].

• Novel Computational Approaches: Emerging techniques include surrogate spatial filtering [3] and deep learning methods using Recurrent Neural Networks (RNNs) to model the nonlinear relationship between ECG and BCG artifacts [23].

The complexity of BCG reduction methodologies is evident in this comparative workflow:

Diagram 3: Complex methodological landscape for BCG artifact reduction. Multiple approaches exist, each with distinct limitations and requirements.

Experimental Evidence and Quantitative Comparisons

Performance Metrics in Method Evaluation

Studies systematically evaluating artifact reduction methods provide compelling quantitative evidence of the greater challenge posed by BCG artifacts. A comprehensive evaluation using hierarchical Bayesian probabilistic modeling found significant differences between methods in their ability to recover neural signals, with the Carbon-Wire Loop (CWL) reference method outperforming software-only approaches [25].

Research on the adaptive Optimal Basis Set (aOBS) method demonstrated its superiority over conventional approaches through several key metrics:

Table 2: Performance Comparison of BCG Artifact Reduction Methods

Method	BCG Residual Intensity (%)	Max Cross-Correlation with ECG	Signal-to-Noise Ratio (SNR)	Source Localization Error
Uncorrected EEG	-	0.180	-	-
AAS	12.51%	0.051	Low	High
ICA	20.63%	0.067	Moderate	Moderate
OBS	9.20%	0.042	Moderate	Moderate
aOBS	5.53%	0.028	High	Low
Reference Layer	-	-	Highest	Lowest

The data in Table 2, synthesized from experimental results [5], shows that even the best software-based methods leave significant BCG residuals (5.53% for aOBS), whereas GA reduction typically achieves near-complete artifact removal [29].

Impact on Neural Signal Analysis

The consequences of imperfect BCG removal extend to fundamental neural signal analyses:

• Event-Related Potentials (ERPs): Residual BCG artifacts can distort waveform morphology and timing, potentially generating spurious findings or obscuring genuine neural effects [20] [5].
• Neural Oscillations: BCG residuals in the alpha (8-12 Hz) and beta (15-30 Hz) bands can masquerade as or obscure genuine neural oscillations, compromising studies of functional brain networks [25] [26].
• Source Localization: Inaccurate BCG removal introduces spatial distortions that propagate through inverse modeling algorithms, reducing the validity of source localization [3].

Experimental evidence demonstrates that the reference layer approach improves contrast-to-noise ratios of alpha waves and visual evoked potentials by 101% and 76%, respectively, compared to OBS [20], highlighting how much neural signal is typically lost or distorted with conventional BCG removal.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Materials and Solutions for BCG Artifact Management

Tool/Reagent	Function/Purpose	Application Notes
Carbon Wire Loops (CWL)	Reference-based artifact recording [25]	Affordable hardware solution; requires setup modification
BCG Reference Layer (BRL)	Records artifact from insulated layer [20]	Reusable cap design; works with standard EEG systems
MR-Compatible EEG Systems	Basic recording in magnetic environment [26]	Must have high dynamic range (>20 bits)
ECG Recording Equipment	Cardiac event detection [5]	Essential for OBS, aOBS, and timing-based methods
Synchronization Interface	Aligns EEG and MRI scanner clocks [31]	Critical for optimal GA removal; reduces BCG variability
aOBS Software	Adaptive artifact template creation [5]	Automates component selection; beat-to-beat adaptation
RNN/Deep Learning Tools	Models nonlinear ECG-BCG relationships [23]	Emerging approach; requires significant training data

The Ballistocardiogram artifact presents a fundamentally greater challenge than the Gradient Artifact in simultaneous EEG-fMRI due to its physiological origins, complex spatio-temporal dynamics, and spectral overlap with neural signals of interest. While the GA is largely a solved problem through Average Artifact Subtraction with scanner synchronization, the BCG artifact continues to require sophisticated, multi-faceted approaches that must carefully balance artifact removal with neural signal preservation.

The persistent challenges in BCG artifact reduction have significant implications for researchers and drug development professionals utilizing simultaneous EEG-fMRI. Imperfect BCG removal can compromise data quality, potentially leading to spurious findings or reduced sensitivity to detect genuine neural effects—particularly critical in pharmaceutical trials where accurate assessment of neural target engagement is essential. Future methodological developments, particularly in reference-based recording and deep learning approaches, show promise for addressing these persistent challenges and finally overcoming the last major artifact barrier in simultaneous EEG-fMRI.

A Methodological Deep Dive: From Classic Algorithms to Next-Generation BCG Removal Techniques

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the millimeter spatial resolution of fMRI, offering unprecedented insights into brain dynamics in both health and disease [33]. However, the technique's significant potential is hampered by substantial artifacts that contaminate EEG signals recorded inside the MRI scanner. Among these, the ballistocardiogram (BCG) artifact persists as one of the most challenging technical obstacles in the field [4] [22].

The BCG artifact arises from multiple physiological phenomena associated with cardiac activity, including pulsatile movements of the scalp and electrodes due to blood flow, head rotation caused by cardiovascular momentum, and the Hall effect from pulsating blood in magnetic fields [25] [22]. With amplitudes typically exceeding 50 μV at 3 Tesla field strength, the BCG artifact often masks genuine neural signals, particularly because its spectral content predominantly falls below 25 Hz, directly overlapping with the frequency range of most clinically and scientifically relevant neural oscillations [25].

To address this persistent challenge, numerous artifact reduction methods have been developed, among which Average Artifact Subtraction (AAS) represents the foundational template-based approach that continues to influence contemporary methodologies [4] [34]. This technical guide examines the core principles of AAS, its methodological implementation, key limitations, and the evolutionary trajectory of BCG artifact reduction strategies that have built upon its template-based foundation.

The Core Principles of Average Artifact Subtraction (AAS)

Historical Development and Fundamental Assumptions

Average Artifact Subtraction emerged as one of the earliest systematic approaches for addressing BCG contamination in simultaneous EEG-fMRI recordings. Originally adapted from gradient artifact reduction methods, AAS was first applied to BCG artifacts by Allen and colleagues in the late 1990s [4] [35]. The method operates on two fundamental assumptions about the nature of the BCG artifact:

• Periodicity: BCG artifacts occur at approximately regular intervals synchronized with cardiac cycles [34].
• Stationarity: The artifact morphology remains relatively stable across successive heartbeats [25].

The core premise of AAS is that, while the artifact demonstrates some variability, its fundamental structure remains consistent enough to be characterized through averaging, whereas genuine neural activity represents random noise that will average toward zero over multiple repetitions [34].

The AAS Algorithm: A Step-by-Step Workflow

The AAS methodology follows a structured pipeline that transforms raw contaminated EEG into artifact-reduced data through template creation and subtraction:

Table 1: Key Steps in the Average Artifact Subtraction Algorithm

Step	Process	Implementation Details
1. Cardiac Event Detection	Identification of heartbeats	Using simultaneous ECG recording with QRS complex detection [34]
2. Epoch Extraction	Segmenting EEG around cardiac events	Typically using a fixed delay (e.g., 210 ms) after R-peak detection [23]
3. Template Creation	Averaging artifact segments	Channel-wise averaging of epochs to create artifact template [34]
4. Template Subtraction	Removing artifact from signal	Subtracting template from each individual artifact occurrence [25]
5. Signal Reconstruction	Generating continuous EEG	Rejoining corrected segments to produce clean continuous data [34]

Figure 1: The Average Artifact Subtraction (AAS) workflow demonstrates the sequential process from raw contaminated EEG to artifact-reduced signals through template-based subtraction.

Practical Implementation Considerations

Successful implementation of AAS requires careful attention to several technical parameters. The EEG sampling frequency must be sufficiently high (typically 1-2 kHz or more) to adequately capture the artifact morphology, and precise alignment of artifact epochs is critical for creating an accurate template [34]. Most implementations use a fixed time delay (approximately 210 ms) between the ECG R-peak and the assumed BCG peak, though this relationship can vary across individuals and recording sessions [23].

The AAS method is typically applied after initial reduction of the larger gradient artifacts induced by MRI sequence operation, as the tremendous amplitude of gradient artifacts would otherwise dominate the averaging process and preclude accurate BCG template creation [34].

Limitations of the AAS Approach

Despite its foundational status and conceptual simplicity, AAS suffers from several significant limitations that restrict its effectiveness in practical research settings.

Theoretical Constraints and Practical Shortcomings

The fundamental assumptions underlying AAS do not always hold true in experimental environments, leading to residual artifacts and signal distortion:

Table 2: Key Limitations of Average Artifact Subtraction

Category	Specific Limitation	Impact on Data Quality
Temporal Variability	Non-stationary artifact morphology	Residual artifacts due to template mismatch [25]
Physiological Complexity	Multiple concurrent artifact sources	Incomplete artifact capture [22]
Temporal Misalignment	Variable ECG-BCG timing	Smearing and residual artifacts [5]
Neural Signal Distortion	Non-random neural activity	Unintended removal of brain signals [25]
Template Selection	Fixed component selection	Under- or over-correction of artifacts [5]

• Temporal Variability and Non-Stationarity: The BCG artifact originates from complex physiological processes including head movement, scalp pulsation, and blood flow effects that exhibit natural variability across cardiac cycles [25] [22]. This violates the stationarity assumption of AAS, resulting in a template that does not perfectly match individual artifact instances, thus leaving residual artifacts after subtraction [25].

• Neural Signal Distortion: When neural activity of interest happens to be time-locked to cardiac cycles (either by chance or due to physiological coupling), it becomes correlated with the artifact template and may be partially subtracted along with the artifact [25]. This problem is particularly acute for event-related potentials and steady-state oscillations that might demonstrate consistent phase relationships to heartbeats in certain experimental paradigms.

• Imperfect Temporal Alignment: AAS typically relies on ECG R-peaks as temporal anchors for artifact epochs, but the relationship between electrical cardiac events and mechanical BCG artifacts is not fixed [5]. The commonly used 210 ms fixed delay does not account for physiological variations in pulse transmission time, leading to misalignment and subsequent template smearing [23].

• Component Selection Challenges: The method provides no objective criteria for determining optimal template complexity, potentially resulting in underfitting or overfitting of artifact representations [5].

Empirical Evidence of Limitations

Experimental comparisons have consistently demonstrated these limitations in practical applications. Studies evaluating AAS against other methods have found significant residual BCG contamination remaining after correction, particularly in the alpha and beta frequency bands where crucial neural oscillations reside [25]. When compared with reference EEG recorded outside the scanner, AAS-corrected signals show reduced fidelity in recovering visual evoked responses and oscillatory activity during motor tasks [25].

One study noted that "simply subtracting the average heartbeat waveform within a predetermined interval can introduce errors into the processed EEG because of the changes in cardiac wave duration and morphology" [35]. This results in residual correlations between the corrected EEG and cardiac activity, potentially generating spurious correlations in subsequent EEG-fMRI integration analyses [5].

Evolution Beyond AAS: Advanced Artifact Reduction Methods

Recognition of AAS limitations has driven the development of more sophisticated approaches that address its fundamental constraints while building upon its template-based foundation.

Optimal Basis Set (OBS) and Adaptive Variants

The Optimal Basis Set (OBS) method extends AAS by applying Principal Component Analysis (PCA) to the artifact epochs, creating a set of basis functions that capture the dominant modes of artifact variation [34] [5]. Instead of subtracting a simple average template, OBS identifies and removes the first several principal components that represent artifact-related variance:

Figure 2: Evolutionary development of AAS-based methods showing how Optimal Basis Set (OBS) and its adaptive variants address fundamental limitations of the original approach.

The adaptive OBS (aOBS) method further advances this approach by incorporating beat-to-beat estimation of the delay between cardiac activity and BCG occurrence, addressing the critical misalignment limitation of standard AAS [5]. This adaptive timing alignment, combined with automated component selection based on explained variance, has been shown to reduce BCG residuals significantly while better preserving neural signals compared to traditional AAS [5].

Independent Component Analysis (ICA) and Hybrid Approaches

Independent Component Analysis (ICA) represents a fundamentally different approach that decomposes multi-channel EEG signals into statistically independent components, allowing for identification and removal of artifact-related sources based on their temporal, spectral, and spatial characteristics [22] [35].

Hybrid methods that combine OBS with ICA (OBS-ICA) leverage the strengths of both approaches: OBS first reduces the bulk of the artifact, while ICA subsequently removes residual contamination [22]. This sequential approach mitigates the risk of ICA struggling with high-amplitude artifacts while providing more precise separation of neural activity from residual artifacts [22].

Reference Layer and Hardware Solutions

Hardware-based approaches using carbon wire loops (CWL) or reference layers provide alternative solutions by directly recording the artifact uncontaminated by neural signals [25] [22]. These systems capture purely artifact-related potentials through additional electrodes or conductive loops isolated from the scalp, generating a reference signal that can be subtracted from the contaminated EEG.

Empirical evaluations have demonstrated the superiority of CWL systems, with studies showing "the CWL system proved superior to the other methods evaluated in improving spectral contrast in the alpha and beta bands and in recovering visual evoked responses" [25]. However, these hardware solutions require additional equipment and setup complexity that may not be feasible in all research environments.

Emerging Machine Learning Approaches

Recent advances have introduced deep learning architectures, particularly Recurrent Neural Networks (RNNs) with Gated Recurrent Units (GRUs), to model the complex nonlinear relationships between ECG and BCG artifacts [27] [23]. These methods can capture the state-dependent dynamics of artifact generation without relying on rigid templates or linear assumptions, potentially offering superior artifact suppression while preserving task-relevant neural activity [23].

The Researcher's Toolkit: Method Comparison and Implementation Guidelines

Comparative Analysis of BCG Artifact Reduction Methods

Table 3: Comprehensive Comparison of BCG Artifact Reduction Techniques

Method	Core Principle	Advantages	Limitations	Best Application Context
AAS	Template subtraction	Simple implementation; Computationally efficient [34]	Residual artifacts; Neural signal distortion [25]	Initial methodology development; Educational purposes
OBS	PCA-based component removal	Captures artifact variability [5]	Subjective component selection [5]	Standard research applications with stable artifacts
aOBS	Adaptive PCA with dynamic alignment	Automatic component selection; Beat-to-beat timing [5]	Increased computational complexity [5]	High-density EEG; Variable heart rate conditions
ICA	Blind source separation	No cardiac reference needed; Handles multiple artifacts [35]	Uncertain component selection; Spatial stationarity assumption [22]	Studies with limited ECG quality; Multi-artifact scenarios
OBS-ICA	Hybrid sequential approach	Reduced residuals; Complementary strengths [22]	Complex pipeline; Potential overcorrection [22]	High-quality publication research; Clinical applications
CWL	Hardware reference subtraction	Direct artifact measurement; Superior performance [25]	Additional equipment required [22]	Resource-rich environments; Critical applications

Experimental Implementation Framework

For researchers implementing BCG artifact reduction methods, the following practical guidelines emerge from the literature:

• Preprocessing Requirements: AAS and its derivatives require high-quality EEG data with minimal saturation artifacts, recorded at high sampling rates (≥1 kHz) to adequately resolve artifact morphology [34]. Gradient artifact removal must precede BCG correction to avoid overwhelming the BCG template with higher-amplitude imaging artifacts [34].

• Quality Control Metrics: Effective implementation should include quantification of BCG residual intensity (post-correction artifact amplitude), cross-correlation with ECG, and signal-to-noise ratio preservation in known neural responses [5].

• Parameter Optimization: Critical parameters requiring empirical optimization include epoch window duration, template complexity (number of OBS components), and temporal alignment strategies, which should be validated for specific experimental paradigms and participant populations [5].

• Validation Approaches: Whenever feasible, method performance should be assessed against reference EEG recordings outside the scanner or through simulated artifact addition to quantify neural signal preservation [25] [22].

Average Artifact Subtraction represents the foundational template-based approach for BCG artifact reduction in simultaneous EEG-fMRI, establishing the critical principle that artifact periodicity can be leveraged to separate contamination from neural signals. However, its limitations—particularly regarding artifact non-stationarity, temporal variability, and neural signal distortion—have driven the development of increasingly sophisticated solutions including OBS, aOBS, ICA-based approaches, and hardware reference systems.

The evolutionary trajectory from simple AAS to adaptive, multivariate, and hybrid methods reflects the growing recognition that BCG artifacts originate from complex, multi-faceted physiological processes that cannot be fully captured by static templates. Future directions will likely involve increased incorporation of subject-specific and state-aware modeling approaches, potentially leveraging machine learning to dynamically adapt to changing artifact characteristics throughout recording sessions.

For the practicing researcher, method selection should be guided by specific research goals, available resources, and data quality requirements, with the understanding that while AAS provides a conceptual foundation, contemporary research demands more advanced solutions to ensure the validity of neural interpretations in simultaneous EEG-fMRI studies.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging approach that combines the millisecond temporal resolution of EEG with the high spatial specificity of fMRI. This integration enables researchers to investigate brain function across complementary spatiotemporal scales, shedding light on neural oscillations during various brain states, event-related potentials, and the neurovascular coupling underlying brain activity [5]. However, the technical challenges of acquiring clean EEG data inside the MRI scanner have significantly hampered the potential of this integrated methodology. The EEG signals recorded during fMRI are contaminated by substantial artifacts, with the ballistocardiogram (BCG) artifact standing as the most persistent and challenging to remove [5] [36].

The BCG artifact arises from multiple cardiac-related phenomena, including head rotation caused by cardiac-induced body movement, pulse-driven expansion of the scalp, and the Hall effect generated by pulsatile blood flow as an electrically conductive fluid moving within the static magnetic field [5] [36]. With an amplitude typically 3-8 times greater than physiological EEG signals, the BCG artifact can completely obscure neural activity of interest, potentially generating spurious correlations or masking true relationships between EEG and fMRI time-courses [5] [24]. Unlike the gradient artifact caused by switching magnetic fields, which can be effectively removed using Average Artifact Subtraction (AAS), the BCG artifact exhibits complex spatio-temporal dynamics with considerable variability in shape, amplitude, and timing across cardiac cycles, electrodes, and subjects [34]. This non-stationary nature has made BCG artifact removal an ongoing research challenge, with the Optimal Basis Set (OBS) method emerging as a particularly effective solution that leverages Principal Component Analysis (PCA) for adaptive subtraction.

The Core Principles of the OBS Method

Fundamental Concepts and Historical Development

The Optimal Basis Set method was first introduced by Niazy et al. (2005) as a significant advancement over simple average artifact subtraction techniques [37] [38]. The fundamental innovation of OBS lies in its recognition that the BCG artifact, while time-locked to cardiac activity, exhibits substantial temporal variations that cannot be captured by a simple average template. Where AAS methods subtract a single average artifact template from each occurrence, OBS employs Principal Component Analysis to identify a set of basis functions that collectively describe the dominant modes of artifact variance across multiple cardiac cycles [38].

The core premise is that PCA applied to EEG segments time-locked to cardiac events (typically identified via ECG R-peaks) identifies the principal components that optimally capture the variance in BCG artifact morphology [5] [38]. By subtracting a linear combination of these principal components—rather than just a simple average—from each artifact occurrence, OBS can adapt to variations in artifact shape while preserving neural signals that are statistically independent from the artifact structure. This approach effectively addresses the key limitation of AAS, which assumes a relatively invariant artifact morphology across occurrences [34].

The Mathematical Foundation

Mathematically, the OBS method operates through a multi-stage decomposition process. For each EEG channel, the data is first epoched into segments time-locked to BCG artifact occurrences (typically identified via ECG R-peaks with an appropriate delay). Let us denote these artifact epochs as ( A_i(t) ) for the i-th occurrence. The method then proceeds through the following computational stages:

• Temporal PCA: The ensemble of artifact epochs ( {A1(t), A2(t), ..., AN(t)} ) is subjected to temporal Principal Component Analysis, which identifies the orthogonal basis functions ( {PC1(t), PC2(t), ..., PCK(t)} ) that optimally explain the variance in artifact morphology across epochs [38].

• Basis Set Selection: A subset of the principal components (typically the first 3-8 components that explain the majority of variance) is selected to form the Optimal Basis Set [5].

• Artifact Reconstruction: For each artifact occurrence, the BCG artifact is reconstructed as a linear combination of the selected basis functions: ( \hat{A}i(t) = \sum{j=1}^{M} c{ij} PCj(t) ), where ( c_{ij} ) are coefficients determined for each occurrence [38].

• Adaptive Subtraction: The reconstructed artifact ( \hat{A}_i(t) ) is subtracted from the original EEG signal in each epoch, leaving a residual signal that ideally contains minimal artifact contamination while preserving neural activity [37].

The number of principal components to include in the basis set represents a critical parameter balancing artifact removal against potential signal distortion. While early implementations typically used fixed numbers (often 3-4 components), advanced implementations employ data-driven criteria based on explained variance or other statistical measures [5].

Comparative Performance of BCG Artifact Removal Methods

Table 1: Quantitative Comparison of BCG Artifact Removal Methods Based on Experimental Studies

Method	BCG Residual Intensity (%)	Max Cross-Correlation with ECG	Key Advantages	Key Limitations
AAS	12.51%	0.051	Simple implementation; Computationally efficient [34]	Assumes stationary artifact; Leaves substantial residuals [5]
ICA	20.63%	0.067	Blind source separation; No need for cardiac timing [5]	Assumes spatial stationarity; Subjective component selection [34]
Standard OBS	9.20%	0.042	Adapts to artifact shape variations; More effective than AAS [5]	Sensitive to peak detection accuracy; Fixed component number [5]
aOBS (Adaptive OBS)	5.53%	0.028	Beat-to-beat delay estimation; Automatic component selection [5]	More complex implementation; Computational demands [5]
PROJIC-OBS	N/A	N/A	Excellent artifact removal when prioritized [36]	Complex pipeline; Parameter sensitivity [36]
PROJIC-AAS	N/A	N/A	Best for physiological signal preservation [36]	Complex pipeline; Parameter sensitivity [36]
Surrogate Methods (PCA-S/ICA-S)	N/A	N/A	Superior source localization; Minimal brain signal distortion [3] [22]	Requires artifact topography knowledge; Semi-automated [3]

Table 2: Advanced OBS Variants and Their Technical Innovations

Method Variant	Core Innovation	Component Selection Approach	Temporal Alignment Handling	Reference
Standard OBS	PCA basis set instead of simple average	Fixed number (typically 3-4 PCs)	Fixed ECG-BCG delay assumption	[38]
aOBS (Adaptive OBS)	Beat-to-beat delay estimation	Automatic based on explained variance	Adaptive delay estimation	[5]
PROJIC-OBS	OBS correction in ICA component space	Projection and mutual information criteria	Combination of ICA and OBS advantages	[36]
Real-time OBS	Modified for online implementation	Fixed basis set with adaptation	Simplified for computational efficiency	[24]
OBS-ICA Combination	Sequential application of both methods	ICA after OBS for residual removal	Two-stage approach to address limitations	[34]

Experimental Protocols and Methodological Implementation

Standard OBS Implementation Protocol

The implementation of the Optimal Basis Set method for BCG artifact removal follows a structured pipeline with clearly defined stages. The following protocol is adapted from the original FMRIB Toolkit implementation and subsequent refinements [34] [38]:

• Gradient Artifact Removal: Before BCG artifact removal, gradient artifacts must be thoroughly removed using AAS or improved methods like FASTR (FMRI Artifact Slice Template Removal), which incorporates PCA-based correction for residual variances [38].

• Cardiac Event Detection: ECG R-peaks are detected using robust algorithms capable of handling MR environment noise. The Christov algorithm or modified versions incorporating complex lead computation and k-Teager energy operators have proven effective [38] [24].

• BCG Event Identification: A fixed delay (typically 210 ms) is applied to R-peaks to approximate BCG artifact onset, though advanced implementations estimate this delay adaptively on a beat-to-beat basis [5] [23].

• Data Epoching: EEG data are segmented into epochs time-locked to the identified BCG events, typically spanning intervals from 100-500 ms relative to the identified onset [38].

• Principal Component Analysis: Temporal PCA is applied to the ensemble of artifact epochs for each channel separately, generating orthogonal basis functions that capture artifact variance patterns [38].

• Basis Set Selection: The first M principal components (typically 3-8, determined by explained variance criteria or fixed thresholds) are selected to form the optimal basis set [5].

• Artifact Reconstruction and Subtraction: For each artifact occurrence, the basis functions are fitted via linear regression and subtracted from the original signal [38].

• Quality Assessment: Residual artifact is quantified using metrics such as BCG residual intensity, cross-correlation with ECG, or spectral power at cardiac frequencies [5].

Diagram 1: The OBS Method Workflow - This diagram illustrates the sequential stages of the Optimal Basis Set method for BCG artifact removal, from initial preprocessing to final quality assessment.

The Adaptive OBS (aOBS) Enhancement

Building upon the standard OBS method, Marino et al. (2018) introduced an adaptive OBS (aOBS) variant that addresses two critical limitations: fixed component selection and assumed timing between cardiac events and BCG occurrences [5]. The aOBS protocol incorporates these key enhancements:

• Beat-to-Beat Delay Estimation: Rather than assuming a fixed delay between ECG R-peaks and BCG events, aOBS estimates this delay adaptively for each cardiac cycle using cross-correlation or similar algorithms, improving artifact alignment before PCA [5].

• Automatic Component Selection: Instead of using a fixed number of principal components, aOBS implements data-driven selection criteria based on explained variance ratios, allowing the number of components to vary across channels and subjects (typically 1-8 components based on experimental data) [5].

• Improved BCG Peak Detection: aOBS can operate using BCG peaks directly identified from the EEG data itself, circumventing potential issues with ECG-based detection in high-field environments [5].

Experimental validation of aOBS demonstrated significantly improved performance compared to standard OBS, with BCG residual intensity reduced to 5.53% compared to 9.20% for standard OBS, and maximum cross-correlation with ECG reduced to 0.028 compared to 0.042 [5].

Table 3: Essential Research Tools and Resources for OBS Implementation

Tool Category	Specific Tools/Software	Key Function	Implementation Considerations
Data Acquisition Hardware	MR-compatible EEG systems (BrainAmp MR Plus, SynAmps2)	EEG data collection in MRI environment	Amplifier synchronization; Sampling rate ≥5000 Hz [23]
Pulse Detection Tools	Modified Christov algorithm; k-Teager energy operator	Robust ECG R-peak detection in MR environment	Handling of high-frequency noise; Adaptive thresholds [24]
Computational Libraries	PCA algorithms (SVD, EVD)	Basis set calculation	Temporal implementation; Component sorting [38]
Software Platforms	EEGLAB with FMRIB Plugin; BrainVoyager; BESA Research; MNE-Python	Integrated pipeline implementation	GRA removal prerequisites; Visualization capabilities [34] [23]
Validation Metrics	BCG residual intensity; Cross-correlation with ECG; SNR; INPS	Method performance quantification	Comparison with ground truth; Statistical testing [5] [24]

Advanced Hybrid Methodologies

Integration with Independent Component Analysis

Recognizing the complementary strengths and limitations of different approaches, researchers have developed hybrid methodologies that integrate OBS with Independent Component Analysis (ICA). These integrated approaches typically follow one of two pathways:

• OBS followed by ICA: The standard OBS method is first applied to remove the bulk of the BCG artifact, followed by ICA decomposition to identify and remove residual artifact components that may have survived the initial correction [34]. This approach benefits from OBS's effectiveness in capturing the primary artifact structure while using ICA to address residuals.

• ICA followed by OBS (PROJIC approaches): EEG data is first decomposed via ICA, after which OBS correction is applied specifically to components identified as BCG-related, before reconstructing the signal from all corrected components [36]. This approach aims to preserve neural signals that might be compromised by direct channel-wise OBS application.

The PROJIC-OBS and PROJIC-AAS methods represent sophisticated implementations of this latter approach, demonstrating superior performance in scenarios where preservation of physiological signals is prioritized alongside effective artifact removal [36].

Diagram 2: Hybrid OBS-ICA Methodologies - This diagram illustrates two predominant approaches for integrating OBS with Independent Component Analysis, each with distinct advantages for different research scenarios.

Emerging Approaches and Future Directions

The evolution of BCG artifact removal continues with several promising research directions building upon the OBS foundation:

• Surrogate Methods: Recent approaches using surrogate source models separate artifact and brain signals by leveraging known spatial distributions, demonstrating superior performance in source localization tasks compared to traditional OBS [3] [22].

• Deep Learning Architectures: Novel methods employing Recurrent Neural Networks (RNNs) with Gated Recurrent Units (GRUs) have been developed to learn the non-linear mapping between ECG and BCG artifacts, showing potential for improved artifact prediction without the linear assumptions inherent in PCA-based methods [23].

• Hardware-Based Solutions: Reference layer approaches and carbon wire loops aim to directly measure artifact signals for subtraction, providing complementary hardware solutions that can potentially be integrated with algorithmic approaches like OBS [3].

