Unlocking the Brain: A Comprehensive Guide to Multimodal fNIRS-EEG Neuroimaging for Research and Drug Development

Abstract

This article provides a detailed introduction to multimodal neuroimaging that integrates electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). Aimed at researchers, scientists, and drug development professionals, it explores the synergistic potential of combining EEG's millisecond temporal resolution with fNIRS's superior spatial localization for brain activity decoding. The content covers foundational principles, methodological approaches for data fusion and analysis, strategies for overcoming technical challenges like artifact removal, and validation through clinical applications in conditions from stroke rehabilitation to epilepsy. By synthesizing the latest advancements, this guide serves as an essential resource for leveraging fNIRS-EEG technology to accelerate neuroscience research and therapeutic development.

The Synergistic Power of fNIRS and EEG: Complementary Principles and Neural Correlates

In the pursuit of comprehending brain function, researchers leverage diverse neuroimaging modalities, each capturing distinct facets of neural activity. Two fundamental principles underpin the most widely used non-invasive techniques: the direct measurement of electrical activity via electroencephalography (EEG) and the indirect assessment of hemodynamic response via functional near-infrared spectroscopy (fNIRS). These methodologies offer complementary insights into brain dynamics; while EEG provides millisecond-level temporal resolution of neuro-electrical events, fNIRS tracks the slower, metabolically coupled blood flow changes with superior spatial specificity [1] [2]. Their integration is a cornerstone of modern multimodal neuroimaging, allowing researchers to correlate the rapid electrophysiological signatures of communication with the localized vascular consequences of energy demand [3]. This whitepaper details the core biophysical principles, experimental methodologies, and analytical frameworks for both techniques, providing a foundational guide for their application in neuroscience research and drug development.

Fundamental Principles of EEG

Electroencephalography (EEG) is a non-invasive technique for recording the brain's spontaneous electrical activity from the scalp surface. Its signal originates from the summed postsynaptic potentials of large, synchronously active ensembles of cortical pyramidal neurons [4] [5].

Neurophysiological Basis

The electrical signals measured by EEG are primarily generated by excitatory and inhibitory postsynaptic potentials (EPSPs and IPSPs) in cortical pyramidal neurons. When a neurotransmitter binds to a postsynaptic neuron, ion channels open, creating a flow of current across the membrane. The summation of these currents from thousands of simultaneously active neurons creates an electrical field strong enough to be detected at the scalp [6] [5]. It is critical to note that the action potentials themselves are too brief and non-synchronous to contribute significantly to the EEG signal; the dominant contribution comes from the slower, graded postsynaptic potentials [6].

• Ionic Mechanisms and Signal Summation: The resting membrane potential of a neuron, typically around -70 mV, is maintained by active ion channels like the sodium-potassium pump. Excitatory neurotransmitters like glutamate promote depolarization (EPSPs) by allowing positive ions into the cell, making the intracellular space more positive. Conversely, inhibitory neurotransmitters like GABA promote hyperpolarization (IPSPs) by allowing negative ions in or positive ions out. For an EEG signal to be detectable on the scalp, synchronous activity over a cortical area of approximately 6 cm² is required, which summation is necessary to overcome the signal attenuation caused by the skull, cerebrospinal fluid (CSF), and other intervening tissues [4] [6].
• The Role of Cortical Dipoles: Pyramidal neurons are oriented perpendicularly to the cortical surface. This organized geometry is crucial for EEG recording. Synaptic activity creates a current dipole—a separation of charge with a source and a sink. The resulting potential is a direct function of the dipole direction. For example, superficial excitatory postsynaptic potentials (EPSPs) create a negative potential at the scalp, while deep EPSPs result in a positive scalp potential [6]. Dipoles oriented parallel to the scalp surface are poorly detected, whereas those oriented radially (perpendicularly) are best captured [6].

Signal Acquisition and Technical Considerations

EEG records electrical potentials using a series of electrodes placed on the scalp according to standardized systems like the 10-20 system. The core technical aspects of signal acquisition are as follows:

• Differential Amplification: EEG employs differential amplifiers that measure the voltage difference between an active "exploring" electrode and a designated "reference" electrode. By convention, when the active electrode is more negative than the reference, the deflection is upward; a more positive active electrode results in a downward deflection [4].
• Signal Attenuation and Artifacts: The electrical signals generated in the cortex are significantly attenuated and spatially smeared as they pass through the cerebrospinal fluid, meninges, skull, and scalp. Furthermore, EEG recordings are susceptible to various biological artifacts, including electrical activity from scalp muscles (EMG), eye movements (EOG), the heart (ECG), and environmental noise. Distinguishing these artifacts from cerebral activity is a critical skill in EEG interpretation [4] [2].

Table 1: Key Characteristics of EEG Signal Acquisition

Feature	Description	Research Implication
Source	Summed excitatory & inhibitory postsynaptic potentials (EPSPs/IPSPs) of cortical pyramidal neurons [6] [5]	Measures input to neurons, not spiking output.
Spatial Resolution	Limited (cm-scale), due to signal volume conduction through tissues [4]	Poor for precise anatomical localization without advanced source modeling.
Temporal Resolution	Excellent (milliseconds) [2]	Ideal for tracking rapid neural dynamics and oscillations.
Key Artifacts	Muscle activity (EMG), eye movements (EOG), cardiac (ECG), environmental noise [4] [2]	Requires robust preprocessing and artifact rejection pipelines.

Fundamental Principles of fNIRS

Functional near-infrared spectroscopy (fNIRS) is a non-invasive optical neuroimaging technique that measures hemodynamic changes in the brain to infer neural activity. It leverages the principle of neurovascular coupling, where localized neural activation triggers a subsequent change in cerebral blood flow and oxygenation [1] [7].

Neurovascular Coupling and Hemodynamic Response

Neurovascular coupling is the fundamental link between neuronal activity and hemodynamic changes. During increased neural firing, there is a local rise in energy consumption, leading to an initial increase in oxygen extraction and a slight rise in deoxygenated hemoglobin (HbR). Within seconds, this triggers a compensatory regional cerebral blood flow (CBF) increase that overcompensates for the demand, leading to a pronounced increase in oxygenated hemoglobin (HbO) and a decrease in HbR in the venous capillaries [1]. This hemodynamic response is similar to the BOLD (Blood-Oxygen-Level-Dependent) signal measured by fMRI but fNIRS has the distinct advantage of measuring HbO and HbR concentrations separately [1] [8].

Biophysical Basis of Light-Tissue Interaction

fNIRS utilizes the relative transparency of biological tissues (including skin, skull, and brain) to light in the near-infrared spectrum (650-900 nm). Within this "optical window," the primary absorbers of light are oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) [1] [8] [7].

• The Modified Beer-Lambert Law (mBLL): Continuous-wave (CW) fNIRS systems, the most common type, quantify changes in hemoglobin concentration using the Modified Beer-Lambert Law. The mBLL relates the attenuation of light to the concentration of chromophores (HbO and HbR), the distance light travels, and a differential pathlength factor (DPF) that accounts for light scattering in tissue. By using at least two different wavelengths of light—typically one below and one above the ~810 nm isosbestic point (where HbO and HbR have equal absorption)—the relative changes in HbO and HbR concentrations can be calculated [8].
• Source-Detector Configuration and Depth Sensitivity: A typical fNIRS setup consists of optical sources (emitting NIR light) and detectors placed on the scalp. The sensitivity of a given source-detector pair (a "channel") to cerebral tissue is a function of their separation, typically 2.5-4 cm for adults. This configuration creates a "banana-shaped" path through the head. Crucially, the measured signal contains hemodynamic information from both the superficial scalp layers and the deeper cerebral cortex. To isolate the brain-specific signal, short-separation channels (e.g., < 1 cm) are used to measure and regress out the confounding systemic physiology from the scalp [1] [9].

Table 2: Key Characteristics of fNIRS Signal Acquisition

Feature	Description	Research Implication
Source	Changes in cerebral blood oxygenation & volume via neurovascular coupling [1] [7]	Indirect measure of neural activity with ~2-6 second delay.
Spatial Resolution	Moderate (1-3 cm), superior to EEG but inferior to fMRI [9]	Suitable for localizing activity to cortical gyri. Improved with high-density (HD) arrays [9].
Temporal Resolution	Slow (seconds) due to hemodynamic delay [2]	Cannot resolve individual neural events; best for block/event-related designs.
Key Artifacts	Systemic physiology (heartbeat, respiration, blood pressure), head motion [1] [2]	Requires short-separation channels and advanced signal processing.

Comparative Analysis: EEG vs. fNIRS

The following table provides a direct, quantitative comparison of the core technical specifications and functional characteristics of EEG and fNIRS, highlighting their complementary nature.

Table 3: Direct Comparison of EEG and fNIRS Measurement Principles

Parameter	EEG (Electrical Activity)	fNIRS (Hemodynamic Response)
Measured Quantity	Electrical potentials (µV) from summed postsynaptic potentials [4] [6]	Optical absorption changes, converted to HbO/HbR concentration (mmol/L) [1] [8]
Temporal Resolution	Excellent (1-5 ms) [2]	Poor (1-2 s) [2]
Spatial Resolution	Poor (cm-scale), limited by volume conduction [4]	Moderate (1-3 cm), improved with high-density arrays (HD-DOT) [9]
Depth Sensitivity	Superficial cortex (pyramidal neurons) [4]	Superficial cortex (1-3 cm depth) [1]
Neurophysiological Basis	Direct measure of neuro-electrical activity [5]	Indirect measure via neurovascular coupling [1]
Primary Signal Origin	Cortical pyramidal neurons [6]	Cortical capillary bed [1]
Key Advantages	Direct neural measurement, excellent temporal resolution, low cost, high portability [4] [2]	Good spatial resolution, less sensitive to motion artifacts, portable, measures HbO/HbR separately [1] [9]
Key Limitations	Poor spatial localization, sensitive to electrical artifacts, limited to cortical surface [4] [2]	Slow temporal response, indirect measure, sensitive to systemic physiology, limited depth penetration [1] [2]
Common Artifacts	Ocular (EOG), muscular (EMG), cardiac (ECG), environmental noise [4] [2]	Cardiac pulse, respiration, blood pressure changes, head motion [1] [2]

Experimental Protocols and Methodologies

Robust experimental design is paramount for generating high-quality, interpretable data in both unimodal and multimodal studies.

Standardized Experimental Paradigms

• Block Design for fNIRS: This is a common paradigm for fNIRS studies due to the slow nature of the hemodynamic response. It involves alternating periods of a task condition (e.g., 20-30 seconds) with periods of a control or rest condition (e.g., 20-30 seconds). The extended blocks allow the hemodynamic response to rise and fall, maximizing the signal-to-noise ratio for detecting activation [1]. This design is suitable for cognitive tasks like the Word-Color Stroop task for prefrontal cortex assessment [9] or language mapping.
• Event-Related Design for EEG: EEG's high temporal resolution makes it ideal for event-related potentials (ERPs). In this design, discrete stimuli or events (e.g., auditory tones, visual images) are presented with varying but sufficiently long inter-stimulus intervals. The EEG activity following each stimulus is time-locked and averaged over many trials to extract the consistent neural response (the ERP) from the background noise. This is used to study components like the P300, which is related to attention and context updating [4].

Detailed Protocol: Multimodal EEG-fNIRS for Motor Imagery

This protocol, adapted from a recent study platform, outlines the steps for a combined EEG-fNIRS investigation of motor imagery, a paradigm relevant for brain-computer interfaces (BCIs) and neurorehabilitation [3].

• Participant Preparation & Sensor Placement: Recruit participants based on study criteria (e.g., right-handedness). Fit the participant with a custom cap that integrates both EEG electrodes and fNIRS optodes. Position the cap so that sensor coverage is focused over the sensorimotor cortices (e.g., C3, Cz, C4 according to the 10-20 system). Apply electrolyte gel for EEG electrodes and ensure good optical contact for fNIRS optodes. For fNIRS, use a layout with both long-separation (e.g., 3 cm) and short-separation (e.g., 0.8 cm) channels.
• Signal Quality Check & Baseline Recording: Verify the impedance of all EEG electrodes is below a threshold (e.g., 10 kΩ). Check the signal quality for fNIRS channels, ensuring light intensity levels are sufficient. Record a 5-minute baseline with the participant at rest (e.g., eyes open, fixating on a cross).
• Task Execution & Neurofeedback (Optional): The participant performs a motor imagery task, such as imagining moving their left hand, without any actual movement. The task can be structured in a block design. In a neurofeedback setting, a real-time processed composite score from both EEG (e.g., event-related desynchronization in the mu/beta band over the right motor cortex) and fNIRS (e.g., increase in HbO in the same region) can be provided to the participant as visual feedback (e.g., a moving ball on a screen) [3].
• Data Synchronization: It is critical to synchronize the clocks of the EEG and fNIRS acquisition systems, either via hardware triggering or a shared software timestamp at the beginning of the experiment. This ensures that events can be precisely aligned during data analysis.
• Post-experiment Procedures: Upon task completion, record the precise locations of the EEG electrodes and fNIRS optodes using a 3D digitizer. This allows for co-registration of the data with individual or standard anatomical brain models, significantly improving the accuracy of spatial localization.

Data Analysis Workflows

The analysis of multimodal data involves distinct but often parallel pipelines for each modality before integration.

• EEG Preprocessing and Analysis: Raw EEG data is filtered (e.g., 0.5-40 Hz bandpass, 50/60 Hz notch). Artifacts are removed using techniques like Independent Component Analysis (ICA) or regression. For ERP analysis, data is epoched around stimulus events, baseline-corrected, and averaged. For oscillatory analysis, time-frequency decomposition (e.g., using wavelets or the Fast Fourier Transform) is performed to compute power in different frequency bands (delta, theta, alpha, beta, gamma) [5].
• fNIRS Preprocessing and Analysis: The raw optical density data is converted to HbO and HbR concentration changes using the mBLL. Short-separation channel regression is applied to remove systemic physiological noise. Bandpass filtering (e.g., 0.01-0.2 Hz) helps isolate the task-related hemodynamic response. Motion artifacts are corrected using algorithms like wavelet-based or PCA-based methods. The processed signals are then analyzed using a General Linear Model (GLM) to identify channels with statistically significant task-related responses [1] [10].
• Multimodal Data Fusion: Fusion of preprocessed EEG and fNIRS data can occur at different levels: 1) Data-level: Concatenating features from both modalities for input into a machine learning model. 2) Model-level: Using joint blind source separation or symmetric decomposition techniques to find latent variables that are represented in both datasets, potentially reflecting underlying neurovascular coupling [2]. 3) Decision-level: Combining the outputs of separate EEG and fNIRS classifiers to make a final decision, for instance, in a BCI application [3].

The Scientist's Toolkit: Essential Research Reagents & Materials

Successful execution of multimodal EEG-fNIRS experiments requires a suite of specialized hardware and software tools.