• Real-Time Implementations: Modified OBS algorithms with simplified basis sets and adaptive mechanisms are being developed for real-time applications, particularly relevant for neurofeedback and clinical monitoring scenarios [24].

The Optimal Basis Set method represents a significant methodological advancement in the challenge of BCG artifact removal in simultaneous EEG-fMRI. By leveraging Principal Component Analysis to capture the temporal variations in BCG artifact morphology, OBS and its adaptive variants address fundamental limitations of simpler template subtraction approaches. The continued evolution of OBS methodology—through improved component selection, adaptive timing estimation, and integration with complementary approaches like ICA—demonstrates its enduring relevance in the multimodal neuroimaging toolkit.

As simultaneous EEG-fMRI applications expand into clinical domains and more complex cognitive neuroscience questions, the importance of robust, validated artifact removal methods only grows. The OBS method, particularly in its enhanced forms, provides researchers with a powerful approach for extracting clean neural signals from the challenging MR environment, enabling more reliable investigation of brain function across spatiotemporal scales. Future developments will likely focus on increasing automation, improving real-time capabilities, and further optimizing the balance between artifact removal and physiological signal preservation.

Independent Component Analysis (ICA) is a computational method for separating a multivariate signal into additive, statistically independent subcomponents. This technique is a fundamental tool in the field of Blind Source Separation (BSS), where the goal is to recover underlying source signals from observed mixtures without prior knowledge of the mixing process or the sources themselves [39]. The power of ICA lies in its ability to exploit the statistical independence of source signals, making it particularly valuable for analyzing complex data where multiple components are mixed together.

In practical terms, ICA addresses the "cocktail party problem" – the challenge of isolating individual speakers' voices from recorded mixtures in a noisy room [39]. Formally, the standard linear ICA model assumes that observed data points x = (x₁, ..., xₘ)ᵀ are linear combinations of n unknown independent components sₖ. The model can be expressed as x = As, where A is an unknown mixing matrix, and s represents the independent sources [39]. The objective is to estimate both A and s from only the observed x, typically by finding a separating matrix W such that s = Wx provides estimates of the original sources.

For ICA to be successful, certain assumptions must be met: the source signals must be statistically independent of each other, and they must have non-Gaussian distributions (with possibly one Gaussian exception) [39]. The statistical independence property is crucial – unlike principal component analysis (PCA), which merely decorrelates signals (removes second-order dependencies), ICA addresses higher-order statistical dependencies to achieve true independence [40].

Core Mathematical Principles of ICA

The ICA Model and Identifiability

The mathematical foundation of ICA rests on a generative model where observations are linear mixtures of independent sources. For an observed M-dimensional random vector x = [x₁, x₂, ⋯, xₘ]ᵀ, the model is:

x = As

Here, s = [s₁, s₂, ⋯, sₙ]ᵀ is an N-dimensional vector whose elements are the random variables representing independent sources, and A is an M×N mixing matrix [40]. Typically, M ≥ N, making A usually of full rank. The goal is to estimate an unmixing matrix W such that:

y = Wx

provides a good approximation to the true sources s [40]. The independent components are identifiable up to permutation and scaling of the sources, meaning ICA cannot determine the original order or magnitude of the sources [39].

Preprocessing: Centering and Whitening

Most ICA algorithms employ critical preprocessing steps to simplify and reduce the complexity of the problem:

• Centering: The observed data x is centered by subtracting its mean, creating a zero-mean signal [39]. This simplifies the subsequent ICA estimation.
• Whitening (or sphering): The centered data is transformed so that its components become uncorrelated and their variances equal unity [41] [39]. Mathematically, whitening creates a new vector z such that E{zzᵀ} = I, where I is the identity matrix. Whitening removes second-order dependencies and prepares the data for the full ICA decomposition [41].

Whitening has a geometrical interpretation: it restores the initial "shape" of the data, after which ICA need only rotate the resulting matrix to find the independent components [41]. From a signal processing perspective, whitening can be achieved using eigenvalue decomposition of the covariance matrix Σₓ = E{xxᵀ}. If Σₓ = EDEᵀ, where E is the orthogonal matrix of eigenvectors and D is the diagonal matrix of eigenvalues, the whitening matrix is D⁻¹/²Eᵀ [39].

Algorithms for ICA Estimation

Several algorithms have been developed to solve the ICA problem, primarily differing in how they define and maximize independence:

• Infomax: This approach, based on information theory, maximizes the mutual information between the inputs and outputs of a neural network [42]. The Infomax principle is implemented in popular toolboxes like EEGLAB for processing electrophysiological data [40] [42].
• FastICA: A widely used fixed-point algorithm that maximizes non-Gaussianity through approximations of negentropy [40]. FastICA is computationally efficient and has found applications in numerous domains.
• JADE (Joint Approximate Diagonalization of Eigenmatrices): This algorithm performs ICA through joint diagonalization of fourth-order cumulant matrices [43]. JADE is particularly valued for its statistical stability and ability to work "off-the-shelf" without parameter tuning.

These algorithms typically work with a fixed nonlinearity or one selected from a small set. Both Infomax and FastICA perform well for symmetric distributions but are less accurate for skewed distributions or sources close to Gaussian [40].

Table 1: Key ICA Algorithms and Their Characteristics

Algorithm	Optimization Principle	Key Features	Common Applications
Infomax [40] [42]	Maximization of mutual information	Information-theoretic approach; often used with natural gradient	EEG artifact removal [42]
FastICA [40]	Maximization of non-Gaussianity via negentropy	Fast convergence; fixed-point iteration	General signal processing
JADE [43]	Joint diagonalization of cumulant matrices	Statistical stability; no parameter tuning	Biomedical signals, communications

ICA in EEG-fMRI and the BCG Artifact Challenge

Simultaneous EEG-fMRI: Opportunities and Artifacts

Simultaneous recording of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) represents a powerful multimodal neuroimaging approach that combines the millisecond temporal resolution of EEG with the millimeter spatial resolution of fMRI [44]. This integration provides unique insights into brain dynamics across different spatial and temporal scales, facilitating research into neural oscillations, cognitive processes, and various neurological disorders [5] [22].

However, EEG signals recorded inside MRI scanners suffer from severe contamination by two main artifacts: gradient artifacts caused by switching magnetic fields during image acquisition, and ballistocardiogram (BCG) artifacts resulting from cardiac-related activities [44] [5]. While gradient artifacts can be effectively removed using average artifact subtraction (AAS) due to their high reproducibility [5], BCG artifacts present a more challenging problem.

Nature and Mechanisms of BCG Artifacts

The BCG artifact is a complex signal distortion with multiple physiological and physical origins. Key mechanisms include:

• Cardiac-induced electrode movement: Pulsatile blood flow causes small but firm movements of EEG electrodes and the scalp as arteries expand and contract between systolic and diastolic phases [42].
• Hall effect of blood flow: As an electrically conductive fluid, pulsatile blood movement in the static magnetic field induces electrical potentials (Hall voltage) on the scalp surface [5] [22].
• Cardiac-related body movements: The entire body experiences small movements tied to the cardiac cycle when in a supine position [42] [22].

The BCG artifact exhibits several challenging characteristics: its amplitude can reach 400 μV in a 3.0T scanner (6-8 times larger than typical EEG signals) [44], and it shows considerable variation in shape, amplitude, and scale over time [44]. Furthermore, the artifact predominantly affects frequency bands important for EEG analysis, particularly alpha frequencies (8-13 Hz) and below [44].

ICA-Based Methodologies for BCG Artifact Removal

Standard ICA Approaches

The application of ICA to BCG artifact removal leverages the statistical independence between neural signals and artifact components. The general procedure involves:

• Data preparation: EEG data is first corrected for gradient artifacts using methods like AAS [5].
• ICA decomposition: ICA algorithm (e.g., Infomax or FastICA) is applied to separate the multichannel EEG data into independent components (ICs) [42].
• Component classification: ICs representing BCG artifacts are identified based on their temporal, spectral, and spatial characteristics [42] [22].
• Artifact removal: Identified BCG components are subtracted from the data, which is then reconstructed using the remaining components [42].

The effectiveness of standard ICA stems from its ability to separate BCG artifacts into a small number of components (typically 5-6) that can be selectively removed [42]. Studies have demonstrated that ICA-based procedures significantly reduce spectral power at frequencies associated with BCG artifacts and outperform simpler template subtraction methods [42].

Advanced and Hybrid Methods

To address limitations of standard ICA, several advanced methodologies have been developed:

• Clustering-constrained ICA (ccICA): This approach combines clustering algorithms with constrained ICA to capture the time-varying features of BCG artifacts [44]. The method first applies clustering to identify BCG artifact features, then uses constrained ICA to separate these artifacts from neural signals. ccICA has demonstrated superior performance compared to traditional methods like AAS, OBS, and basic cICA in both simulated and real EEG data [44].

• Optimal Basis Set (OBS) and adaptive OBS (aOBS): OBS extends AAS by applying principal component analysis (PCA) to EEG segments time-locked to cardiac events, using the first few principal components for adaptive artifact removal [5]. The adaptive OBS (aOBS) method further improves this by estimating the delay between cardiac activity and BCG occurrence on a beat-to-beat basis, ensuring more accurate alignment of BCG occurrences [5]. aOBS automatically determines which PCA components are artifact-related based on explained variance criteria [5].

• Surrogate Methods: These approaches use surrogate source models to separate artifact-related signals from brain signals with minimal distortion [22]. Artifact topographies are obtained either through PCA on an averaged artifact template (PCA-S) or by manual selection of artifact components using ICA (ICA-S) [22]. These methods have shown significant improvements in source localization accuracy compared to established techniques like OBS and BSS [22].

Table 2: Comparison of BCG Artifact Removal Methods

Method	Key Principle	Advantages	Limitations
AAS [42] [5]	Average artifact subtraction using ECG R-peaks	Simple implementation; computationally efficient	Cannot handle shape variability; residual artifacts
Standard ICA [42]	Blind source separation into independent components	Handles complex artifact morphology; no reference needed	Subjective component selection; potential signal loss
OBS [5]	PCA on artifact epochs + subtraction	Adapts to artifact shape variations	Fixed component number; sensitive to timing accuracy
aOBS [5]	Adaptive detection + PCA component selection	Automatic component selection; better alignment	Complex implementation
ccICA [44]	Clustering + constrained ICA	Captures time-varying features; improved accuracy	Computationally intensive
Surrogate Methods [22]	Spatial filtering with surrogate models	Minimal brain signal distortion; improved source analysis	Requires additional modeling

Experimental Protocols and Implementation

Data Acquisition and Preprocessing

For simultaneous EEG-fMRI experiments focused on BCG artifact removal, standard protocols include:

• EEG Acquisition: High-density EEG systems (typically 64 channels) with MR-compatible amplifiers, sampling rates of 1,000-10,000 Hz, synchronized with the MRI scanner clock [44] [22]. Additional electrodes are placed for electrocardiogram (ECG) recording to detect cardiac events.
• fMRI Acquisition: Standard echo-planar imaging (EPI) sequences on 3T scanners with parameters such as TR=2000 ms, TE=30-35 ms, flip angle=90°, and appropriate spatial resolution [44] [22].
• Gradient Artifact Removal: Initial preprocessing using AAS methods with improved realignment to address scanner drift [5].

Experimental paradigms vary from resting-state recordings to task-based designs. For validation, simulated auditory evoked potentials are often superimposed on real resting-state data to quantitatively assess artifact removal performance [22].

ICA Implementation Workflow

A standardized workflow for ICA implementation in BCG artifact removal includes:

• Data Centering and Whitening: Preprocessing to make signals zero-mean and uncorrelated [41].
• Algorithm Selection and Application: Choosing an appropriate ICA algorithm (Infomax, FastICA, or JADE) based on data characteristics and applying it to the preprocessed data [40].
• Component Identification: Classifying components as neural activity or artifacts based on:
  • Temporal properties: Correlation with cardiac events [42]
  • Spatial patterns: Topographic distributions [22]
  • Spectral characteristics: Power distribution across frequency bands [42]
• Artifact Removal and Reconstruction: Projecting data back to sensor space while excluding artifact-related components [42].

The implementation can be facilitated by toolboxes such as EEGLAB [40] or custom MATLAB toolboxes that combine EEG and fMRI analysis in the same module [45].

Diagram 1: BCG Artifact Removal Workflow. This flowchart illustrates the standard processing pipeline for removing BCG artifacts from simultaneous EEG-fMRI data using ICA.

Table 3: Essential Research Tools for ICA in EEG-fMRI Research

Tool/Resource	Function/Purpose	Implementation Notes
EEGLAB [40]	MATLAB toolbox for EEG processing including ICA	Provides Infomax ICA; extensive visualization capabilities
JADE Algorithm [43]	ICA via joint diagonalization of cumulant matrices	Robust performance; available in MATLAB, Python, R
FMRIB Plug-in [44]	Toolbox for EEG-fMRI processing	Implements AAS, OBS, and QRS detection algorithms
BESA Research [22]	Commercial software for EEG/MEG analysis	Includes surrogate methods for BCG artifact removal
64-Channel EEG Systems [22]	High-density EEG acquisition in MRI environment	Enables better spatial sampling for ICA decomposition
MR-Compatible Amplifiers [44]	EEG signal acquisition in magnetic field	Essential for simultaneous EEG-fMRI recording

Comparative Analysis and Performance Metrics

Quantitative Performance Measures

The effectiveness of ICA-based BCG artifact removal is typically evaluated using several quantitative metrics:

• BCG Residual Intensity: Measures the remaining artifact power after correction. Studies show aOBS achieves the lowest residuals (5.53% on average) compared to AAS (12.51%), ICA (20.63%), and OBS (9.20%) [5].
• Cross-Correlation with ECG: Quantifies the residual temporal relationship between cleaned EEG and cardiac activity. aOBS demonstrates the lowest maximum cross-correlation values (0.028) compared to pre-correction values (0.180) [5].
• Signal-to-Noise Ratio (SNR): Assesses the preservation of neural signals of interest, particularly important for event-related potentials [22].
• Source Localization Error: Evaluates the impact on source analysis accuracy, with surrogate methods showing significant improvements over traditional approaches [22].

Applications in Neuroimaging Research

Effective BCG artifact removal enables various applications of simultaneous EEG-fMRI:

• Resting-State Networks: Studying the relationship between neural oscillations and BOLD fluctuations during rest [5].
• Event-Related Potentials: Investigating the neural generators of cognitive processes with combined temporal and spatial precision [22].
• Epilepsy Research: Localizing epileptogenic zones by combining EEG spike information with fMRI localization [44].
• Neurofeedback: Developing closed-loop systems that use EEG features to guide fMRI neurofeedback protocols [44].

Diagram 2: Basic ICA Concept. This diagram illustrates the fundamental ICA model where observed mixtures (x) are separated into estimated sources (y) using a separating matrix (W).

Independent Component Analysis provides a powerful framework for blind source separation that has proven particularly valuable for addressing the challenging problem of BCG artifacts in simultaneous EEG-fMRI research. While standard ICA methods effectively separate neural signals from cardiac-related artifacts, advanced approaches like ccICA, aOBS, and surrogate methods offer improved performance by addressing the non-stationary nature of BCG artifacts. The continued refinement of ICA methodologies promises to enhance the utility of simultaneous EEG-fMRI across basic neuroscience and clinical applications, ultimately providing clearer insights into brain function through multimodal integration.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the millimeter spatial resolution of fMRI, enabling unprecedented insights into brain dynamics across spatiotemporal scales [4] [3]. This integration is particularly valuable for investigating neurological disorders, cognitive processes, and drug effects on brain function. However, the technique faces a significant technical challenge: the ballistocardiogram (BCG) artifact, a persistent noise source that contaminates EEG signals recorded inside the MRI scanner [4] [36]. The BCG artifact arises from multiple physiological mechanisms including cardiac-induced head movements, pulsatile scalp expansion, and the Hall effect of conductive blood flow within the static magnetic field, creating complex artifacts that are time-locked to the heartbeat but exhibit considerable non-stationarity in both time and space [3] [36] [5].

Despite two decades of methodological development, no single artifact removal method has proven entirely sufficient. Traditional approaches include Average Artifact Subtraction (AAS), which creates and subtracts an averaged artifact template; the Optimal Basis Set (OBS) method, which uses principal components to model artifact variance; and Independent Component Analysis (ICA), which separates artifact components from neural signals through blind source separation [3] [11] [5]. Individually, each method has notable limitations: AAS cannot adequately handle BCG variability, OBS requires careful component selection and assumes consistent artifact timing, and ICA struggles with the spatial non-stationarity of BCG sources and introduces subjectivity in component selection [36] [5] [24]. The limitations of these individual approaches have motivated the development of hybrid methodologies, particularly the combination of OBS and ICA, which leverages their complementary strengths to achieve superior artifact reduction while better preserving neurological signals of interest for research and clinical applications [3] [36].

Theoretical Foundation: Limitations of Standalone Methods

The Optimal Basis Set (OBS) Approach

The OBS method, introduced by Niazy et al. (2005), employs Principal Component Analysis (PCA) to model the temporal variability of the BCG artifact [3] [5]. For each EEG channel, segments time-locked to cardiac events (typically R-peaks detected from ECG) are extracted and assembled into a matrix. PCA decomposes this matrix to create an optimal basis set comprising the principal components that capture the dominant patterns of artifact variance [5]. This basis set is then fitted to and subtracted from each artifact occurrence in the continuous EEG data. The primary advantage of OBS lies in its ability to adapt to temporal variations in artifact morphology through component-based reconstruction.

However, OBS faces two significant challenges. First, the selection of how many principal components to remove lacks a universal standard; while the first 3-4 components are typically used, this fixed number may not be optimal across all channels, subjects, or recording sessions [5]. Second, OBS performance is highly dependent on the accurate alignment of BCG events, typically identified from ECG signals with an assumed fixed delay. The inherent variability between the electrical cardiac event (ECG) and the mechanical BCG manifestation introduces misalignment that reduces the efficacy of PCA decomposition [5]. This limitation has prompted the development of adaptive OBS variants that perform beat-to-beat delay estimation, demonstrating improved artifact capture [5].

Independent Component Analysis (ICA) Framework

ICA is a blind source separation technique that decomposes multi-channel EEG data into statistically independent components (ICs) based on higher-order statistics [36] [24]. The fundamental assumption is that BCG artifacts and neural signals constitute spatially fixed but statistically independent sources that mix linearly at the electrodes. After decomposition, components identified as artifact-related are removed, and the remaining components are back-projected to sensor space to reconstruct cleaned EEG [24].

The critical limitation of ICA stems from its underlying assumptions. The BCG artifact originates from multiple physiological mechanisms that may not be statistically independent, and its spatial topography varies over time due to slight head movements and physiological changes [36] [5]. This spatial non-stationarity violates the instantaneous mixing assumption of standard ICA. Additionally, ICA requires subjective component selection based on visual inspection or correlation with reference signals, introducing variability and expertise dependency [24]. While constrained ICA (cICA) approaches incorporating prior information about artifact characteristics have been developed, they still face challenges in completely separating artifacts from neural signals with similar temporal or spectral properties [24].

The Rationale for Hybrid Integration

The complementary strengths and weaknesses of OBS and ICA create a compelling rationale for their integration. OBS excels at capturing the temporally structured, channel-specific aspects of the BCG artifact but may inadvertently remove neural signals that project onto the same basis vectors [5]. ICA effectively separates spatially stationary sources but struggles with spatially non-stationary artifacts like the BCG [36]. By combining these approaches, the hybrid methodology leverages OBS for initial artifact reduction, creating a partially cleaned signal that better satisfies the spatial stationarity assumptions of ICA, which then targets residual artifacts with different spatial characteristics [3] [36]. This sequential processing strategy has demonstrated superior performance in preserving neural signals while achieving more complete artifact removal, as evidenced by multiple validation studies [3] [36] [1].

Hybrid Methodologies: Implementation Frameworks

PROJIC-OBS Framework

The PROJIC-OBS framework represents an advanced hybrid approach that reverses the traditional processing order [36]. This method begins with ICA decomposition of the raw EEG data to obtain independent components (ICs). Rather than immediately rejecting artifact-related components, it applies a novel projection-based method (PROJIC) to identify BCG-related components using features derived from the artifact's temporal structure and spatial distribution. The OBS correction is then applied specifically to these BCG-related ICs to remove the artifact while preserving the neural information within each component. Finally, all components (corrected and uncorrected) are back-projected to sensor space.

Table 1: Key Steps in PROJIC-OBS Implementation

Step	Process	Description	Parameters
1	Data Preparation	Gradient artifact removal, band-pass filtering (0.5-70 Hz)	TR information, filter cutoff frequencies
2	ICA Decomposition	Separation of EEG into independent components	Infomax or FastICA algorithm, number of components
3	PROJIC Selection	Identification of BCG-related components using spatial and temporal features	Correlation thresholds, template matching
4	OBS Correction	Application of Optimal Basis Set to selected components	Number of principal components, alignment window
5	Signal Reconstruction	Back-projection of all components to sensor space	-

This approach demonstrated particular efficacy in preserving event-related potentials (ERPs), with studies showing a 32-62% reduction in standard error across trials compared to standalone methods [36]. The component-specific application of OBS minimizes the impact on neural activity that might be contained within components with mixed neural and artifactual content.

OBS-ICA Sequential Processing

The more traditional OBS-ICA approach implements a sequential pipeline where OBS serves as the initial preprocessing step before ICA decomposition [3] [36]. In this method, the raw EEG first undergoes standard OBS correction to remove the bulk of the BCG artifact. The partially cleaned data then undergoes ICA decomposition, during which components representing residual BCG artifacts are identified and removed. The rationale is that the initial OBS step reduces the artifact amplitude and variability, resulting in a signal that better conforms to ICA's stationarity assumptions and facilitating more accurate identification of residual artifact components.

Experimental comparisons have shown that this sequential approach effectively reduces BCG residuals that persist after standalone OBS processing [36]. However, a potential limitation exists in that the initial OBS correction may alter the spatial characteristics of both artifacts and neural signals, potentially complicating the subsequent ICA decomposition and component selection. The optimal number of OBS components to remove in the first stage requires careful consideration, as overly aggressive removal may distort neural signals while insufficient removal leaves excessive artifact for ICA to handle effectively [36] [5].

Surrogate-Based Hybrid Methods

Surrogate methods represent a sophisticated variant that combines elements of both OBS and ICA within a spatial filtering framework [3]. These approaches use surrogate source models consisting of regional dipole sources distributed throughout the brain to describe most EEG signals of neural origin. The artifact topographies needed for separation can be obtained either through PCA (similar to OBS) or through manual selection of artifact components using ICA.

In the PCA-based surrogate method (PCA-S), principal components are computed from an averaged artifact template and used to create artifact spatial filters. These are combined with the surrogate brain model to separate artifacts from neural activity with minimal distortion [3]. Validation studies using simulated auditory evoked potentials added to real resting-state EEG-fMRI data showed that PCA-S and its ICA-based counterpart (ICA-S) outperformed traditional OBS, ICA, and their combinations in terms of signal-to-noise ratio improvement and source localization accuracy [3].

Experimental Protocols & Validation Metrics

Benchmarking Experimental Design

Rigorous evaluation of hybrid OBS-ICA methods requires carefully designed experiments that enable quantification of both artifact removal efficacy and neural signal preservation. The most compelling validation approaches incorporate one of two paradigms: (1) Simulated Artifact Addition, where clean EEG recorded outside the scanner is combined with real BCG artifacts recorded inside the scanner, creating a ground truth comparison [3] [24]; or (2) Task-Based Designs, where known neurophysiological phenomena (e.g., visual evoked potentials, auditory oddball responses) are recorded simultaneously with fMRI, enabling assessment of how well biologically plausible signals are preserved after artifact removal [3] [36] [5].

In simulated artifact studies, clean EEG is typically recorded outside the MRI environment, often during specific cognitive tasks or sensory stimulation protocols to provide rich neural content. Simultaneously, resting-state EEG is recorded inside the MRI scanner from the same subject to capture authentic BCG artifacts. The artifacts are then extracted and added to the clean outside EEG, creating a controlled dataset with known neural signals [24]. This approach enables direct comparison between the original clean EEG and the artifact-corrected results, providing precise quantification of recovery quality.

For real simultaneous recordings, visual and auditory paradigms are particularly valuable as they generate robust, well-characterized event-related potentials (ERPs) with predictable timing and topography [36] [5]. The preservation of these known neural signatures after artifact correction provides strong evidence for the method's ability to retain signals of interest while removing artifacts.

Quantitative Performance Metrics

Table 2: Key Performance Metrics for Hybrid Method Validation

Metric Category	Specific Metric	Description	Interpretation
Artifact Reduction	BCG Residual Intensity	RMS of average BCG waveform after correction	Lower values indicate better artifact removal
	INPS (Improvement in Normalized Power Spectrum)	Power reduction at cardiac frequency bands	Higher values indicate better artifact suppression
	Cross-Correlation with ECG	Maximum correlation between EEG and reference ECG	Lower values indicate better decorrelation
Signal Preservation	ERP Signal-to-Noise Ratio (SNR)	Ratio of signal power to noise power in ERP components	Higher values indicate better neural signal preservation
	Inter-trial Variability	Standard error across trials in ERP analysis	Lower values indicate more reliable signal extraction
	Source Localization Error	Difference in dipole location between inside and outside scanner	Lower values indicate better spatial accuracy
Overall Performance	Sample Entropy	Complexity measure of reconstructed signal	Preservation of original signal complexity
	Similarity Index	Correlation between recovered and original signals	Higher values indicate better overall reconstruction

These metrics collectively assess different aspects of performance, with the optimal balance depending on the specific research application. Studies focusing on oscillatory neural activity may prioritize spectral purity, while ERP research emphasizes temporal precision and SNR [36] [1].

Recent Comparative Evidence

Recent comprehensive evaluations demonstrate the superior performance of hybrid approaches. Abreu et al. (2016) conducted an extensive comparison of multiple BCG correction methods using data acquired at both 3T and 7T field strengths [36]. Their results indicated that PROJIC-OBS and PROJIC-AAS outperformed other methods across multiple metrics, with PROJIC-OBS achieving the best performance when priority was given to artifact removal, while PROJIC-AAS excelled when physiological signal preservation was prioritized. The study reported 32% (at 3T) and 62% (at 7T) reductions in standard error across trials for PROJIC-OBS, indicating substantially improved ERP quality after correction.

Another study by Marino et al. (2018) introduced an adaptive OBS (aOBS) method that automatically identifies BCG-related components and optimizes artifact template alignment [5]. When compared to standard OBS, AAS, and ICA, aOBS demonstrated significantly lower BCG residuals (5.53% vs. 9.20-20.63%) and reduced cross-correlation with ECG (0.028 vs. 0.042-0.067), while simultaneously improving ERP SNR and reducing inter-trial variability. This adaptive approach addresses key limitations of traditional OBS and could potentially be integrated with ICA in future hybrid implementations.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Reagents and Equipment for EEG-fMRI Artifact Research

Item	Function/Application	Technical Specifications
MR-Compatible EEG Systems	Simultaneous data acquisition with minimal interference	64+ channels, synchronized sampling, high dynamic range amplifiers
Carbon Fiber Wire Loops	Hardware-based BCG artifact reference recording	Conductive loops placed around head, record artifact field
Reference Layer Caps	Integrated artifact sensing with additional electrode layer	Specialized caps with dual layers for signal and reference
ECG/EOG Monitoring	Reference signal acquisition for artifact template creation	MR-compatible electrodes, shielded cabling
Synchronization Interface	Temporal alignment of EEG and fMRI data streams	TTL pulse integration, precise clock synchronization
High-Density EEG Montages	Improved spatial sampling for source separation	128-256 channels, optimized electrode positioning
Software Toolboxes	Implementation of artifact removal algorithms	FMRIB Plug-in, EEGLAB, BESA, custom MATLAB scripts

The combination of specialized hardware and software tools enables more effective implementation of hybrid artifact removal strategies. For instance, reference layer caps and carbon fiber loops provide additional artifact reference signals that can be incorporated into adaptive filtering approaches, complementing the software-based OBS-ICA methods [46]. These hardware solutions record the artifact field directly, offering a reference that can guide component selection in ICA or validate OBS template construction.

Implementation Guidelines & Parameter Optimization

Successful implementation of hybrid OBS-ICA methods requires careful attention to parameter selection and algorithmic details. For the OBS component, the number of principal components must be optimized—too few components leave significant residuals, while too many remove neural signals [5]. Adaptive approaches that automatically determine the optimal number of components based on explained variance thresholds (e.g., 95-98% cumulative variance) generally outperform fixed-number approaches [5]. Additionally, accurate BCG peak detection is critical; rather than relying solely on ECG R-peaks with assumed delays, directly detecting BCG peaks from the EEG data after gradient artifact removal improves temporal alignment and subsequent PCA effectiveness [5].

For the ICA component, the choice of decomposition algorithm (Infomax vs. FastICA) and the number of components to estimate require consideration. Infomax ICA often performs better for BCG artifact separation due to its sensitivity to sub-Gaussian distributions common in artifact components [24]. The number of ICs should be substantially less than the number of channels to avoid overfitting while still capturing the major sources of variance. For component selection, automated approaches based on correlation with reference signals or template matching outperform visual inspection alone, particularly in the PROJIC framework which uses spatial and temporal features for identification [36].

When implementing the complete hybrid pipeline, processing order significantly impacts results. The PROJIC-OBS approach, which applies OBS correction selectively to ICA-identified components, generally produces superior signal preservation compared to sequential OBS-ICA processing, particularly for ERP studies [36]. However, this comes at increased computational cost and implementation complexity. For practical applications, researchers should match the method selection to their specific goals—PROJIC-OBS for maximal signal preservation in ERP research, and sequential OBS-ICA for efficient artifact reduction in oscillatory activity studies.

The integration of OBS and ICA represents a significant advancement in addressing the persistent challenge of BCG artifacts in simultaneous EEG-fMRI. By leveraging the complementary strengths of these approaches, hybrid methods achieve more complete artifact reduction while better preserving neurologically meaningful signals compared to standalone techniques. The PROJIC-OBS framework, which applies OBS correction selectively to ICA-identified components, has demonstrated particular efficacy for ERP studies, while adaptive OBS implementations offer improved template alignment and component selection.

Future methodological developments will likely focus on three areas: (1) increased automation through machine learning approaches for component classification and parameter optimization, reducing expertise barriers and improving reproducibility [14]; (2) real-time implementation enabling closed-loop experiments and neurofeedback applications [46]; and (3) deep learning architectures that learn the nonlinear mapping between artifact references and contaminated EEG, potentially surpassing traditional linear methods [27] [14]. As these techniques mature and become more accessible in standard analysis packages, they will enhance the reliability and interpretability of simultaneous EEG-fMRI across basic neuroscience, clinical research, and pharmaceutical development applications.

For researchers implementing these methods, we recommend a systematic validation approach incorporating multiple quantitative metrics tailored to specific research questions. Combining objective performance measures with visual inspection of cleaned data provides the most comprehensive assessment of method efficacy. As the field moves toward standardized evaluation frameworks and reporting guidelines, cross-study comparisons will become more meaningful, accelerating methodological refinements and ultimately strengthening the scientific insights derived from this powerful multimodal imaging technique.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) is a powerful neuroimaging technique that combines the high temporal resolution of EEG with the high spatial resolution of fMRI. This integration provides unparalleled insights into brain dynamics by capturing both rapid electrical neural events and their associated hemodynamic responses within a single experimental session [27] [22]. However, the quality of EEG signals recorded inside the MRI scanner is severely compromised by significant artifacts, among which the ballistocardiogram (BCG) artifact presents the most formidable challenge for analysis [22] [5].

The BCG artifact is a complex signal distortion predominantly caused by cardiac-induced head movements and pulsatile blood flow within the strong static magnetic field [22]. Unlike simpler artifacts, the BCG exhibits non-stationary spatio-temporal dynamics and propagates through the EEG signals in a manner non-linearly dependent on the electrocardiogram (ECG) [27]. Traditional removal methods, including Optimal Basis Set (OBS) and Independent Component Analysis (ICA), have shown limitations in effectively separating BCG artifacts from neural signals without distorting brain activity of interest [22] [5].

This technical guide explores three advanced computational frameworks—Deep Learning (exemplified by BCGNet), Harmonic Regression, and Surrogate Methods—that represent emerging frontiers in addressing the persistent challenge of BCG artifact removal. These approaches leverage sophisticated mathematical principles and machine learning architectures to overcome limitations of conventional techniques, enabling more accurate artifact suppression while preserving neural signals critical for neuroscientific discovery and clinical applications.

The BCG Artifact: Characteristics and Challenges

The BCG artifact originates from multiple physical phenomena related to cardiac activity. When a subject lies in the MRI scanner, each heartbeat generates microscopic head movements due to ballistic forces from cardiac contraction and blood pulsation [22]. These movements within the strong static magnetic field induce electrical currents that contaminate EEG recordings [5]. Additional contributors include the Hall effect, where pulsatile blood flow as a conductive fluid generates potential changes across the scalp, and pulsation artifacts near electrodes adjacent to blood vessels [22].

The BCG artifact exhibits several characteristics that complicate its removal:

• Spatio-temporal variability: The artifact's topography and timing vary across subjects, recording sessions, and even within a single recording [22] [5]
• Non-linear propagation: The relationship between ECG and BCG is non-linear, making linear decomposition methods suboptimal [27]
• Non-stationarity: BCG characteristics change over time due to factors like head position shifts, blood pressure variations, and physiological state changes [22]
• Spectral overlap: BCG artifacts often occupy similar frequency bands as neural signals of interest, particularly for event-related potentials [22]

These properties necessitate advanced processing techniques that can adapt to the complex, dynamic nature of the BCG artifact while preserving the integrity of underlying neural signals.

Deep Learning Approaches: BCGNet and Beyond

Theoretical Foundations

Deep learning approaches for BCG artifact removal leverage neural networks to learn the complex, non-linear mapping between reference signals (typically ECG) and the resulting BCG artifacts in EEG recordings. Unlike traditional methods that rely on linear assumptions, deep learning models can capture the intricate temporal dependencies and spatial patterns characteristic of BCG contamination [27].

The fundamental principle involves training a neural network to model the transformation f in the equation: EEG_measured(t) = EEG_true(t) + f(ECG(t)) + ε where EEG_measured is the recorded signal, EEG_true is the desired neural activity, f(ECG(t)) represents the BCG artifact as a function of the ECG signal, and ε encompasses other noise sources [27].

BCGNet Architecture and Implementation

BCGNet employs a recurrent neural network (RNN) architecture, specifically designed to handle sequential data with long-term dependencies. The RNN learns to predict BCG artifacts based on current and historical ECG inputs, enabling reconstruction and subsequent subtraction of the artifact from contaminated EEG signals [27].

Table 1: Key Components of BCGNet Architecture

Component	Specification	Function
Network Type	Recurrent Neural Network (RNN)	Models temporal dependencies in ECG-BCG relationship
Input Layer	ECG signal timestamps	Receives raw or preprocessed ECG input
Hidden Layers	Gated Recurrent Units (GRUs) or LSTM	Learns complex temporal relationships
Output Layer	EEG channel dimensions	Generates BCG artifact prediction
Training Objective	Mean Squared Error (MSE)	Minimizes difference between predicted and actual BCG

The implementation typically involves these stages:

• Data Preparation: Segment simultaneous EEG-ECG recordings into training examples
• Preprocessing: Apply bandpass filtering and normalization to both EEG and ECG signals
• Network Training: Optimize network parameters using backpropagation through time
• Artifact Removal: Subtract the network's BCG prediction from measured EEG signals
• Validation: Assess performance using metrics like signal-to-noise ratio improvement and artifact power reduction [27]

Performance and Applications

Research demonstrates that BCGNet significantly outperforms traditional methods like OBS in both artifact reduction and preservation of neural signals. Studies report larger average power reduction of BCG artifacts at critical frequencies while simultaneously improving task-relevant EEG classification accuracy [27]. The method shows particular promise for cognitive neuroscience studies investigating event-related potentials during simultaneous fMRI acquisition, where preserving the temporal characteristics of neural responses is crucial.

Harmonic Regression Methods

Mathematical Framework

Harmonic regression employs Fourier series to model periodic components in time series data. For BCG artifact removal, this approach leverages the fact that cardiac-related artifacts exhibit pseudo-periodic characteristics tied to the heart rate [47].

The fundamental model represents a signal y(t) as: y(t) = β₀ + Σ[βₖ cos(2πfₖt) + γₖ sin(2πfₖt)] + ε(t) where β₀ is the DC component, βₖ and γₖ are Fourier coefficients for frequency fₖ, and ε(t) represents non-periodic components including neural signals [47].

Implementation for BCG Removal

In the context of BCG artifact removal, harmonic regression is applied as follows:

• Cardiac Frequency Identification: Detect R-peaks in simultaneously recorded ECG to determine fundamental cardiac frequency and its harmonics
• Regression Model Fitting: Estimate Fourier coefficients for the identified frequencies using the contaminated EEG signal
• Artifact Reconstruction: Generate BCG artifact estimate using the fitted model
• Artifact Subtraction: Remove the reconstructed artifact from the original signal [47]

This approach effectively captures the periodic structure of BCG artifacts while being computationally efficient. The method can be particularly effective when combined with adaptive filtering techniques that account for heart rate variability throughout the recording session.

Surrogate Methods for Spatial Filtering

Theoretical Basis

Surrogate methods in BCG artifact removal leverage spatial filtering techniques to separate artifact and neural components based on their distinct topographic distributions across the scalp [22]. The core assumption is that measured EEG signals represent a superposition of brain activity and artifact components with fixed but different spatial topographies [22].

The method is grounded in the principle that if the spatial distributions (topographies) of artifact sources are known, they can be used to construct a surrogate model that separates artifact from brain signals with minimal distortion [22].

Implementation Approaches

Two primary implementations have been developed for BCG artifact removal:

1. PCA-based Surrogate Method (PCA-S)

• Artifact topographies are derived by applying Principal Component Analysis (PCA) to averaged BCG artifact templates [22]
• The first few principal components that explain maximal variance in the artifact template are selected as artifact spatial basis functions [22]
• A surrogate brain model comprising regional dipole sources distributed throughout the brain is used to represent neural activity [22]
• The algorithm then separates activity into artifact and neural components by leveraging their distinct spatial distributions [22]

2. ICA-based Surrogate Method (ICA-S)

• Independent Component Analysis (ICA) is applied to identify artifact-related components [22]
• Manual selection of artifact components is performed based on their temporal and spatial characteristics [22]
• Selected components are used to construct artifact spatial topographies for the surrogate model [22]
• Similar to PCA-S, spatial filtering is applied to separate artifact and neural signals [22]

Surrogate Method Workflow for BCG Artifact Removal

Performance Evaluation

Studies comparing surrogate methods with established approaches (OBS, BSS, OBS-ICA) demonstrate superior performance in both artifact removal and neural signal preservation. Using resting-state data from 55 subjects with simulated auditory evoked potentials, PCA-S and ICA-S showed highly significant improvements in source localization accuracy and signal-to-noise ratio compared to traditional methods [22].

Table 2: Performance Comparison of BCG Artifact Removal Methods

Method	BCG Residual (%)	Signal Distortion	Computational Complexity	Preservation of Neural Signals
AAS	12.51	Moderate	Low	Moderate
ICA	20.63	High	Medium	Low
OBS	9.20	Low-Medium	Low	Moderate
aOBS	5.53	Low	Low-Medium	High
PCA-S	<5.53	Very Low	Medium	Very High
ICA-S	<5.53	Very Low	Medium	Very High
BCGNet (RNN)	Not Reported	Very Low	High	Very High

Comparative Analysis and Implementation Guidelines

Method Selection Framework

Choosing the appropriate BCG removal method depends on multiple factors including research objectives, available computational resources, and technical expertise:

• For maximum artifact reduction: Adaptive OBS (aOBS) and surrogate methods (PCA-S/ICA-S) demonstrate lowest BCG residuals [22] [5]
• For optimal neural signal preservation: Surrogate methods and BCGNet show superior performance in preserving task-relevant neural activity [27] [22]
• For computational efficiency: Traditional OBS and AAS offer faster processing times with reasonable artifact reduction [5]
• For dynamic artifact characteristics: BCGNet and aOBS adapt better to non-stationary BCG properties throughout recording sessions [27] [5]

Integrated Approaches

Combining multiple methods in a sequential pipeline often yields superior results compared to individual approaches:

• Initial preprocessing with adaptive OBS for bulk artifact reduction
• Secondary processing with surrogate spatial filtering for refined separation
• Final refinement with deep learning approaches for residual artifact removal

This integrated strategy leverages the complementary strengths of different methodologies while mitigating their individual limitations.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Tools for BCG Artifact Investigation

Tool/Resource	Function/Purpose	Example Applications
High-Density EEG Systems (64+ channels)	Captures detailed spatial information essential for spatial filtering methods	Provides sufficient spatial sampling for PCA-S and ICA-S methods [22]
MRI-Compatible ECG	Records reference cardiac signal for artifact template creation	Essential for OBS, aOBS, and BCGNet implementations [5]
BESA Research Software	Implements surrogate spatial filtering methods	Used in validation studies for PCA-S and ICA-S approaches [22]
FieldTrip Toolbox (MATLAB)	Open-source platform for implementing custom artifact removal algorithms	Flexible framework for method development and comparison [48]
Surrogate Data Toolbox (MatLab)	Implements various surrogate data generation methods	Useful for testing nonlinearity and validating artifact separation [49]
Graph Neural Networks (GNNs)	Handles geometrically variable data structures	Potential application for subject-specific BCG modeling [50]
Carbon Wire Loop Systems	Hardware-based BCG reduction	Provides reference artifact signal without brain activity [22]

Emerging Trends

The field of BCG artifact removal continues to evolve with several promising directions:

• Hybrid architectures combining deep learning with spatial filtering techniques
• Transfer learning approaches allowing models trained on one subject to generalize to others
• Real-time implementation enabling artifact removal during data acquisition
• Multi-modal integration incorporating information from additional sensors (e.g., motion tracking, photoplethysmography)

The emerging frontiers of deep learning (BCGNet), harmonic regression, and surrogate methods represent significant advances in addressing the persistent challenge of BCG artifacts in simultaneous EEG-fMRI. Each approach offers distinct advantages: deep learning excels at modeling non-linear dynamics, harmonic regression effectively captures periodic structure, and surrogate methods provide superior spatial separation with minimal neural signal distortion.

As these methodologies continue to mature and integrate, researchers gain increasingly powerful tools for unlocking the full potential of simultaneous EEG-fMRI. This progress promises to enhance our understanding of brain function by providing cleaner neural signals with combined high temporal and spatial resolution, ultimately advancing both basic neuroscience and clinical applications.

Conceptual Relationships Among BCG Artifact Removal Methods

The simultaneous acquisition of Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI) represents a powerful multimodal neuroimaging technique, offering the high temporal resolution of EEG alongside the high spatial resolution of fMRI. However, the utility of this technique is significantly challenged by the presence of prominent artifacts in the EEG data, chief among them the ballistocardiogram (BCG) artifact [4]. This artifact arises from cardiac-related phenomena, including head movements and pulsatile motion of scalp vessels within the static magnetic field, which induce electrical currents in EEG electrodes that can obscure neural signals [23] [14].

While numerous software-based methods exist for BCG artifact reduction, hardware-based solutions offer a distinct approach by directly measuring the artifact's source for subsequent subtraction. Among these, Carbon-Wire Loops (CWLs) and Reference Layer Systems have proven effective. These methods aim to provide a reference signal that is highly correlated with the BCG artifact but independent of cerebral activity, enabling cleaner EEG signal recovery without the need for complex post-processing algorithms that risk distorting neural data [22]. This guide provides a technical examination of these hardware solutions, their implementation, and their performance within the context of simultaneous EEG-fMRI research.

Understanding the BCG Artifact

The BCG artifact is a complex signal distortion with multiple physical origins. Its primary generators are:

• Cardiac-Induced Head Movement: The abrupt ejection of blood from the heart during each cardiac cycle causes minute, rhythmic movements of the subject's head while in a supine position. Movement of the EEG electrodes within the scanner's powerful static magnetic field induces electrical currents [23] [5].
• Pulsatile Scalp Movement: Localized expansion and contraction of scalp blood vessels with each heartbeat also contribute to electrode movement [14].
• The Hall Effect: As blood, an electrically conductive fluid, pulsates through the magnetic field, it generates a electrical potential, which adds to the measured artifact [22] [5].

Unlike the gradient artifact caused by switching MRI gradients, the BCG artifact is non-stationary, showing variability in shape over time and across different EEG channels. Its temporal relationship with the cardiac cycle makes it particularly challenging to remove without affecting concurrent brain activity, such as epileptic discharges or event-related potentials [22] [14].

Carbon-Wire Loops (CWLs)

Principle and Hardware Setup

Carbon-Wire Loops are a hardware solution designed to record the BCG artifact directly. The core principle is that CWLs, when placed on the scalp, are affected by the same head movements and ballistocardiogram forces as the EEG electrodes but do not record any brain electrical activity. The signals from the loops thus serve as approximate templates of the pure artifact, which can be subtracted from the contaminated EEG channels [51].

A typical experimental setup involves:

• Loop Construction: CWLs are often fabricated from a carbon filament wire with a polyvinyl chloride (PVC) jacket, approximately 1 mm in diameter. A single loop with a total length of 7.5 meters and a resistance of around 1125 Ohms is common [51].
• Physical Configuration: Multiple loops (often five or six) of about 8 cm in diameter are placed over the patient's head on top of the EEG electrodes, covering the sides, back, and top. The wires of these loops are braided together to minimize magnetic induction. The loops are secured using bandages [51].
• Signal Acquisition: The CWLs require a dedicated MR-compatible bipolar amplifier (e.g., a BrainAmp ExG MR 16ch amplifier). The signals from the loops are recorded simultaneously with the EEG and electrocardiogram (ECG) data [51].

Table 1: Research Reagent Solutions for Carbon-Wire Loop Setup

Item Name	Specifications / Function
Carbon Filament Wire	Microscopic carbon strands in a PVC jacket; 1 mm diameter; ~150 Ohm/m resistance; forms the sensing loop [51].
MR-Conditional Amplifier	A bipolar amplifier (e.g., BrainAmp ExG MR) safe for use in the MRI environment; amplifies the weak signals from the CWLs [51].
Braided Wire Housing	The wires from multiple loops are braided; reduces magnetic induction and cross-talk between loops [51].
Secure Fastening Bandages	Non-metallic, MRI-safe bandages; used to fix the CWLs in position on the subject's scalp [51].

Signal Processing and Workflow

The correction algorithm is based on the idea that the BCG artifact in any EEG channel can be modeled as a linear combination of the signals recorded by the CWLs and their time-shifted versions. The time shifts account for the non-instantaneous propagation of pressure waves through blood vessels [51]. Two prominent algorithms for this are:

• Recursive Least Squares (RLS) Algorithm: This method, proposed by Masterton et al. (2007), uses a recursive filter to continuously estimate and update the weights of the linear combination linking the CWL signals to the EEG artifact. While effective, its recursive nature carries a theoretical risk of divergence, which could amplify noise in the EEG signal [51].
• Windowed Root Mean Square (RMS) Minimization: This approach, proposed by van der Meer et al. (2016), estimates the weights by minimizing the RMS error in partially overlapping, short-duration windows. This non-iterative method is faster and avoids the risk of filter divergence, making it a preferred option in some settings [51].

The following workflow diagram illustrates the typical data processing steps involved in CWL artifact removal:

Performance and Practical Considerations

Research from the Montreal Neurological Institute and Hospital indicates that CWL correction is effective in a majority of cases. In a study of 21 patients, CWL-based correction alone was sufficient in about 75% of subjects [51].

Key advantages of CWLs include:

• Superiority to Software Methods: The corrected EEG is often visually cleaner than that processed with standard Average Artifact Subtraction, as CWLs also effectively reduce movement artifacts unrelated to the heartbeat [51].
• ECG-Independence: The method does not rely on the quality of the ECG signal for QRS complex detection, avoiding a common point of failure in template-based software methods [51].

However, some limitations exist:

• Risk of False Positives: Incomplete artifact correction can occasionally generate sharp transients that resemble interictal epileptiform discharges, requiring careful review by an expert [51].
• Workflow Integration: Researchers must be prepared to apply different correction methods (CWL, AAS, or a combination) on a subject-by-subject basis, as the most effective method can vary [51].

Reference Layer Systems

Principle and Hardware Setup

The reference layer approach, also known as the "reference layer adaptive noise cancelation," involves integrating an additional layer of conductive material within the EEG cap itself. This layer, typically made of a thin, flexible material such as carbon-loaded rubber, is positioned between the scalp and the EEG electrodes. It acts as a physical reference that passively captures the fluctuating electrical fields induced by head movement and pulsatile effects within the magnetic field [22].

The core principle is that this reference layer picks up the same BCG artifact as the EEG electrodes but is, ideally, blind to the neural potentials generated by the brain. The signal from this layer can then be used as a common reference for all EEG electrodes or fed into adaptive filtering algorithms to subtract the artifact from each channel.

Implementation and Workflow

While the provided search results note reference layer systems as a viable hardware-based solution [22], they offer less specific technical detail compared to CWLs. The general workflow, as inferred from the principle and other hardware methods, can be summarized as follows:

• Specialized Cap: An EEG cap with an integrated conductive reference layer is used.
• Signal Recording: The potential of the reference layer is recorded continuously alongside all EEG channels.
• Adaptive Filtering: The reference signal is used as an input to an adaptive filter (e.g., a recursive least squares filter) that models its relationship to the artifact component in each EEG channel.
• Artifact Subtraction: The output of the adaptive filter—the estimated artifact—is subtracted from each EEG channel.

The following diagram outlines this generalized signal processing pathway:

Comparative Analysis of BCG Artifact Removal Techniques

The performance of hardware-based methods must be contextualized against common software-based approaches. A 2025 systematic evaluation of BCG artifact removal techniques provides quantitative metrics for comparison [1].

Table 2: Performance Comparison of Key BCG Artifact Removal Methods

Method	Type	Key Performance Metrics	Advantages	Disadvantages
Carbon Wire Loops (CWL)	Hardware	N/A (Qualitative assessment shows sufficient correction in ~75% of subjects [51])	Corrects both BCG and movement artifacts; independent of ECG signal quality [51].	Requires additional hardware setup; risk of algorithm divergence or false positives [51].
Reference Layer	Hardware	Effective BCG reduction, improved source localization [22].	Integrated into cap design; provides a direct physical measurement of the artifact.	Can be expensive; complex setup; not applicable to previously recorded data [22].
Average Artifact Subtraction (AAS)	Software	Best signal fidelity (MSE = 0.0038, PSNR = 26.34 dB) [1].	Simple and easy to implement [1] [4].	Enhances movement artifacts; requires high-quality ECG; fails with poor QRS detection [51].
Optimal Basis Set (OBS)	Software	Highest structural similarity (SSIM = 0.72) [1].	Adapts to artifact shape variability better than AAS [5].	Performance depends on number of components removed; hampered by ECG-BCG delay variability [5].
Independent Component Analysis (ICA)	Software	Sensitive to frequency-specific patterns in dynamic network graphs [1].	Does not require an ECG reference; data-driven [4].	Component selection requires expertise; risk of neural signal loss; non-stationarity can violate model assumptions [1] [5].
Deep Learning (e.g., BCGNet, BCGGAN)	Software	Larger BCG power reduction at critical frequencies; improves task-relevant EEG classification [27] [23] [14].	Models non-linear ECG-BCG relationships; no additional hardware needed [27] [14].	Requires large datasets for training; "black box" nature; performance depends on training data quality [23] [14].

The table shows that no single method is universally superior. The choice depends on the research priorities: hardware methods offer a direct physical approach to artifact capture, while advanced software methods like deep learning show promise in handling the artifact's non-linearity without extra equipment.

Hardware-based solutions like Carbon-Wire Loops and Reference Layer Systems provide robust and effective means for mitigating the pervasive BCG artifact in simultaneous EEG-fMRI. Their primary strength lies in directly measuring the artifact's physical cause, allowing for cleaner signal recovery that is less dependent on the assumptions of software-based algorithms.

For researchers integrating these solutions, the following guidelines are recommended:

• For Maximum Artifact Fidelity: Consider CWLs or a reference layer system, especially in studies where preserving high-frequency neural signals or dealing with significant subject movement is critical.
• For Clinical Epilepsy Studies: CWLs have a proven track record in revealing interictal epileptiform discharges, but a workflow that includes a review of raw signals is essential to avoid false positives [51].
• Hardware vs. Software: Weigh the benefits of hardware solutions against their cost, setup complexity, and the availability of subject-specific optimization. In many cases, a combination of a simple hardware setup followed by a carefully chosen software correction may yield the best results.
• Reporting Standards: When publishing EEG-fMRI studies, clearly detail the hardware used (including CWL/resource layer specifications) and the specific artifact removal algorithms and parameters applied. This ensures reproducibility and allows for a better comparison of findings across different research centers [4].

The ongoing development of both hardware and software methods, including deep learning approaches that do not require additional hardware, continues to advance the field. The optimal approach will often be a tailored one, leveraging the strengths of both hardware and software to achieve the cleanest possible EEG data for a given experimental context.

Troubleshooting BCG Correction: Parameter Optimization, Pitfalls, and Best Practices

Simultaneous Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI) represents a powerful multimodal neuroimaging technique, combining EEG's millisecond-level temporal resolution with fMRI's detailed spatial mapping of brain activity [1]. However, the utility of this combined approach is significantly compromised by artifacts introduced into the EEG data by the MRI environment. The ballistocardiogram (BCG) artifact remains one of the most challenging technical obstacles in this field. This artifact arises from cardiac-induced body movements—specifically, the pulsatile motion of blood in the scalp and head within the strong static magnetic field, causing minute movements of EEG electrodes and leading to potential differences that contaminate the neural signals [52] [23]. The BCG artifact is characterized by its complex spatiotemporal dynamics, varying significantly across channels, subjects, and over time, with its main influence typically occurring 150–500 ms post-QRS complex of the cardiac cycle [52]. Its morphology and amplitude are directly proportional to the strength of the magnetic field, making it particularly problematic in high-field MRI scanners [52]. Unlike the more stereotypical gradient artifact caused by switching magnetic fields, the BCG's variability complicates its removal and poses a fundamental challenge for accurate peak detection and timing.

Fundamental Differences: ECG R-Peaks vs. BCG J-Peaks

The core of the peak detection problem lies in the physiological and temporal distinctions between the electrical event captured by electrocardiography (ECG) and the mechanical force measured by ballistocardiography (BCG).

Physiological Origins and Temporal Relationships

The R-peak in an ECG signal represents the electrical depolarization of the ventricles, marking the initiation of ventricular contraction [53]. In contrast, the J-peak in a BCG signal corresponds to the mechanical force resulting from the ejection of blood into the aortic arch and the subsequent change in flow direction, creating a measurable momentum [53]. This physiological sequence results in a consistent temporal delay between the electrical trigger (R-peak) and the mechanical force (J-peak), known as the R-J interval (RJI). According to established literature, this interval typically varies between 180 ms and 240 ms and can change slowly over time [53]. Furthermore, certain activities such as paced respiration can induce hemodynamic changes that affect the RJI by 150–300 ms [53]. The inter-subject variability in this interval is influenced by factors including body mass, heart size, body placement relative to the sensor, and overall physiological state [53].

Table 1: Key Characteristics of ECG R-Peaks vs. BCG J-Peaks

Feature	ECG R-Peaks	BCG J-Peaks
Physiological Origin	Electrical ventricular depolarization	Mechanical force from blood ejection
Waveform Sharpness	Typically sharp and distinct	Smoother, less distinct from background signal
Detection Challenge	Relatively straightforward	Low signal-to-noise ratio, high variability
Temporal Relationship	Precedes J-peak	Follows R-peak by 180-240 ms (R-J interval)
Inter-Subject Variability	Lower	Significant due to physiological and sensor factors

The Signal-to-Noise Ratio Challenge

BCG signals inherently suffer from a lower signal-to-noise ratio (SNR) compared to ECG, making J-peak detection particularly challenging. As noted in research, "J-peaks in BCG occur closer to the heart and are sharper in its waveform than PPG, which measures pulses at the skin or wrist" [53]. However, the current analysis of BCG data during sleep remains challenging, "mainly due to low signal-to-noise ratio, physical movements, as well as high inter- and intra-individual variability" [53]. This SNR challenge is quantified in BCG artifact removal research through specific equations, such as one estimating SNR using subensemble averages: (SNR = \frac{(E1 - E2)^2}{4 \times \frac{1}{N} \sum{i=1}^{N} (E1i - E2_i)^2}), where E1 and E2 represent subensemble averages of successive BCG signal segments [54].

Quantitative Impact of Timing Jitter on Downstream Applications

The precision of heartbeat detection directly influences the reliability of derived physiological parameters and their application in clinical and research settings. Timing jitters—small, unpredictable variations in the temporal locations of detected heartbeats—can significantly impact the accuracy of heart rate variability (HRV) analysis and subsequent applications such as sleep staging.