Table 4: Essential Materials and Tools for EEG-fNIRS Research

Item Category	Specific Examples & Specifications	Primary Function
Integrated EEG-fNIRS System	Custom caps with embedded EEG electrodes and fNIRS optodes; synchronized acquisition hardware (e.g., from Brain Products, Artinis, NIRx) [3]	Enables simultaneous, temporally aligned recording of electrical and hemodynamic activity.
fNIRS Optode Configurations	Sparse Arrays: ~30mm spacing for broad coverage. High-Density (HD) Arrays: Multiple overlapping source-detector pairs (e.g., 15-40mm) for improved spatial resolution and 3D image reconstruction via HD-DOT [9]. Short-Separation Detectors: ~8mm from a source.	Measures cortical hemodynamics. HD arrays improve localization. Short-separation channels enable removal of scalp blood flow contamination [1] [9].
Electrophysiology Solutions	Electrolyte gel or saline solution; abrasive skin preparation gel; Ag/AgCl or gold cup electrodes.	Ensures low electrical impedance between the scalp and EEG electrode, critical for high-quality signal acquisition.
3D Digitizer	Polhemus Patriot, Structure Sensor	Records the precise 3D locations of EEG electrodes and fNIRS optodes on the head for anatomical co-registration with MRI data, drastically improving spatial accuracy [2].
Stimulation & Feedback Software	Presentation, PsychToolbox; custom BCI/Neurofeedback software (e.g., BCILAB, OpenVibe) [3]	Presents experimental paradigms (visual/auditory stimuli) and provides real-time feedback to participants during neurofeedback protocols.
Data Analysis Software	EEG: EEGLAB, Brainstorm, MNE-Python. fNIRS: HOMER3, NIRS Toolbox, AtlasViewer [8]. Multimodal Fusion: Custom scripts in MATLAB or Python.	Provides toolboxes for signal processing, statistical analysis, source localization, and visualization of neuroimaging data.
Quality Control Tools	Impedance checkers for EEG; power meters for fNIRS source output.	Verifies the integrity of the hardware setup and signal quality before and during data collection.

EEG and fNIRS are powerful neuroimaging techniques rooted in the measurement of two distinct but interrelated physiological processes: electrical neural signaling and hemodynamic-metabolic coupling. EEG offers an unrivalled, direct view into the brain's millisecond-scale electrical dynamics, while fNIRS provides a more localized, indirect measure of the consequent blood flow changes. The integration of these modalities creates a more comprehensive picture of brain function, mitigating the limitations of either technique used in isolation. As hardware becomes more wearable and data fusion algorithms more sophisticated, multimodal EEG-fNIRS is poised to become a cornerstone for neuroscience research in naturalistic settings, with significant potential for advancing our understanding of brain disorders and accelerating the development of novel therapeutics and neurotechnologies.

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a paradigm shift in non-invasive neuroimaging, effectively bridging the critical gap between temporal and spatial resolution in brain research. This multimodal approach synergistically combines EEG's millisecond-scale temporal precision with fNIRS's millimeter-scale spatial localization, enabling researchers to capture both the electrical dynamics and hemodonic underpinnings of neural activity simultaneously. This technical guide examines the core principles, methodological frameworks, and experimental applications of EEG-fNIRS integration, providing researchers and drug development professionals with comprehensive insights into its transformative potential for understanding brain function and dysfunction. By leveraging the complementary strengths of these technologies, scientists can now investigate complex neural circuits and networks with unprecedented spatiotemporal precision, opening new frontiers in cognitive neuroscience, clinical diagnostics, and therapeutic development.

The fundamental challenge in non-invasive neuroimaging has historically been the trade-off between temporal and spatial resolution. Electroencephalography (EEG) measures electrical activity generated by neuronal populations through electrodes placed on the scalp, providing exceptional temporal resolution on the millisecond scale, which is essential for capturing the rapid dynamics of neural processing [11]. However, EEG signals are spatially blurred as they pass through the skull and other tissues, resulting in limited spatial resolution typically on the centimeter scale, making it difficult to precisely localize neural activity sources [11] [12].

Conversely, functional near-infrared spectroscopy (fNIRS) measures hemodynamic responses associated with neural activity by detecting changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) concentrations using near-infrared light, providing superior spatial resolution at the millimeter scale [13]. This technique benefits from more localized measurements but is constrained by the slow nature of hemodynamic responses, which evolve over seconds, thus offering poor temporal resolution compared to EEG [11].

The integration of EEG and fNIRS creates a powerful multimodal imaging approach that overcomes these individual limitations. This synergy allows researchers to simultaneously capture both the "when" of neural processing through EEG and the "where" through fNIRS, providing a more complete picture of brain function than either modality could deliver alone [14] [13]. This complementary relationship is particularly valuable for studying complex cognitive processes, developing brain-computer interfaces, and investigating neurological disorders where both the timing and location of neural events are critical for understanding underlying mechanisms.

Fundamental Principles and Technical Specifications

Neurophysiological Basis of EEG and fNIRS Signals

The complementary nature of EEG and fNIRS stems from their measurement of fundamentally different but neurophysiologically linked aspects of brain activity. EEG directly records electrical potentials generated by postsynaptic dendritic currents of synchronously firing pyramidal neurons. These signals propagate through various tissues including cerebrospinal fluid, skull, and scalp before being detected by surface electrodes, with different frequency bands (delta, theta, alpha, beta, gamma) reflecting distinct brain states and cognitive processes [12].

Simultaneously, fNIRS measures hemodynamic changes resulting from neurovascular coupling—the mechanism by which neural activity triggers localized increases in cerebral blood flow. Active neurons consume oxygen, leading to an initial slight increase in deoxygenated hemoglobin followed by a substantial increase in oxygenated hemoglobin that overshoots metabolic demands. fNIRS systems typically use light sources emitting at two wavelengths (695 nm and 830 nm in the Hitachi ETG-4100 system) to distinguish between HbO and HbR concentrations based on their distinct absorption spectra [15] [13].

The connection between these signals forms the foundation of multimodal integration: the electrical activity captured by EEG triggers the hemodynamic responses measured by fNIRS. This relationship enables researchers to connect rapid neural processing with its metabolic consequences, providing complementary insights into brain function that neither modality alone can deliver [14].

Technical Implementation and System Integration

Successful EEG-fNIRS integration requires careful consideration of technical implementation. EEG systems typically employ 16 to 128 electrodes arranged according to the international 10-20 system or higher-density variants, with electrode impedances kept below 5-10 kΩ for optimal signal quality [15]. fNIRS configurations vary from sparse optode arrangements to high-density arrays with source-detector separations of approximately 2.5-3.0 cm for cortical measurements, with shorter separations (0.5-1.0 cm) used to account for superficial physiological noise [11].

Hardware integration can be achieved through separate but synchronized systems or unified platforms with common processing units. The latter approach provides more precise temporal synchronization, which is crucial for analyzing the relationship between fast EEG dynamics and slower hemodynamic responses [12]. Customized headgear solutions incorporating both EEG electrodes and fNIRS optodes have been developed using 3D printing or thermoplastic materials to ensure stable positioning and proper scalp contact across diverse head shapes and sizes [12].

Table 1: Technical Specifications of EEG and fNIRS Modalities

Parameter	EEG	fNIRS
Measured Signal	Electrical potentials from neuronal activity	Hemodynamic changes (HbO, HbR)
Temporal Resolution	Millisecond scale (~1-100 ms) [11]	Seconds (~1-5 s) [11]
Spatial Resolution	Centimeters (~1-3 cm) [11] [12]	Millimeters (~5-10 mm) [11]
Depth Sensitivity	Superficial and deep sources (with volume conduction)	Cortical surface (up to 1.5-2.0 cm) [11]
Key Artifacts	Ocular, muscle, cardiac, line noise	Motion, cardiac, respiratory, blood pressure [14]
Portability	High (wearable systems available)	High (wearable systems available) [13]

Quantitative Performance Advantages of Multimodal Integration

Research has consistently demonstrated that combined EEG-fNIRS approaches yield significant improvements in brain activity characterization compared to unimodal applications. Simulation studies using the ICBM152 brain atlas have shown that neuronal sources separated by only 2.3-3.3 cm and 50 ms can be accurately recovered using joint EEG-fNIRS reconstruction, while remaining indistinguishable to either modality alone [11]. This represents a substantial advancement in spatiotemporal resolution that enables researchers to dissect fine-scale neural dynamics previously beyond the reach of non-invasive techniques.

In brain-computer interface applications, multimodal integration has dramatically improved classification accuracy across various paradigms. Studies implementing feature-level fusion of EEG and fNIRS data have achieved classification accuracies up to 98.38% for distinguishing brain states induced by preferred versus neutral music, significantly outperforming unimodal approaches [16]. Similarly, advanced deep learning architectures like the Multimodal DenseNet Fusion model have demonstrated superior performance in classifying cognitive and motor imagery tasks by effectively leveraging both temporal richness from EEG and spatial specificity from fNIRS [17].

The table below summarizes key quantitative improvements demonstrated across various experimental paradigms:

Table 2: Performance Metrics of EEG-fNIRS Integration Across Applications

Application Domain	Performance Metric	Unimodal Performance	Multimodal Performance
Source Localization	Minimum separable distance-temporal separation	Indistinguishable at 2.3-3.3 cm and 50 ms [11]	Accurately resolved [11]
Music Preference Classification	Classification accuracy	Not reported	98.38% [16]
Motor Imagery Tasks	Classification accuracy	78.21-92% (varies by modality and method) [17]	Improved by 5-15% over best unimodal [17]
Cognitive Task Classification	Classification accuracy	Not reported	87-92% [17]

Methodological Approaches and Experimental Protocols

Joint Source Reconstruction Protocol

The integration of EEG and fNIRS data for enhanced source localization follows a structured computational pipeline. The process begins with the creation of a head model, typically using a segmented MRI atlas such as ICBM152, which is converted into a tetrahedral mesh with distinct tissue compartments (scalp, skull, CSF, brain) [11]. For the ICBM152 atlas, this procedure typically yields approximately 96,593 nodes and 512,627 tetrahedrons, with source locations restricted to the outer surface of the brain compartment (approximately 15,255 locations) [11].

The forward models for both modalities are then computed. The EEG leadfield matrix is calculated using a boundary element method or finite element approach, incorporating appropriate conductivity values for different tissue types (e.g., scalp:skull:CSF:brain ratio of 1:80:1/5:1) [11]. The fNIRS forward model is derived using photon migration models (e.g., Monte Carlo simulations or analytical solutions to the diffusion equation) to characterize light propagation through tissue [11].

Inverse solutions are computed using algorithms that leverage the complementary strengths of both modalities. One effective approach utilizes the restricted maximum likelihood (ReML) framework, where the high-spatial-resolution fNIRS reconstruction serves as a spatial prior to constrain the high-temporal-resolution EEG reconstruction [11]. This method has been shown to successfully reconstruct multiple temporally overlapping neuronal sources activated with separations as brief as 50-60 ms, significantly outperforming single-modality approaches [11].

Multimodal Experimental Paradigms

Several well-established experimental protocols effectively leverage EEG-fNIRS integration across cognitive domains:

Motor Execution, Observation, and Imagery Paradigm: This protocol investigates the Action Observation Network (AON) through three conditions: (1) Motor Execution (ME) - participants physically grasp and move an object using their right hand upon an audio cue; (2) Motor Observation (MO) - participants observe an experimenter performing the same action; (3) Motor Imagery (MI) - participants mentally rehearse the action without movement [15]. Each trial begins with a fixation period (10-15s) followed by the condition-specific task (5-10s), with simultaneous EEG-fNIRS recording throughout. This paradigm has revealed consistent activation in the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus across all three conditions, validating shared neural mechanisms while highlighting subtle differences in activation patterns [15].

Semantic Category Decoding Protocol: This approach examines neural representations of semantic categories (e.g., animals vs. tools) during mental imagery tasks. Participants perform four distinct cognitive tasks after viewing object images: (1) Silent naming - internally generating the object's name; (2) Visual imagery - mentally visualizing the object; (3) Auditory imagery - imagining sounds associated with the object; (4) Tactile imagery - imagining the feeling of touching the object [18]. Each mental task lasts 3-5 seconds with simultaneous EEG-fNIRS recording, enabling investigation of how different sensory modalities contribute to semantic representation.

Personalized Music Listening Study: This protocol investigates brain responses to personalized musical stimuli. Participants first complete a questionnaire to identify their preferred music, while neutral music (typically unfamiliar relaxation music) serves as a control [16]. During the experiment, participants listen to both music types in randomized order while EEG and fNIRS data are synchronously collected. The paradigm typically includes 2-5 minute music presentation blocks interspersed with resting baseline periods, allowing analysis of both transient and sustained neural and hemodynamic responses to emotionally salient versus neutral auditory stimuli [16].

Signaling Pathways and Neural Hemodynamic Coupling

The relationship between electrical neural activity and hemodynamic responses forms the fundamental biological basis for EEG-fNIRS integration. This neurovascular coupling involves complex signaling pathways between neurons, astrocytes, and blood vessels that translate rapid electrical events into slower hemodynamic changes.

This neurovascular coupling pathway illustrates how the rapid electrical activity measured by EEG (millisecond domain) triggers metabolic and signaling processes that ultimately generate the hemodynamic responses measured by fNIRS (second domain). The integration of these complementary signals provides a more complete picture of brain activity than either modality can provide alone, capturing both the initiating neural events and their metabolic consequences [14] [19].

Data Processing and Fusion Methodologies

Multimodal Data Processing Pipeline

The analysis of simultaneous EEG-fNIRS data requires coordinated processing streams that account for the distinct characteristics of each modality while enabling meaningful integration. The following workflow outlines a standardized processing approach:

Advanced Fusion Techniques

Several sophisticated data fusion methods have been developed specifically for EEG-fNIRS integration:

Structured Sparse Multiset Canonical Correlation Analysis (ssmCCA): This technique identifies shared components across EEG and fNIRS datasets by maximizing correlations between linear combinations of features from both modalities while imposing structured sparsity constraints to improve interpretability [15]. Applied to motor execution, observation, and imagery tasks, ssmCCA successfully identified consistent activation in the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus across all conditions—patterns that were less distinct in unimodal analyses [15].

Multimodal DenseNet Fusion (MDNF): This deep learning approach transforms EEG data into 2D time-frequency representations using Short-Time Fourier Transform, extracts features using DenseNet architectures, and fuses them with fNIRS-derived spectral entropy features [17]. This method has demonstrated superior classification accuracy (87-92%) across cognitive tasks including n-back, discrimination/selection response, word generation, and motor imagery [17].

Improved Normalized-ReliefF Feature Selection: This feature-level fusion method normalizes multimodal features from both modalities, applies an enhanced ReliefF algorithm to select the most discriminative features, and integrates them for classification tasks [16]. This approach achieved 98.38% accuracy in distinguishing brain activity evoked by preferred versus neutral music, significantly outperforming unimodal classification [16].

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of EEG-fNIRS research requires careful selection of equipment, software, and analytical tools. The following table details essential components for establishing a multimodal neuroimaging laboratory:

Table 3: Essential Research Tools for EEG-fNIRS Investigations

Tool Category	Specific Examples	Function & Purpose
fNIRS Systems	Continuous-Wave (CW) systems (e.g., Cortivision Photon Cap) [13]	Measures relative changes in HbO and HbR concentrations; preferred for portability and cost-effectiveness
EEG Systems	Versatile EEG systems (e.g., Bitbrain Versatile EEG 16/32ch) [13]	Records electrical brain activity with high temporal resolution
Integrated Platforms	Customized EEG-fNIRS helmets [12]	Enables precise co-registration of modalities; 3D-printed or thermoplastic solutions optimize fit
Data Acquisition Software	Lab Streaming Layer (LSL) protocol [13]	Synchronizes multiple data streams with precise timing
Analytical Frameworks	Structured Sparse Multiset CCA (ssmCCA) [15]	Identifies shared neural components across modalities
Computational Tools	Variational Bayesian Multimodal Encephalography [11]	Provides statistical framework for joint source reconstruction
Head Modeling Resources	ICBM152 Brain Atlas [11]	Standardized anatomical reference for source localization

The integration of EEG and fNIRS technologies represents a significant advancement in non-invasive neuroimaging, effectively bridging the critical divide between temporal and spatial resolution in brain research. By combining millisecond-scale electrical measurements with millimeter-scale hemodynamic monitoring, this multimodal approach provides unprecedented insights into brain function across diverse domains including cognitive neuroscience, clinical diagnostics, and therapeutic development.