Effect on Heart Rate Variability (HRV) Parameters

Heartbeat intervals (HBIs) vary over time, and this variance is quantified as HRV features, which are traditionally based on R-R intervals from ECG [55]. When BCG is used instead of ECG, the resulting HBIs (J-J intervals) are numerically different from their ECG-based counterparts (R-R intervals), leading to what are termed HBI and HRV errors [55]. These discrepancies stem from multiple sources: physiological factors (the variable R-J interval), sensing modality differences, and crucially, peak detection algorithm performance [55]. Research has demonstrated that these timing jitters directly propagate into errors in calculated HRV parameters, potentially compromising their clinical and research utility.

Table 2: Impact of HBI Error on Sleep Staging Performance

HBI Error (ms)	Sleep-Scoring Error Increase	Clinical Implications
Up to 60 ms	Increase from 17% to 25%	Moderate reduction in staging reliability
> 60 ms	Potentially >25%	Significant compromise for diagnostic use
Minimal (reference)	Baseline ~17%	Acceptable for many research applications

Consequences for Sleep Staging Accuracy

Sleep staging represents a critical application where HRV features play a crucial role. Studies have investigated the relationship between HBI errors and sleep-staging accuracy, revealing that increased timing jitter directly correlates with reduced staging performance [55]. One key finding indicates that "at an HBI error range of up to 60 ms, the sleep-scoring error could increase from 17% to 25%" based on specific scenarios examined [55]. This quantitative relationship provides researchers with a benchmark for assessing the required accuracy of BCG-based heartbeat detection algorithms intended for sleep monitoring applications. The implications extend to other HRV-dependent applications, including long-term health monitoring, sleep quality assessment, and various clinical diagnostics where precise timing of cardiac events is paramount.

Methodologies for BCG Peak Detection and Artifact Removal

Various methodological approaches have been developed to address the challenges of BCG peak detection and artifact removal, ranging from classical signal processing techniques to advanced machine learning algorithms.

Classical Signal Processing Approaches

Traditional methods for BCG analysis often adapt approaches originally designed for ECG signal processing. The Pan-Tompkins algorithm, initially developed for ECG R-peak detection, has been modified for BCG applications, typically involving sequential processing steps such as bandpass filtering, cubic functions, low-pass filtering, derivatives, and absolute value transformations to enhance J-peak prominence [53]. Another established approach utilizes the Continuous Wavelet Transform (CWT) with Morlet wavelets, capitalizing on their similarity to BCG waveforms [54]. Research protocols describe identifying optimal scales for CWT (e.g., scales 27-31) where heartbeat information is most differentiable, with scale 30 often providing the best performance across participants [54]. Template-matching methods and frequency-domain analyses have also been employed, though their effectiveness is limited by the BCG's variability and low SNR [53].

Modern Machine Learning and Deep Learning Solutions

Recent advances have introduced sophisticated machine learning approaches to overcome the limitations of classical methods. Supervised deep learning setups now model discrete reference heartbeat events using symmetric, continuous kernel functions to create surrogate signals, which deep learning models then approximate to detect target heartbeats [53]. This approach has demonstrated superior accuracy in heartbeat estimation, achieving a mean absolute error (MAE) of 1.1 seconds in 64-second windows and 1.38 seconds in 8-second windows [53]. For BCG artifact removal in EEG-fMRI, recurrent neural networks (RNNs), particularly gated recurrent units (GRUs), have been configured to learn nonlinear mappings between ECG and BCG-corrupted EEG, effectively suppressing artifacts without requiring additional hardware [23]. Generative Adversarial Networks (GANs) represent another innovation, with models like BCGGAN designed to transform BCG-corrupted EEG signals into clean EEG without paired examples, eliminating the need for reference signals like ECG [14]. Additional deep learning architectures, including denoising autoencoders, have shown significant performance improvements, achieving root-mean-squared error (RMSE) of 0.0218 ± 0.0152 and signal-to-noise ratio (SNR) gain of 14.63 dB [56].

Diagram 1: BCG J-Peak Detection Workflow

Experimental Protocols for BCG Analysis

Direct BCG Period Determination Protocol

A multi-scale peak detection method enables automatic cardioballistic artifact period determination directly from EEG-fMRI data without ECG reference [57]. The protocol involves several stages: (1) ICA is applied to EEG data after MRI artifact removal to separate BCG artifact components (CB-ICs); (2) CB-ICs are automatically identified based on peak contrast values (mean amplitude of peaks between 1-4 standard deviations), selecting components with C > 2.3 for optimal detection; (3) A two-step band-pass filtering process—first to estimate period based on the fundamental frequency of heart beats (identified via smoothed spectrum analysis), and second to smooth the signal between 0.1Hz and the MR slice acquisition frequency (e.g., 17Hz); (4) Peak selection based on amplitude, rise, slope, and past period variations; (5) Adjustment of detected peak locations using reference frequency thresholds [57]. This method achieved impressive performance metrics on a large dataset (281 resting scans from 48 subjects, totaling 39.98 hours), with precision of 0.996, recall of 0.994, and F1-score of 0.995 [57].

Deep Learning Model Training Protocol

For BCG artifact removal using deep learning, a comprehensive training protocol has been established: (1) Simultaneous EEG-fMRI data collection using standard equipment (e.g., 3T Siemens Prisma scanner with 64-channel BrainAmp MR Plus system); (2) EEG pre-processing including low-pass filtering (cutoff at 70 Hz), gradient artifact removal using FASTR algorithm, resampling to 500 Hz, and high-pass filtering at 0.25 Hz; (3) Data preparation involving further resampling to 100 Hz and channel normalization to zero-mean; (4) Model training using recurrent neural networks (RNNs) with gated recurrent units (GRUs) to learn nonlinear mappings between ECG and BCG-corrupted EEG; (5) Evaluation against traditional methods like Optimal Basis Set (OBS) at individual subject level and investigation of cross-subject generalization [23]. This approach demonstrates larger average power reduction of BCG at critical frequencies while simultaneously improving task-relevant EEG classification performance [23].

Diagram 2: BCG Artifact Removal Approaches

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Materials for BCG Artifact Investigation

Item	Function/Application	Specifications/Alternatives
EEG-fMRI System	Simultaneous data acquisition	32-64 channel MR-compatible EEG system (e.g., Brain Products) with fMRI scanner
BCG Sensors	Measuring cardiac mechanical forces	IMUs, electromechanical films, piezoelectric, hydraulic- or pneumatic-pressure sensors
ECG Recording Equipment	Reference signal acquisition	MR-compatible ECG electrodes and amplifiers
Signal Processing Software	Data analysis and algorithm development	MATLAB with EEGLAB, FMRIB Plugin, Python with MNE, scikit-learn
Deep Learning Frameworks	Advanced artifact removal models	TensorFlow, PyTorch for implementing RNNs, GANs, autoencoders
Template-Based Removal Tools	Standard BCG artifact correction	Average Artifact Subtraction (AAS), Optimal Basis Set (OBS) algorithms
Blind Source Separation Tools	Component-based artifact removal	Independent Component Analysis (ICA) implementations (Infomax, FastICA)

The peak detection problem, encompassing the challenges of accurately identifying BCG J-peaks versus ECG R-peaks and managing timing jitter, remains a significant focus in simultaneous EEG-fMRI research. The physiological distinction between electrical cardiac events (R-peaks) and mechanical forces (J-peaks), combined with the variable R-J interval, creates fundamental challenges for precise heartbeat detection. Timing jitters in BCG peak detection directly impact downstream applications, particularly HRV analysis and sleep staging, where errors as small as 60 ms can increase sleep-scoring errors from 17% to 25% [55]. While classical signal processing approaches provide foundational methods, modern deep learning solutions—including surrogate modeling, RNNs, GANs, and denoising autoencoders—demonstrate superior performance in both J-peak detection and BCG artifact removal [53] [23] [56]. The development of direct BCG period determination methods that eliminate dependency on ECG signals further simplifies experimental setups and enables large-scale data processing [57]. As research advances, the refinement of these methodologies will continue to enhance the accuracy and reliability of BCG analysis, strengthening its utility in both clinical and research applications within neuroscience and beyond.

The ballistocardiogram (BCG) artifact remains one of the most significant challenges in simultaneous EEG-fMRI research. This artifact, caused by cardiac-related motions such as scalp pulsation and head movement within the strong magnetic field, can obscure neural signals of interest and compromise data integrity [3] [4]. Among the various correction techniques, the Optimal Basis Set (OBS) method has established itself as a fundamental approach for BCG artifact removal. At the core of implementing OBS effectively lies a critical parameter selection problem: determining the optimal number of principal components to remove from the EEG signal [5] [36].

This technical guide provides an in-depth examination of strategies for optimizing this key parameter, framed within the broader context of BCG artifact correction methodologies. We synthesize evidence from recent research to present structured decision frameworks, quantitative evaluation metrics, and advanced adaptive techniques that enable researchers to make informed, data-driven decisions about component selection. Proper optimization of this parameter is essential for balancing two competing objectives: maximizing artifact removal while minimizing distortion of physiological brain signals [5] [36].

The OBS Method: Foundation and Component Selection Challenge

The Optimal Basis Set method operates on a fundamental principle: representing the complex BCG artifact using a set of basis functions derived via Principal Component Analysis (PCA). The algorithm functions by segmenting the continuous EEG signal into epochs time-locked to cardiac events (typically R-peaks detected from ECG). PCA is then applied to these artifact-rich epochs to extract principal components that capture the dominant patterns of BCG artifact variation across the scalp [5] [58]. The central implementation challenge resides in determining how many of these components constitute the "optimal basis set" for artifact representation and subsequent subtraction.

Table 1: Core Components of the OBS Methodology

Component	Function	Implementation Considerations
Cardiac Event Detection	Identifies timing of BCG artifact occurrences for epoching	Typically uses ECG R-peaks; alternative methods use BCG peaks from EEG itself [5]
Principal Component Analysis (PCA)	Decomposes artifact epochs into orthogonal components	Captures spatio-temporal variations of BCG artifact across channels and time [5] [36]
Basis Set Selection	Determines which components represent artifact	Critical step: Too few components leave residuals; too many distort neural signals [5]
Artifact Reconstruction & Subtraction	Removes artifact from original signal	Linear combination of selected components fitted to and subtracted from each artifact occurrence [58]

The fundamental challenge in OBS implementation stems from the inherent trade-off between artifact removal efficacy and neural signal preservation. Selecting too few components results in incomplete artifact removal, leaving residual BCG contamination that can generate spurious correlations in subsequent analyses. Conversely, selecting too many components risks removing genuine neural activity along with the artifact, potentially distorting event-related potentials and other signals of interest [5] [36]. This balance is particularly crucial for drug development research, where precise characterization of neural signals is paramount.

Established Approaches and Their Limitations

Traditional Fixed-Component Approaches

Early implementations of OBS frequently employed a fixed-component approach, typically removing the first 3-4 principal components from each channel based on initial validation studies [5] [36]. This method offered simplicity and standardization across studies but failed to account for important sources of variability that affect optimal component selection.

Key limitations of the fixed-component approach include:

• Inter-subject variability: Differences in head size, anatomy, and electrode-scalp contact affect BCG artifact characteristics [5]
• Inter-channel variability: Artifact manifestation differs across brain regions and electrode locations [59]
• Recording condition effects: Magnetic field strength (3T vs. 7T) and EEG system setup influence artifact complexity [60] [36]

Visual Inspection and Qualitative Assessment

Many researchers have employed qualitative assessment of corrected EEG traces and averaged artifact templates to guide component selection. This approach involves visually inspecting the residual artifact after subtraction with different component numbers and selecting the threshold where BCG artifacts appear minimized without introducing signal distortion [58]. While providing valuable direct feedback, this method introduces subjectivity and depends heavily on researcher experience, potentially compromising reproducibility across studies.

Quantitative Frameworks for Component Selection

Recent research has developed more rigorous, quantitative frameworks for determining the optimal number of OBS components. These approaches leverage specific metrics to systematically evaluate the trade-off between artifact removal and signal preservation.

Variance-Based Selection Criteria

The adaptive OBS (aOBS) method automates component selection by employing a variance-explained threshold across subjects and channels. This data-driven approach determines the number of components needed to capture a pre-defined percentage of cumulative variance in the artifact template [5]. Implementation studies have revealed that the optimal number of components can vary significantly, ranging from 1 to 8 components depending on the subject and specific channel being analyzed [5].

Table 2: Component Selection Methods and Their Performance

Method	Selection Criteria	Components Typically Removed	Advantages	Limitations
Fixed-Component OBS	Pre-defined number (e.g., 3-4)	3-4	Simple, standardized	Fails to account for inter-subject and inter-channel variability [5] [36]
Variance-Based aOBS	Cumulative variance threshold	1-8 (varies by subject/channel)	Data-driven, adaptive	Requires parameter tuning; may overfit in high-noise conditions [5]
PROJIC-OBS	Component classification via projection	Selective removal of BCG-related components	Preserves neural signal better	Complex implementation; requires additional processing steps [36]
Cross-Correlation Methods	Residual correlation with ECG	Variable	Direct measure of artifact removal efficacy	Does not directly assess neural signal preservation [5]

Performance Metrics for Method Evaluation

Researchers have developed sophisticated evaluation pipelines to quantitatively assess different component selection strategies. These metrics help determine the optimal balance between artifact removal and physiological signal preservation:

• BCG Residual Intensity: Measures remaining artifact power after correction [5]
• Cross-Correlation with ECG: Quantifies residual cardiac-related activity in corrected EEG [5]
• Signal-to-Noise Ratio (SNR) of ERPs: Assesses preservation of neural signals of interest [5] [36]
• Inter-trial Variability: Measures consistency of evoked responses across trials [5] [36]
• Source Localization Error: Evaluates impact on downstream analysis applications [3]

Studies employing these metrics have demonstrated that adaptive selection methods can achieve significantly better outcomes than fixed-component approaches. For instance, aOBS has been shown to reduce BCG residuals to approximately 5.5% compared to 9.2% for traditional OBS and 12.5% for average artifact subtraction (AAS) methods [5].

Advanced Hybrid Methodologies

Surrogate-Based Source Modeling

Surrogate source modeling approaches represent a significant advancement in BCG artifact correction. These methods use spatial filtering techniques that incorporate biophysical head models to separate artifact-related signals from brain activity with minimal distortion [3]. Rather than relying solely on component counting, these approaches create artifact topographies using PCA (PCA-S) or manual selection of ICA components (ICA-S). Studies have demonstrated that these surrogate methods, particularly PCA-S, provide substantial improvements for subsequent source localization analysis compared to traditional OBS approaches [3].

PROJIC Framework and Component Classification

The PROJIC (PROJection onto Independent Components) framework introduces a novel approach for identifying BCG-related independent components before applying OBS correction in the component space [36]. This hybrid methodology – PROJIC-OBS – has demonstrated superior performance when priority is given to artifact removal, while PROJIC-AAS (using Average Artifact Subtraction) performs better when emphasizing physiological signal preservation [36]. These approaches help mitigate the component selection problem by first classifying components as artifact-related or neural-related before applying correction.

Experimental Protocols for Parameter Optimization

For researchers implementing OBS methods, we recommend the following experimental protocol for determining the optimal number of components in specific experimental contexts:

Protocol 1: Systematic Component Sweep

• Preprocessing: Apply gradient artifact correction using established methods (e.g., AAS) [58]
• Cardiac Event Detection: Identify BCG occurrences using ECG R-peaks or BCG peaks from EEG [5]
• Template Creation: Epoch EEG data around cardiac events and create average artifact templates
• PCA Decomposition: Apply PCA to artifact templates across channels
• Iterative Reconstruction: Reconstruct and subtract artifacts using different numbers of components (1 to k)
• Performance Assessment: Calculate BCG residual intensity and neural signal preservation for each k
• Optimal Point Selection: Identify the value k that maximizes artifact removal while minimizing neural signal distortion

Protocol 2: Validation Using Known Neural Signals

• Data Collection: Acquire simultaneous EEG-fMRI during a paradigm generating robust, known ERPs [5] [36]
• Reference Data: Collect outside-scanner EEG for the same paradigm to establish ground truth neural responses [36]
• Component Sweep: Apply OBS correction with varying component numbers
• Fidelity Assessment: Compare inside-scanner ERPs (after correction) to outside-scanner ground truth
• Optimal Selection: Choose component count that produces ERPs most similar to ground truth

This protocol was used effectively in a visual stimulation study, where aOBS demonstrated clearer spatial topography aligned with occipital region activation compared to fixed-component approaches [5].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for OBS Optimization Research

Tool/Resource	Function in OBS Optimization	Implementation Example
High-Density EEG Systems (64+ channels)	Captures spatial distribution of BCG artifact for improved PCA decomposition	64-channel MRI-compatible EEG system [3]
Synchronized ECG Recording	Provides precise cardiac timing for accurate artifact epoching	MRI-compatible ECG with pulse oximeter [5]
PCA/Linear Algebra Libraries	Performs decomposition of artifact templates into principal components	MATLAB PCA functions; Python scikit-learn [5]
Visual Stimulation Apparatus	Generates robust ERPs for validation of neural signal preservation	MRI-compatible visual presentation systems [5]
Reference Layer Artifact Subtraction Hardware	Provides alternative artifact correction for method comparison	Carbon wire loops; insulated electrode systems [59] [61]

Optimizing the number of components in OBS-based BCG artifact correction represents a critical step in ensuring the validity of simultaneous EEG-fMRI research. The evidence overwhelmingly indicates that adaptive, data-driven approaches outperform fixed-component strategies across multiple performance metrics. Researchers should implement systematic evaluation protocols that quantitatively assess both artifact removal efficacy and neural signal preservation specific to their experimental context.

Future methodological developments will likely focus on fully automated optimization of this parameter, potentially leveraging machine learning approaches to dynamically adjust component selection across subjects, channels, and even temporal segments within recordings. Such advancements will further enhance the reliability and accessibility of simultaneous EEG-fMRI for both basic neuroscience and applied drug development research.

Visual Workflows

Optimizing OBS Component Selection Workflow

Impact of Component Selection on Correction Outcomes

Simultaneous EEG-fMRI is a powerful neuroimaging technique that combines the high temporal resolution of electroencephalography (EEG) with the high spatial resolution of functional magnetic resonance imaging (fMRI). However, EEG signals recorded inside the MRI scanner are contaminated by significant artifacts, with the ballistocardiogram (BCG) artifact representing one of the most challenging to remove. The BCG artifact is a complex signal distortion originating from multiple cardiac-related phenomena: head rotation in the static magnetic field, the Hall effect of pulsatile blood flow, and pulse-driven scalp expansion [36] [22]. Without effective correction, these artifacts can obscure genuine neural activity and lead to false interpretations in both clinical and research settings.

Independent Component Analysis (ICA) has emerged as a primary tool for BCG artifact removal due to its ability to separate mixed signals into statistically independent components (ICs). The fundamental challenge, however, lies in accurately identifying which components represent artifacts versus neural signals. Incorrect selection can result in two problematic outcomes: (1) residual artifacts contaminating the EEG data, or (2) accidental removal of physiological signals of interest [62] [26]. This component selection challenge is particularly acute in drug development studies, where preserving authentic neural signatures is crucial for assessing pharmacological effects on brain function.

The Complexity of BCG Artifacts and ICA Assumptions

The BCG artifact exhibits several characteristics that complicate ICA-based removal. Unlike gradient artifacts, BCG artifacts are highly non-stationary, with shape, amplitude, and scale varying over time due to factors such as blood pressure changes and slight head movements [44] [36]. This non-stationarity violates the fundamental statistical assumptions of conventional ICA, which expects source signals to remain constant over time.

Additionally, BCG artifacts originate from multiple concurrent physiological processes that may not be perfectly statistically independent from each other or from neural sources [22]. This source dependency challenges the core ICA model. The artifact also demonstrates considerable inter-subject variability, making universal correction templates ineffective and necessitating individualized approaches [44].

These complexities mean that conventional ICA application often results in either residual artifact contamination or excessive removal of neural signals. As noted by Abreu et al., "BCG artifact correction invariably yields residual artifacts and/or deterioration of the physiological signals of interest" without advanced methodological adaptations [62].

Quantitative Comparison of ICA Component Selection Methods

Table 1: Performance Comparison of BCG Artifact Removal Methods

Method	Key Principle	Artifact Reduction Efficacy	Neural Signal Preservation	Best Application Context
PROJIC-OBS [62]	ICA decomposition followed by OBS correction of BCG-related ICs	High (32-62% SE reduction)	Moderate	Priority on artifact removal
PROJIC-AAS [62]	ICA decomposition followed by AAS correction of BCG-related ICs	Moderate (18-61% SE reduction)	High	Priority on signal preservation
ccICA [44]	Clustering algorithm to capture features + constrained ICA	High error reduction in simulated data	High INPS values	Handling time-varying BCG features
aOBS [5]	Adaptive detection of BCG peaks + automated component selection	Low BCG residuals (5.53%)	Clear spatial topography	High-density EEG setups
Surrogate Methods [22]	Spatial filtering using artifact topographies from PCA/ICA	Substantial improvement in source localization	Minimal distortion of brain activity	Source analysis studies

Table 2: Performance Metrics Across Methodologies

Method	Signal-to-Noise Improvement	Residual Artifact Level	Computational Complexity	Implementation Accessibility
AAS [5]	Moderate	12.51%	Low	High
OBS [5]	Moderate-High	9.20%	Medium	High
Standard ICA [5]	Variable	20.63%	Medium-High	Medium
aOBS [5]	High	5.53%	Medium	Medium
PROJIC Variants [62]	High	Low-Moderate	High	Low

Advanced Methodologies for Improved Component Selection

PROJIC Framework and Variants

The PROJIC (PROJection onto Independent Components) framework introduces a novel approach for selecting BCG-related independent components before applying correction. The methodology employs three distinct strategies:

• PROJIC: BCG-related ICs are identified and completely removed from the back-reconstruction of the EEG.
• PROJIC-OBS: BCG-related ICs are corrected using an Optimal Basis Set framework before back-projection.
• PROJIC-AAS: BCG-related ICs are corrected using Average Artifact Subtraction before back-projection [62].

The key innovation lies in the evaluation pipeline, which quantitatively assesses both artifact removal and physiological signal preservation. This allows researchers to flexibly weight the importance given to neural signal preservation based on their specific research goals. In experimental validation, these methods achieved significant reductions in standard error across trials: 26-66% for PROJIC, 32-62% for PROJIC-OBS, and 18-61% for PROJIC-AAS at different field strengths [62].

Clustering-Constrained ICA (ccICA)

The ccICA method addresses the time-varying nature of BCG artifacts by incorporating a clustering algorithm to capture the artifacts' features before applying constrained ICA. This approach specifically accounts for variations in BCG artifact shape, amplitude, and scale over time [44].

The experimental implementation demonstrated superior performance compared to traditional methods. In simulated data analysis, the error in signal amplitude (Er) computed by ccICA was significantly lower than AAS, OBS, and standard cICA methods (p < 0.005). In vivo data analysis showed the Improvement of Normalized Power Spectrum (INPS) calculated by ccICA was substantially higher than these comparison methods [44].

Adaptive Optimal Basis Set (aOBS)

The aOBS method enhances traditional OBS by incorporating adaptive detection of BCG occurrences and automated selection of principal components. The algorithm operates through two sequential steps:

• Detection of BCG occurrences: ECG peaks are detected, followed by identification of BCG peaks in gradient artifact-corrected EEG data.
• BCG artifact reconstruction and subtraction: For each channel, EEG data are epoched based on BCG peaks, PCA is applied, and components are selected based on explained variance criteria [5].

This adaptive approach automatically identifies artifactual components based on signal features rather than fixed thresholds. In validation studies, aOBS demonstrated significantly lower BCG residuals (5.53%) compared to AAS (12.51%), ICA (20.63%), and traditional OBS (9.20%) [5].

Surrogate Spatial Filtering

Surrogate methods utilize spatial filtering to separate artifact signals from brain activity based on their distinct spatial distributions. The approach can be implemented using two variations:

• PCA-Surrogate (PCA-S): Artifact topographies are obtained through principal component analysis performed on an averaged artifact template.
• ICA-Surrogate (ICA-S): Artifact topographies are identified through manual selection of artifact components using independent component analysis [22].

This method assumes that artifact and brain signals can be separated if their spatial distributions are known. The brain signals are estimated using a surrogate model consisting of regional dipole sources distributed throughout the brain. In evaluations focusing on source localization accuracy, surrogate methods outperformed established approaches including OBS, BSS, and OBS-ICA [22].

Diagram 1: Workflow of Advanced ICA Component Selection Methods for BCG Artifact Removal

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Materials for BCG Artifact Removal Studies

Item	Specification	Research Function	Example Implementation
MR-Compatible EEG System	64+ channels; Amplifier synchronized with MRI clock [22]	Records neural activity without interference from magnetic fields	SynAmps2 system (Neuroscan) used with 3T Siemens scanner
ECG Recording Electrode	Additional electrode placed on chest [44]	Captures cardiac signal for R-peak detection and artifact template creation	Single Ag/AgCl electrode integrated with EEG cap
Carbon Wire Loops [26]	Motion sensors placed on head	Records motion-induced currents for reference subtraction	Custom-built loops attached to EEG cap
PCA/ICA Software Tools	FMRIB Plug-in (University of Oxford) [44]	Implements OBS, AAS, and related algorithms for artifact removal	Used for parameter optimization across subject groups
BESA Research Software [22]	Version 7.1 with surrogate spatial filtering	Implements PCA-S and ICA-S methods for artifact reduction	Used for source localization accuracy assessment
Visual Stimulation Apparatus	MR-compatible display system	Generates event-related potentials for method validation	Presents face images for N170 ERP component [44]

Experimental Protocols for Method Validation

Protocol for PROJIC Method Evaluation

Data Acquisition Parameters:

• Simultaneous EEG-fMRI data acquired using different setups at both 3T and 7T field strengths
• EEG sampling rate: 1,000 Hz with 64 Ag/AgCl electrodes positioned according to international system
• fMRI parameters: TR = 2,000 ms, TE = 35 ms, flip angle = 90°, FOV = 230 × 180 mm², 24 slices per volume [62]

Evaluation Pipeline:

• Implement novel evaluation pipeline assessing both artifact removal and physiological signal preservation
• Apply group-level parameter optimization for each algorithm
• Assess impact on event-related potentials through relative reduction of standard error across trials
• Compare with state-of-the-art BCG correction methods (AAS, OBS) [62]

Protocol for ccICA Validation

Experimental Design:

• Acquire EEG-fMRI data from 10 healthy subjects (age range: 20-25; 7 males, 3 females)
• Subjects lie relaxed in scanner with eyes closed and head fixed
• Additional task-EEG data collected from 3 subjects to extract ERP "N170" component
• Visual stimulation paradigm: 56 trials with duration of 4s each [44]

Analysis Framework:

• Apply ccICA to simulated data constructed by adding BCG artifacts to conventional EEG
• Test method on in vivo EEG-fMRI data from healthy subjects
• Calculate error in signal amplitude (Er) for simulated data analysis
• Compute Improvement of Normalized Power Spectrum (INPS) for in vivo data [44]

Protocol for aOBS Performance Assessment

Visual Stimulation Paradigm:

• Implement visual stimulation during simultaneous EEG-fMRI acquisition
• Use high-density EEG montage for comprehensive spatial sampling [5]

Quantitative Metrics:

• Calculate BCG residual intensity post-correction
• Compute maximum cross-correlation between EEG signals and ECG signal
• Assess event-related potentials in terms of signal-to-noise ratio (SNR)
• Evaluate inter-trial variability across methods [5]

Implications for Drug Development and Future Directions

In pharmaceutical research, where fMRI increasingly contributes to understanding drug mechanisms, preserving authentic neural signals is paramount. The component selection challenge in ICA directly impacts the reliability of EEG biomarkers used in clinical trials. As regulatory agencies like the FDA and EMA emphasize the need for qualified biomarkers in drug development, robust artifact removal methods become essential for generating reproducible results [63].

Future methodological developments should focus on integrating hardware-based solutions with algorithmic approaches. Reference layer systems and carbon wire loops show promise in capturing artifact signals directly, potentially simplifying the component selection challenge [26] [22]. Additionally, machine learning approaches for automated component classification may further reduce the subjective elements currently present in IC selection.

The field would benefit from standardized evaluation frameworks and benchmark datasets to facilitate direct comparison of emerging methodologies. Such standardization is particularly crucial for drug development applications, where methodological consistency across multiple research sites is essential for regulatory acceptance [63] [26].

The component selection challenge in ICA represents a critical methodological frontier in simultaneous EEG-fMRI research. While BCG artifacts remain problematic, advanced approaches such as PROJIC, ccICA, aOBS, and surrogate methods demonstrate significant progress in selectively removing artifacts while preserving neural signals. The continuing refinement of these methodologies holds particular importance for drug development, where accurately measuring pharmacodynamic effects on brain function depends on extracting authentic neural signatures from artifact-contaminated recordings. Through careful implementation of these advanced component selection strategies, researchers can maximize the unique potential of simultaneous EEG-fMRI to illuminate brain function in both health and disease.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) is a powerful multimodal neuroimaging technique that combines the high temporal resolution of EEG with the high spatial resolution of fMRI. This integration offers unparalleled potential for studying human brain function by capturing rapid neural events and their underlying hemodynamic correlates within a precise anatomical framework [23] [22]. The technique has found significant applications in diverse areas, including studies of attentional orienting, perceptual decision making, value-based decision making, reward processing, resting-state networks, and the localization of epileptic events [23]. However, the quality of EEG signals recorded inside the MRI scanner is severely compromised by two predominant artifacts: the gradient artifact (GA), caused by the rapid switching of magnetic field gradients during image acquisition, and the ballistocardiogram (BCG) artifact, which is the focus of this discussion [23] [5].

The BCG artifact is a large-amplitude contamination of the EEG signal, predominantly caused by cardiac-induced motion. Its primary generators include cardiac-driven head movements within the static magnetic field, local pulsatile movements of the scalp due to blood flow through scalp vessels, and, to a lesser extent, the Hall effect produced by pulsatile blood flow as a conductive fluid [20] [22]. In a 3T MRI system, the amplitude of the BCG artifact can exceed 200 μV, often dwarfing the underlying neural EEG signals by tens of times [20]. Unlike the highly reproducible gradient artifact, the BCG artifact exhibits complex spatiotemporal dynamics and morphological variability both across different cardiac cycles and over the duration of a recording session [23] [22]. This time-varying nature, or non-stationarity, is a central challenge. The BCG artifact does not form clean, stereotypical repeats around the cardiac cycle, as its waveform can vary significantly from channel to channel and its temporal relation to the electrocardiogram (ECG) is intricate and non-linear [23]. This non-stationarity arises from various factors, including subtle changes in head position within the MRI scanner, fluctuations in blood pressure, and alterations in electrode-skin contact impedance [22]. Consequently, constructing a single, static template that accurately describes all cardiac events is not feasible, making the BCG artifact a more persistent obstacle to clean EEG data than the gradient artifact [20]. This whitepaper delves into the core strategies that have been developed to address the critical challenge of time-varying BCG morphology.

Methodological Strategies for Time-Varying BCG

A range of software and hardware-based methods have been developed to tackle the BCG artifact, with their efficacy heavily dependent on how well they handle its non-stationary characteristics. The following sections and Table 1 provide a comparative overview of these key methodologies.

Table 1: Comparison of BCG Artifact Removal Methods Addressing Non-Stationarity

Method Category	Examples	Key Mechanism for Handling Non-Stationarity	Reported Performance Advantages	Limitations
Template-Based & Linear Decomposition	Optimal Basis Set (OBS) [5]	Uses PCA on epoched data to create an adaptive basis set, better capturing variance than a fixed average.	More robust than AAS; widely used and integrated into processing pipelines.	Assumes linearity; fixed number of components may not suit all dynamics; performance hampered by imperfect ECG-BCG alignment.
Adaptive Extensions of OBS	Adaptive OBS (aOBS) [5]	Beat-to-beat estimation of delay between ECG and BCG peak; automatic selection of PCA components based on explained variance.	Lower BCG residuals (5.53%) vs OBS (9.20%) and AAS (12.51%); improved reduction of EEG-ECG cross-correlation.	Algorithmic complexity is increased; requires validation of the adaptive peak detection.
Blind Source Separation	Independent Component Analysis (ICA) [22]	Statistically decomposes data into independent components; artifactual components are manually or automatically identified and removed.	Does not require cardiac timing information; can separate sources with overlapping topographies.	ICA's stationarity assumption is violated by time-varying BCG topography; challenging to reliably identify artifact components.
Hardware-Based Referencing	BCG Reference Layer (BRL) [20]	Dedicated electrodes on an insulated conductive layer record only BCG artifact, providing a reference signal free of neural EEG.	Improved CNR for alpha-wave (101%), VEP (76%), and AEP (75%) over OBS using 160 reference electrodes.	Requires specialized equipment/caps; insensitive to the Hall effect component of the BCG; not applicable to existing datasets.
Deep Learning	BCGNet (RNN with GRU) [23]	Models the non-linear and state-dependent mapping between the ECG and the BCG-corrupted EEG using recurrent neural networks.	Greater BCG power reduction at critical frequencies; improved task-relevant EEG classification compared to OBS.	Requires large datasets for training; "black box" nature; computational cost and complexity.
Surrogate Spatial Filtering	PCA-Surrogate, ICA-Surrogate [22]	Uses a surrogate source model of the brain to separate artifact spatial topographies (from PCA/ICA) from brain signals, minimizing distortion.	Superior source localization accuracy; higher signal-to-noise ratio and lower inter-trial variability in ERPs.	Relies on accuracy of the surrogate brain model; can be computationally intensive.

Adaptive Template-Based and Decomposition Methods

Early and fundamental software approaches for BCG removal are template-based. The Average Artifact Subtraction (AAS) method involves epoching the EEG data around detected cardiac events (e.g., R-peaks from the ECG), creating an average artifact template for each channel, and subtracting this template from each individual occurrence [20] [5]. However, AAS is inherently limited by the non-stationarity of the BCG, as the average template fails to capture beat-to-beat variability. The Optimal Basis Set (OBS) method was developed to address this limitation. OBS extends AAS by applying Principal Component Analysis (PCA) to the epoched data. Instead of a single average, the first several principal components form a basis set that adaptively captures the main modes of BCG variance, which is then subtracted from the data [23] [5]. A key challenge in OBS is selecting the number of components to remove, as using too few leaves residuals, while using too many risks removing neural activity [20] [5].

The Adaptive Optimal Basis Set (aOBS) method represents a significant evolution designed explicitly for non-stationarity. aOBS introduces two key innovations [5]. First, it dynamically estimates the delay between the ECG R-peak and the subsequent BCG peak on a beat-to-beat basis, ensuring superior alignment of BCG occurrences before PCA is applied. Second, it implements an automated data-driven criterion for selecting the number of principal components to remove, moving beyond the typical fixed number of three or four. This adaptability allows aOBS to better track the changing BCG morphology, resulting in substantially lower BCG residuals compared to standard AAS, ICA, and OBS methods [5].

Blind Source Separation and Hybrid Approaches

Blind Source Separation (BSS), particularly using Independent Component Analysis (ICA), offers a different philosophy. ICA decomposes the multi-channel EEG data into statistically independent components (ICs), with the assumption that artifacts and brain signals will separate into different ICs [22]. The artifactual components are then identified and removed before reconstructing the cleaned EEG. However, the fundamental assumption of ICA—that sources are linearly mixed and stationary—is challenged by the time-varying topography of the BCG artifact, which can lead to components containing mixed neural and artifactual signals [5] [22].

To leverage the strengths of different methods, hybrid approaches have been developed. The OBS-ICA method, for instance, involves applying OBS first to reduce the bulk of the BCG artifact, followed by ICA to identify and remove any residual BCG components that survived the initial cleaning [22]. Conversely, the ICA-OBS method involves performing ICA first, then applying OBS to each independent component to clean it of any remaining BCG influence before reconstruction [5]. While these combinations can be powerful, they also risk propagating the limitations of both methods if not applied carefully, potentially leading to distortion of the neural signal [22].

Hardware-Based and Advanced Model-Based Approaches

Hardware-based solutions provide an alternative strategy by directly measuring the artifact. The BCG Reference Layer (BRL) method employs a specialized cap where a set of "BCG electrodes" is placed on a thin conductive layer that is mechanically attached to but electrically insulated from the scalp. These electrodes record the BCG artifact in the absence of neural signals, providing a nearly pure reference signal that can be used for template-based subtraction [20]. This method effectively handles non-stationarity arising from both global head movements and local scalp movements without distorting the underlying EEG, as the reference signal itself captures the variability.

More recently, advanced model-based and deep learning approaches have emerged. Surrogate Spatial Filtering uses a predefined source model of the brain to separate the measured signal into brain and artifact components based on their spatial topographies. The artifact topographies can be derived from PCA or ICA, but the key difference from standard BSS is the use of the brain model to prevent neural signal distortion during artifact removal, leading to improved source localization [22]. Deep learning, specifically Recurrent Neural Networks with Gated Recurrent Units (RNNs-GRU), represents a paradigm shift by learning the non-linear and state-dependent mapping between the ECG (input) and the BCG-corrupted EEG (output). This model, termed BCGNet, can capture complex temporal dynamics that linear methods like OBS cannot, offering superior artifact suppression and enhanced preservation of task-relevant neural signals [23] [27].

Experimental Protocols for Key Methodologies

Protocol for Adaptive Optimal Basis Sets (aOBS)

The aOBS method can be implemented based on the workflow detailed by [5]. The following is a generalized protocol:

• Data Preprocessing: Begin with EEG data that has already undergone standard preprocessing, including gradient artifact removal using a tool like FASTR [23] [5]. Resample the data to a manageable rate (e.g., 500 Hz) and apply appropriate band-pass filtering (e.g., 0.25-70 Hz).
• BCG Peak Detection: Instead of relying solely on ECG R-peaks with a fixed delay, identify the BCG peaks directly from the gradient-artifact-corrected EEG data. This can be done by band-pass filtering a central EEG channel (e.g., 8-15 Hz) and using a peak detection algorithm to find the dominant positive peak following each R-peak. This step adaptively locks onto the actual BCG event for each heartbeat.
• Data Epoching and PCA: For each EEG channel, epoch the data into segments time-locked to the identified BCG peaks. Perform Principal Component Analysis (PCA) on the matrix of these epoched segments.
• Automatic Component Selection: Calculate the cumulative explained variance for the principal components. Automatically select the number of components, k, required to exceed a predetermined threshold of explained variance (e.g., 90-95%). This k can vary across channels and subjects, adapting to the local complexity of the artifact.
• Artifact Reconstruction and Subtraction: For each epoch, reconstruct the BCG artifact by projecting the data onto the first k principal components and then back-projecting. Subtract this reconstructed artifact from the original epoch. Finally, concatenate the cleaned epochs to form the continuous, artifact-reduced EEG signal.

Protocol for Deep Learning-Based Removal (BCGNet)

The protocol for implementing the BCGNet deep learning approach, as described by [23], involves the following stages:

• Subject and Data Acquisition: Collect simultaneous EEG-fMRI data from subjects (e.g., 19 subjects as in the original study) performing a paradigm, such as an auditory oddball task. Record 63-channel EEG at a high sampling rate (e.g., 5 kHz) along with ECG, inside the MRI scanner.
• Data Preprocessing: Preprocess the raw EEG data by importing it into a processing environment like EEGLAB. Apply low-pass filtering (e.g., 70 Hz cutoff), remove gradient artifacts using a method like FASTR, and then resample the data to a lower frequency (e.g., 100 Hz) to reduce computational load. This preprocessed data is termed "BCG corrupted EEG" (BCE).
• Network Architecture and Training: Design a deep recurrent neural network using Gated Recurrent Units (GRUs). The input to the network is the ECG signal, and the target output is the BCE signal. The objective of the network is to learn the non-linear mapping function that predicts the BCG artifact component. The model is trained by minimizing the difference between its prediction and the actual BCE signal.
• Artifact Subtraction and Evaluation: Once trained, use the network to generate a prediction of the BCG artifact for the data. Subtract this predicted artifact from the original BCE to obtain the cleaned EEG. Evaluate the performance by comparing the power reduction in BCG-dominated frequency bands and the improvement in the classification accuracy of task-relevant EEG features (e.g., P300 ERP from the oddball task) against established methods like OBS.

Protocol for BCG Reference Layer (BRL)

The implementation of the BRL method, according to [20], requires both hardware setup and a processing pipeline:

• Hardware Setup: Utilize a specialized EEG cap that incorporates a permanent, reusable conductive reference layer insulated from the scalp. This cap includes two sets of electrodes: standard "scalp electrodes" that contact the scalp to record EEG + BCG, and "BCG electrodes" that contact only the reference layer to record the BCG artifact in isolation.
• Data Acquisition: Record data simultaneously from all scalp electrodes and BCG reference electrodes during fMRI acquisition. The number of reference electrodes can vary (e.g., 20 to 160).
• Signal Modeling and Subtraction: For each scalp electrode, model the recorded signal as a linear combination of the true EEG and the BCG artifact. Use the signals from multiple BCG reference electrodes to create a patient-specific and session-specific estimate of the BCG artifact at each scalp electrode. This is typically done by finding a spatial filter (e.g., via linear regression) that best matches the reference signals to the artifact in the scalp channels.
• Reconstruction: Subtract the estimated BCG artifact from the recorded signal at each scalp electrode to obtain the cleaned EEG.

Visualization of Method Workflows

Workflow for Adaptive Optimal Basis Set (aOBS)

Diagram 1: The aOBS method uses adaptive BCG peak detection and automated component selection to handle non-stationarity.

Workflow for Deep Learning BCG Removal (BCGNet)

Diagram 2: BCGNet uses a deep RNN to model the complex, non-linear relationship between the ECG and BCG artifact.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Materials and Tools for BCG Artifact Research

Item	Function / Description	Key Consideration
MRI-Compatible EEG System (e.g., BrainAmp MR Plus)	Records EEG data inside the high-magnetic-field environment without causing interference or safety hazards.	Must be specifically designed for MRI use; includes specialized amplifiers and cables.
High-Density EEG Cap (e.g., 64-channel)	Provides extensive scalp coverage for capturing the spatial dynamics of both neural signals and BCG artifacts.	Crucial for methods like ICA and source localization; electrode impedances should be kept low (< 20 kΩ).
ECG Recording Electrode	Records the electrocardiogram, which serves as the timing reference for cardiac-locked artifacts.	Essential for template-based methods (AAS, OBS, aOBS) and for deep learning approaches like BCGNet.
BCG Reference Layer Cap	A specialized cap with a conductive layer and dedicated electrodes to record the BCG artifact in isolation.	Required for hardware-based BRL methods; eliminates need for software template estimation.
Sandbags / Head Stabilization	Used to stabilize the head and EEG amplifiers against scanner vibrations and minimize gross movement.	Reduces movement-induced artifacts, though does not eliminate BCG.
Software Tools (EEGLAB, FMRIB Plug-in, BESA, MNE-Python)	Provide ecosystems for data preprocessing, standard artifact removal (e.g., FASTR for GA), and implementation of advanced methods (OBS, ICA, custom scripts).	Flexibility for implementing and comparing different algorithms is key for research.
Computational Resources (GPU)	Accelerates the training of deep learning models (e.g., BCGNet's RNN) and the execution of computationally intensive methods like ICA.	Becomes critical for large datasets and complex non-linear models.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the high spatial resolution of fMRI. This integration enables unprecedented insights into brain dynamics across different spatiotemporal scales [3] [26]. However, the utility of this combined methodology is severely compromised by the presence of significant artifacts in the EEG data recorded inside the MR environment. Among these, the ballistocardiogram (BCG) artifact persists as one of the most challenging technical obstacles [26] [20].

The BCG artifact arises from multiple cardiac-related phenomena: head movements caused by cardiac pulsations, pulsatile scalp movements from blood flow in scalp arteries, and the Hall effect generated by the flow of conductive blood in the static magnetic field [20] [14]. In a 3T MRI system, the amplitude of BCG artifacts can exceed 200 μV—tens of times larger than neural EEG signals—making effective artifact removal essential for meaningful data interpretation [20]. The core challenge lies in the fundamental trade-off: overly aggressive artifact removal distorts neural signals, while insufficient removal leaves artifact residuals that generate spurious correlations between EEG and fMRI time-courses [64] [5].

This technical guide examines current methodologies for BCG artifact removal, focusing specifically on strategies to manage the critical trade-off between artifact suppression and neural signal preservation. We provide a comprehensive analysis of quantitative performance metrics, detailed experimental protocols, and practical implementation guidelines to assist researchers in optimizing their artifact removal pipelines for specific research applications.

BCG Artifact Removal Methodologies

Traditional and Modern Removal Approaches

Multiple methodological approaches have been developed to address the BCG artifact challenge, each with distinct mechanisms and limitations concerning the preservation of neural signals.

Template-Based Methods represent the earliest approach to BCG removal. The Average Artifact Subtraction (AAS) method creates a template by averaging EEG segments time-locked to cardiac events, then subtracts this template from the contaminated signal [20] [14]. While straightforward to implement, AAS assumes stationarity of the BCG artifact and often leaves significant residuals due to the artifact's inherent variability [5]. The Optimal Basis Set (OBS) method extends AAS by applying Principal Component Analysis (PCA) to artifact epochs and using the first several principal components as an adaptive template, better accounting for artifact shape variations [20] [5].

Blind Source Separation Methods, particularly Independent Component Analysis (ICA), separate EEG data into statistically independent components, allowing for manual or automated identification and removal of artifact-related components before signal reconstruction [3] [14]. The effectiveness of ICA depends critically on correct component selection, with misclassification potentially leading to substantial loss of neural information [14].

Hardware-Based Solutions utilize additional sensors to directly measure artifact signals. The BCG Reference Layer (BRL) approach employs additional electrodes placed on a conductive layer insulated from the scalp to record reference signals containing primarily BCG artifacts, which are then subtracted from the contaminated EEG [20]. While effective, these methods require specialized equipment and setup procedures [20].

Deep Learning Approaches represent the most recent innovation. Generative Adversarial Networks (GANs) have been applied to learn the mapping from BCG-contaminated to clean EEG without requiring paired examples [14]. The BCGGAN model employs a modular training strategy to optimize the generator network specifically for BCG removal, demonstrating improved performance over traditional methods [14].

Hybrid and Adaptive Methods

Recognizing the limitations of individual approaches, researchers have developed hybrid methods that combine strengths from multiple techniques. The OBS-ICA approach applies OBS before or after ICA decomposition to address residual artifacts [3] [26]. Similarly, the adaptive OBS (aOBS) method incorporates beat-to-beat estimation of the delay between cardiac activity and BCG occurrence, along with an automated criterion for selecting artifact-related components based on explained variance [5].

Table 1: Performance Comparison of BCG Artifact Removal Methods

Method	BCG Residual Intensity	Max Cross-Correlation with ECG	Signal Preservation Quality	Computational Complexity
AAS	12.51%	0.051	Moderate	Low
ICA	20.63%	0.067	Variable (component-dependent)	Medium
OBS	9.20%	0.042	Good	Medium
aOBS	5.53%	0.028	Excellent	Medium-High
BRL	N/A	N/A	Excellent	High (hardware required)
BCGGAN	N/A	N/A	Excellent	High

Table 2: Signal Quality Metrics Across Removal Methods [64]

Method	MSE	PSNR (dB)	SSIM	SNR
AAS	0.0038	26.34	0.68	15.22
OBS	0.0042	25.82	0.72	14.85
ICA	0.0051	24.94	0.65	13.97
OBS+ICA	0.0040	25.99	0.71	15.05

Quantitative Performance Evaluation

Artifact Removal Efficacy

The performance of BCG artifact removal methods can be quantified using multiple metrics assessing both artifact suppression and signal preservation. BCG residual intensity measures the percentage of artifact power remaining after correction, with lower values indicating better removal. Studies demonstrate that aOBS achieves significantly lower residual intensity (5.53%) compared to AAS (12.51%), ICA (20.63%), and standard OBS (9.20%) [5]. The maximum cross-correlation between EEG and ECG provides another key metric, with aOBS achieving the lowest values (0.028) compared to other methods [5].

Signal fidelity metrics offer complementary insights into neural signal preservation. The Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR) quantify the difference between processed signals and ground truth references, with AAS demonstrating superior performance (MSE=0.0038, PSNR=26.34dB) according to recent evaluations [64]. The Structural Similarity Index (SSIM) assesses structural preservation, where OBS achieves the highest scores (0.72) [64].

Impact on Functional Connectivity

The choice of artifact removal method significantly influences derived functional connectivity measures. Different methods produce distinct frequency-specific patterns in network topology, particularly affecting beta and gamma bands [64]. ICA, while weaker in traditional signal metrics, demonstrates sensitivity to frequency-specific patterns in dynamic graphs, whereas hybrid methods like OBS+ICA produce the most statistically significant differentiation across frequency band pairs [64].

Table 3: Network Topology Metrics After Artifact Removal [64]

Method	Connection Strength	Clustering Coefficient	Global Efficiency	Dynamic Range
AAS	0.58	0.42	0.63	0.38
OBS	0.61	0.45	0.66	0.41
ICA	0.54	0.38	0.59	0.42
OBS+ICA	0.63	0.47	0.68	0.45

Experimental Protocols for Method Validation

Protocol for aOBS Method Evaluation

The adaptive Optimal Basis Set (aOBS) method represents a significant advancement in template-based approaches. The following protocol outlines its implementation and validation:

Data Acquisition Requirements:

• High-density EEG recordings (64+ channels) acquired simultaneously with fMRI
• ECG recording synchronized with EEG data
• Visual stimulation paradigm (e.g., checkerboard reversal) for evoked response validation
• fMRI parameters: TR=3s, TE=30ms, slice thickness=3mm, no gap [5]

Step 1: Gradient Artifact Removal

• Apply established methods (e.g., AAS) to remove gradient artifacts from raw EEG
• Verify removal quality through visual inspection and quantitative metrics [5]

Step 2: BCG Peak Detection

• Identify R-peaks in the simultaneously recorded ECG signal
• Use this information to guide BCG peak identification in gradient-corrected EEG
• Alternatively, extract BCG peaks directly from EEG data when ECG quality is poor [5]

Step 3: Adaptive Epoch Analysis

• Segment EEG data into epochs based on identified BCG peaks
• For each channel, apply PCA to the epoched data
• Automatically select components for the artifact basis using variance-based criteria
• The number of components typically ranges between 1-8 across subjects and channels [5]

Step 4: Artifact Reconstruction and Subtraction

• Reconstruct BCG artifact epoch by epoch through linear combination of artifact basis
• Subtract reconstructed artifact from original EEG data
• Validate performance using quantitative metrics and visual inspection [5]

Validation Metrics:

• BCG residual intensity calculation
• Cross-correlation with ECG signal
• Event-Related Potential (ERP) signal-to-noise ratio analysis
• Inter-trial variability assessment [5]

Protocol for Deep Learning Approach (BCGGAN)

The BCGGAN method employs a generative adversarial network framework specifically designed for BCG artifact removal:

Network Architecture:

• Generator network with modular design for transformation learning
• Discriminator network for distinguishing between clean and artifact-contaminated signals
• Modular training strategy with additional constraints to optimize individual modules [14]

Training Strategy:

• Use unpaired training approach (CycleGAN framework)
• No requirement for simultaneously recorded clean and contaminated EEG data
• Optimization of generator parameters through adversarial training [14]

Implementation Considerations:

• No additional hardware or reference signals required
• Compatibility with standard EEG systems
• Demonstrated effectiveness across multiple evaluation metrics including time-domain, frequency-domain, and EO/EC effect preservation [14]

Diagram Title: BCG Artifact Removal Methodology Workflow

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of BCG artifact removal requires both computational tools and specialized hardware components. The following table outlines essential resources for establishing an effective EEG-fMRI artifact removal pipeline.

Table 4: Research Reagent Solutions for BCG Artifact Management

Item Name	Function/Purpose	Implementation Considerations
MRI-Compatible EEG Systems (e.g., Neuroscan SynAmps2)	Acquires EEG data inside MR environment with minimal artifact induction	Must use carbon fiber or non-metallic components to ensure safety and reduce artifacts [3]
BCG Reference Layer (BRL)	Measures artifact signals without neural activity for template creation	Requires specialized cap design with insulated conductive layer; reusable implementation available [20]
Carbon Fiber Wire Loops	Motion detection for hardware-based artifact correction	Provides reference signals for global head movement; less sensitive to local scalp movements [26] [20]
PCA/OBS Software Module	Implements template-based artifact removal	Standard component in EEG-fMRI processing toolboxes; automated component selection enhances performance [5]
ICA Algorithms (Infomax, FastICA)	Blind source separation for artifact component identification	Multiple algorithms available; component selection requires expertise to avoid neural signal loss [3] [64]
Deep Learning Frameworks (BCGGAN)	Learns artifact transformation without paired examples	No additional hardware required; modular training strategy improves generator performance [14]
Visual Stimulation Paradigms	Validation of neural signal preservation through evoked responses	Checkerboard reversal or similar protocols establish ground truth for method evaluation [5]

Diagram Title: Method Selection Based on Research Priority

Effective management of the trade-off between BCG artifact removal and neural signal preservation requires careful method selection based on specific research requirements. For studies prioritizing maximum neural signal preservation, hardware-based approaches (BRL) or advanced algorithms (aOBS, BCGGAN) demonstrate superior performance, though with increased implementation complexity. When maximum artifact removal is the primary objective, traditional OBS or optimized AAS methods provide robust solutions, particularly in high-motion conditions. For general-purpose applications with unknown signal characteristics, hybrid approaches (OBS+ICA) or adaptive methods (aOBS) offer the most balanced performance.

Validation remains crucial regardless of method selection. Researchers should implement comprehensive evaluation protocols assessing both artifact suppression (BCG residuals, cross-correlation with ECG) and neural signal integrity (ERP quality, connectivity measures). As EEG-fMRI applications continue to expand across clinical and cognitive neuroscience domains, ongoing method development promises increasingly sophisticated solutions to this fundamental challenge in multimodal neuroimaging.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the high spatial resolution of fMRI. This integration provides unprecedented insights into human brain function, enabling researchers to correlate electrical brain activity with hemodynamic responses [3] [65]. However, the utility of simultaneous EEG-fMRI is fundamentally limited by significant artifacts introduced into EEG recordings during fMRI acquisition, with the ballistocardiogram (BCG) artifact being one of the most challenging to remove effectively [3] [36].

The BCG artifact is a complex signal distortion originating from multiple physiological phenomena related to cardiac activity. When a subject is within the strong static magnetic field of an MRI scanner, cardiac-induced movements generate electrical currents that contaminate EEG signals [3]. Primary mechanisms include: pulse-driven head rotation, the Hall effect from pulsatile blood flow (as blood is an electrically conductive fluid), and scalp expansion from arterial pulsation [36] [5]. The resulting artifact exhibits complex spatiotemporal dynamics, with waveforms that vary considerably across channels, subjects, and over time [14] [5]. Unlike the gradient artifact (induced by switching magnetic fields), which is highly reproducible and relatively straightforward to remove, the BCG artifact is non-stationary and exhibits variable timing and morphology, making its correction particularly challenging [34].

Effective BCG artifact removal is crucial because residuals can obscure genuine neural signals or generate spurious correlations between EEG and fMRI data, potentially compromising scientific conclusions and clinical applications [5]. This guide provides a comprehensive, step-by-step pipeline for robust preprocessing of simultaneous EEG-fMRI data, with particular emphasis on advanced BCG artifact correction techniques.

Understanding the BCG Artifact: Physiological Origins and Characteristics

The BCG artifact manifests in EEG recordings as periodic deflections time-locked to the cardiac cycle but with variable timing and morphology across different electrode locations. Understanding its physiological origins is essential for developing effective removal strategies. The artifact arises from three primary mechanisms that operate simultaneously:

First, cardiac-induced head movement occurs when the head experiences tiny rotational movements in the strong static magnetic field due to the cardiac pulse. As described by the Maxwell equations, any movement of a conductor (including the head) in a magnetic field induces electrical currents [3] [36]. Second, the Hall effect generates potentials as conductive blood flows rhythmically through the scalp's blood vessels in the presence of the strong magnetic field [36] [5]. Third, pulse-driven scalp expansion creates electrode movement relative to the scalp surface, further contributing to the artifact [36].

The complex nature of the BCG artifact presents several challenges for removal algorithms. Its amplitude typically exceeds true EEG signals by 3-4 times, often obscuring neural activity of interest [34]. The artifact demonstrates spatial variability, with different topographic distributions across the scalp [3]. It also exhibits temporal non-stationarity, meaning its characteristics change over the duration of a recording due to factors like head position shifts, blood pressure changes, or physiological state alterations [3] [5]. Furthermore, the relationship with cardiac timing is not fixed; while BCG artifacts are approximately time-locked to heartbeats, the precise delay between the ECG R-peak and BCG artifact varies across beats and individuals [5].

Table: Key Characteristics of BCG Artifacts in Simultaneous EEG-fMRI

Characteristic	Description	Impact on Removal
Amplitude	3-4 times larger than typical EEG signals	Can completely obscure neural signals of interest
Spatial Distribution	Varies across electrode locations	Requires multi-channel correction approaches
Temporal Stability	Non-stationary; changes over time	Limits effectiveness of static template methods
Relationship to ECG	Approximately time-locked with variable delay	Complicates precise artifact alignment
Frequency Content	Overlaps with physiological EEG bands	Prevents simple frequency-based filtering

Comprehensive Preprocessing Pipeline

Initial Data Preparation and Gradient Artifact Removal

Before addressing BCG artifacts, EEG data require substantial preprocessing to remove other contamination sources. The initial step involves importing raw data at the native sampling rate (typically 5-10 kHz) without decimation, as high temporal resolution is crucial for effective artifact removal [34]. The data should then be visually inspected for saturation intervals where extreme artifact amplitudes cause signal clipping; these segments require interpolation or exclusion [34].