Future developments in this field will likely focus on several key areas: enhanced hardware integration with more compact and wearable systems enabling real-world applications; advanced analytical techniques incorporating machine learning and deep learning approaches for more sophisticated data fusion; and expanded clinical applications particularly in monitoring neurological disorders, assessing therapeutic efficacy, and developing closed-loop intervention systems [20] [12]. As these technologies continue to evolve, EEG-fNIRS integration is poised to become an increasingly indispensable tool for unraveling the complexities of brain function and dysfunction, ultimately advancing both fundamental neuroscience and clinical practice.

Neurovascular coupling (NVC) describes the fundamental physiological process whereby neuronal activity elicits regulated changes in local cerebral blood flow. This mechanism forms the biological basis for correlating electrophysiological signals from electroencephalography (EEG) with hemodynamic signals from functional near-infrared spectroscopy (fNIRS). This whitepaper provides an in-depth technical examination of NVC, detailing the integrated biological pathways, presenting quantitative data from multimodal studies, and outlining standardized experimental protocols for concurrent EEG-fNIRS research. The framework presented herein is essential for interpreting multimodal neuroimaging data within a cohesive physiological context, offering critical insights for researchers and drug development professionals investigating brain function and neurological disorders.

Neurophysiological Foundations of Neurovascular Coupling

Neurovascular coupling (NVC) represents the sophisticated biological mechanism that orchestrates cerebral blood flow to meet the metabolic demands of active neurons. This process creates a direct functional relationship between the electrical activity captured by EEG and the hemodynamic responses measured by fNIRS. Neuronal activation, particularly through glutamatergic neurotransmission, triggers a complex signaling cascade involving astrocytes, interneurons, and vascular smooth muscle cells, ultimately leading to vasodilation of local arterioles and an increase in cerebral blood flow [21].

This hemodynamic response is characterized by a local increase in oxygenated hemoglobin (HbO) and a decrease in deoxyhemoglobin (HbR), which fNIRS detects optically. Simultaneously, the synaptic and postsynaptic electrical activity underlying this process generates the electrical potentials measured by EEG. The temporal relationship between these signals is crucial: electrical activity occurs on a millisecond scale, while the hemodynamic response unfolds over seconds, creating a predictable delay that can be quantified through cross-modal correlation analyses [21] [19]. This temporal coupling, despite the different time scales, provides a robust foundation for integrating the complementary information from EEG and fNIRS.

Disruptions in NVC have been implicated in various neurological conditions, including Alzheimer's disease, traumatic brain injury, and neurodegenerative disorders [22] [23]. Consequently, concurrent EEG-fNIRS measurements offer a powerful approach for investigating both normal brain function and pathophysiological states, providing a window into the integrity of the neurovascular unit.

Analytical Frameworks for Multimodal Integration

Correlation-Based Fusion Methods

Canonical Correlation Analysis (CCA) and its extensions provide a powerful statistical framework for identifying shared latent variables between high-dimensional EEG and fNIRS datasets. The standard CCA seeks linear combinations of variables from each modality that maximize their cross-correlation. However, traditional CCA performs poorly when the number of features exceeds the number of observations, a common scenario in neuroimaging [24].

To address this limitation, Structured Sparse Multiset CCA (ssmCCA) incorporates sparsity constraints through regularization techniques like the graph-guided fused LASSO. This approach performs feature selection while preserving spatial relationships between brain regions, effectively mitigating overfitting and enhancing the interpretability of results. The ssmCCA algorithm optimizes an objective function that maximizes the correlation between canonical variates from multiple modalities while imposing structured sparsity penalties, resulting in more robust and neurobiologically plausible integrations of electrical and hemodynamic data [24] [15].

Component Analysis and Feature Fusion

Task-Related Component Analysis (TRCA) offers an alternative approach by extracting reproducible components that maximize inter-trial covariance. This method effectively separates task-related neural activity from physiological noise and external disturbances, enhancing the signal-to-noise ratio for both EEG and fNIRS signals [22]. By applying TRCA to each modality independently before investigating their coupling, researchers can achieve more precise characterization of neural patterns underlying specific cognitive or motor tasks.

For classification applications, feature-level fusion strategies such as the improved Normalized-ReliefF algorithm have demonstrated superior performance in distinguishing brain states compared to unimodal approaches. This method normalizes multi-modal features from EEG and fNIRS before performing feature selection and fusion, effectively leveraging the complementary information contained in both modalities to achieve classification accuracies exceeding 98% in some applications [16].

Table 1: Quantitative Findings from Key EEG-fNIRS Studies

Study Focus	Participants	Key Finding	Neurovascular Correlation
Cognitive-Motor Interference [25] [22]	16 healthy adults	Decreased NVC in dual-task vs. single-task in theta, alpha, and beta rhythms	Negative correlation (r = -0.68 to -0.72)
Action Observation Network [24] [15]	21 adults (16 right-handed)	Left inferior parietal lobe activation during execution, observation, and imagery	Maximum cross-correlation: 0.74 (EEG-fNIRS)
Auditory Intensity Processing [21]	33 (Exp 1) & 31 (Exp 2) adults	HbO increased & HbR decreased with auditory intensity in auditory cortex	Spearman's ρ: 0.45 (Left auditory cortex with N1 amplitude)
Music Perception [16]	9 healthy adults	Preferred music evoked stronger prefrontal activation than neutral music	Significant cross-correlation (p<0.01) during preferred music listening

Experimental Evidence and Quantitative Relationships

Cognitive-Motor Interference

Studies investigating cognitive-motor interference (CMI) have provided compelling evidence for task-dependent modulation of NVC. When participants simultaneously perform a motor task (grip force tracking) and a cognitive task (number detection), the divided attention demanded by this dual-task paradigm leads to significantly decreased neurovascular coupling between fNIRS and EEG signals across theta, alpha, and beta frequency bands compared to single-task conditions [25] [22]. This finding suggests that cognitive overload disrupts the normal temporal coordination between neuronal electrical activity and subsequent hemodynamic responses, providing a neural correlate for the behavioral interference effect observed in dual-task performance.

Action Observation Network

Research on the action observation network (AON) utilizing ssmCCA for data fusion has identified the left inferior parietal region as consistently active during motor execution, observation, and imagery. This region showed robust activation patterns detected by both EEG and fNIRS, with cross-correlation values reaching 0.74 between modalities [24] [15]. The multimodal approach demonstrated superior localization of AON activity compared to unimodal analyses, highlighting how integrated EEG-fNIRS can reveal neural correlates that might be missed when using either technique in isolation.

Sensory Processing and Clinical Applications

In auditory processing, studies have demonstrated intensity-dependent amplitude changes where increases in tone intensity produce corresponding enhancements in EEG ERP components (N1, P2) and fNIRS hemodynamic responses (increased HbO, decreased HbR) [21]. Correlation analyses revealed specific relationships between left auditory cortex activity and N1 amplitude, particularly for deoxyhemoglobin concentrations, providing evidence for neurovascular coupling during basic sensory processing.

In clinical populations, retired rugby players with a history of multiple concussions showed blunted hemodynamic responses during a "Where's Wally?" NVC test compared to controls, with significantly smaller increases in HbO in the left middle frontal gyrus (-0.015 ± 0.258 μM vs. -0.160 ± 0.311 μM) [23]. This finding suggests long-term alterations in neurovascular function following repetitive mild traumatic brain injury, demonstrating the clinical relevance of NVC assessment.

Table 2: Hemodynamic and Electrical Response Characteristics

Parameter	EEG Response	fNIRS Response	Temporal Relationship
Primary Signal Origin	Neuronal electrical activity (postsynaptic potentials)	Hemodynamic changes (HbO/HbR)	EEG precedes fNIRS by 2-6 seconds
Temporal Resolution	Millisecond range (high)	~0.1-1.0 second (moderate)	Delay due to hemodynamic response time
Spatial Resolution	~1-3 cm (low)	~1-2 cm (moderate)	Complementary when integrated
Motor Task Activation	Alpha/Beta desynchronization over sensorimotor cortex	HbO increase in contralateral motor cortex	Cross-correlation: r = 0.54-0.82 [19]
Cognitive Task Activation	Frontal theta increase, alpha decrease	Prefrontal HbO increase	Strongest coupling in dorsolateral PFC

Methodological Protocols for Multimodal Assessment

Concurrent EEG-fNIRS Recording Setup

Equipment Configuration: A synchronized EEG-fNIRS system requires careful integration of both modalities. The EEG system typically employs 16-128 electrodes arranged according to the international 10-20 system, while the fNIRS system utilizes optodes positioned over regions of interest (e.g., prefrontal, motor, parietal cortices) with source-detector distances of 25-35 mm to ensure adequate cortical penetration [15] [19]. To minimize interference, fNIRS optodes can be embedded within the EEG cap, with careful attention to ensuring proper scalp contact for both systems. Synchronization pulses should be sent between systems to align data streams with millisecond precision.

Signal Acquisition Parameters: EEG should be recorded with a sampling rate ≥500 Hz to capture relevant neural oscillations, while fNIRS typically acquires data at 10-25 Hz using continuous-wave systems operating at two or more wavelengths (e.g., 695 nm and 830 nm) to distinguish HbO and HbR concentrations [15] [26]. Simultaneous monitoring of systemic physiological parameters (heart rate, blood pressure, respiration) is recommended to account for non-neural influences on hemodynamic signals.

Experimental Paradigms for NVC Assessment

Cognitive-Motor Dual-Task: This protocol involves comparing single tasks (motor-only or cognitive-only) with dual-task conditions. For example, participants may perform: (1) a single motor task (e.g., grip force tracking), (2) a single cognitive task (e.g., number detection), and (3) simultaneous execution of both tasks [25] [22]. Each condition should be presented in randomized blocks with adequate rest periods between trials to allow hemodynamic responses to return to baseline.

Action Observation Paradigm: To investigate the AON, participants complete three conditions: (1) Motor Execution - physically performing an action (e.g., grasping and moving a cup), (2) Motor Observation - watching an experimenter perform the same action, and (3) Motor Imagery - mentally rehearsing the action without movement [15]. Blocks of each condition (e.g., 20-30 seconds) are alternated with baseline rest periods, with auditory cues signaling condition transitions.

Auditory Intensity Processing: This paradigm presents tones of varying intensities (e.g., 70-95 dB) in randomized order, with each tone typically lasting 50-500 ms and separated by variable inter-stimulus intervals [21]. The design should include sufficient trials per intensity level (e.g., 50+ repetitions) to obtain robust ERPs and hemodynamic responses.

Data Processing and Analysis Pipeline

Preprocessing: EEG data should be filtered (e.g., 0.5-40 Hz bandpass), corrected for artifacts (e.g., ocular, muscular), and re-referenced. fNIRS data requires conversion of optical density changes to hemoglobin concentrations using the modified Beer-Lambert law, followed by bandpass filtering (e.g., 0.01-0.2 Hz) to remove physiological noise and drift [26].

NVC Quantification: Time-locked EEG responses (ERPs or time-frequency representations) are correlated with fNIRS hemodynamic responses using cross-correlation analysis, typically with the EEG signal lagging behind the fNIRS signal [21] [19]. Advanced methods like ssmCCA or TRCA can then be applied to extract task-related components and identify maximal correlations between modalities, providing quantitative measures of neurovascular coupling strength.

Diagram 1: Experimental Workflow for EEG-fNIRS NVC Analysis

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Materials for EEG-fNIRS NVC Research

Item	Specification	Function/Purpose
EEG System	16-128 channels; sampling rate ≥500 Hz; compatible with fNIRS integration	Records electrical brain activity with high temporal resolution
fNIRS System	Continuous-wave; ≥2 wavelengths (e.g., 695 & 830 nm); 10-25 Hz sampling	Measures hemodynamic changes via HbO/HbR concentration
Synchronization Module	Hardware/software trigger box with sub-millisecond precision	Aligns EEG and fNIRS data streams for precise temporal correlation
EEG-fNIRS Cap	Integrated cap with openings for EEG electrodes and fNIRS optodes	Ensures stable positioning and co-registration of both modalities
Conductive EEG Gel	Hypoallergenic, chloride-free	Ensures good electrical contact between electrodes and scalp
fNIRS Optode Spacers	Various sizes (e.g., 25-35 mm source-detector distance)	Controls penetration depth and spatial resolution of fNIRS signals
3D Digitizer	Magnetic or optical system (e.g., Polhemus Fastrak)	Records precise 3D positions of EEG electrodes and fNIRS optodes
Physiological Monitor	ECG, respiration, blood pressure, capnography	Records systemic physiological changes that affect fNIRS signals

Diagram 2: Neurovascular Coupling Signaling Pathway

The biological link between EEG and fNIRS signals, established through neurovascular coupling, provides a robust foundation for multimodal investigations of brain function. The integrated analysis of electrical and hemodynamic activity offers complementary insights with high spatiotemporal resolution, enabling more comprehensive characterization of neural processes than either modality alone. The standardized experimental frameworks and analytical approaches outlined in this whitepaper provide researchers with validated methodologies for quantifying NVC across diverse cognitive, motor, and clinical applications. As these multimodal techniques continue to evolve, they hold significant promise for advancing our understanding of brain function in both healthy states and neurological disorders, ultimately contributing to improved diagnostic approaches and therapeutic interventions.

Human neuroscience is undergoing a significant transformation, moving from conventional laboratory settings toward embracing the complexity of natural environments [2]. This shift toward an ecologically valid depiction of human brain function promises new scientific insights into neuronal development, health, and aging, while simultaneously driving innovation in medicine and psychiatry [2]. At the forefront of this transition are two non-invasive neuroimaging techniques: electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). Independentally, each modality offers valuable insights into brain function; however, their integration creates a synergistic platform that overcomes the limitations inherent in either approach alone [27] [13]. This multimodal approach harnesses the complementary strengths of electrophysiological and hemodynamic signals, providing a more comprehensive window into the dynamic interplay of brain networks [2] [28]. The fusion of EEG and fNIRS is particularly compelling due to three interconnected advantages: superior portability for real-world application, significant cost-effectiveness compared to established alternatives, and unparalleled tolerance for naturalistic experimental paradigms. These characteristics collectively position EEG-fNIRS as a transformative tool for both basic neuroscience research and clinical applications, from brain-computer interfaces (BCIs) and neurorehabilitation to drug development and functional assessment in neurological disorders [28] [29].

Core Technical Advantages of EEG and fNIRS

The power of combining EEG and fNIRS stems from their fundamental complementarity. EEG and fNIRS measure distinct yet related physiological phenomena, providing two different perspectives on brain activity that, when combined, offer a more complete picture than either could alone [30] [13].