The most critical initial processing step is gradient artifact (GA) removal, which arises from the rapid switching of magnetic field gradients during fMRI acquisition. The GA is substantially larger than EEG signals (100 times greater) but highly reproducible, making average artifact subtraction (AAS) effective [34]. The recommended approach includes:

• Slice-wise template creation: Detect each slice acquisition trigger and create artifact templates by averaging across multiple occurrences [34].
• High-temporal resolution processing: Upsample EEG data using cubic spline interpolation to accurately reconstruct GA morphology [34].
• Adaptive subtraction: Subtract the template from each artifact occurrence, with local adjustments for temporal variations [34].

Successful GA removal is essential before addressing BCG artifacts, as GA residuals would significantly complicate subsequent BCG correction [34]. After GA removal, data are typically downsampled to 500-1000 Hz for more efficient processing [23].

Cardiac Event Detection

Accurate identification of cardiac events is fundamental to most BCG artifact removal methods. This process can utilize either a dedicated electrocardiogram (ECG) channel or the EEG data itself:

ECG-based detection is preferred when a clean ECG recording is available. The process involves:

• QRS complex identification: Detect R-peaks in the ECG signal using algorithms like Pan-Tompkins [5].
• Pulse verification: Remove falsely detected or missed beats through visual inspection or automated thresholding [5].

EEG-based detection is necessary when no dedicated ECG channel is available or its quality is poor. This approach:

• Leverages multi-channel information: Combines data from all EEG channels to identify consistent pulse-related patterns [34].
• Employs iterative optimization: Refines pulse detection by maximizing cross-correlation across channels and minimizing temporal variability [34].
• Generates heart pulse indicators (HPI): Creates functions with well-contrasted peaks indicating each heart pulse [34].

The adaptive Optimal Basis Set (aOBS) method enhances this process by directly detecting BCG peaks in gradient-corrected EEG data, thereby estimating the variable delay between cardiac activity and BCG occurrence on a beat-to-beat basis [5].

BCG Artifact Removal Methods

Traditional Correction Approaches

Several well-established methods form the foundation of BCG artifact correction:

Average Artifact Subtraction (AAS) represents the simplest approach, where an artifact template is created by averaging EEG segments time-locked to cardiac events. This template is then subtracted from each artifact occurrence [36] [5]. While straightforward, AAS often leaves significant residuals because it cannot account for BCG shape variations across cardiac cycles [5].

Optimal Basis Set (OBS) extends AAS by applying Principal Component Analysis (PCA) to artifact-locked EEG segments. Instead of subtracting a simple average, OBS uses the first few principal components that capture the primary variance of BCG artifacts as a basis for subtraction [3] [5]. The standard approach typically removes the first 3-4 components, though this fixed number may not be optimal for all datasets [5].

Independent Component Analysis (ICA) employs blind source separation to decompose EEG data into statistically independent components. BCG-related components are identified and removed before signal reconstruction [36] [34]. The challenge lies in accurately identifying artifact components, as misclassification can remove neural signals or leave artifact residuals [5].

Advanced and Hybrid Methods

Adaptive Optimal Basis Set (aOBS) addresses key limitations of standard OBS by incorporating two innovations [5]. First, it performs beat-to-beat delay estimation between ECG events and BCG occurrences, improving artifact alignment. Second, it implements automatic component selection based on explained variance criteria rather than using a fixed number of components. Studies show aOBS reduces BCG residuals to approximately 5.5% on average, compared to 9.2% for OBS and 12.5% for AAS [5].

PROJIC-based approaches combine ICA with projection techniques, offering three variants [36]. PROJIC removes BCG-related independent components entirely during reconstruction. PROJIC-OBS applies OBS correction to BCG-related components before back-projection, while PROJIC-AAS uses AAS for component correction. These methods allow flexible trade-offs between artifact removal and neural signal preservation [36].

Deep Learning methods represent the most recent advancement. BCGNet uses recurrent neural networks (RNNs) with gated recurrent units (GRUs) to learn nonlinear mappings between ECG and BCG-corrupted EEG [27] [23]. BCGGAN employs generative adversarial networks (GANs) in a modular training framework to transform BCG-corrupted EEG into clean EEG without requiring paired examples [14]. These approaches show promise in handling the nonlinear relationship between ECG and BCG artifacts.

Table: Comparison of BCG Artifact Removal Methods

Method	Key Principle	Advantages	Limitations
AAS	Template averaging and subtraction	Simple implementation	High residual artifacts due to BCG variability
OBS	PCA-based basis set subtraction	Accounts for some BCG variance	Fixed component number; sensitive to misalignment
ICA	Blind source separation	Utilizes spatial information	Component selection challenge; spatial stationarity assumption
aOBS	Adaptive delay estimation + automated component selection	Handles BCG variability; data-driven	More computationally intensive
PROJIC-OBS	ICA component selection + OBS correction	Excellent artifact removal	Complex implementation; potential signal distortion
PROJIC-AAS	ICA component selection + AAS correction	Optimal physiological signal preservation	May leave more artifact residuals
BCGNet	RNN-based nonlinear mapping	Models ECG-BCG relationship; improves single-trial analysis	Requires high-quality ECG; extensive training data
BCGGAN	GAN-based unpaired translation	No reference signal needed; effective artifact removal	Training instability; complex implementation

The following diagram illustrates the decision process for selecting an appropriate BCG artifact removal method based on data characteristics and research objectives:

Quality Control and Validation

Assessing Artifact Removal Effectiveness

Robust quality control is essential to verify BCG artifact removal without excessive distortion of neural signals. Multiple quantitative metrics should be employed:

BCG residual intensity measures remaining artifact power after correction, calculated as the percentage of artifact power remaining compared to pre-correction levels. The adaptive OBS method achieves approximately 5.5% residuals, compared to 9.2% for standard OBS and 20.6% for ICA [5].

Cross-correlation with ECG evaluates the temporal relationship between corrected EEG and reference ECG. Effective methods reduce maximum cross-correlation values from approximately 0.180 (pre-correction) to 0.028 (aOBS), 0.042 (OBS), or 0.051 (AAS) [5].

Spectral metrics include the Improvement in Normalized Power Spectrum (INPS), which quantifies power reduction at cardiac fundamental frequency and harmonics [36]. Additionally, the ratio of artifact band power to neural band power can assess selective artifact suppression.

Evaluating Neural Signal Preservation

Preserving physiological EEG signals during BCG removal is equally important. Several approaches validate signal integrity:

Event-Related Potentials (ERPs) analysis quantifies how BCG correction affects task-related neural responses. Metrics include signal-to-noise ratio (SNR) and inter-trial variability [36] [5]. PROJIC-AAS demonstrates particularly strong performance in preserving ERPs, achieving 18-61% reduction in standard error across trials depending on magnetic field strength [36].

Eyes Open/Closed (EO/EC) paradigm validates the preservation of well-characterized physiological phenomena, particularly posterior alpha rhythm modulation. Effective BCG removal should maintain the characteristic alpha power increase during eyes-closed conditions [14].

Resting-state oscillations analysis examines whether correction methods distort natural brain rhythm topography and spectral properties, which is crucial for studies investigating neural oscillations in simultaneous EEG-fMRI [3].

The following workflow diagram outlines the complete quality assessment process for BCG artifact correction:

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful simultaneous EEG-fMRI research requires specific hardware and software tools optimized for the challenging MR environment. The following table details essential components of the experimental toolkit:

Table: Essential Research Toolkit for Simultaneous EEG-fMRI Studies

Category	Item	Specification/Function
EEG Hardware	MR-Compatible EEG System	64+ channels; SynAmps2 (Neuroscan) or BrainAmp MR Plus (Brain Products) [3] [23]
	EEG Electrodes	Ag/AgCl with high-abrasion resistance; impedance < 20 kΩ [23]
	Carbon Fiber Wires	Reduce induction of currents from magnetic fields [3]
Additional Hardware	MR-Compatible Amplifier	Battery-powered with fiber optic transmission [23]
	Sandbag Stabilization	Minimize amplifier vibration from scanner [23]
	ECG Electrode	Single channel for cardiac timing reference [5]
Software Tools	EEG Processing Suite	EEGLAB with FMRIB Plugin [23]
	BCG Correction Implementation	BESA Research, BrainVoyager QX EMEG Suite [3] [34]
	Custom Algorithm Platform	MATLAB, Python with MNE library [23]
Reference Materials	ECG QRS Detector	Robust detection in high-static field environments [5]
	Standardized ERP Paradigms	Auditory oddball, visual evoked potentials for validation [3] [23]

Experimental Protocols and Methodological Details

Implementing the Adaptive OBS (aOBS) Method

The aOBS method represents a significant advancement over traditional OBS by addressing its key limitations. Implementation involves two sequential stages:

Stage 1: BCG Occurrence Detection

• ECG peak detection: Identify R-peaks in the simultaneously recorded ECG using robust QRS detection algorithms optimized for the MR environment [5].
• BCG peak refinement: Use gradient-corrected EEG data to refine the timing of BCG events, estimating the variable delay between ECG and BCG on a beat-to-beat basis [5].
• Pulse verification: Apply iterative procedures to maximize cross-correlation across pulses while minimizing temporal variability [34].

Stage 2: BCG Artifact Reconstruction and Subtraction

• Epoching: Extract EEG segments time-locked to the refined BCG peaks for each channel [5].
• PCA decomposition: Apply principal component analysis to the epoched data to derive spatial and temporal components [5].
• Automated component selection: Use a data-driven criterion based on explained variance to determine how many components to remove, typically ranging from 1-8 across subjects and channels [5].
• Artifact reconstruction and subtraction: Reconstruct the BCG artifact using selected components and subtract from the original signal [5].

Deep Learning Implementation (BCGNet)

For researchers implementing the BCGNet deep learning approach, the following protocol details the essential steps:

Network Architecture and Training:

• Network configuration: Implement a deep recurrent neural network using gated recurrent units (GRUs) to capture temporal dependencies [23].
• Input preparation: Use BCG-corrupted EEG and simultaneously recorded ECG as input-output pairs for training [23].
• Data preprocessing: Resample data to 100 Hz due to computational constraints and normalize each channel to zero mean and unit variance [23].
• Training objective: Optimize the network to learn the nonlinear mapping between ECG and BCG artifacts present in EEG signals [23].

Validation Approach:

• Single-trial analysis: Assess method performance by comparing single-trial EEG classification accuracy before and after BCG removal [23].
• Power spectrum examination: Verify reduction of artifact power at critical frequencies related to cardiac cycles [23].
• Generalization testing: Evaluate model performance across subjects to ensure robustness [23].

This guide has presented a comprehensive pipeline for robust preprocessing of simultaneous EEG-fMRI data, with particular emphasis on the challenging problem of BCG artifact removal. The field has evolved from simple template subtraction methods to sophisticated adaptive and machine learning approaches that better account for the complex, non-stationary nature of these artifacts. Current evidence suggests that no single method is universally superior; rather, the choice depends on specific research goals, data characteristics, and available resources.

Future methodological developments will likely focus on several promising directions. Hybrid approaches that combine the strengths of multiple techniques may offer superior performance, such as using deep learning for initial artifact estimation followed by traditional methods for refinement [14] [23]. Real-time processing capabilities would expand clinical applications, particularly in epilepsy monitoring [23]. Improved hardware solutions, including advanced electrode materials and motion detection systems, may reduce artifact magnitude at the source [3]. Finally, standardized validation frameworks with open-source benchmarking datasets would accelerate method development and comparative evaluation [66].

As simultaneous EEG-fMRI continues to advance our understanding of brain function, robust preprocessing pipelines remain foundational to generating reliable, interpretable results. The methods and protocols outlined in this guide provide researchers with practical tools to address the unique challenges of this powerful multimodal imaging technique.

Benchmarking Performance: Validation Metrics and Comparative Analysis of BCG Removal Methods

Simultaneous electroencephalography and functional magnetic resonance imaging (EEG-fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the high spatial resolution of fMRI. However, the utility of this integrated approach is fundamentally limited by the ballistocardiogram (BCG) artifact, a complex, non-stationary artifact induced by cardiac-related head movements and pulsatile blood flow within the strong static magnetic field of the MRI scanner [67] [36]. The development of effective BCG artifact removal algorithms is an active area of research, necessitating a robust and standardized validation framework to assess both artifact removal efficacy and physiological signal preservation. This technical guide establishes such a framework, detailing key metrics, experimental protocols, and methodological considerations essential for rigorous evaluation within the specific context of BCG artifact contamination in simultaneous EEG-fMRI research.

Core Validation Metrics: A Dual-Focus Approach

A comprehensive validation framework must balance two competing objectives: effective artifact suppression and maximal preservation of underlying neural signals. The metrics can be categorized accordingly for clarity and precision in reporting.

Table 1: Metrics for Assessing BCG Artifact Removal Efficacy

Metric Category	Specific Metric	Description	Interpretation
Artifact Amplitude	Root Mean Square (RMS) of BCG waveform [36]	Quantifies the average power of the residual artifact post-correction.	Lower values indicate more effective artifact reduction.
	Peak-to-Peak (PTP) values [36]	Measures the amplitude difference between the highest and lowest points in the artifact waveform.	Lower values indicate better suppression of the largest artifact deflections.
Spectral Purity	Improvement in Normalized Power Spectrum (INPS) [36]	A ratio expressing the loss in normalized spectral power after correction around cardiac harmonic frequencies.	Higher values indicate greater reduction of artifact-specific spectral components.
	Cross-Correlation with ECG [5]	Measures the temporal correlation between the corrected EEG signal and the reference ECG signal.	Lower correlation coefficients indicate less residual cardiac-locked contamination.
Spatio-Temporal Residue	BCG Residual Intensity [5]	Computes the intensity of the artifact remaining in the EEG data after correction.	Lower intensity values signify a cleaner signal.

Table 2: Metrics for Assessing Physiological Signal Preservation

Metric Category	Specific Metric	Description	Interpretation
Event-Related Potentials (ERPs)	Signal-to-Noise Ratio (SNR) [5] [36]	Measures the ratio of the power of the ERP to the power of the background noise.	Higher SNR indicates better preservation of the task-evoked neural response.
	Inter-trial Variability [67] [36]	Assesses the consistency of the ERP waveform across multiple trials.	Lower variability suggests more reliable signal recovery and less distortion.
	Relative Reduction of Standard Error (SE) [36]	Quantifies the reduction in uncertainty of the averaged ERP estimate.	A higher reduction implies a more robust and well-defined ERP.
Spontaneous Oscillations	Spectral Power Fidelity [67]	Compares the spectral power in classic frequency bands (delta, theta, alpha, beta, gamma) in corrected data vs. a clean reference.	A closer match indicates minimal distortion of ongoing brain oscillations.
	Functional Reactivity (e.g., Alpha Power) [67]	Evaluates the preservation of expected spectral changes, such as the increase in posterior alpha power during eyes closure versus eyes opening.	Successful preservation shows the method retains biologically meaningful signal dynamics.

Experimental Protocols for Validation

The Gold-Standard Paradigm: Eyes-Closure/Eyes-Opening (EC-EO)

This paradigm is highly recommended for its ability to elicit a robust and well-characterized modulation of posterior alpha power (8-13 Hz), providing a reliable benchmark for assessing signal preservation [67].

• Procedure: Participants are instructed to alternately close and open their eyes in blocks (e.g., 30-60 seconds per condition) while simultaneous EEG-fMRI data is acquired.
• Data Acquisition:
  • Inside-scanner EEG-fMRI: Record EEG inside the MRI scanner concurrently with fMRI.
  • Reference EEG: Record EEG outside the scanner with an identical setup to obtain BCG-artifact-free data.
• Validation Analysis:
  • Apply the BCG artifact correction method to the inside-scanner EEG.
  • For both corrected and reference EEG, calculate the absolute power in the alpha band from a posterior electrode cluster (e.g., Pz, P3, P4, O1, O2) during EC and EO blocks.
  • Compute the EC/EO alpha power ratio. A method that successfully preserves physiological signals will show a significantly higher alpha ratio in the corrected data, closely matching the ratio observed in the reference outside-scanner EEG [67].

Event-Related Potential (ERP) Validation

This protocol tests a method's ability to preserve transient, time-locked neural responses.

• Procedure: Implement a simple sensory or cognitive task known to generate a specific ERP component (e.g., visual oddball task for P300) during simultaneous EEG-fMRI acquisition.
• Data Acquisition:
  • Acquire inside-scanner EEG-fMRI data during the task.
  • Acquire reference EEG data outside the scanner during the same task.
• Validation Analysis:
  • Correct the inside-scanner EEG and epoch the data around the event of interest.
  • Average the epochs to obtain the ERP for both corrected and reference datasets.
  • Quantify and compare the amplitude and latency of the key ERP component (e.g., P300). Use metrics like SNR and inter-trial variability to objectively assess the quality and reliability of the recovered ERP [5] [36].

Quantitative Signal Quality Indexing

Adopting a standardized, quantitative index for signal quality, such as the Signal Quality Index (SQI) algorithm used in fNIRS, can objectify assessments [68]. This algorithm rates signal quality on a numeric scale (e.g., 1 to 5) based on the strength of the cardiac component, which is a key indicator of good sensor-scalp coupling. While developed for fNIRS, its principle of using an objective, automated metric to eliminate researcher bias is directly applicable to the goal of standardizing BCG artifact correction validation.

Visualization of the Validation Workflow

The following diagram illustrates the integrated workflow for validating a BCG artifact correction method, incorporating the key metrics and protocols described above.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for BCG Artifact Research

Item / Tool	Function / Description	Relevance to Validation
High-Density MR-Compatible EEG Cap	A specialized cap with electrodes (e.g., 32-64 channels) designed for safe use inside the MRI environment, minimizing artifacts and risks.	The primary tool for data acquisition; channel count and placement impact spatial analysis of artifact correction [67].
Electromechanical Film (EMFi) Sensor	A high-sensitivity sensor that can detect tiny pressure changes, used in some setups for acquiring reference BCG signals [69].	Can be used for alternative BCG monitoring or as part of hardware-based correction approaches.
ECG/EOG Recording Setup	Essential for acquiring the electrocardiogram (ECG) to identify cardiac events and electrooculogram (EOG) to monitor eye movements.	ECG is crucial for AAS, OBS, and related methods. EOG helps control for ocular artifacts in signal preservation tests [5] [14].
Average Artifact Subtraction (AAS)	A software method that creates and subtracts an average artifact template from the EEG signal, aligned to cardiac events [67] [36].	A foundational and widely used algorithm that serves as a common baseline for comparison against newer methods.
Optimal Basis Set (OBS)	A method that uses Principal Component Analysis (PCA) on artifact epochs to create a tailored basis set for artifact reconstruction and subtraction [5] [36].	A standard against which novel algorithms are often benchmarked for performance.
Independent Component Analysis (ICA)	A blind source separation technique that decomposes EEG data into independent components, allowing for manual or automatic identification and removal of BCG-related components [67] [36].	A powerful but subjective method; its performance is a key point of comparison for automated approaches.
Generative Adversarial Network (GAN)	A deep learning model (e.g., BCGGAN) that learns to map BCG-corrupted EEG to clean EEG without requiring paired examples [14].	Represents the cutting-edge in data-driven correction; requires rigorous validation using the proposed framework.

Comparative Analysis of BCG Correction Methods

The following diagram provides a high-level taxonomy and comparison of the primary BCG artifact correction methods discussed in the literature, highlighting their core characteristics and relationships.

The performance of these methods varies, with studies showing that no single method is universally superior across all metrics. For instance, the adaptive Optimal Basis Set (aOBS) has been shown to achieve lower BCG residual intensity (5.53%) compared to AAS (12.51%), ICA (20.63%), and standard OBS (9.20%) [5]. Furthermore, ICA-based correction approaches have demonstrated a better ability to preserve the functional reactivity of posterior alpha power during an EC-EO task compared to other methods [67]. Emerging deep learning approaches, such as BCGGAN, show promise in effectively removing the BCG artifact without requiring an ECG reference signal, while also retaining useful physiological information [14]. This underscores the necessity of the proposed multi-metric validation framework to guide researchers in selecting the most appropriate method for their specific application.

Simultaneous Electroencephalography and functional Magnetic Resonance Imaging (EEG-fMRI) represents a powerful multimodal neuroimaging technique that combines the millisecond temporal resolution of EEG with the millimeter spatial resolution of fMRI. However, the utility of this integration is critically challenged by the ballistocardiogram (BCG) artifact, a complex, cardiac-induced signal that contaminates EEG recordings inside the MRI scanner. The BCG artifact arises from multiple physiological phenomena including cardiac-induced head movement, pulse-driven scalp expansion, and the Hall effect of pulsatile blood flow in the static magnetic field [13] [20]. With amplitudes often exceeding 200 μV in 3T scanners—tens of times larger than neural EEG signals—the BCG artifact obscures brain activity and compromises analysis [20]. Consequently, developing and validating effective BCG artifact removal methods is paramount. This whitepaper provides an in-depth technical guide to the quantitative evaluation of these methods, focusing on the core metrics of Residual Artifact Levels and Signal-to-Noise Ratio (SNR), to inform researchers, scientists, and drug development professionals in their analytical endeavors.

Quantitative Comparison of BCG Artifact Removal Methods

The performance of BCG artifact removal techniques is quantified using a suite of metrics that evaluate both the effectiveness of artifact suppression and the preservation of underlying neural signals. The table below summarizes the quantitative performance of established and emerging methods.

Table 1: Quantitative Performance of BCG Artifact Removal Methods

Method	Category	Key Metric & Performance	Residual Artifact Level	Impact on Neural Signals
AAS (Average Artifact Subtraction) [1]	Template Subtraction	Best Signal Fidelity (MSE = 0.0038, PSNR = 26.34 dB)	Moderate	Can distort non-stationary neural signals [22]
OBS (Optimal Basis Set) [1]	Template Subtraction	Highest Structural Similarity (SSIM = 0.72)	Lower than AAS	Improved preservation over AAS, but may remove brain activity correlated with artifact [13]
aOBS (adaptive OBS) [13]	Template Subtraction	Average BCG Residual = 5.53%; Max Cross-Correlation with ECG = 0.028	Lowest among template methods	Automatically estimates components for removal, improving signal preservation
ICA (Independent Component Analysis) [1]	Blind Source Separation	Sensitive to frequency-specific patterns in dynamic graphs	Higher (20.63% residual)	Risk of misclassifying neural components as artifact, leading to signal loss
PROJIC-OBS [36]	Hybrid (ICA + OBS)	Outperforms others when priority is artifact removal	Low	Better preservation of physiological signals compared to OBS alone
PROJIC-AAS [36]	Hybrid (ICA + AAS)	Best for intermediate trade-offs; 18-61% reduction in ERP standard error	Low	Superior physiological signal preservation
BRL (Reference Layer) [20]	Hardware-Based	Improved CNR of Alpha, VEP, and AEP by 101%, 76%, and 75% (vs. OBS)	Very Low	Removes artifact without software-based signal distortion; requires special equipment
BCGNet (RNN) [23]	Deep Learning	Larger average power reduction at critical frequencies vs. OBS	Low	Improves task-relevant EEG classification; models non-linear ECG-BCG relationship
BCGGAN (GAN) [14]	Deep Learning	Effectively removes BCG without ECG reference; retains physiological information	Low	Does not require additional hardware or reference signals

Beyond the tabulated metrics, the contrast-to-noise ratio (CNR) provides a critical measure for assessing the detectability of neural signals. Hardware-based approaches, such as the BCG reference layer (BRL), demonstrate substantial CNR improvements. One study showed that BRL improved the CNR of alpha-wave, visual, and auditory evoked potentials by 101%, 76%, and 75% respectively compared to traditional OBS, even when using only 20 reference electrodes [20]. Furthermore, the preservation of event-related potentials (ERPs) is a vital benchmark. Hybrid methods like PROJIC-AAS have been shown to reduce the standard error across trials by 18% (at 3T) to 61% (at 7T), indicating a significant enhancement in the reliability of recovered neural responses [36].

Table 2: Summary of Key Performance Evaluation Metrics

Metric	Description	Interpretation
Mean Squared Error (MSE) [1]	Average squared difference between cleaned and ideal signal	Lower values indicate better signal fidelity.
Peak Signal-to-Noise Ratio (PSNR) [1]	Ratio between maximum possible signal power and corrupting noise power	Higher values (in dB) indicate a better-cleaned signal.
Structural Similarity Index (SSIM) [1]	Measures perceived quality and structural integrity of the signal	Values closer to 1.0 indicate better preservation of original structure.
BCG Residual Intensity [13]	Quantitative measure of remaining artifact power after correction	Lower values indicate more effective artifact removal.
Cross-Correlation with ECG [13]	Measures residual temporal locking between EEG and cardiac cycle	Lower values indicate more effective decoupling from the artifact.
Contrast-to-Noise Ratio (CNR) [20] [70]	Measure of the detectability of a signal (e.g., alpha wave) against noise	Higher values indicate easier detection of the signal of interest.
Reduction in ERP Standard Error [36]	Decrease in trial-to-trial variability of an Event-Related Potential	Higher reduction indicates more reliable recovery of neural responses.

Detailed Experimental Protocols for Method Evaluation

To ensure the reproducibility and rigorous benchmarking of BCG artifact removal methods, researchers employ standardized experimental protocols and evaluation pipelines. The following section details key methodologies cited in comparative studies.

Protocol for Evaluating Signal Quality and Connectivity (2025 Study)

A 2025 study adopted a holistic framework to evaluate AAS, OBS, and ICA, alongside hybrid methods (AAS+ICA, OBS+ICA) [1].

• Data Acquisition: Utilizing simultaneous EEG-fMRI recordings, the protocol emphasizes the importance of a well-defined paradigm (e.g., resting-state or task-based) for subsequent analysis.
• Signal Quality Assessment: The cleaned EEG data is evaluated using a multifaceted set of metrics:
  • Mean Squared Error (MSE) & Peak Signal-to-Noise Ratio (PSNR): Quantify signal fidelity.
  • Structural Similarity Index (SSIM): Assesses the preservation of the signal's structural information.
  • Power Spectral Density (PSD): Examines the preservation of frequency components.
• Functional Connectivity Analysis: EEG and fMRI time series are combined based on correlation to construct functional brain networks. Graph theory metrics—such as Connection Strength (CS), Clustering Coefficient (CC), and Global Efficiency (GE)—are then computed for static and dynamic graphs across different EEG frequency bands to assess the impact of artifact removal on network topology.

Protocol for the Adaptive OBS (aOBS) Method

The aOBS method enhances traditional OBS through a two-step sequential process designed to improve artifact characterization [13]:

• Detection of BCG Occurrences: Electrocardiogram (ECG) peaks are first detected. This information is then used to guide the identification of BCG peaks in the gradient-artifact-corrected EEG data. Alternatively, BCG peaks can be directly extracted from the EEG data itself, making the method more robust to variability in the ECG-BCG delay.
• BCG Artifact Reconstruction and Subtraction:
  • For each EEG channel, data are epoched based on the identified BCG peaks.
  • Principal Component Analysis (PCA) is applied to the epoched data.
  • An automated selection criterion, based on explained variance, is used to determine which principal components (PCs) are BCG artifact-related. The number of selected components can vary across subjects and channels (e.g., ranging from 1 to 8).
  • The BCG artifact is reconstructed epoch-by-epoch via a linear combination of the selected artifact basis and is then subtracted from the original EEG data.

Protocol for Deep Learning-Based Removal (BCGNet)

The BCGNet approach employs a recurrent neural network (RNN) with Gated Recurrent Units (GRUs) to model the non-linear relationship between the ECG and the BCG artifact [23]:

• Data Pre-processing: Raw EEG data is imported, low-pass filtered (e.g., cutoff at 70 Hz), and subjected to Gradient Artifact removal using algorithms like FASTR. Data is then resampled (e.g., to 500 Hz) and high-pass filtered (e.g., at 0.25 Hz) to reduce drift, resulting in the "BCG corrupted EEG" (BCE).
• Network Training and Application: The BCE and synchronized ECG data are fed into the deep learning model. The RNN is trained to learn the non-linear mapping from the ECG to the BCG artifact present in the EEG. Once trained, the network predicts the artifact waveform, which is then subtracted to yield the cleaned EEG.

Protocol for Hybrid PROJIC Method Evaluation

A comprehensive evaluation pipeline was proposed for the PROJIC family of methods, which prioritizes both artifact removal and physiological signal preservation [36].