Table 1: Fundamental Complementary Characteristics of EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity from synchronized neuronal firing [30]	Hemodynamic response (changes in HbO and HbR) via near-infrared light [30] [13]
Signal Source	Post-synaptic potentials of cortical neurons [30]	Changes in cerebral blood flow and oxygenation linked to neural activity [28] [31]
Temporal Resolution	High (milliseconds) [27] [30]	Low (seconds), constrained by hemodynamic response [27] [30]
Spatial Resolution	Low, due to signal dispersion through skull and scalp [27] [30]	Moderate, better than EEG, but limited to cortical surface [27] [30]
Depth of Measurement	Cortical surface [30]	Outer cortex (~1–3 cm deep) [31] [30]

This complementarity directly addresses the core challenge in neuroimaging: no single modality can simultaneously capture brain activity with high spatial and temporal precision. EEG's millisecond-level temporal resolution allows researchers to track the rapid dynamics of brain electrical activity, making it ideal for studying fast cognitive processes, sensory perception, and event-related potentials [30]. Conversely, fNIRS provides superior spatial localization of hemodynamic changes within surface cortical areas, similar to fMRI but with greater portability [27] [31]. Furthermore, because they rely on different physical principles—electricity versus light—the two modalities are largely "blind" to each other's artifacts, making their combined analysis highly complementary for distinguishing neural activity from non-neural physiological noise [2].

Figure 1: The Complementary Data Fusion Workflow. EEG and fNIRS capture different physiological manifestations of the same underlying neural activity. Their fusion compensates for the inherent limitations of each standalone modality, resulting in a more robust and comprehensive spatiotemporal profile of brain function.

Key Advantage 1: Portability and Wearability

The portability of both EEG and fNIRS systems constitutes a primary advantage over bulky, fixed-site neuroimaging technologies like fMRI and PET. Recent technological advancements have led to the development of wearable, fiberless, continuous-wave fNIRS systems and similarly wearable EEG headsets [2] [32]. These systems are increasingly lightweight, wireless, and designed for user comfort, enabling brain imaging outside the static confines of the laboratory [13] [32].

This portability directly facilitates long-term monitoring and studies in ecologically valid settings. For example, researchers can conduct measurements in classrooms, homes, workplaces, or during rehabilitation sessions, capturing brain function in the context of real-life activities [30] [29]. The integration of both modalities into a single, wearable cap system further enhances this potential. While early integration helmets faced challenges with consistent optode placement and comfort, recent approaches utilize 3D printing to create customized headgear or cryogenic thermoplastic sheets that can be molded to an individual's head, improving both signal quality and wearability [27]. This hardware synergy allows for the simultaneous capture of electrical and hemodynamic brain activity while participants engage in a wide range of natural behaviors, from walking and talking to interacting with complex interfaces [29] [32].

Key Advantage 2: Cost-Effectiveness

When establishing a neuroimaging research program or clinical service, the financial considerations are substantial. EEG and fNIRS present a significantly more accessible cost profile compared to high-end modalities like fMRI, MEG, and PET [27] [30].

Table 2: Cost and Accessibility Comparison of Neuroimaging Modalities

Modality	Relative Cost	Infrastructure & Operational Requirements	Notes
fMRI	Very High	Requires magnetic shielding, cryogenic cooling, and dedicated space; immobile [28] [31]	High maintenance and operational costs.
MEG	Very High	Requires magnetically shielded room and specialized infrastructure [27]	Extremely expensive and complex to maintain.
PET	Very High	Requires a cyclotron on-site or nearby for radioisotope production [28] [27]	Involves radioactive tracers, limiting repeated use.
EEG	Low [28] [30]	Minimal infrastructure; portable and can be used in various environments [28] [30]	Lower-cost systems available; minimal consumables.
fNIRS	Moderate [30]	Portable; requires minimal room preparation [31] [30]	Higher than EEG, but far lower than fMRI/MEG/PET.

The lower entry and operational costs of EEG and fNIRS democratize access to functional brain imaging for a wider range of research institutions, clinics, and private enterprises [27]. This affordability enables larger-scale studies, longer monitoring periods, and the deployment of these technologies in resource-constrained settings, including developing countries and smaller medical practices [28]. Furthermore, the cost-effectiveness of a bimodal EEG-fNIRS setup is not merely additive but multiplicative in terms of data value; the investment in two moderately priced systems yields a synergistic data product whose richness often surpasses that of a single, more expensive modality [29] [33].

Key Advantage 3: Naturalistic Application and Motion Tolerance

Perhaps the most transformative advantage of combined EEG-fNIRS is its applicability in naturalistic scenarios and its relative tolerance to motion artifacts. This stands in stark contrast to fMRI, which requires strict participant immobilization, and MEG, which is highly sensitive to head movement [31] [15].

fNIRS is notably more robust to movement artifacts than fMRI and, to a certain extent, EEG [30]. This is because optical signals are less affected by motion-induced electromagnetic fields. While EEG is susceptible to motion artifacts, particularly from muscle activity, advanced processing techniques like Independent Component Analysis (ICA) have proven effective in isolating and removing these artifacts [2] [32]. The simultaneous recording of both signals offers a unique opportunity for cross-modal artifact correction; for instance, fNIRS data can help identify and correct motion-related artifacts in EEG recordings, thereby improving the overall signal quality in dynamic experiments [13].

This motion tolerance unlocks a vast array of previously impossible or highly challenging experimental paradigms:

• Live Motor Tasks: Studying brain activity during actual motor execution, observation, and imagery in face-to-face interactions [15].
• Gait and Rehabilitation: Monitoring cortical activity during robot-assisted gait training (RAGT) and other rehabilitation exercises in stroke and brain-injured patients [28] [32].
• Neuroergonomics: Investigating brain function in real-world settings such as driving simulators, pilot cockpits, and other complex work environments to assess mental workload, drowsiness, and cognitive performance [29].
• Developmental and Pediatric Research: Enabling brain imaging in infants and children, populations that are typically unable to remain still for long periods in a scanner [30] [13].

Experimental Protocols and Validation

The practical validation of the EEG-fNIRS fusion advantage is evident across numerous experimental protocols. A prominent example is the investigation of the Action Observation Network (AON) during motor execution, observation, and imagery.

Protocol: Motor Execution, Observation, and Imagery [15]

• Objective: To elucidate the differences in neural activity across motor execution (ME), motor observation (MO), and motor imagery (MI) using simultaneous fNIRS-EEG.
• Setup: A 24-channel fNIRS system is embedded within a 128-electrode EEG cap. Participants sit face-to-face with an experimenter.
• Task:
  • ME: Participant grasps and moves a cup upon an audio cue.
  • MO: Participant observes the experimenter perform the same action.
  • MI: Participant mentally imagines performing the action without moving.
• Data Fusion & Findings: Unimodal analysis showed activated regions that did not fully overlap between fNIRS and EEG. However, applying a fusion algorithm called structured sparse multiset Canonical Correlation Analysis (ssmCCA) consistently identified activation in the left inferior parietal lobe across all three conditions. This demonstrates the multimodal approach's power to pinpoint shared neural mechanisms that single-modality studies might miss [15].

Another critical application is in the domain of Brain-Computer Interfaces (BCIs). Studies consistently show that hybrid BCIs leveraging both EEG and fNIRS features achieve higher classification accuracy than unimodal systems. For instance, one study on motor imagery and mental arithmetic tasks achieved remarkable accuracy rates of 95.86% and 95.80%, respectively, by using a dual-stream deep learning model (E-FNet) to integrate the modalities, significantly outperforming EEG alone [33]. This performance boost is attributed to the complementary nature of the signals, which enhances the system's robustness and information transfer rate [29] [33].

The Scientist's Toolkit: Essential Research Reagents and Materials

Implementing a successful multimodal EEG-fNIRS research program requires careful selection of hardware, software, and analytical tools.

Table 3: Essential Components for a Multimodal EEG-fNIRS Laboratory

Item / Solution	Function / Description	Examples / Technical Notes
Continuous-Wave (CW) fNIRS System	Measures relative changes in HbO and HbR concentration. Most common, portable, and cost-effective type of fNIRS [13].	Systems like the Cortivision Photon Cap [13]. Can be integrated into EEG caps.
Wearable EEG System	Records electrical brain activity with high temporal resolution.	Systems like the Bitbrain Versatile EEG [13]. High-density systems (e.g., 128-channel) allow for better source localization.
Integrated Acquisition Helmet	Physically co-locates EEG electrodes and fNIRS optodes with stable geometry. Critical for spatial coregistration.	Custom 3D-printed helmets [27], cryogenic thermoplastic sheets [27], or modified elastic EEG caps with pre-defined fNIRS openings [30].
Synchronization Hardware/Software	Ensures precise temporal alignment of EEG and fNIRS data streams.	Use of TTL pulses, parallel ports, or shared acquisition software (e.g., Lab Streaming Layer - LSL) [30] [13].
Data Fusion & Analysis Software	Preprocesses and fuses multimodal datasets for joint analysis.	Algorithms include Joint ICA (jICA), Canonical Correlation Analysis (CCA), and structured sparse multiset CCA (ssmCCA) [30] [15]. Machine learning frameworks (e.g., E-FNet) [33].
3D Digitizer	Records the precise spatial locations of EEG electrodes and fNIRS optodes on the scalp.	Essential for accurate spatial analysis and source modeling (e.g., Fastrak by Polhemus) [15].
Short-Separation Channels	Special fNIRS source-detector pairs with very short distances (<1 cm).	Measures systemic physiological noise from the scalp, which can be regressed out from standard channels to improve brain signal quality [2] [32].

Figure 2: Multimodal Experimental Workflow. A generalized workflow for conducting EEG-fNIRS studies, from hardware setup and synchronized data acquisition through modality-specific preprocessing and final data fusion and interpretation.

The integration of EEG and fNIRS represents a paradigm shift in neuroimaging, moving the field decisively toward a future where high-fidelity brain activity monitoring is possible in the complex, dynamic settings of everyday life. The trio of key advantages—portability, cost-effectiveness, and naturalistic application—makes this multimodal approach uniquely powerful. By harnessing the complementary spatiotemporal profiles and physiological bases of electrical and hemodynamic signals, researchers and clinicians can achieve a more nuanced and comprehensive understanding of brain function and its pathologies. As hardware becomes increasingly wearable and data fusion algorithms more sophisticated, the potential applications in cognitive neuroscience, clinical diagnosis, neurorehabilitation, drug development, and neuroergonomics will continue to expand. The fusion of EEG and fNIRS is not merely a technical achievement but a fundamental enabler of the next generation of brain science, conducted in the real world, for the benefit of real people.

The quest to understand the neural underpinnings of human brain function has been significantly advanced by neuroimaging technologies. Within this landscape, multimodal neuroimaging, particularly the combined use of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), has emerged as a powerful paradigm for probing the brain's dynamic activity across diverse domains from basic motor execution to higher-order cognitive processing [27]. This integration leverages the complementary strengths of each modality: EEG provides millisecond-level temporal resolution of neuro-electrical activity, while fNIRS offers better spatial resolution through hemodynamic measurements of cortical blood flow, linking neural firing to subsequent vascular responses via neurovascular coupling (NVC) [2] [15]. This whitepaper synthesizes current research to elucidate the established neural correlates of motor and cognitive functions revealed through EEG-fNIRS studies, detailing experimental protocols, analytical frameworks, and the core reagents essential for this rapidly advancing field.

Neural Signatures of Motor and Cognitive Processes

Research using simultaneous EEG-fNIRS recording has delineated characteristic neural patterns associated with various motor and cognitive states. The following table summarizes the established neural correlates across different functional domains.

Table 1: Established Neural Correlates in Motor Execution and Cognitive Processing

Functional Domain	EEG Correlates	fNIRS Correlates	Key Brain Regions	Multimodal Fusion Insights
Motor Execution (ME)	Decreased alpha (8-12 Hz) and beta (13-30 Hz) power (ERD) [34] [35]	Increased HbO concentration in contralateral motor areas [15] [36]	Primary Motor Cortex (M1), Premotor Cortex, Supplementary Motor Area (SMA) [15] [36]	Simultaneous electrical activation and hemodynamic response confirm motor network engagement [15]
Motor Imagery (MI)	Decreased alpha and beta power (ERD), similar to execution but with a more distributed pattern [34] [35]	Increased HbO in SMA and prefrontal regions; weaker M1 activation than during ME [3] [15]	SMA, Prefrontal Cortex, Parietal Lobules [34] [15]	Dissociation between primary motor (EEG) and secondary motor (fNIRS) activation highlights shared planning but absent execution [34] [15]
Motor Observation (MO)	Modulations in mu rhythm (8-13 Hz) [15]	Increased HbO in parietal and temporal regions [15]	Superior Temporal Sulcus, Inferior Parietal Lobule [15]	Fused data pinpoints shared hub in the Action Observation Network (AON), e.g., left inferior parietal lobe [15]
Cognitive-Motor Dual-Task	Decreased power in theta, alpha, and beta bands compared to single tasks [22]	Prefrontal cortex (PFC) hyperactivity; disrupted hemodynamic response in motor areas [22]	Prefrontal Cortex, Premotor Cortex [22]	Decreased neurovascular coupling strength indicates divided attention and neural resource competition [22]
Working Memory (n-back)	Frontal midline theta power increase; alpha/beta power decrease [37]	Increased HbO in dorsolateral and frontal polar PFC [37]	Dorsolateral Prefrontal Cortex (DLPFC) [37]	Cross-modal attention models show improved state decoding versus unimodal approaches [37]

Experimental Protocols in Multimodal Research

The Motor Execution, Imagery, and Observation Paradigm

A foundational protocol for investigating the Action Observation Network (AON) involves the simultaneous recording of EEG and fNIRS during three core conditions: Motor Execution (ME), Motor Imagery (MI), and Motor Observation (MO) [15].

• Participant Setup: Participants are fitted with a custom integration cap holding both EEG electrodes (e.g., 128-channel) and fNIRS optodes. Optode positions are digitized relative to cranial landmarks (nasion, inion) for precise co-registration. The participant sits facing an experimenter across a table [15].
• Task Protocol:
  • Motor Execution (ME): An audio cue prompts the participant to grasp, lift, and move a cup approximately two feet toward themselves using their right hand.
  • Motor Imagery (MI): The same audio cue prompts the participant to vividly imagine performing the cup-moving task without any physical movement.
  • Motor Observation (MO): An audio cue signals the participant to watch carefully as the experimenter performs the cup-moving task.
• Data Acquisition: EEG data and fNIRS data (recording changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) at 10 Hz) are recorded simultaneously throughout all conditions. The block includes multiple trials with rest periods [15].
• Analysis: Unimodal analysis (EEG and fNIRS separately) is followed by multimodal fusion using methods like structured sparse multiset Canonical Correlation Analysis (ssmCCA), which identifies brain regions where both electrical and hemodynamic activity are consistently correlated [15].

The Cognitive-Motor Dual-Task Protocol

This protocol is designed to investigate Cognitive-Motor Interference (CMI) by examining the neural effects of performing a cognitive task concurrently with a motor task [22].

• Tasks:
  • Single Motor Task (SMT): Participants perform an isometric grip force tracking task, matching a target force displayed on a screen.
  • Single Cognitive Task (SCT): Participants perform a mental arithmetic task or a number detection task without any motor component.
  • Dual Task (DT): Participants perform the SMT and SCT simultaneously.
• Recording: Bimodal EEG-fNIRS signals are recorded throughout. fNIRS optodes and EEG electrodes are focused over the prefrontal cortex and sensorimotor areas [22].
• Analysis Framework: A key innovation is the use of Task-Related Component Analysis (TRCA). TRCA is applied to both EEG and fNIRS signals to extract components that are maximally reproducible across trials, enhancing the signal-to-noise ratio. The correlation (neurovascular coupling) between these task-related components from EEG and fNIRS is then computed and compared across conditions [22].

Signaling Pathways and Workflows

From Neural Firing to Hemodynamic Response

The following diagram illustrates the neurovascular coupling process, the fundamental physiological link connecting the electrical activity measured by EEG to the hemodynamic response measured by fNIRS.

Multimodal Experimental Workflow

A standardized workflow for a typical simultaneous EEG-fNIRS experiment, from setup to data fusion, is depicted below.

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful execution of multimodal EEG-fNIRS research requires a suite of specialized hardware, software, and analytical tools. The following table details these essential components.

Table 2: Essential Research Reagents and Solutions for EEG-fNIRS Research

Tool Category	Specific Examples	Function & Rationale
Integrated Hardware	Custom integration caps (elastic fabric, 3D-printed, cryogenic thermoplastic) [27]	Provides stable, co-registered placement of EEG electrodes and fNIRS optodes on the scalp, which is critical for data quality and spatial alignment.
fNIRS System	Continuous-wave systems (e.g., Hitachi ETG-4100) with sources at 695 nm & 830 nm [15]	Emits near-infrared light into the scalp and measures intensity changes to calculate concentration changes in oxygenated (HbO) and deoxygenated (HbR) hemoglobin.
EEG System	High-density amplifier systems (e.g., 128-channel Electrical Geodesics) [15]	Measures electrical potentials on the scalp with high temporal resolution to capture neural oscillations (e.g., alpha, beta, theta).
Digitization System	3D magnetic space digitizer (e.g., Polhemus Fastrak) [15]	Records the precise 3D locations of fNIRS optodes and EEG electrodes relative to cranial landmarks, enabling accurate co-registration to brain anatomy.
Artifact Handling	Short-separation fNIRS channels, Independent Component Analysis (ICA) for EEG [2]	Critical for removing confounding signals: short-separation channels regress out scalp blood flow; ICA removes ocular and muscle artifacts from EEG.
Fusion Algorithms	Structured Sparse Multiset CCA (ssmCCA) [15], Task-Related Component Analysis (TRCA) [22]	Advanced data-driven methods to fuse EEG and fNIRS data, identifying latent components that are maximally correlated across modalities and reproducible across trials.
Experimental Control	Presentation or Psychtoolbox (MATLAB)	Software for precise stimulus delivery and synchronization of task events with multimodal brain data acquisition.

The established neural correlates from motor execution to cognitive processing, unveiled through multimodal EEG-fNIRS research, underscore the profound complexity and dynamic nature of human brain function. The convergence of electrophysiological and hemodynamic data provides a more complete and validated picture than any single modality could achieve, confirming shared networks for action simulation, identifying distinct patterns for execution and imagery, and quantifying the neural cost of cognitive-motor interference. The continued refinement of integrated hardware, robust artifact handling techniques, and sophisticated data fusion algorithms will further solidify the role of multimodal neuroimaging. This approach is poised to drive significant advancements in clinical applications, from developing targeted neurorehabilitation strategies for stroke patients to creating more robust brain-computer interfaces, ultimately bridging the gap between laboratory research and real-world brain health monitoring and intervention.

From Data Acquisition to Real-World Applications: Implementing Multimodal fNIRS-EEG Systems

Multimodal neuroimaging that integrates electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a powerful approach for investigating brain function across spatiotemporal scales. EEG captures millisecond-level fluctuations in electrical brain activity, providing excellent temporal resolution, while fNIRS measures hemodynamic changes associated with neural activity with better spatial localization than EEG [28] [2]. This complementary nature makes combined EEG-fNIRS particularly valuable for research areas including brain-computer interfaces (BCI), cognitive neuroscience, neurorehabilitation, and social interaction studies via hyperscanning [2] [38] [39].

The fundamental link between these modalities is neurovascular coupling (NVC), the mechanism that relates transient neural activity to subsequent hemodynamic changes [2]. However, achieving robust hardware integration and precise synchronization presents significant technical challenges that must be addressed to ensure data quality and interpretability. This guide details the current methodologies and protocols for successful EEG-fNIRS integration.

Hardware Integration Approaches

Integrated vs. Separate Systems

Two primary configurations exist for combining EEG and fNIRS equipment: integrated commercial systems and custom separate-system setups.

• Integrated Commercial Systems: These are increasingly available, where EEG electrodes and fNIRS optodes are embedded within a single head cap [2]. Examples include systems from vendors like NIRx and others. This approach offers streamlined setup, inherent synchronization, and minimized physical interference between components.
• Custom Separate-System Setups: Many laboratories continue to integrate separate EEG and fNIRS systems. This offers flexibility in component selection but requires careful attention to cap design, component placement, and synchronization.

Physical Integration and Cap Design

The physical co-location of EEG electrodes and fNIRS optodes on the scalp requires careful planning to minimize cross-talk and signal interference.

• Component Arrangement: The layout must account for the different spatial requirements of each modality. fNIRS requires specific source-detector separations (typically 3-4 cm for long channels measuring cerebral signals, and ~0.8 cm for short channels to regress out superficial physiological noise) [40]. EEG electrode placement typically follows the international 10-20 system or its high-density variants. Figure 1 illustrates a conceptual layout for a combined cap.
• Minimizing Interference: fNIRS optodes must not electrically interfere with EEG electrodes. This often requires using non-metallic, plastic housings for fNIRS optodes. Furthermore, the pressure from EEG electrodes can affect local blood flow, potentially contaminating the fNIRS signal; thus, strategic placement is crucial [2].

Figure 1: Conceptual hardware setup for an integrated EEG-fNIRS system. The diagram shows how both systems connect to a unified cap and are synchronized via a central trigger. Key components include EEG electrodes, fNIRS sources and detectors for long channels (measuring cerebral tissue), and short-separation detectors (for regressing out superficial confounds) [40].

Incorporating Peripheral Physiology

Systemic physiological processes (e.g., cardiac pulsation, respiration, blood pressure changes) are major confounds for fNIRS signals and can also affect EEG [40] [2]. Therefore, modern multimodal setups often extend beyond EEG-fNIRS to include Systemic Physiology Augmented fNIRS (SPA-fNIRS). Dedicated modules, such as the NIRxWINGS2, can be seamlessly added to setups like the NIRSport2 to capture:

• Cardiac Activity: via Electrocardiography (ECG) or Pulse Oximetry (PPG) for heart rate.
• Respiration: via a chest belt sensor.
• Autonomic Arousal: via Electrodermal Activity (EDA) or Skin Conductance.
• Other Signals: Electromyography (EMG) and skin temperature [40].

Integrating these signals allows for more robust denoising of fNIRS data and enables the investigation of rich brain-body dynamics [40].

Synchronization Protocols

Precise temporal alignment of EEG and fNIRS data streams is critical for analyzing their neurovascular relationship. Even millisecond-level drifts can compromise analysis.

Synchronization Methods

Table 1 compares the common methods for synchronizing EEG and fNIRS systems.

Table 1: Comparison of EEG-fNIRS Synchronization Methods

Method	Principle	Typical Accuracy	Ease of Implementation	Best For
Hardware Triggering	A central device (e.g., Arduino, NI DAQ) sends TTL pulses to both systems to mark events or a common clock.	Sub-millisecond to milliseconds	Moderate (requires wiring and programming)	Most experimental studies requiring precise trial-based analysis [18] [38].
Software Synchronization	A software command sent via network (TCP/IP) to both systems to start recording simultaneously.	Milliseconds to tens of milliseconds	Easy (may be built into commercial software)	Resting-state or block-design studies where exact trial onset precision is less critical.
Integrated System Clock	Systems share a common master clock within a single integrated hardware unit.	Highest (sub-millisecond)	Easiest (handled automatically by the manufacturer)	All applications, especially those requiring the highest temporal fidelity. Available in integrated commercial systems [40].
Post Hoc Alignment	Aligning data based on a shared, recorded physiological signal (e.g., ECG, PPG).	Variable (depends on signal)	Difficult, less precise	Typically a backup or validation method rather than a primary synchronization strategy.

The most robust and recommended method for task-based experiments is hardware triggering or using an integrated system clock. For example, in a motor imagery study, a trigger signal should be sent from the stimulus presentation computer to both the EEG amplifier and fNIRS console at the exact moment the "Start Imagining" cue appears on the screen [38].

Data Recording and Workflow

A standardized experimental workflow ensures consistency and data integrity. Figure 2 outlines a generalized protocol for a multimodal experiment.

Figure 2: Generalized workflow for a synchronized EEG-fNIRS experiment. The protocol emphasizes a critical synchronization check before data collection begins. Data from all modalities, including peripheral physiology and the stimulus log, are recorded in parallel and stored for subsequent preprocessing and multimodal fusion analysis.

The Scientist's Toolkit: Research Reagent Solutions

Table 2 details key hardware and software components essential for a multimodal EEG-fNIRS research laboratory.

Table 2: Essential Materials and Equipment for EEG-fNIRS Research

Item	Function	Technical Specification Examples
fNIRS Console	Measures hemodynamic activity by emitting near-infrared light and detecting its attenuation after passing through brain tissue.	NIRSport2 (NIRx), NIRScout (NIRx) [38], MedelOpt+ (BIOPAC) [41]. Systems typically have 8-64 source-detector channels.
EEG Amplifier & Cap	Measures electrical potentials on the scalp surface generated by neuronal activity.	Neuroscan SynAmps2 [38], ActiChamp (Brain Products). 64-channel caps are common for full-brain coverage.
Peripheral Physiology Module	Acquires systemic physiological signals to aid in denoising fNIRS and studying brain-body interaction.	NIRxWINGS2 [40], BIOPAC MP200 with BioNomadix [41]. Measures ECG, PPG, RESP, EDA, EMG.
Synchronization Hardware	Generates and routes TTL pulses to mark event onsets across all recording devices.	Arduino, National Instruments DAQ card, or specialized trigger boxes.
Stimulus Presentation Software	Presents experimental paradigms and sends synchronization triggers.	PsychoPy, E-Prime, Presentation.
Data Acquisition & Synchronization Software	Records and synchronizes data from all modalities.	Native software (e.g., NIRx Aurora [40], Neuroscan Acquire) or custom LabVIEW/Matlab scripts.
Multimodal Data Analysis Suite	Preprocesses, fuses, and analyzes combined EEG-fNIRS data.	MATLAB with toolboxes (Homer2, NIRS Brain AnalyzIR, EEGLAB), MNE-Python, FieldTrip.

Experimental Protocol Example: Motor Imagery

A detailed protocol from a published open dataset illustrates the application of these principles [38].

• Objective: To collect a simultaneous EEG-fNIRS dataset during multiple motor imagery (MI) tasks of different upper-limb joints for BCI applications.
• Participants: 18 right-handed, healthy subjects.
• Equipment:
  • EEG: 64-channel cap with Neuroscan SynAmps2 amplifier, sampled at 1000 Hz. Reference: left mastoid.
  • fNIRS: NIRScout system (NIRx) with 8 sources and 8 detectors placed over the left motor cortex, arranged in a grid (24 channels), sampled at 7.8125 Hz.
  • Synchronization: Trigger signals from the stimulus PC were sent to both acquisition systems.
• Experimental Paradigm:
  • Each trial (18-20 s) began with a fixation cross (2 s).
  • A text and video cue appeared for 2 s, indicating the specific MI task (e.g., hand open/close, wrist flexion).
  • The "Start Imagining" cue was displayed for 4 s, during which subjects performed kinesthetic motor imagery.
  • A "Rest" period of 10-12 s (randomized) followed to allow the hemodynamic response to return to baseline.
• Data Validity: The quality of the multimodal data was confirmed through observed event-related desynchronization in EEG beta rhythms and activation of the motor cortex in fNIRS, consistent with motor imagery [38].

Successful multimodal neuroimaging with EEG and fNIRS hinges on meticulous hardware integration and robust synchronization protocols. The field is moving towards increasingly wearable, integrated systems that facilitate studies in naturalistic environments [2]. The addition of peripheral physiological monitoring is becoming standard practice for improving signal quality and exploring the embodied brain.

Future advancements will rely on the development of more sophisticated data-driven fusion algorithms, such as multilayer network models [42] and symmetric source-decomposition techniques, to fully leverage the complementary information in the concurrent, precisely synchronized electrophysiological and hemodynamic data streams.

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a paradigm shift in multimodal neuroimaging, offering a complementary window into brain dynamics by combining EEG's millisecond-scale temporal resolution with fNIRS's spatially localized hemodynamic response monitoring. This technical guide provides an in-depth examination of three core data fusion strategies—concatenation, model-based, and source-decomposition approaches—framed within the context of advancing EEG-fNIRS research. We detail methodological frameworks, implementation protocols, and analytical considerations for each fusion paradigm, supported by empirical validation methods and practical toolkits for the research community. The fusion of these modalities is particularly valuable for brain-computer interface (BCI) development, neurorehabilitation, and cognitive monitoring applications where both rapid neural dynamics and metabolic processes provide critical insights into brain function.

Multimodal neuroimaging has emerged as a powerful approach to overcome the inherent limitations of individual neuroimaging techniques. EEG measures electrical activity generated by synchronized neuronal firing with exceptional temporal resolution (milliseconds), enabling real-time tracking of neural dynamics. Conversely, fNIRS measures hemodynamic changes associated with neural activity through near-infrared light absorption by hemoglobin, providing better spatial localization and sensitivity to metabolic processes linked to neural function via neurovascular coupling [2]. This complementary nature makes EEG and fNIRS ideal partners for fusion approaches, particularly in naturalistic settings where traditional neuroimaging methods like fMRI face limitations [43].

The fundamental challenge in multimodal fusion stems from the divergent nature of the signals: EEG captures direct neural electrical activity with high temporal precision but limited spatial resolution, while fNIRS measures indirect hemodynamic responses with slower temporal characteristics but improved spatial localization [2]. Furthermore, each modality possesses distinct artifact profiles and physiological confounds—EEG is susceptible to ocular and muscle artifacts, while fNIRS is influenced by systemic physiological processes like blood pressure changes and scalp hemodynamics [2]. Effective fusion strategies must accommodate these differences to extract meaningful neurophysiological relationships.

Data fusion in this context refers to the systematic integration of multiple data sources to produce more consistent, accurate, and useful information than that provided by any individual data source [44]. In EEG-fNIRS research, this encompasses a spectrum of approaches ranging from simple concatenation of features to sophisticated blind source separation techniques that identify latent components representing shared neural processes. The selection of an appropriate fusion strategy depends on multiple factors including research objectives, signal characteristics, and computational resources.

Concatenation-Based Fusion Approaches

Fundamental Principles and Implementation

Concatenation-based fusion, also known as feature-level fusion, involves combining preprocessed features from multiple modalities into a unified representation vector for subsequent analysis. This approach represents one of the most straightforward and widely implemented fusion strategies, particularly in classification contexts such as brain-computer interfaces [45]. The fundamental premise involves transforming raw EEG and fNIRS data into modality-specific features, then concatenating these features into a combined feature vector that preserves information from both sources.

The technical implementation typically follows a structured pipeline. For EEG data, time-domain features (e.g., band power, statistical moments) or frequency-domain features (e.g., power spectral density) are extracted from specific frequency bands of interest (alpha, beta, gamma). For fNIRS, hemoglobin concentration changes (oxygenated, deoxygenated, or total) are calculated using the modified Beer-Lambert law, with subsequent feature extraction focusing on temporal characteristics (e.g., mean, slope, variance) of the hemodynamic response [38]. These feature sets are then normalized to account for inter-modality differences in scale and dynamic range before concatenation into a single feature vector.