• Data Acquisition with Different Setups: Data is acquired using multiple setups (e.g., 3 T and 7 T scanners) to demonstrate wide applicability.
• Method Application: The PROJIC method involves decomposing the EEG using ICA and then using a novel projection approach to select BCG-related independent components (ICs). Three correction strategies are then explored: simply removing these ICs (PROJIC), or correcting them using OBS (PROJIC-OBS) or AAS (PROJIC-AAS) before back-projecting all ICs.
• Performance Assessment: The pipeline flexibly weights the importance of artifact removal versus signal preservation. The impact on Event-Related Potentials (ERPs) is assessed via the relative reduction of the standard error (SE) across trials, providing a direct measure of neural signal reliability.

Visualizing Workflows and Method Relationships

The following diagrams illustrate the logical workflows of key artifact removal methods and their hierarchical relationships, providing a clear conceptual understanding.

The Scientist's Toolkit: Key Research Reagents and Materials

Successful execution of BCG artifact removal experiments requires specific hardware, software, and analytical tools. The following table details essential components of the research toolkit.

Table 3: Essential Research Toolkit for BCG Artifact Removal Studies

Tool Category	Specific Item / Solution	Function & Application
EEG Acquisition Hardware	MRI-Compatible EEG System (e.g., BrainAmp MR Plus) [23]	Records EEG data inside the MRI scanner with specialized amplifiers to prevent saturation and ensure high SNR.
EEG Acquisition Hardware	High-Density EEG Montages (64+ channels) [13]	Provides large scalp coverage, improving spatial sampling for source separation methods like ICA.
Reference Signal Hardware	ECG Electrode [23]	Provides the electrocardiogram signal crucial for timing cardiac events and for template-based (AAS, OBS) and some deep learning (BCGNet) methods.
Reference Signal Hardware	BCG Reference Layer (BRL) [20]	A conductive layer with dedicated electrodes, electrically insulated from the scalp, to directly measure BCG artifacts uncontaminated by neural signals.
Reference Signal Hardware	Carbon Fiber Wire Loops / Piezoelectric Sensors [22]	External motion detection devices used to measure head movement and generate reference signals for artifact subtraction.
Software & Algorithms	FASTR Algorithm [23]	An integrated tool (e.g., in EEGLAB's FMRIB plugin) for the removal of Gradient Artifacts, a necessary pre-processing step before BCG correction.
Software & Algorithms	Optimal Basis Set (OBS) [13] [1]	A PCA-based algorithm available in toolboxes like FMRIB's for adaptive BCG artifact template subtraction.
Software & Algorithms	Independent Component Analysis (ICA) [1] [36]	Blind source separation algorithms (e.g., Infomax, FastICA) implemented in software like EEGLAB for decomposing EEG and identifying artifact components.
Software & Algorithms	Real-Time Processing Platforms (e.g., EEG-LLAMAS) [1]	Open-source software for low-latency, real-time BCG artifact removal, enabling closed-loop EEG-fMRI paradigms.
Data Analysis & Evaluation	Graph Theory Metrics [1]	Computational tools (e.g., in MATLAB, Python) to calculate network properties like Clustering Coefficient and Global Efficiency from functional connectivity graphs.
Data Analysis & Evaluation	Signal Quality Metrics [1]	Scripts to compute quantitative evaluation metrics such as MSE, PSNR, and SSIM for objective performance comparison.

The rigorous quantitative comparison of BCG artifact removal methods is a cornerstone of valid and reliable simultaneous EEG-fMRI research. As evidenced by the data, no single method is universally superior; each presents a unique trade-off between residual artifact level and physiological signal preservation. Template-based methods like AAS and OBS provide a solid foundation, with adaptive versions (aOBS) offering improved performance. Blind source separation (ICA) and hybrid approaches (PROJIC) offer powerful alternatives but require careful component selection to avoid neural signal loss. Emerging deep learning techniques (BCGNet, BCGGAN) show great promise in modeling the artifact's complex non-linearity, while hardware-based solutions (BRL) can fundamentally avoid software-related distortions. The choice of method must be guided by the specific research question, the relative priority of artifact removal versus signal integrity, and the available experimental resources. By leveraging standardized quantitative metrics and protocols outlined in this guide, researchers can make informed decisions, ensuring that the full spatio-temporal potential of simultaneous EEG-fMRI is realized in neuroscience and drug development.

The simultaneous acquisition of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) represents a powerful multimodal neuroimaging approach that combines high temporal resolution from EEG with high spatial resolution from fMRI [33]. This technique enables researchers to investigate brain dynamics with unprecedented detail, capturing rapid neural events through event-related potentials (ERPs) and oscillatory activity while precisely localizing their neural generators via the blood oxygenation level-dependent (BOLD) signal [33]. However, the integrity of neural signals derived from EEG during simultaneous EEG-fMRI recordings is critically threatened by the ballistocardiogram (BCG) artifact, a complex contamination originating from cardiac-induced head movements and pulsatile blood flow in the strong static magnetic field [27] [3].

The BCG artifact presents a formidable challenge for ERP research because its amplitude often exceeds that of neural signals of interest, particularly for cognitive ERP components such as the P300, N400, and earlier endogenous components [71] [72]. This artifact exhibits non-stationary properties, varies across subjects and recording sessions, and demonstrates complex spatio-temporal dynamics that complicate its removal [3] [5]. The recovery of genuine neural signals from BCG-contaminated EEG data requires sophisticated artifact removal techniques that can effectively suppress artifacts while preserving the integrity of neural signals [14] [5]. This technical guide provides a comprehensive overview of current methodologies for assessing and recovering ERP and oscillatory signals in simultaneous EEG-fMRI research, with particular emphasis on BCG artifact mitigation strategies, their quantitative performance, and practical implementation considerations for researchers and drug development professionals.

The BCG Artifact Challenge in Simultaneous EEG-fMRI

Origin and Characteristics of the BCG Artifact

The BCG artifact arises from multiple physical phenomena associated with cardiac activity. As described by the Maxwell equations, microscopic head movements in the strong static magnetic field of the MRI scanner generate electrical currents that contaminate EEG recordings [3]. The primary mechanisms include: (1) cardiac-induced head motion in the supine position, (2) pulsatile expansion of scalp vessels near electrodes, (3) Hall effect from pulsatile blood flow (an electrically conductive fluid), and (4) rotational movements of the head caused by cardiac activity [3] [22]. These combined effects result in a complex artifact pattern that is time-locked to the cardiac cycle but exhibits considerable non-stationarity in both topography and temporal evolution [5].

The BCG artifact is particularly problematic for ERP research due to its overlapping frequency content with neural signals of interest and its variable morphology across cardiac cycles [14]. Unlike the gradient artifact caused by switching magnetic fields, which can be effectively removed using average artifact subtraction due to its highly reproducible nature, the BCG artifact requires more sophisticated removal approaches [3] [5]. The artifact's amplitude can be an order of magnitude larger than endogenous ERP components, potentially obscuring genuine neural activity or generating spurious correlations between EEG and fMRI time-courses [5].

Impact on ERP and Oscillatory Signal Integrity

The BCG artifact significantly compromises the integrity of neural signals in multiple dimensions. For ERPs, which are typically characterized by low signal-to-noise ratio (SNR) and require averaging across numerous trials, residual BCG contamination can distort component morphology, latency, and topography [71] [72]. This is particularly critical for cognitive ERP components such as the P300, which reflects cognitive processes like context updating and decision-making [72]. Similarly, the accurate characterization of neural oscillations—including alpha, beta, and gamma rhythms—is complicated by BCG residuals that may mimic or obscure genuine oscillatory patterns [3].

The problem extends to source localization accuracy, where residual BCG artifacts can significantly displace estimated generator locations or create false sources [3]. This has profound implications for studies seeking to integrate ERP temporal precision with fMRI spatial resolution, as inaccurate source localizations may lead to erroneous neurophysiological interpretations [3] [72]. Furthermore, in drug development research where ERP parameters may serve as biomarkers for treatment efficacy, uncompensated BCG artifacts could confound outcome measures and compromise clinical trial results [71].

Quantitative Assessment of BCG Artifact Removal Methods

Table 1: Performance Comparison of Primary BCG Artifact Removal Methods

Method	BCG Residual Intensity	Max Cross-Correlation with ECG	Signal Preservation	Computational Complexity
AAS	12.51%	0.051	Moderate	Low
ICA	20.63%	0.067	Variable	Medium
OBS	9.20%	0.042	Moderate to High	Medium
aOBS	5.53%	0.028	High	Medium to High
GAN-based	Not reported	Not reported	High	High
RNN-based	Not reported	Not reported	High	High

Table 2: Deep Learning Approaches for BCG Artifact Removal

Method	Architecture	Reference Signal	Key Advantages	Limitations
RNN-based [27]	Recurrent Neural Network	ECG	Models nonlinear ECG-BCG relationships	Requires ECG reference
BCGGAN [14]	Generative Adversarial Network	None	No additional hardware; handles unpaired data	Complex training strategy
EEGNet [71]	Compact Convolutional Neural Network	None	Adaptable for single-trial ERP quantification	Requires simulated training data

Table 3: ERP Component Recovery After BCG Artifact Removal

ERP Component	Affected Parameters	Sensitivity to BCG Residuals	Optimal Removal Method
P300/P3b [72]	Amplitude, latency, topography	High	aOBS, Surrogate Methods
N400 [71]	Amplitude, latency jitter	High	Neural Network Approaches
Early Sensory Components	Amplitude, latency	Medium	OBS, aOBS
Endogenous Emitted Potentials [72]	All parameters	Very High	Surrogate Methods, aOBS
Oscillatory Activity	Power, phase, connectivity	Medium to High	GAN-based Methods

Methodological Approaches for BCG Artifact Removal

Traditional and Adaptive Signal Processing Methods

Optimal Basis Set (OBS) and Adaptive OBS (aOBS) The OBS method represents a significant advancement over simple average artifact subtraction (AAS) by employing principal component analysis (PCA) on EEG segments time-locked to cardiac events [5]. The first several principal components, which capture the dominant BCG artifact variance, are subtracted from the EEG data. The adaptive OBS (aOBS) method enhances this approach by estimating the delay between cardiac activity and BCG occurrence on a beat-to-beat basis, ensuring more accurate alignment of BCG events [5]. Furthermore, aOBS automatically determines the number of PCA components to remove based on explained variance criteria, addressing a significant limitation of traditional OBS which typically removes a fixed number of components (often 3-4) regardless of inter-subject or inter-channel variability [5]. Implementation of aOBS involves: (1) detection of BCG peaks from gradient-corrected EEG data or synchronized ECG, (2) epoching EEG data around BCG peaks, (3) PCA decomposition of epoched data, (4) automatic selection of artifact components based on variance criteria, and (5) reconstruction and subtraction of the BCG artifact from the original EEG [5].

Independent Component Analysis (ICA) and Surrogate Methods ICA approaches separate EEG signals into statistically independent components, which are then manually or automatically classified as neural or artifactual [3]. The BCG artifact components are removed before signal reconstruction. A significant limitation of ICA is the difficulty in establishing unanimous criteria for component selection and the potential for removing neural activity mixed with artifactual components [3]. Surrogate methods combine spatial filtering with artifact topography information derived either from PCA (PCA-S) or manual ICA component selection (ICA-S) [3] [22]. These methods use a surrogate source model consisting of regional dipole sources distributed throughout the brain to describe genuine EEG activity, enabling separation of artifact and brain signals with minimal distortion [3]. Studies have demonstrated that surrogate methods, particularly PCA-S, yield superior performance in source localization accuracy compared to traditional approaches [3].

Deep Learning Approaches

Generative Adversarial Networks (GANs) The BCGGAN framework represents a novel approach that formulates BCG artifact removal as an unpaired signal translation problem [14]. This method employs a modular training strategy that optimizes generator network parameters through successive constraint layers, enhancing the local representation capability of the model. The BCGGAN does not require additional reference signals (e.g., ECG) or specialized hardware, making it applicable to existing datasets [14]. The generator network learns to transform BCG-contaminated EEG signals into clean EEG through an adversarial training process against a discriminator network that distinguishes between real clean EEG and generated signals. This approach has demonstrated superior performance in preserving physiological information such as the eyes-open/eyes-closed (EO/EC) alpha rhythm contrast while effectively suppressing BCG artifacts [14].

Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs) RNN-based approaches leverage the temporal dependencies in BCG artifacts by learning nonlinear mappings between ECG references and BCG-contaminated EEG [27]. These networks are particularly suited for capturing the complex, time-varying relationship between cardiac activity and the resulting artifact. Alternatively, CNN-based architectures like EEGNet have been adapted for artifact removal and single-trial ERP quantification [71]. These compact convolutional networks can learn spatial and temporal filters that separate neural activity from artifacts without requiring explicit reference signals. For single-trial ERP analysis, neural networks have demonstrated superior performance in quantifying latency jitter and amplitude variability compared to traditional methods like Woody filtering or spatiotemporal beamforming [71].

Experimental Protocols for ERP Recovery Validation

Protocol for Validating BCG Removal Using Omitted-Target Paradigms

The omitted-target paradigm represents a powerful approach for validating BCG artifact removal methods because it elicits purely endogenous ERPs (emitted potentials) without contamination from exogenous sensory components [72]. This protocol involves presenting subjects with a standard oddball sequence but replacing the target stimulus with its omission, creating a purely cognitive ERP related to target expectation and processing.

Stimuli and Task Design:

• Implement a visual oddball paradigm with two conditions: standard oddball and omitted-target oddball
• Use square-wave gratings with different orientations as frequent non-target stimuli (probability: 0.80)
• For the standard oddball: include rare target stimuli (probability: 0.20) with different orientation
• For the omitted-target oddball: replace rare targets with stimulus omissions (probability: 0.20)
• Set stimulus-onset asynchrony (SOA) to 2 seconds
• Instruct participants to mentally count rare targets/omissions and report the total after each run

Data Acquisition Parameters:

• Simultaneous EEG-fMRI recording using MRI-compatible systems
• EEG sampling rate: ≥5,000 Hz (10,000 Hz recommended) [3]
• 64-channel EEG cap with current-limiting resistors for safety
• fMRI parameters: TR=2-3 s, TE=30 ms, slice thickness=3 mm
• Synchronize EEG sampling clock with MRI scanner

Validation Metrics:

• Source localization accuracy of endogenous components (P3b, pP2)
• Topographical consistency across subjects
• Anticipatory slow wave preservation (Bereitschaftspotential, prefrontal negativity)
• Statistical comparison of activation patterns with standard oddball task

Protocol for Assessing Single-Trial ERP Quantification

Single-trial ERP analysis provides crucial information about trial-to-trial variability that is lost in conventional averaging approaches [71]. This protocol evaluates the performance of BCG artifact removal methods in preserving single-trial features.

Experimental Tasks:

• Attentive oddball task with phoneme stimuli
• Semantic sentence congruity task
• Counterbalanced task order across participants
• Sufficient trial numbers (≥80 per condition) to support single-trial analysis

Neural Network Training for Single-Trial Analysis:

• Adapt EEGNet or convolutional LSTM architectures for ERP component quantification
• Generate simulated training data with known latency jitter and amplitude variability
• Train networks to estimate single-trial latencies and amplitudes for target components (P300, N400)
• Compare performance against traditional methods (Woody filtering, spatiotemporal beamforming)

Evaluation Criteria:

• Correlation between estimated latencies and reaction times
• Amplitude recovery accuracy compared to simulated ground truth
• Topography and morphology preservation of ERP components
• Latency variability measures across subject populations (e.g., aging, clinical groups)

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 4: Essential Materials for Simultaneous EEG-fMRI ERP Research

Item	Specification	Function	Technical Considerations
MRI-Compatible EEG System	64+ channels with amplifier	Neural signal acquisition	Must include current-limiting resistors for safety; synchronization with MRI clock
EEG Electrode Cap	Ag/AgCl electrodes with carbon fiber leads	Signal transduction	Carbon fiber reduces heating; appropriate sizing for stable electrode placement
Conductive Electrode Gel	MRI-compatible formulation	Signal conduction	Non-ferromagnetic; stable impedance over recording duration
ECG Recording Equipment	MRI-compatible electrodes	Cardiac reference	Essential for BCG artifact removal methods requiring cardiac timing
Visual Presentation System	MRI-compatible goggles	Stimulus delivery	Capable of precise timing synchronization with EEG-fMRI acquisition
Artifact Removal Software	Custom or commercial (e.g., Analyzer-2)	Signal processing	Should support multiple removal methods (OBS, ICA, etc.) for comparison
Data Analysis Platform	MATLAB, Python, or BESA Research	Signal analysis	Must handle large multimodal datasets; support for custom algorithm implementation

The recovery of ERP and oscillatory signals from BCG-contaminated EEG data requires careful methodological consideration and validation. Traditional approaches like aOBS and surrogate methods provide robust artifact removal with high signal preservation, particularly valuable for source localization studies [3] [5]. Emerging deep learning methods offer powerful alternatives, especially for single-trial analysis and when reference signals are unavailable [27] [71] [14]. The selection of an appropriate artifact removal strategy should be guided by research objectives, available resources, and required signal integrity. For clinical applications and drug development research, where ERP parameters may serve as critical biomarkers, implementation of rigorous validation protocols using omitted-target paradigms and single-trial analyses is essential to ensure neural signal integrity and meaningful physiological interpretation [71] [72].

Simultaneous Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI) provides a powerful multimodal framework for investigating brain dynamics across temporal and spatial scales. However, the magnetic field generated during fMRI acquisition introduces significant artifacts in EEG data, particularly the ballistocardiogram (BCG) artifact, which arises from cardiac-related motion and poses a substantial challenge for accurate signal interpretation [64] [4]. While traditional evaluation of BCG artifact removal methods has focused primarily on signal-level metrics, emerging research demonstrates that these preprocessing choices fundamentally reshape the topological interpretation of functional brain networks [64] [1]. This technical review examines how different BCG correction techniques affect not only signal quality but also the integrity of functional connectivity and graph-theoretical network properties derived from simultaneous EEG-fMRI data.

The BCG artifact represents one of the most persistent challenges in simultaneous EEG-fMRI, with amplitude often exceeding neural signals by 3-4 times and exhibiting complex non-stationary patterns synchronized with the cardiac cycle [36] [58]. Effective removal is complicated by the artifact's temporal variability and spatial distribution across electrodes, making complete elimination without neural signal loss particularly challenging [4] [5]. Understanding how these methodological decisions propagate through analysis pipelines to influence final network representations is essential for reliable interpretation of brain connectivity in basic neuroscience and drug development research.

Core Methodologies for BCG Artifact Removal

Fundamental Algorithms and Principles

Three primary algorithmic approaches dominate BCG artifact removal, each with distinct theoretical foundations and implementation considerations:

• Average Artifact Subtraction (AAS): This template-based method constructs an average artifact waveform from multiple cardiac cycles and subtracts it from the EEG signal [64] [58]. While computationally efficient and straightforward to implement, AAS assumes relatively stable artifact morphology over time, making it less effective against the BCG artifact's inherent non-stationarity [64] [5].

• Optimal Basis Set (OBS): Using Principal Component Analysis (PCA), OBS captures the dominant temporal variations in BCG artifact structure across cardiac cycles [64] [58]. This approach creates a basis set of principal components that represent the artifact's variance, which are then fitted and subtracted from each artifact occurrence. OBS better accommodates morphological variations in BCG artifacts compared to AAS [64] [5].

• Independent Component Analysis (ICA): This blind source separation technique decomposes EEG signals into statistically independent components, allowing for identification and removal of BCG-related components before signal reconstruction [64] [58]. The effectiveness of ICA depends critically on accurate component selection, which remains challenging due to the spatial non-stationarity of BCG sources [36] [24].

Hybrid and Advanced Approaches

Recent methodological advances include hybrid pipelines and adaptive algorithms that address specific limitations of traditional approaches:

• PROJIC Framework: This family of methods combines ICA with a novel component selection approach (PROJection onto Independent Components) followed by OBS or AAS correction in the component space before back-projection [36]. Evaluation shows PROJIC-OBS excels when prioritizing artifact removal, while PROJIC-AAS better preserves physiological signals [36].

• Adaptive OBS (aOBS): This extension incorporates beat-to-beat estimation of the delay between cardiac activity and BCG occurrence, enabling more precise artifact alignment before PCA decomposition [5]. The approach automatically identifies artifact-related components based on explained variance, addressing the arbitrary component selection in traditional OBS [5].

• Clustering-Constrained ICA (ccICA): This method integrates clustering algorithms to capture the time-varying features of BCG artifacts before applying constrained ICA, showing improved performance in both simulated and real data compared to traditional approaches [24].

Table 1: Performance Comparison of Primary BCG Artifact Removal Methods

Method	Key Principle	Strengths	Limitations
AAS	Template averaging and subtraction	Computational efficiency; Simple implementation	Poor handling of non-stationary artifacts; Residual artifacts
OBS	PCA-based basis set extraction	Accommodates artifact variability; Better preservation of neural signals	Component selection arbitrariness; Performance depends on precise cardiac timing
ICA	Blind source separation	Spatial filtering; Does not require cardiac timing information	Component selection subjectivity; Spatial non-stationarity issues; Potential neural signal loss
Hybrid (OBS+ICA)	Sequential application of OBS then ICA	Combines strengths of both methods; Reduces residuals	Increased complexity; Potential signal distortion from sequential processing

Quantitative Performance Evaluation: From Signal Fidelity to Network Integrity

Signal-Level Metrics and Method-Specific Performance

Comprehensive evaluation of BCG removal techniques requires multiple metrics assessing different aspects of signal quality preservation. Recent research implementing a multifaceted assessment framework reveals method-specific performance patterns [64] [1]:

• AAS demonstrates superior performance in traditional signal fidelity metrics, achieving the lowest Mean Squared Error (MSE = 0.0038) and highest Peak Signal-to-Noise Ratio (PSNR = 26.34 dB) in comparative studies [64] [1].

• OBS excels in structural similarity preservation, attaining the highest Structural Similarity Index (SSIM = 0.72), indicating better maintenance of the original signal's structural information [64] [1].

• ICA shows comparatively weaker performance on standard signal metrics but demonstrates unique sensitivity to frequency-specific patterns, particularly in dynamic connectivity analyses [64] [73].

• Hybrid Methods (particularly OBS+ICA) show significant benefits, producing the lowest p-values across frequency band pairs (e.g., theta-beta and delta-gamma) in statistical comparisons of connectivity patterns [64] [1].

Table 2: Quantitative Performance Metrics Across BCG Removal Methods

Performance Metric	AAS	OBS	ICA	OBS+ICA
Mean Squared Error (MSE)	0.0038	0.0047	0.0059	0.0041
Peak Signal-to-Noise Ratio (PSNR)	26.34 dB	25.18 dB	23.98 dB	25.92 dB
Structural Similarity Index (SSIM)	0.68	0.72	0.63	0.71
Spectral Correlation Coefficient	0.89	0.92	0.85	0.93
Improvement in Normalized Power Spectrum	22.4%	28.7%	19.3%	31.2%

Frequency-Specific and Dynamic Effects

Different artifact removal approaches exhibit distinct frequency-dependent impacts on EEG signals, with particularly pronounced effects in higher frequency bands [64] [73]:

• Beta and Gamma Bands: Show the strongest differentiation between methods under dynamic conditions, with hybrid approaches (OBS+ICA) demonstrating superior preservation of connectivity patterns in these frequency ranges critical for cognitive processing [64] [1].

• Alpha Band: Particularly vulnerable to contamination from BCG artifacts and susceptible to unintended removal during artifact correction, especially with ICA-based approaches [24].

• Dynamic Connectivity: Time-varying connectivity analyses reveal more pronounced frequency-specific effects than static approaches, with method selection significantly influencing detected network reconfigurations [64] [73].

Impact on Functional Connectivity and Network Topology

Graph-Theoretical Metrics and Network Organization

The choice of BCG removal method substantially influences derived functional network properties, with implications for neuroscientific interpretation [64] [1] [73]:

• Connection Strength: Systematic differences emerge across methods, with AAS typically yielding strongest overall connections but potential overestimation due to residual artifacts, while ICA produces more conservative estimates [64].

• Clustering Coefficient: Nodal segregation metrics vary significantly, particularly in hub regions, with OBS-based approaches generally preserving small-world architecture more effectively [64] [1].

• Global Efficiency: Integration metrics show method-dependent patterns, with hybrid approaches balancing preservation of neural connectivity while effectively removing artifactual connections [64] [73].

• Network Hubs: The identification of highly connected network hubs differs across artifact removal pipelines, potentially altering interpretations of brain network organization [64].

Method-Specific Topological Profiles

Each artifact removal approach produces distinctive topological signatures in resulting functional networks [64] [1]:

• AAS: Generates networks with higher overall connection density and stronger small-world properties, though potentially inflated by residual artifacts [64].

• OBS: Produces intermediate connection densities with preserved modular organization and more biologically plausible hub distributions [64] [1].

• ICA: Yields sparser networks with reduced connection strengths but potentially superior specificity for true neural connections [64] [73].

• Hybrid Methods: Combine topological features of constituent approaches, with OBS+ICA showing particularly balanced performance across multiple graph metrics [64] [1].

Dynamic Connectivity and Temporal Network Properties

Beyond static connectivity, BCG removal methods differentially impact time-varying network properties [64] [73]:

• Temporal Stability: OBS and hybrid methods demonstrate superior stability in dynamic connectivity patterns compared to ICA, which shows higher trial-to-trial variability [64].

• State Transitions: Method selection influences detected transitions between brain network states, particularly during task conditions or paradigm shifts [64] [73].

• Frequency-Band Interactions: Artifact removal approaches differentially modulate cross-frequency coupling estimates, with hybrid methods (OBS+ICA) showing enhanced detection of theta-beta and delta-gamma interactions [64] [1].

Experimental Protocols and Methodological Implementation

Standardized Evaluation Pipeline

Robust assessment of BCG artifact removal efficacy requires a comprehensive evaluation pipeline incorporating multiple validation approaches [36]:

• Residual Artifact Quantification: Calculate root mean square of BCG waveforms and peak-to-peak values after correction, with lower values indicating better artifact removal [36] [5].

• Physiological Signal Preservation: Assess event-related potentials (ERPs) or steady-state responses before and after correction, measuring signal-to-noise ratio improvement and inter-trial variability reduction [36] [5].

• Spectral Integrity Evaluation: Compare power spectral density before and after correction, particularly in frequency bands of interest, using metrics like Improvement in Normalized Power Spectrum (INPS) [36] [24].

• Cross-Correlation Assessment: Compute maximum cross-correlation between EEG signals and ECG reference, with successful correction producing lower correlation values [5].

Implementation Considerations and Parameter Optimization

Successful application of BCG removal methods requires careful attention to implementation details [58] [5]:

• Cardiac Event Detection: Accurate identification of QRS complexes is fundamental, with recommended use of adaptive threshold algorithms applied to band-pass filtered ECG signals (7-40 Hz) [58] [24].

• Component Selection: For OBS, systematic determination of principal components to retain (typically 3-8 based on explained variance criteria) rather than fixed numbers across datasets [5].

• Temporal Alignment: Implementation of dynamic time warping or similar approaches to address variable BCG artifact duration, particularly for OBS-based methods [58].

• Quality Metrics Integration: Incorporation of multiple quantitative indicators (MSE, PSNR, SSIM, spectral correlation) for comprehensive method evaluation [64] [1].

Table 3: Research Reagent Solutions for BCG Artifact Investigation

Tool/Category	Specific Examples	Function/Application
Software Platforms	BrainVoyager EMEGSuite, FMRIB Plugin, EEG-LLAMAS	Implementation of AAS, OBS, and real-time artifact removal algorithms
Analysis Toolboxes	EEGLAB, FieldTrip, Custom MATLAB scripts	ICA implementation and component selection
Hardware Solutions	MR-compatible EEG systems, Piezoelectric sensors, Carbon-fiber electrodes	Hardware-based artifact reduction at acquisition
Evaluation Metrics	MSE, PSNR, SSIM, INPS, Graph theory metrics	Quantitative assessment of method performance
Visualization Tools	BrainNet Viewer, Circos plots, Custom connectivity displays	Topological representation of network changes

The selection of BCG artifact removal methodologies extends far beyond conventional signal quality metrics to fundamentally shape functional connectivity patterns and network topological properties. Method-specific profiles emerge across analytical dimensions: AAS excels in signal fidelity but may inflate connection strengths; OBS preserves structural similarity and produces biologically plausible network architectures; ICA demonstrates unique sensitivity to frequency-specific patterns despite weaker signal metrics; while hybrid approaches (particularly OBS+ICA) offer balanced performance across multiple evaluation domains [64] [1] [73].

Future methodological development should prioritize (1) adaptive algorithms that dynamically adjust to BCG artifact variability, (2) standardized evaluation frameworks incorporating connectivity and network metrics alongside traditional signal measures, and (3) open-source implementations ensuring reproducibility and accessibility across the research community [4] [5]. For researchers and drug development professionals, explicit reporting of artifact removal methodologies and parameter selections is essential for valid interpretation of functional connectivity findings and cross-study comparisons [4]. As simultaneous EEG-fMRI continues to illuminate brain network dynamics in health and disease, acknowledging and addressing the methodological dependencies of BCG artifact removal on functional network inference remains paramount for scientific advancement.

This technical guide synthesizes empirical evidence on the performance of various ballistocardiogram (BCG) artifact removal methods for simultaneous EEG-fMRI. Effective BCG artifact removal is crucial for ensuring signal fidelity and the accurate interpretation of brain network dynamics in multimodal research [1] [26].

Quantitative Performance Comparison of BCG Artifact Removal Methods

The table below summarizes key performance metrics from empirical studies evaluating different BCG artifact removal techniques.

Table 1: Empirical Performance of BCG Artifact Removal Methods

Method	Reported Performance Metrics	Key Empirical Findings	Study Context
Average Artifact Subtraction (AAS)	Best signal fidelity (MSE = 0.0038, PSNR = 26.34 dB) [1]. BCG residual intensity of 12.51% [5].	Effective for stable artifacts but struggles with temporal variability, leaving residuals [1] [74] [5].	Evaluation on signal quality and functional connectivity [1].
Optimal Basis Set (OBS)	Highest structural similarity (SSIM = 0.72) [1]. BCG residual intensity of 9.20% [5].	Outperforms AAS by modeling artifact variability via PCA; performance highly dependent on the correct number of Principal Components (PCs) [74] [5].	Comparison focused on evoked response SNR [74].
Independent Component Analysis (ICA)	BCG residual intensity of 20.63% [5]. Lower signal metrics but sensitive to frequency-specific patterns in dynamic graphs [1].	Effectiveness limited by non-stationary spatial topography of BCG; component selection is critical and can lead to neural signal loss if done incorrectly [1] [74] [5].	Assessment of signal quality and network topology [1].
Adaptive OBS (aOBS)	Lowest BCG residual intensity (5.53%) and cross-correlation with ECG (0.028) [5].	Automatically identifies BCG peaks and selects PCs based on explained variance, outperforming standard OBS, AAS, and ICA [5].	Validation on high-density EEG during simultaneous fMRI with visual stimulation [5].
OBS + ICA Hybrid	Produced the lowest p-values in dynamic graph analysis across frequency bands (e.g., theta-beta, delta-gamma) [1].	Combines strengths of both methods; shows benefits for connectivity analysis but may introduce distortion if not applied carefully [1] [22].	Analysis of static and dynamic functional connectivity graphs [1].
Deep Learning (BCGNet)	Larger average power reduction at critical frequencies and improved task-relevant EEG classification compared to OBS [23].	Uses RNN (GRU) to learn non-linear mapping between ECG and BCG-corrupted EEG; does not require QRS complex detection [23].	Evaluation on EEG-fMRI data from an auditory oddball paradigm [23].
Denoising Autoencoder (DAR)	RMSE of 0.0218 ± 0.0152, SSIM of 0.8885 ± 0.0913, and SNR gain of 14.63 dB; outperformed PCA, ICA, AAS, and wavelet thresholding [56].	A deep learning model that learns direct mapping from noisy to clean EEG signals; generalizes well to unseen subjects [56].	Trained and tested on the CWL EEG-fMRI dataset [56].

Detailed Experimental Protocols and Methodologies

Protocol 1: Evaluating the Impact on Functional Connectivity

A 2025 study adopted a holistic framework to evaluate AAS, OBS, and ICA, alongside hybrid methods (AAS+ICA, OBS+ICA) [1].

• Data Acquisition: Simultaneous EEG-fMRI recordings were performed.
• Preprocessing: Standard gradient artifact removal was first applied.
• BCG Removal: Each method (AAS, OBS, ICA, hybrids) was applied to the same dataset.
• Signal-Level Evaluation: Methods were assessed using multiple metrics: Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) [1].
• Connectivity Analysis: EEG and fMRI time-series were combined via correlation to generate static and dynamic brain graphs for different frequency bands. Graph theory metrics (Connection Strength, Clustering Coefficient, Global Efficiency) were used to analyze effects on network topology [1].
• Key Finding: The choice of artifact removal method significantly impacted the topological interpretation of functional brain networks, with dynamic analyses showing pronounced frequency-specific effects [1].

Protocol 2: Individual Optimization of OBS

A 2015 study compared OBS and AAS, focusing on the critical parameter of the number of Principal Components (PCs) for OBS [74].

• Paradigm: Auditory and visual evoked potentials were recorded from subjects inside an MR scanner.
• Processing: Data was processed using different pipelines (Brain Vision Analyzer, EEGlab) and combinations of AAS and OBS.
• Optimization: The number of PCs for OBS was optimized at three levels: for the entire group, for individual subjects, and for individual recording runs.
• Evaluation: Performance was quantified by the Signal-to-Noise Ratio (SNR) of the recovered evoked responses.
• Key Finding: Individually optimizing the number of PCs for each subject and run significantly improved the SNR of evoked responses compared to using a fixed number for a group. The optimal number varied widely across datasets (e.g., 3-25 PCs) [74].

Protocol 3: Validation of Deep Learning (BCGNet)

A 2020 study introduced BCGNet, a deep learning approach using Gated Recurrent Units (GRUs) [23].

• Task & Acquisition: 19 subjects performed an auditory oddball paradigm during simultaneous EEG-fMRI in a 3T scanner. EEG was recorded at 5 kHz.
• Preprocessing: Standard GA removal was performed using FASTR, followed by resampling and filtering.
• Network Training: The GRU-based network was trained to learn the non-linear mapping between the simultaneously recorded ECG channel (input) and the BCG-corrupted EEG (target).
• Artifact Removal: The trained network was used to predict the BCG artifact waveform from the ECG, which was then subtracted from the EEG.
• Evaluation: Performance was compared against OBS based on the power spectrum of the cleaned EEG and the accuracy of single-trial classification of event-related potentials.
• Key Finding: BCGNet achieved greater power reduction of the BCG artifact at critical frequencies and improved task-relevant EEG classification compared to OBS [23].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Materials and Tools for EEG-fMRI BCG Artifact Research

Item Name	Function / Application	Relevance to BCG Artifact Research
MR-Compatible EEG Amplifier (e.g., BrainAmp MR Plus)	Records EEG signals inside the high magnetic field environment.	Essential for data acquisition; specialized hardware minimizes initial artifact intensity and ensures safety [23].
64+ Channel EEG Caps	High-density electrode configuration for scalp potential recording.	Provides rich spatial information that can be leveraged by spatial filtering methods (ICA, OBS) for better artifact separation [5] [22].
Electrocardiogram (ECG) Electrode	Records the electrical activity of the heart.	Crucial for AAS, OBS, and BCGNet to identify cardiac events (QRS complexes) and time-lock the artifact removal process [23] [5].
Carbon Fiber Wire Loops / Motion Sensors	Record motion-induced currents independently of brain activity.	Used in hardware-based solutions as a reference signal for BCG artifact subtraction [26] [22].
FASTR Algorithm	An implementation for removing Gradient and BCG artifacts, often available as a plugin for EEGLAB.	A widely used tool that incorporates AAS and OBS methods, providing a standard benchmark for comparing new techniques [23] [74].
EEGLAB / BESA Research Toolboxes	Software environments for advanced EEG signal processing and analysis.	Provide platforms for implementing and testing various artifact removal pipelines, including ICA and surrogate methods [23] [74] [22].

Experimental Workflow for Method Evaluation

The diagram below illustrates a generalized experimental workflow for evaluating BCG artifact removal methods, integrating common elements from the cited protocols.

Guidelines for Method Selection Based on Research Goals and Data Type

Simultaneous Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI) provides a powerful multimodal framework for investigating brain dynamics across complementary temporal and spatial scales [1] [4]. However, the interpretation of these integrated datasets is critically compromised by the ballistocardiogram (BCG) artifact, a persistent contaminant generated by cardiac-related movements and pulsatile blood flow within the scanner's static magnetic field [23] [22]. This artifact exhibits complex spatio-temporal dynamics, with amplitude and morphology varying across channels, subjects, and time, while its frequency content overlaps substantially with neural signals of interest [4] [52]. Effective BCG artifact removal is therefore not merely a preprocessing step but a fundamental determinant of data integrity, influencing subsequent analyses from simple signal interpretation to complex network connectivity metrics [1] [5]. The selection of an appropriate removal strategy must be guided by specific research objectives, data characteristics, and analytical requirements to balance artifact suppression with neural signal preservation. This guide provides a structured framework for method selection based on empirical evidence and technical considerations, equipping researchers with practical criteria for optimizing their EEG-fMRI studies.

Comparative Performance of BCG Artifact Removal Methods

Quantitative Method Performance

Table 1: Performance characteristics of established BCG artifact removal methods

Method	Key Principle	Best Suited Research Applications	Strengths	Limitations
AAS [1] [58]	Template subtraction using averaged artifact waveforms	Initial preprocessing; studies where signal fidelity is prioritized over connectivity analysis	Highest signal fidelity (MSE = 0.0038, PSNR = 26.34 dB) [1]; computational efficiency; simplicity	Limited handling of BCG variability; template misalignment issues [5]
OBS [1] [58] [5]	PCA-derived components capture artifact variations	Studies requiring balance between artifact removal and structural preservation; visual evoked potentials	High structural similarity (SSIM = 0.72) [1]; adapts to shape variations; improved residue reduction vs. AAS	Requires QRS detection; fixed component selection may remove neural signals [75]
ICA [1] [58] [22]	Blind source separation into statistically independent components	Exploratory analyses; dynamic network studies; frequency-specific investigations	Sensitivity to frequency-specific patterns in dynamic graphs [1]; no need for cardiac reference	Manual component selection expertise; potential neural signal loss; spatial stationarity assumption [58]
aOBS [5]	Adaptive detection of BCG peaks with automated component selection	High-density EEG; studies with variable heart-rate; automated processing pipelines	Lowest BCG residuals (5.53%) and cross-correlation with ECG (0.028) [5]; automated component selection	Increased computational complexity; newer method with less established track record
Hybrid (OBS+ICA) [1] [22]	Sequential application of OBS then ICA	High-precision studies; source localization; clinical applications requiring minimal distortion	Superior dynamic connectivity results (lowest p-values across band pairs) [1]; reduced residuals	Complex implementation; potential error propagation; longer processing time

Advanced and Emerging Methods

Table 2: Specialized and emerging BCG artifact removal approaches

Method	Key Principle	Best Suited Research Applications	Strengths	Limitations
Deep Learning (BCGNet) [23] [27]	RNN/GRU modeling of nonlinear ECG-EEG mappings	Real-time applications; task-relevant classification; high-artifact datasets	Superior power reduction at critical frequencies; improved single-trial classification [23]	Black-box nature; extensive training data required; limited interpretability
Surrogate Methods (PCA-S/ICA-S) [22]	Spatial filtering using artifact topographies with surrogate brain models	Source localization; ERP studies; minimal distortion requirements	Excellent source localization accuracy; preserves signal topography [22]	Requires specialized software (BESA); semi-automated component selection
Pulse-Triggered Stimulation [52]	Stimulus presentation during BCG-free intervals	Event-related potential studies; auditory/visual paradigms	Avoids artifact contamination at acquisition; complements post-processing methods [52]	Limited to specific experimental designs; increases protocol complexity
DRPE [76]	Direct recording of BCG/EEG bases with optimization-based reconstruction	Continuous EEG recordings; resting-state studies; alpha rhythm investigation	7x improvement over OBS in continuous signals [76]; incorporates prior knowledge	Complex experimental setup; requires additional recording sessions

Experimental Protocols for Key Methodologies

Optimal Basis Set (OBS) Implementation

The OBS method extends AAS by employing Principal Component Analysis (PCA) to capture temporal variations in BCG artifact morphology [58] [5]. The standard implementation involves: (1) Cardiac Event Detection: Identify QRS complexes from simultaneously recorded ECG or extract BCG peaks directly from EEG channels [5]. (2) Epoching: Segment EEG data into epochs time-locked to cardiac events, typically using a window of -200 to +500 ms relative to each R-peak [5]. (3) PCA Decomposition: Apply PCA to the ensemble of artifact epochs for each channel separately, generating an optimal basis set that captures the dominant variance patterns in the BCG artifacts [58]. (4) Component Selection: Select the first K principal components (typically 3-6) that explain the majority of artifact-related variance [5] [75]. (5) Artifact Reconstruction and Subtraction: Reconstruct the artifact waveform for each epoch using the selected components and subtract from the original signal [58].

Critical implementation details include the potential application of non-linear time warping (NLTW) to correct for duration variability between heartbeats before PCA [58], and careful selection of the number of components to avoid under- or over-correction [75]. The adaptive OBS (aOBS) variant introduces beat-to-beat delay estimation between cardiac activity and BCG occurrence, along with automated component selection based on explained variance criteria, significantly improving performance over standard OBS [5].

Independent Component Analysis (ICA) Implementation

ICA approaches separate multichannel EEG data into statistically independent components, requiring: (1) Data Decomposition: Apply ICA algorithms (Infomax or FastICA are most common for BCG removal) to the continuous, multichannel EEG data after gradient artifact correction [1] [22]. (2) Component Identification: Visual inspection or automated algorithms identify BCG-related components based on temporal correlation with cardiac rhythms, stereotypical topography, or waveform characteristics [22]. (3) Component Removal: Reconstruct EEG signals excluding the artifactual components [58].

A significant challenge is the manual expertise required for accurate component classification, though automated approaches based on correlation thresholds or machine learning are emerging [22]. The assumption of spatial stationarity for BCG sources may be violated in practice, potentially splitting BCG artifacts across multiple components and complicating identification [58].

Hybrid and Advanced Method Protocols

OBS-ICA Hybrid Method: This sequential approach applies OBS followed by ICA to address residual artifacts [1] [22]. Implementation involves: (1) Perform standard OBS correction; (2) Apply ICA to OBS-corrected data; (3) Identify and remove any residual BCG components that persist after OBS; (4) Reconstruct final cleaned EEG. This method combines the strength of OBS in capturing the primary artifact with ICA's ability to identify residual elements, potentially offering superior performance for dynamic connectivity analyses [1].

Deep Learning Approach (BCGNet): This emerging methodology utilizes gated recurrent units (GRUs) in a deep architecture to model nonlinear relationships between ECG and BCG-corrupted EEG [23] [27]. The protocol includes: (1) Data Preparation: Preprocess EEG with standard gradient artifact removal and filter (0.25-70 Hz); (2) Network Training: Train the RNN to predict BCG waveforms from concurrent ECG signals; (3) Artifact Subtraction: Subtract predicted BCG from recorded EEG to recover cleaned signals. This method shows particular promise for improving single-trial analysis and operates without additional hardware requirements [23].

BCG Artifact Removal Decision Workflow

Method Selection Framework Based on Research Goals

Decision Matrix for Method Selection

Method Selection Based on Multiple Factors

Application-Specific Recommendations

Event-Related Potential Studies: For auditory or visual evoked potentials, OBS or aOBS methods are generally preferred due to their balance of artifact reduction and signal preservation [5] [75]. The adaptive variant (aOBS) demonstrates particularly strong performance for visual stimulation paradigms, with clearer spatial topographies in occipital regions [5]. Pulse-triggered stimulation presents an innovative acquisition-based approach that can be combined with post-processing methods for additional improvement in ERP estimation [52].

Resting-State and Continuous EEG: For studies of ongoing neural oscillations or resting-state networks, ICA-based methods often outperform template approaches due to their ability to handle the non-stationary characteristics of continuous recordings [76]. The Direct Recording Prior Encoding (DRPE) method shows remarkable effectiveness for continuous data, demonstrating nearly sevenfold improvement over OBS in separating BCG and EEG components [76].

Functional Connectivity and Network Analysis: When investigating brain networks using graph theory metrics, hybrid approaches (OBS+ICA) provide superior results, particularly for dynamic connectivity analyses across frequency bands [1]. These methods yield the lowest p-values for frequency band pairs in dynamic graphs and minimize distortions in network topology interpretation [1].

Source Localization Applications: For studies requiring precise spatial localization of neural generators, surrogate methods (PCA-S or ICA-S) offer significant advantages, providing highly accurate source reconstruction with minimal distortion of neural signals [22]. These approaches explicitly incorporate spatial information through surrogate source models, outperforming established methods on criteria including source localization error and signal-to-noise ratio [22].

Single-Trial Analysis: When analyzing trial-by-trial variations in neural responses, deep learning approaches (BCGNet) show promising results, improving task-relevant EEG classification while achieving substantial BCG power reduction [23] [27]. The nonlinear modeling capabilities of RNNs effectively capture the complex relationship between ECG and BCG artifacts without requiring manual component selection.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential materials and software tools for BCG artifact research

Item	Function/Purpose	Implementation Notes
MR-Compatible EEG Systems (BrainAmp MR Plus, SynAmps2) [23] [22]	Acquire EEG data within high magnetic field environments	64-channel configurations typical; impedance maintenance <20 kΩ critical; amplifier stabilization with sandbags recommended [23]
ECG/Electrodes [23] [5]	Record cardiac reference for artifact removal timing	Single dedicated channel sufficient; proper placement vital for clean QRS detection
Insulating Materials (plastic barriers, saline-dampened paper) [76]	Create BCG-only reference signals for specialized methods	Used in DRPE method; enables direct BCG subspace characterization [76]
PCA-S [22]	Automated artifact topography identification for surrogate methods	Provides fully automated component selection; excellent for source localization [22]
ICA-S [22]	Manual component identification for surrogate methods	Allows expert curation; superior for challenging datasets with atypical artifacts [22]
FASTR Algorithm [23]	Integrated gradient and BCG artifact removal	Part of FMRIB EEGLAB plugin; implements OBS approach [23]
BrainVoyager EMEG Suite [58]	Comprehensive EEG-fMRI processing environment	Implements AAS, OBS, with non-linear time warping options [58]
BESA Research Software [22]	Source analysis and surrogate method implementation	Required for PCA-S/ICA-S approaches; excellent for source localization [22]
EEGLAB [23]	Flexible MATLAB-based processing environment	Extensible with plugins; supports multiple ICA algorithms and custom scripts [23]

The selection of an appropriate BCG artifact removal method constitutes a critical methodological decision that directly influences the validity and interpretability of simultaneous EEG-fMRI findings. Rather than applying a one-size-fits-all approach, researchers should carefully align their method selection with specific research goals, data characteristics, and analytical requirements. Traditional methods like AAS and OBS provide solid performance for many applications, while newer approaches including aOBS, surrogate methods, and deep learning offer specialized advantages for particular research contexts. Hybrid methods that combine the strengths of multiple approaches often yield superior results for complex analyses like dynamic functional connectivity. As the field advances toward more sophisticated applications including real-time processing and clinical translation, continued refinement of these methods will further enhance our ability to extract meaningful neural information from the challenging environment of simultaneous EEG-fMRI acquisition.

Conclusion

The effective removal of the BCG artifact remains a critical, yet solvable, prerequisite for harnessing the full potential of simultaneous EEG-fMRI. No single method is universally superior; the choice depends on a careful trade-off between the aggressiveness of artifact removal and the preservation of underlying neural signals. Template-based methods like AAS and OBS offer robustness, while ICA and hybrid approaches can address complex residuals at the potential cost of increased complexity. Emerging techniques, particularly deep learning models like BCGNet and sophisticated harmonic regression, show great promise in modeling the artifact's non-linear nature without requiring external references. Future directions point towards the increased use of machine learning, real-time correction algorithms for closed-loop paradigms, and the development of standardized, open-source validation frameworks. For the biomedical research community, mastering these artifact correction strategies is indispensable for generating reliable, high-fidelity data that can accurately illuminate brain function and drive discoveries in drug development and clinical neuroscience.

References

• 1. Enhancing Signal and Network Integrity: Evaluating BCG ... [https://www.mdpi.com/1424-8220/25/22/7036]
• 2. Evaluation and comparison of most prevalent artifact ... [https://www.frontiersin.org/journals/neuroimaging/articles/10.3389/fnimg.2022.968363/full]
• 3. EEG-fMRI: Ballistocardiogram Artifact Reduction by Surrogate ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8960642/]
• 4. Artifact Reduction in Simultaneous EEG-fMRI - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC7991907/]
• 5. Adaptive optimal basis set for BCG artifact removal in ... [https://www.nature.com/articles/s41598-018-27187-6]
• 6. Automated pipeline for EEG artifact reduction (APPEAR ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10696919/]
• 7. NeuXus open-source tool for real-time artifact reduction in ... [https://www.sciencedirect.com/science/article/pii/S1053811923005049]
• 8. A hybrid model for detecting motion artifacts in ... [https://pubmed.ncbi.nlm.nih.gov/40702570/]
• 9. A hybrid model for detecting motion artifacts in ... [https://biomedical-engineering-online.biomedcentral.com/articles/10.1186/s12938-025-01426-0]
• 10. Data‐driven beamforming technique to attenuate ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8288107/]
• 11. Removal of BCG artefact from concurrent fMRI-EEG ... [https://www.sciencedirect.com/science/article/abs/pii/S0165027017303011]
• 12. Cardiovascular function and ballistocardiogram: a relationship ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6752973/]
• 13. Adaptive optimal basis set for BCG artifact removal in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5995808/]
• 14. Ballistocardiogram artifact removal in simultaneous EEG ... [https://www.sciencedirect.com/science/article/abs/pii/S0165027022000255]
• 15. Motion-robust anterior–posterior imaging ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809423007401]
• 16. Ballistocardiogram artifact removal in simultaneous EEG ... [https://arxiv.org/abs/2011.01710]
• 17. Reference-Free Removal of EEG-fMRI Ballistocardiogram ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4761527/]
• 18. When Is Simultaneous Recording Necessary? A Guide for ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2021.636424/full]
• 19. Reference-free removal of EEG-fMRI ballistocardiogram ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811915005935]
• 20. Ballistocardiogram Artifact Removal with a Reference ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4126606/]
• 21. Adaptive optimal basis set for BCG removal in simultaneous... artifact [https://www.nature.com/articles/s41598-018-27187-6?error=cookies_not_supported&code=1c4cbde5-4ff9-4d3f-9487-53ea0c7029fc]
• 22. EEG-fMRI: Ballistocardiogram Artifact Reduction by ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.842420/full]
• 23. Ballistocardiogram artifact reduction in simultaneous EEG ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7808341/]
• 24. Clustering-Constrained ICA for Ballistocardiogram Artifacts ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2018.00059/full]
• 25. Evaluation and comparison of most prevalent artifact reduction methods for EEG acquired simultaneously with fMRI - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10406266/]
• 26. Frontiers | Artifact Reduction in Simultaneous EEG-fMRI: A Systematic Review of Methods and Contemporary Usage [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2021.622719/full]
• 27. Ballistocardiogram Artifact Reduction in Simultaneous EEG ... [https://pubmed.ncbi.nlm.nih.gov/32746037/]
• 28. Avoiding the Ballistocardiogram (BCG) Artifact of EEG data ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811909705750]
• 29. Gradient Artefact Correction and Evaluation of the EEG ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4942699/]
• 30. Reducing the Gradient Artefact in Simultaneous EEG-fMRI by ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3095086/]
• 31. Evaluating gradient artifact correction of EEG data acquired ... [https://www.sciencedirect.com/science/article/abs/pii/S0730725X07002159]
• 32. Properties of the ballistocardiogram artefact as revealed by ... [https://www.sciencedirect.com/science/article/abs/pii/S0167876007001584]
• 33. Simultaneous EEG-fMRI: What Have We Learned and What ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8952790/]
• 34. Simultaneous EEG-fMRI: FMRI Artifact Detection and ... [https://www.brainvoyager.com/bv/doc/UsersGuide/EMEGSuite/SimultaneousEEG-fMRIArtifactDetectionAndRemoval.html]
• 35. Complete artifact removal for EEG recorded during ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811906009761]
• 36. Ballistocardiogram artifact correction taking into account ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811916002391]
• 37. Removal of FMRI environment artifacts from EEG data ... [https://pubmed.ncbi.nlm.nih.gov/16150610/]
• 38. Removal of FMRI environment artifacts from EEG data ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811905004726]
• 39. Independent component analysis [https://en.wikipedia.org/wiki/Independent_component_analysis]
• 40. Mining EEG-fMRI using independent component analysis - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC2693483/]
• 41. ICA for dummies - Arnaud Delorme [https://arnauddelorme.com/ica_for_dummies/]
• 42. ICA-based procedures for removing ballistocardiogram ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811904005683]
• 43. Blind source separation and Independent component ... [http://www2.iap.fr/users/cardoso/compsep_classic.html]
• 44. Clustering-Constrained ICA for Ballistocardiogram Artifacts ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5816921/]
• 45. Independent Component Analysis with Functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10111109/]
• 46. Online Reduction of Artifacts in EEG of Simultaneous ... [https://link.springer.com/article/10.1007/s10548-017-0606-7]
• 47. Harmonic regression of Landsat time series for modeling ... [https://research.fs.usda.gov/treesearch/55701]
• 48. Surrogate data testing [https://en.wikipedia.org/wiki/Surrogate_data_testing]
• 49. Surrogate data for hypothesis testing of physical systems [https://www.sciencedirect.com/science/article/pii/S0370157318301340]
• 50. Deep Learning-based surrogate models for parametrized ... [https://arxiv.org/abs/2308.01602]
• 51. Carbon Wire Loops (CWLs) in EEG-fMRI studies [https://pressrelease.brainproducts.com/cwl-ux-montreal/]
• 52. Avoiding the ballistocardiogram (BCG) artifact of EEG data ... [https://www.sciencedirect.com/science/article/abs/pii/S0165027009006074]
• 53. Surrogate modelling of heartbeat events for improved J-peak ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11294145/]
• 54. BCG J-Peaks Detection for Heart Rate Calculation [https://bio-protocol.org/exchange/minidetail?id=10200155&type=30]
• 55. Effects of Ballistocardiogram Peak Detection Jitters on the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10007206/]
• 56. Deep Learning for Gradient and BCG Artifacts Removal in ... [https://www.arxiv.org/abs/2507.22263]
• 57. Multi-scale peak detection method for an automatic ... [https://cds.ismrm.org/protected/17MProceedings/PDFfiles/5414.html]
• 58. Simultaneous EEG-fMRI: FMRI Artifact Detection and ... [http://brainvoyager.com/bv/doc/UsersGuide/EMEGSuite/SimultaneousEEG-fMRIArtifactDetectionAndRemoval.html]
• 59. Removing ballistocardiogram (BCG) artifact from full-scalp ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2014.00218/full]
• 60. An optimized framework for simultaneous EEG-fMRI at 7T ... [https://pubmed.ncbi.nlm.nih.gov/41200234/]
• 61. Artefact reduction in simultaneous EEG-fMRI [https://cds.ismrm.org/protected/20MProceedings/PDFfiles/3880.html]
• 62. Ballistocardiogram artifact correction taking into account ... [https://pubmed.ncbi.nlm.nih.gov/27012501/]
• 63. The role of fMRI in drug development - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC5931333/]
• 64. Enhancing Signal and Network Integrity: Evaluating BCG ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12655990/]
• 65. EEG-Informed fMRI: A Review of Data Analysis Methods [https://pmc.ncbi.nlm.nih.gov/articles/PMC5808233/]
• 66. The multiverse of data preprocessing and analysis in graph ... [https://www.sciencedirect.com/science/article/pii/S0149763424003154]
• 67. Preservation of EEG spectral power features during ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9816433/]
• 68. Signal quality index: an algorithm for quantitative ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7687963/]
• 69. Non-invasive monitoring of cardiac function through ... [https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2023.1201722/full]
• 70. On the Definition of Signal-To-Noise Ratio and Contrast-To ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0077089]
• 71. Single-trial ERP Quantification Using Neural Networks - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10522773/]
• 72. “Hit the missing stimulus”. A simultaneous EEG-fMRI study ... [https://www.nature.com/articles/s41598-019-39812-z]
• 73. Evaluating BCG Artifact Removal Techniques in Simultaneous ... [https://sciety-labs.elifesciences.org/articles/by?article_doi=10.21203/rs.3.rs-7553208/v1]
• 74. Comparison of BCG artifact removal methods for evoked ... [https://www.sciencedirect.com/science/article/abs/pii/S0165027015000709]
• 75. Comparison of BCG artifact removal methods for evoked ... [https://pubmed.ncbi.nlm.nih.gov/25721269/]
• 76. Separation and reconstruction of BCG and EEG signals ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2014.00163/full]