A recent implementation in industrial quality monitoring demonstrated the efficacy of this approach, where one-dimensional image data characterized by gray-level co-occurrence matrix parameters were fused with structured process data through vector concatenation [45]. The fused features were then processed using kernel principal component analysis to capture underlying variations, achieving a monitoring accuracy of 98.57% for the training set and 96.67% for the test set, outperforming single-source models [45]. This demonstrates the power of concatenation to leverage complementary information from diverse data sources.

Experimental Protocol and Considerations

Protocol for EEG-fNIRS Concatenation Fusion:

• Data Acquisition: Collect simultaneous EEG and fNIRS recordings using synchronized systems. For EEG, standard pre-processing including filtering (e.g., 0.5-45 Hz), artifact removal (e.g., ocular, muscle), and potentially re-referencing. For fNIRS, convert raw optical densities to hemoglobin concentrations, apply band-pass filtering (e.g., 0.01-0.2 Hz) to remove physiological noise, and motion artifact correction [38].

• Feature Extraction: For EEG, segment data into epochs relative to experimental events and extract features such as band power from relevant frequency bands. For fNIRS, segment the hemodynamic response and extract temporal features including mean amplitude, peak value, time-to-peak, and area under the curve.

• Feature Normalization: Apply z-score normalization or min-max scaling to both feature sets independently to address scale differences between modalities: X_normalized = (X - μ) / σ

• Vector Concatenation: Combine the normalized feature vectors from both modalities into a single feature vector: F_fused = [F_EEG, F_fNIRS]

• Dimensionality Reduction: Apply Principal Component Analysis (PCA) or other dimensionality reduction techniques to address the curse of dimensionality: F_reduced = PCA(F_fused)

• Model Training: Utilize the reduced fused features for classifier training (e.g., Support Vector Machines, Random Forests) or regression analysis.

Table 1: Typical Feature Sets for EEG-fNIRS Concatenation

Modality	Feature Category	Specific Features	Temporal Characteristics
EEG	Spectral Power	Delta (1-4 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz)	Instantaneous (milliseconds)
EEG	Time-Domain	Event-Related Potentials (Amplitude, Latency)	Event-locked (milliseconds)
fNIRS	Hemodynamic	HbO, HbR, HbT concentration changes	Delayed (seconds)
fNIRS	Temporal	Mean, Slope, Variance, Peak value	Slow fluctuations (seconds)

The primary advantage of concatenation-based fusion lies in its simplicity and preservation of modality-specific information. However, this approach faces challenges with high-dimensional feature spaces which can lead to overfitting, particularly with limited training samples. Additionally, the direct concatenation assumes independence between features and may not explicitly model the underlying neurovascular coupling relationship between EEG and fNIRS signals [43].

Model-Based Fusion Approaches

Theoretical Framework and Variants

Model-based fusion approaches employ predefined mathematical models to describe the relationship between neural electrical activity and hemodynamic responses, typically leveraging the neurovascular coupling mechanism. Unlike concatenation methods that simply combine features, model-based approaches impose physiological constraints through their mathematical structure, potentially offering more biologically plausible integrations of EEG and fNIRS data.

The General Linear Model (GLM) framework represents one of the most established model-based approaches, originally developed for fMRI analysis and adapted for fNIRS-EEG fusion. In this framework, the hemodynamic response measured by fNIRS is modeled as a linear combination of regressors derived from EEG features plus error terms. For instance, EEG-derived measures of neural activity (e.g., power in specific frequency bands or event-related potential components) can be convolved with a hemodynamic response function to predict the fNIRS signal [2]. The model can be expressed as:

Y_fNIRS = X_EEG * β + ε

Where Y_fNIRS is the fNIRS hemoglobin concentration, X_EEG is the design matrix containing EEG-derived regressors, β represents the regression coefficients quantifying the relationship strength, and ε is the error term.

Another sophisticated approach involves dynamic causal modeling (DCM) or Bayesian frameworks that model the directed interactions between neural populations and their hemodynamic consequences. These methods allow for testing specific hypotheses about how electrical neural activity gives rise to hemodynamic responses and can incorporate physiological priors to constrain solutions. For naturalistic imaging scenarios where precise stimulus timing is unavailable, model-based approaches face limitations unless combined with data-driven elements [2].

Experimental Implementation

Protocol for GLM-Based EEG-fNIRS Fusion:

• Preprocessing: Independently preprocess EEG and fNIRS data following standard pipelines. For EEG, this includes filtering, artifact removal, and epoching. For fNIRS, process raw light intensity measurements to obtain oxygenated (HbO) and deoxygenated (HbR) hemoglobin concentrations.

• Regressor Construction: Extract relevant features from EEG data such as time-frequency representations or event-related potential amplitudes. Convolve these features with a canonical hemodynamic response function (HRF) to generate predictors for the fNIRS signal.

• Model Specification: Construct the design matrix with EEG-derived regressors, additional confound regressors (e.g., physiological noises, motion artifacts), and constant terms.

• Parameter Estimation: Solve the linear model using ordinary least squares or maximum likelihood estimation to obtain parameter estimates β that quantify the relationship between EEG features and fNIRS responses.

• Model Validation: Assess model goodness-of-fit using metrics like R², F-tests, or cross-validation. Compare with alternative models to determine the most parsimonious explanation for the data.

Table 2: Comparison of Model-Based Fusion Approaches

Method	Theoretical Basis	Key Parameters	Advantages	Limitations
General Linear Model (GLM)	Linear regression	β coefficients, variance explained	Simple implementation, interpretable results	Assumes linear relationship, requires timing information
Dynamic Causal Modeling (DCM)	Bayesian inference, neurovascular coupling	Effective connectivity, hemodynamic parameters	Models directed interactions, biologically plausible	Computationally intensive, complex implementation
Joint Bayesian Framework	Probability theory, Bayesian estimation	Posterior distributions, uncertainty quantification	Naturally handles noise, flexible for multiple data types	Requires prior distributions, computationally demanding

Model-based approaches are particularly valuable when investigating neurovascular coupling mechanisms or when strong a priori hypotheses exist about the relationship between specific neural events and hemodynamic responses. However, they typically require precise knowledge about stimulus timing and may fall short when latent complex physiological relationships and coupling mechanisms are unknown [2]. Furthermore, the assumption of a fixed hemodynamic response function across brain regions and individuals may limit their accuracy in certain applications.

Source-Decomposition Fusion Approaches

Theoretical Foundations

Source-decomposition approaches, particularly those based on blind source separation (BSS) techniques, offer a powerful data-driven alternative for multimodal fusion by identifying latent components that represent shared sources of variance across modalities. These methods do not require explicit models of the neurovascular coupling or precise stimulus timing, making them particularly valuable for exploring complex interactions in naturalistic settings [43].

Independent Component Analysis (ICA) and its multivariate extension Independent Vector Analysis (IVA) represent the most prominent BSS techniques in multimodal neuroimaging. ICA decomposes data from a single modality into statistically independent components, with the underlying assumption that observed signals are linear mixtures of independent sources [46]. The generative model for ICA can be expressed as:

x(v) = As(v)

Where x(v) represents the observed data at sample point v, A is the mixing matrix, and s(v) contains the independent sources. For multimodal fusion, Joint ICA (jICA) extends this concept by assuming that multiple modalities share the same mixing matrix but have modality-specific source maps [46].

IVA further generalizes ICA to multiple datasets by exploiting statistical dependence across datasets, specifically designed to handle grouped data where dependencies exist between corresponding components across modalities [46]. Transposed IVA (tIVA), a more recent extension, enables fusion of datasets with different dimensionalities and properties by applying the decomposition along different dimensions [46]. These approaches can exploit different types of diversity in the data—including non-Gaussianity, sample dependence, and higher-order statistics—to achieve useful decompositions.

Implementation Methodologies

Protocol for Source-Decomposition Fusion Using ICA/IVA:

• Data Preprocessing and Dimension Reduction: Preprocess EEG and fNIRS data according to standard pipelines. Perform Principal Component Analysis (PCA) on each modality separately to reduce dimensionality and determine the signal subspace, while removing Gaussian noise.

• Data Organization and Grouping: Organize the dimension-reduced data into appropriate matrices for decomposition. For IVA, group corresponding components across modalities that should be linked in the decomposition.

• Model Selection and Application: Based on data characteristics and research questions, select an appropriate BSS algorithm (ICA, jICA, or IVA). Apply the chosen algorithm to extract independent components or vectors from the multimodal data.

• Component Identification and Validation: Identify components of interest (neural vs. artifact) based on their temporal, spatial, and spectral characteristics. Validate components through statistical comparisons with task paradigms or between experimental conditions.

• Result Interpretation: Interpret the resulting components in the context of neurophysiology, considering the relationship between electrical and hemodynamic components that are linked in the decomposition.

Table 3: Source-Decomposition Methods for Multimodal Fusion

Method	Model Assumptions	Diversity Exploited	Modality Constraints	Typical Applications
Independent Component Analysis (ICA)	Statistical independence of sources	Non-Gaussianity	Single modality analysis	Artifact removal, feature extraction
Joint ICA (jICA)	Shared mixing matrix across modalities	Statistical independence	Same number of components across modalities	Identifying joint features in multi-modal data
Independent Vector Analysis (IVA)	Component dependence across datasets	Higher-order statistics, sample dependence	Linked components across modalities	Fully multivariate fusion of EEG-fNIRS
Transposed IVA (tIVA)	Generalization of IVA model	All available statistical information	Different dimensionalities accepted	Fusion of heterogeneous data types

Source-decomposition methods are particularly effective for identifying latent neural processes that manifest in both electrical and hemodynamic responses, enabling the discovery of complex neurovascular coupling relationships without strong a priori assumptions [46]. These approaches can simultaneously separate neural signals from various artifacts and confounds that affect both modalities differently, providing a powerful tool for data cleaning and feature extraction in naturalistic environments where experimental control is limited [43].

Comparative Analysis and Application Guidelines

Performance Considerations and Method Selection

Each fusion approach offers distinct advantages and limitations, making them suitable for different research scenarios and questions. The selection of an appropriate fusion strategy depends on multiple factors including research objectives, data characteristics, and computational resources.

Concatenation-based approaches excel in classification tasks where the goal is to maximize discrimination between conditions or groups by leveraging complementary information from both modalities [45]. Their straightforward implementation makes them accessible, but they may not provide direct insights into neurophysiological mechanisms. Model-based approaches are ideal for testing specific hypotheses about neurovascular coupling relationships, particularly in controlled experiments with well-defined timing [2]. Source-decomposition methods offer the greatest flexibility for exploratory analysis in naturalistic settings or when the relationship between modalities is complex and not well understood [46] [43].

Empirical evidence from the literature demonstrates that multimodal fusion generally outperforms unimodal approaches. For instance, in motor imagery classification, fused EEG-fNIRS approaches typically achieve 5-15% higher classification accuracy compared to unimodal methods [38]. Similarly, in clinical applications such as distinguishing Parkinson's disease, fused modalities provide improved discrimination compared to single modalities [43].

Table 4: Decision Matrix for Fusion Method Selection

Research Scenario	Recommended Approach	Rationale	Implementation Considerations
BCI Classification	Concatenation-based	Maximizes feature information for discrimination	Address high dimensionality with regularization
Neurovascular Coupling Investigation	Model-based (GLM/DCM)	Tests specific physiological hypotheses	Requires precise timing information
Naturalistic Paradigms	Source-decomposition (IVA)	No need for timing information; handles unknown relationships	Computationally intensive; requires expertise in interpretation
Artifact Removal	Source-decomposition (ICA)	Identifies and removes artifact components	Careful component selection needed to preserve neural signals
Small Sample Sizes	Model-based	Reduced risk of overfitting compared to concatenation	Simple models preferred with limited data

Experimental Validation and Practical Considerations

Validating fusion results presents unique challenges, particularly for source-decomposition approaches where ground truth is generally unavailable. Multiple validation strategies should be employed, including:

• Task-Related Validation: Assessing whether extracted components or models show expected modulation with experimental conditions or task paradigms.

• Cross-Modal Consistency: Evaluating whether fused results demonstrate spatiotemporal correspondence with known neuroanatomy and neurophysiology.

• Reproducibility Assessment: Testing the stability of results across cross-validation folds, subjects, or sessions.

• Comparison with Established Findings: Relating results to previously established neural signatures or networks identified in the literature.

Practical considerations for implementing fusion analyses include the critical importance of temporal synchronization between modalities, which should be ensured at the hardware level or through post hoc alignment algorithms. Additionally, differences in spatial coverage and resolution must be addressed through appropriate co-registration and interpolation methods. The handling of disparate temporal scales (milliseconds for EEG vs. seconds for fNIRS) represents another key consideration, often addressed through multiscale analysis frameworks or by extracting appropriate temporal features from each modality.

The Scientist's Toolkit

Essential Research Reagents and Computational Tools

Successful implementation of multimodal fusion requires both specialized software tools and methodological resources. The following table summarizes key resources for EEG-fNIRS fusion research:

Table 5: Essential Resources for EEG-fNIRS Fusion Research

Resource Category	Specific Tools/Methods	Function/Purpose	Implementation Notes
Preprocessing Tools	EEGLAB, MNE-Python, NIRS-KIT	Standardized preprocessing pipelines	Handle modality-specific artifacts
Fusion Algorithms	ICA/IVA implementations, GLM frameworks	Implement core fusion methodologies	Customize parameters for specific applications
Visualization Software	Spaco, SpacoR [47]	Enhanced categorical data visualization	Optimize color assignments for spatial data
Validation Datasets	Public multimodal datasets [38]	Method benchmarking and validation	Ensure appropriate task paradigms and data quality
Computational Environments	Python, R, MATLAB	Flexible implementation and customization	Leverage specialized toolboxes and packages

Experimental Workflow Visualization

The following diagram illustrates a generalized workflow for implementing and validating multimodal fusion approaches:

Generalized Workflow for Multimodal Fusion

Emerging Trends and Future Directions

The field of EEG-fNIRS fusion continues to evolve rapidly, with several emerging trends shaping future research directions. Deep learning approaches are increasingly being applied to fusion problems, with architectures specifically designed to handle multimodal data and learn complex nonlinear relationships between electrical and hemodynamic brain activities [43]. The development of wearable, integrated EEG-fNIRS systems is facilitating naturalistic brain imaging outside conventional laboratory settings, creating new opportunities and challenges for fusion methodologies [2].

There is growing interest in adaptive fusion algorithms that can adjust their parameters based on signal quality or task demands, particularly for real-time applications such as brain-computer interfaces [48]. The integration of additional modalities, including physiological measurements (e.g., EKG, respiration) and behavioral data, represents another frontier for multimodal fusion, potentially providing more comprehensive models of brain-body interactions in health and disease.

As these methodologies advance, standardization of validation frameworks and reporting standards will be crucial for comparing results across studies and translating fusion approaches to clinical applications. The creation of shared, annotated datasets with ground truth information [38] will accelerate method development and benchmarking, fostering continued innovation in this rapidly advancing field.

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a transformative approach in clinical neuroimaging, leveraging the complementary strengths of each modality. EEG records the brain's electrical activity with millisecond temporal resolution, providing exquisite detail on neural dynamics, while fNIRS measures hemodynamic changes in cortical blood flow with better spatial specificity than EEG, offering insights into the metabolic demands of active brain regions [2] [49]. This synergy is anchored in neurovascular coupling (NVC), the fundamental physiological process where neural activity triggers subsequent hemodynamic changes [49]. The combination allows researchers to capture a more complete picture of brain function by simultaneously observing the rapid electrical signals and their slower hemodynamic consequences.

Multimodal EEG-fNIRS is particularly valuable in clinical populations due to several practical advantages. Both technologies are non-invasive, portable, and cost-effective compared to modalities like fMRI, and they demonstrate superior tolerance to motion artifacts [2] [31]. This enables brain monitoring in naturalistic settings and with patient groups who cannot remain perfectly still, such as those undergoing motor rehabilitation or children. Furthermore, wearable, fiberless systems have advanced significantly, facilitating continuous brain imaging in everyday environments [2]. These characteristics make EEG-fNIRS an exceptionally powerful tool for investigating neurological and psychiatric disorders, monitoring disease progression, and evaluating the efficacy of therapeutic interventions.

Technical Foundations & Data Fusion Methodologies

Signal Characteristics and Complementarity

The power of multimodal EEG-fNIRS stems from the innate differences in the signals each modality captures, which are linked through neurovascular coupling.

Table 1: Comparative Characteristics of EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (functional Near-Infrared Spectroscopy)
Measured Quantity	Electrical potential from synaptic activity	Concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR)
Temporal Resolution	Excellent (Milliseconds)	Good (Seconds)
Spatial Resolution	Limited (Centimeters)	Better than EEG (1-3 Centimeters)
Portability	High (Wearable systems available)	High (Wearable systems available)
Key Artifacts	Ocular & muscle activity (EOG/EMG)	Systemic physiology (cardiac, respiration), scalp blood flow
Primary Strength	Tracking rapid neural dynamics	Localizing cortical activation

EEG captures postsynaptic potentials from populations of neurons, reflecting synchronous neural firing in real-time. In contrast, fNIRS relies on the absorption properties of near-infrared light by hemoglobin in the blood. When a brain region becomes active, the subsequent increase in cerebral blood flow leads to a rise in oxygenated hemoglobin (HbO) and a slight decrease in deoxygenated hemoglobin (HbR), which fNIRS detects with a delay of several seconds [2] [31]. The relationship between these two signals is governed by neurovascular coupling, making their simultaneous measurement a powerful proxy for investigating underlying brain physiology and its disruptions in clinical conditions [49].

Data Fusion Approaches and Advanced Analysis

Robust machine-learning methods are essential for combining these disparate signals. Fusion strategies can be broadly categorized as follows:

• Data-Level Fusion (Concatenation): This approach involves combining the raw or pre-processed features from EEG and fNIRS into a single, high-dimensional dataset for model input [2]. While straightforward, it requires careful handling of the different temporal scales and signal-to-noise ratios of each modality.
• Model-Based Fusion: These methods use structured frameworks to model the relationship between the modalities. For instance, the EEG–fNIRS Representation learning Model (EFRM) is a novel framework that uses a masked autoencoder (MAE) to extract modality-specific features and contrastive learning to identify a shared representation between EEG and fNIRS in an unsupervised manner [49]. This approach has demonstrated high performance in classification tasks even with few labeled samples.
• Decision-Level Fusion: In this strategy, separate classifiers are trained on EEG and fNIRS data, and their outputs are combined at the final decision stage (e.g., through voting or weighted averaging) [2]. This method is flexible but may fail to capture more complex, latent interactions between the signals.

A significant challenge in data processing is artifact removal. While EEG artifact correction (e.g., for EOG and EMG) is relatively advanced, confounder correction in fNIRS often remains limited to filtering or basic motion artifact removal, with underutilization of auxiliary signals like short-separation channels to filter out scalp blood flow [2]. Overcoming these challenges is an active area of research, crucial for improving the reliability of multimodal findings.

Diagram 1: Neurovascular coupling linking EEG and fNIRS signals.

Clinical Application 1: Stroke Rehabilitation

Motor recovery after stroke is a complex process involving neuroplasticity, where the brain reorganizes its structure and function. Multimodal EEG-fNIRS provides a unique window into these changes, enabling the evaluation of rehabilitation outcomes and the development of targeted therapies.

Experimental Protocols for Motor Rehabilitation

A typical protocol involves patients performing motor tasks, such as finger tapping or grasping, with their affected and unaffected limbs. EEG and fNIRS sensors are placed over the primary motor cortices (M1) of both hemispheres. The paradigm includes:

• Task Blocks: Repeated periods of motor execution (e.g., 20-30 seconds) interspersed with rest blocks of equal or longer duration.
• Baseline Recording: A resting-state period is recorded before the task to establish a baseline for both EEG oscillatory power and hemoglobin concentrations.
• Simultaneous Recording: EEG and fNIRS data are collected concurrently throughout the experiment. fNIRS optodes are often arranged in a high-density grid over the motor and premotor areas to enable better spatial localization, sometimes as high-density diffuse optical tomography (HD-DOT) [2].

Key Findings and Biomarkers

Studies using this protocol have identified reliable biomarkers of recovery. In the affected hemisphere, fNIRS often shows a reduced amplitude and slower rise of the HbO signal during movement of the affected hand compared to the unaffected side. EEG frequently reveals pathological changes in oscillatory patterns, such as a attenuated Event-Related Desynchronization (ERD) in the beta band (13-30 Hz) over the ipsilesional motor cortex, which is associated with motor planning and execution [2] [49]. The correlation between these EEG and fNIRS measures—for instance, between beta-band ERD and the HbO response—serves as an indicator of intact neurovascular coupling in the recovering brain [49]. Furthermore, the laterality of cortical activation, as measured by fNIRS, can shift during recovery, and the restoration of a more normal lateralized pattern is often correlated with better motor outcomes.

Table 2: EEG-fNIRS Biomarkers in Stroke Motor Rehabilitation

Biomarker	Modality	Observation in Stroke	Clinical Correlation
HbO Amplitude	fNIRS	Attenuated in ipsilesional motor cortex during affected hand movement	Indicator of residual cortical metabolic capacity
Beta ERD	EEG	Reduced desynchronization over ipsilesional motor cortex	Impaired motor planning and execution
Inter-hemispheric Balance	fNIRS/EEG	Increased activation in contralesional (unaffected) hemisphere	Often associated with poorer recovery in severe stroke
NVC Correlation	EEG-fNIRS Fusion	Weakened correlation between beta ERD and HbO response	Disrupted neurovascular coupling

Clinical Application 2: Epilepsy Monitoring

Epimultimodal EEG-fNIRS has emerged as a critical tool for localizing epileptogenic foci and understanding the hemodynamic changes that accompany epileptic activity, particularly in cases where EEG alone is inconclusive.

Experimental Protocols for Seizure Monitoring

Monitoring protocols are designed to capture both interictal (between seizures) and ictal (during a seizure) activity. Patients are monitored in controlled clinical settings or long-term monitoring units.

• Longitudinal Recording: Continuous or prolonged simultaneous EEG-fNIRS recording is performed over hours or days to capture spontaneous seizure events.
• Provocation Paradigms: In some cases, protocols may include phiotic stimulation or hyperventilation to provoke epileptiform activity, though this is done under strict medical supervision.
• High-Density Setup: A dense array of EEG electrodes is used for precise electrical localization. fNIRS optodes are arranged to provide coverage over the suspected brain regions, often the temporal or frontal lobes, to map the accompanying hemodynamic changes.

Key Findings and Diagnostic Utility

The combination of modalities provides a more complete pathophysiological picture. The high temporal resolution of EEG pinpoints the exact onset of an epileptic spike or seizure, while fNIRS reveals the hemodynamic response that follows. A typical finding is a regional increase in HbO and a decrease in HbR in the epileptogenic zone during an ictal event, reflecting the intense metabolic demand and blood flow increase associated with the seizure [31]. This "hemodynamic focus" can be precisely time-locked to the electrical onset from the EEG. This is especially valuable for localizing the seizure onset zone in patients with drug-resistant epilepsy being evaluated for surgical resection. Furthermore, fNIRS can help distinguish epileptic seizures from non-epileptic events that may appear similar on surface EEG. The lack of a characteristic hemodynamic response in the cortex can rule out an epileptic origin.

Diagram 2: Workflow for epilepsy focus localization with EEG-fNIRS.

Clinical Application 3: Neurological and Psychiatric Disorder Classification

Beyond stroke and epilepsy, multimodal EEG-fNIRS shows significant promise in classifying a wider range of brain disorders, including Alzheimer's disease, schizophrenia, and depression, by identifying unique neural "fingerprints."

Experimental Protocols for Disorder Classification

Protocols for classification often use standardized cognitive tasks to probe the neural networks implicated in specific disorders.

• Resting-State: The patient simply rests with eyes open or closed, allowing measurement of intrinsic brain network connectivity and baseline oscillatory/hemodynamic activity.
• Cognitive Tasks: Tasks such as the N-back test (working memory), Stroop test (executive function), or emotional recognition paradigms are used. For example, in a study on cognitive motivation and memory, participants viewed images and decided whether to remember them while EEG-fNIRS was recorded [50].
• Passive Stimulation: Auditory or visual stimuli may be presented to probe sensory processing deficits.

Key Findings and Classification Performance

Machine learning models trained on fused EEG-fNIRS features have demonstrated superior classification accuracy compared to single-modality approaches.

• Alzheimer's Disease (AD): Studies have shown reduced functional connectivity in the default mode network from resting-state fNIRS, combined with a slowing of the EEG background rhythm (increased theta power, decreased beta power). A bimodal deep learning model (CNN-GRU) integrating these features has been shown to improve the classification accuracy between AD, Mild Cognitive Impairment (MCI), and healthy controls [49] [51].
• Schizophrenia: During working memory tasks, patients often show hypoactivation of the prefrontal cortex (PFC) as measured by fNIRS (reduced HbO), alongside EEG gamma-band abnormalities that reflect impaired neural synchronization. The EFRM model, which learns shared and modality-specific representations, can leverage these complementary features for high-accuracy classification even with limited labeled data [49].
• Depression: A common finding is lateralized prefrontal dysfunction, with some studies reporting reduced activity in the left PFC (measured by fNIRS) and altered alpha-asymmetry patterns in EEG. These features can serve as potential biomarkers for diagnosing major depressive disorder and monitoring response to treatment.

Table 3: EEG-fNIRS Features for Disorder Classification

Disorder	EEG Features	fNIRS Features	Fusion Model Performance
Alzheimer's Disease	Slowing (↑Theta, ↓Beta), ↓Functional Connectivity	↓Prefrontal HbO during cognition, ↓Resting-state connectivity	Bimodal CNN-GRU outperforms single-modality in AD vs MCI vs HC classification [49]
Schizophrenia	↓Gamma band synchronization, Auditory ERPs	↓Prefrontal HbO during working memory tasks	EFRM model achieves high accuracy with few labeled samples by leveraging shared domain [49]
Depression	Frontal Alpha Asymmetry	Lateralized ↓Prefrontal HbO (e.g., left PFC)	Multimodal feature concatenation improves classification accuracy vs. EEG or fNIRS alone [2]
Motor Imagery (BCI)	Sensorimotor Rhythms (ERD/ERS)	Prefrontal & Motor Cortex HbO changes	Early fusion models minimize data loss, enhancing BCI control accuracy [49] [52]

The Scientist's Toolkit: Essential Research Reagents & Materials

Successful execution of multimodal EEG-fNIRS research requires a suite of reliable hardware, software, and methodological components.

Table 4: Essential Reagents and Tools for Multimodal EEG-fNIRS Research

Item	Function/Description	Example/Note
Wearable EEG System	Records electrical brain activity. High-density systems improve source localization.	Systems with 32+ channels; often integrated with fNIRS in hybrid caps [2].
Wearable fNIRS System	Measures cortical hemodynamics via near-infrared light. HD-DOT arrays improve spatial resolution.	Continuous-wave (CW) systems are common; portable, fiberless devices enable naturalistic studies [2] [31].
Integrated EEG-fNIRS Caps	Holds electrodes and optodes in stable, co-registered positions.	Customizable caps based on international 10-20 system; crucial for signal quality [2].
Short-Separation Channels	fNIRS source-detector pairs with very short distances (<1 cm).	Measures systemic artifacts from scalp; used as regressors to clean cerebral fNIRS signal [2].
MRI-Compatible fNIRS	Allows for simultaneous fMRI-fNIRS acquisition.	Used for spatial validation of fNIRS signals and studying deep brain structures [31].
Data Fusion Software Platform	Pre-processes, synchronizes, and fuses multimodal data.	Custom scripts (Python, R) or platforms like MNE-Python, Homer2, NIRS-KIT.
Representation Learning Model (EFRM)	Self-supervised model for learning shared & specific features from unlabeled data.	Enables high classification performance with minimal labeled data [49].
Synthetic Multimodal Datasets	Publicly available or computer-generated benchmark datasets.	Addresses data scarcity; used for method validation and development [2].

Visualization Standards and Reporting Guidelines

Effective visualization of multimodal EEG-fNIRS data is critical for accurate interpretation and communication of findings. Adherence to the following standards is recommended:

• Color Palette Selection: Avoid the misuse of the "rainbow" color scheme, which can introduce perceptual distortions and lead to biased interpretation of neurophysiological data [52]. Instead, use perceptually uniform colormaps (e.g., viridis, magma) for representing continuous data like time-frequency power or HbO concentration. For categorical data (e.g., different patient groups), use a limited palette of easily distinguishable colors (max 5-7) [51].
• Spatially-Aware Colorization: When visualizing spatial data (e.g., activation maps on a brain template), employ protocols like Spaco (Spatially Aware Colorization) that calculate the degree of interlacement (DOI) between categories and align it with a color contrast matrix. This ensures that neighboring spatial categories (e.g., adjacent brain regions) are assigned contrasting colors, avoiding perceptual ambiguity [47].
• Accessibility: Always select color combinations with sufficient contrast and verify that they are distinguishable to individuals with color vision deficiencies (e.g., avoid red-green contrasts). Use online contrast checkers and simulators during the design phase [53] [51].

Multimodal EEG-fNIRS has firmly established its clinical value in providing a comprehensive, portable, and robust window into brain function across a spectrum of neurological and psychiatric conditions. The synergy between EEG's temporal precision and fNIRS's spatial and metabolic information, grounded in the principle of neurovascular coupling, offers biomarkers for diagnosis, monitoring, and predicting recovery in stroke, epilepsy, and other brain disorders.

Future advancements in this field will be driven by several key developments. Hardware innovation will focus on more compact, fully integrated, and MRI-compatible systems that facilitate easier data acquisition in diverse environments [31]. Methodological progress will involve the creation of standardized data fusion protocols and the wider adoption of powerful unsupervised and self-supervised learning models, like EFRM, to overcome the challenge of limited labeled clinical data [2] [49]. Finally, a push toward large-scale, publicly available datasets and validation studies in more naturalistic clinical settings will be crucial for translating these promising research tools into standardized clinical practice, ultimately improving patient diagnosis and care.

Brain-Computer Interfaces (BCIs) represent a transformative technological frontier, establishing a direct communication pathway between the brain and external devices [54]. By translating neuronal information into commands capable of controlling external software or hardware, BCIs bypass conventional neuromuscular pathways, offering revolutionary potential for restoring communication and control to individuals with severe motor disabilities [55] [56]. This whitepaper explores the core principles of BCI technology, with particular emphasis on the integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) as a multimodal approach that provides unprecedented insights into brain function for enhancing communication and control systems. Framed within a broader thesis on multimodal neuroimaging, this technical guide examines the operational principles, current applications, experimental methodologies, and future directions of BCIs, providing researchers and clinicians with a comprehensive resource for advancing this rapidly evolving field.

The fundamental objective of a BCI is to measure and analyze specific features of brain activity that reflect the user's intent and to translate these features in real-time into device control commands [56]. This process creates a closed-loop control system where the user's brain signals serve as the input, and the BCI system provides feedback, enabling the user to learn and optimize their brain signal generation to achieve more accurate control. For individuals with conditions such as amyotrophic lateral sclerosis (ALS), brainstem stroke, or locked-in syndrome, BCIs offer the potential to restore basic communication capabilities, significantly improving quality of life and independence [57] [56].

Neurophysiological Fundamentals of BCI Signals

BCI systems operate by detecting and interpreting various types of brain signals, each with distinct characteristics, advantages, and limitations. These signals can be broadly categorized into electrophysiological and hemodynamic responses, reflecting different aspects of neural activity with complementary temporal and spatial properties [56].

Electrophysiological signals, measured primarily through EEG, capture the electrical activity generated by synchronized firing of cortical neurons, particularly pyramidal cells aligned perpendicular to the scalp [58]. EEG provides direct measurement of neural electrical potentials with exceptional temporal resolution on the millisecond scale, making it ideal for tracking rapid cognitive processes and developing responsive BCIs [58] [27]. However, EEG signals suffer from limited spatial resolution due to the blurring effect of the skull and scalp, making precise localization of neural activity challenging.

Hemodynamic responses, measured through fNIRS, monitor changes in cerebral blood oxygenation associated with neural activity through neurovascular coupling [58] [27]. fNIRS uses near-infrared light to measure concentration changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) in the outer layers of the cortex [58]. While fNIRS offers better spatial resolution than EEG for surface cortical areas, its temporal resolution is limited to seconds due to the slow nature of the hemodynamic response [58].

Table 1: Comparison of Primary Non-Invasive Neuroimaging Modalities for BCI Applications

Feature	EEG	fNIRS	fMRI
What It Measures	Electrical activity of neurons	Hemodynamic response (blood oxygenation)	Blood oxygenation level dependent (BOLD) signal
Temporal Resolution	High (milliseconds)	Low (seconds)	Very low (seconds)
Spatial Resolution	Low (centimeter-level)	Moderate (better than EEG)	High (millimeter-level)
Depth of Measurement	Cortical surface	Outer cortex (~1–2.5 cm deep)	Whole brain
Portability	High (wearable systems available)	High (portable and wearable formats)	Low (requires fixed scanner)
Sensitivity to Motion	High – susceptible to movement artifacts	Low – more tolerant to subject movement	Very high – requires immobilization
Best Use Cases	Fast cognitive tasks, ERP studies, real-time BCI	Naturalistic studies, child development, sustained cognitive states	Precise spatial localization, deep brain structures

Multimodal Integration of EEG and fNIRS

The integration of EEG and fNIRS into a dual-modality imaging system represents a significant advancement in neuroimaging, overcoming the limitations of individual techniques while substantially enhancing neural signal detection precision [27]. This multimodal approach capitalizes on the complementary strengths of each modality: EEG provides exceptional temporal resolution for capturing rapid neural dynamics, while fNIRS offers superior spatial localization for cortical areas [58] [27]. Together, they provide a more comprehensive picture of brain function by simultaneously capturing electrical activity and hemodynamic responses.

Technical implementation of simultaneous EEG-fNIRS recording requires careful consideration of hardware integration and signal synchronization. There are two primary methods for integrating these modalities: (1) using separate acquisition systems synchronized via a host computer, and (2) employing a unified processor to simultaneously acquire and process both signals [27]. While the first approach is simpler to implement, the second method achieves more precise synchronization and is more widely used for concurrent fNIRS-EEG detection [27].

The design of joint acquisition helmets is particularly important for successful multimodal integration. Current approaches include integrating EEG electrodes and fNIRS probes on a shared substrate material, arranging them separately while maintaining spatial co-registration, or directly integrating fNIRS fiber optics into existing EEG caps [27]. Customized helmets using 3D printing or cryogenic thermoplastic sheets have shown promise in addressing individual head-size variations and ensuring consistent probe-to-scalp contact pressure [27].

Figure 1: Framework for simultaneous fNIRS-EEG multimodal setup

Data fusion techniques for integrating EEG and fNIRS signals include joint Independent Component Analysis (jICA), canonical correlation analysis (CCA), and structured sparse multiset CCA (ssmCCA) [58] [15]. These methods enable researchers to identify relationships between the electrophysiological and hemodynamic data, providing insights into neurovascular coupling and generating more robust features for BCI classification [15] [42]. The ssmCCA approach, in particular, has demonstrated effectiveness in fusing electrical and hemodynamic responses to pinpoint brain regions consistently detected by both modalities [15].

Current Clinical Applications and Research

Communication Restoration

Significant advances in BCI technology have focused on restoring communication abilities for individuals with severe motor impairments. Invasive BCIs utilizing microelectrode arrays implanted directly in the motor cortex have demonstrated remarkable success in decoding attempted speech with high accuracy [59] [60]. Recent research has progressed from decoding attempted speech movements to investigating inner speech (also called inner monologue), which involves imagining speech without any physical movement [60].

Stanford University researchers have developed a BCI that detects inner speech from speech-impaired patients, representing a crucial step toward restoring rapid communication [60]. This approach offers potential advantages for users with paralysis, as attempting physical speech can be slow and fatiguing. The system uses machine learning algorithms trained to recognize repeatable patterns of neural activity associated with phonemes—the smallest units of speech—which are then stitched into complete sentences [60].

Several neurotechnology companies are advancing clinical trials of implanted BCI systems for communication restoration. Paradromics has received FDA approval for a first long-term clinical trial of its BCI device, which features an active area of roughly 7.5 millimeters in diameter with thin, stiff, platinum-iridium electrodes that penetrate the surface of the cerebral cortex [59]. The initial trial will involve implanting electrode arrays in the area of the motor cortex that controls articulatory movements (lips, tongue, and larynx) to restore communication through real-time speech synthesis [59]. Similarly, Johns Hopkins University is conducting clinical trials of the CortiCom system, which consists of up to 128 electrodes surgically implanted on the surface of the brain to help improve communication for patients with muscular weakness from ALS, brainstem stroke, and other causes [57].

Motor Rehabilitation

BCI technology shows considerable promise in motor rehabilitation for patients with neurological injuries or diseases. Multimodal fNIRS-EEG systems have been particularly valuable in studying the Action Observation Network (AON)—a shared neural network recruited during motor execution, observation, and imagery [15]. Understanding the neural mechanisms underlying these cognitive-motor processes is essential for developing effective rehabilitation strategies.

Research using simultaneous fNIRS-EEG recordings has elucidated differences in neural activity across motor execution (ME), motor observation (MO), and motor imagery (MI) conditions [15]. Unimodal analyses revealed differentiated activation between conditions, with fNIRS identifying activation in the left angular gyrus, right supramarginal gyrus, and right superior and inferior parietal lobes, while EEG detected activity in bilateral central, right frontal, and parietal regions [15]. However, when using fused fNIRS-EEG data with ssmCCA, researchers consistently found activation over the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus during all three conditions, suggesting that this multimodal approach identifies a shared neural region associated with the AON [15].

Table 2: BCI Approaches for Communication and Motor Rehabilitation

Application	BCI Type	Neural Signals	Target Population	Performance Metrics
Speech Restoration	Invasive (intracortical)	Single-neuron activity from speech motor cortex	ALS, brainstem stroke, locked-in syndrome	Decoding phonemes and sentences from neural patterns [59] [60]
Inner Speech Decoding	Invasive (microelectrode arrays)	Neural activity during imagined speech	Severe paralysis with speech impairments	Differentiation of intended speech without physical movement [60]
Motor Rehabilitation	Non-invasive (EEG-fNIRS)	Hemodynamic and electrical activity from AON	Stroke, spinal cord injury, motor disorders	Activation of shared neural regions during ME, MO, MI [15]
Environmental Control	Non-invasive (EEG)	Event-related potentials, sensorimotor rhythms	Severe neuromuscular disabilities	Control of computer cursor, smart devices [56]

Experimental Protocols and Methodologies

Protocol for Multimodal fNIRS-EEG Motor Study

A representative experimental protocol for studying motor cognition using simultaneous fNIRS-EEG recording involves three primary conditions: motor execution (ME), motor observation (MO), and motor imagery (MI) [15]. This protocol utilizes a live-action paradigm where participants and experimenters perform actions in front of each other, creating an ecologically valid context for studying AON activity.

Equipment Setup:

• A 24-channel continuous-wave fNIRS system (e.g., Hitachi ETG-4100) measuring changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) concentration at two wavelengths (695 nm and 830 nm) with a sampling rate of 10 Hz.
• A 128-electrode EEG system (e.g., Electrical Geodesics) embedded within an elastic cap, with the fNIRS probe integrated into the same cap.
• Proper digitization of fNIRS optodes using a 3D-magnetic space digitizer (e.g., Fastrak, Polhemus) to account for variance in cap positioning and inter-subject sizing differences.

Experimental Procedure:

• Participants sit face-to-face with an experimenter across a table.
• After equipment setup and explanation of procedures, participants practice for approximately 5 minutes.
• During ME condition, a pre-recorded audio command prompts the participant to grasp, lift, and move a cup approximately two feet toward themselves using their right hand.
• During MO condition, an audio cue prompts the participant to watch the experimenter perform the same cup-moving action.
• During MI condition, participants imagine themselves performing the cup-moving action following an audio cue without any physical movement.
• Conditions are presented in randomized or counterbalanced order with adequate rest periods between trials.

Data Processing Pipeline:

• Preprocessing: Separate preprocessing pipelines for fNIRS and EEG data, including filtering, artifact removal, and motion correction.
• Feature Extraction: For fNIRS—calculation of HbO and HbR concentration changes; for EEG—extraction of time-frequency features and event-related potentials.
• Data Fusion: Application of structured sparse multiset CCA (ssmCCA) to identify relationships between fNIRS and EEG data and pinpoint consistently activated brain regions.

Figure 2: Experimental workflow for multimodal fNIRS-EEG study

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Materials and Equipment for Multimodal fNIRS-EEG Research

Item	Function	Specifications/Examples
fNIRS System	Measures hemodynamic responses by detecting changes in hemoglobin concentrations using near-infrared light.	Continuous-wave systems (e.g., Hitachi ETG-4100); measures HbO and HbR at 695 nm and 830 nm wavelengths [15].
EEG System	Records electrical activity from the scalp with high temporal resolution.	High-density systems (e.g., 128-channel Electrical Geodesics); captures millisecond-level neural dynamics [58] [15].
Integrated Cap	Provides stable platform for co-registered fNIRS optodes and EEG electrodes.	Elastic fabric caps with predefined openings; 3D-printed customized helmets; cryogenic thermoplastic sheets [27].
3D Digitizer	Records precise spatial locations of fNIRS optodes and EEG electrodes on head.	Magnetic space digitizers (e.g., Fastrak, Polhemus); essential for accurate spatial localization and co-registration [15].
Synchronization Hardware	Ensures temporal alignment of fNIRS and EEG data streams.	Unified processors; external trigger systems (TTL pulses); shared clock systems for precise synchronization [27].
Data Fusion Software	Implements algorithms for integrating fNIRS and EEG data.	Custom MATLAB/Python scripts for ssmCCA, jICA, canonical correlation analysis [15] [42].

Future Directions and Challenges

The future of BCI technology lies in addressing current limitations while advancing toward more sophisticated, reliable, and accessible systems. Key areas of development include the creation of fully implantable, wireless systems that eliminate external hardware and reduce infection risk [59] [60]. Several companies are working on such hardware, with availability expected within the next few years [60]. Additional advancements focus on improving signal quality through higher electrode density and more biocompatible materials, such as flexible neural interfaces that minimize tissue damage and improve long-term recording stability [55] [54].

The integration of artificial intelligence and machine learning represents another critical frontier for BCI development. Advanced decoding algorithms are essential for translating complex neural patterns into accurate device control, particularly for challenging applications like inner speech decoding [54] [60]. Researchers are also exploring brain regions outside the motor cortex that might contain higher-fidelity information for specific applications, such as regions traditionally associated with language or hearing for speech BCIs [60].

From a clinical perspective, the development of personalized digital prescription systems that deliver customized therapeutic strategies via digital platforms holds significant promise for optimizing rehabilitation outcomes [54]. These systems would adapt BCI parameters and training protocols based on individual user characteristics and progress, creating tailored interventions for maximum efficacy.

Despite these promising directions, significant challenges remain. For invasive BCIs, issues of long-term stability, tissue response, and signal consistency over time require continued attention [59] [56]. For non-invasive multimodal systems, technical challenges include further improving hardware integration, reducing costs, enhancing real-time monitoring capabilities, and developing more sophisticated data fusion techniques [27]. Additionally, ethical considerations surrounding neural privacy, agency, and equitable access must be proactively addressed as BCI technology advances [54] [60].

Brain-Computer Interfaces represent a rapidly advancing field with transformative potential for enhancing communication and control systems, particularly for individuals with severe neuromuscular disabilities. The integration of multimodal neuroimaging approaches, specifically the combination of EEG and fNIRS, provides a powerful framework for overcoming the limitations of individual techniques while gaining a more comprehensive understanding of brain function. This technical guide has outlined the fundamental principles, current applications, methodological protocols, and future directions of BCI technology, with particular emphasis on its role in communication restoration and motor rehabilitation.

As research progresses, continued interdisciplinary collaboration among neuroscientists, engineers, clinicians, and computer scientists will be essential for addressing technical challenges and translating laboratory advances into clinically viable solutions. The development of more sophisticated implantable systems, advanced decoding algorithms, and personalized approaches will further expand the capabilities and applications of BCI technology. Through these concerted efforts, BCIs hold the promise of restoring communication and control to those with severe disabilities, ultimately enhancing independence and quality of life for countless individuals worldwide.

Biomarkers are defined, measurable characteristics of normal biological processes, pathogenic processes, or responses to an exposure or intervention, including therapeutic interventions. In drug development, they serve as vital tools for diagnosing diseases, assessing patient risk, monitoring disease status, watching for treatment side effects, and measuring patient response to drugs [61]. The appropriate application of validated biomarkers benefits drug development and regulatory assessments by providing a window into the body's inner workings that complements clinical assessments of how a patient feels, functions, or survives [62] [61].

The U.S. Food and Drug Administration (FDA) categorizes biomarkers based on their specific application through a formal Context of Use (COU) definition, which describes the biomarker's specified purpose in drug development [62]. The Biomarkers, EndpointS, and other Tools (BEST) resource establishes standardized categories that include susceptibility/risk, diagnostic, monitoring, prognostic, predictive, pharmacodynamic/response, and safety biomarkers [62]. A single biomarker may fulfill multiple roles depending on its application; for instance, Hemoglobin A1c serves both to diagnose diabetes (diagnostic) and monitor long-term glycemic control (response biomarker) [62].

For central nervous system (CNS) disorders, biomarkers face particular challenges due to the complexity of the brain and limitations in accessing neural tissue. Traditional drug development for CNS applications has been costly and fraught with high failure rates [63]. The emergence of neurophysiological and neuroimaging biomarkers, particularly electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), offers promising tools for bridging the gap between preclinical and clinical stages in translational neuroscience [63] [64]. These non-invasive measurement techniques provide objective insights into brain activity and drug effects, enabling researchers to track changes in neural circuits and detect early responses to experimental drugs [63].

Biomarker Validation and Regulatory Framework

Fit-for-Purpose Validation

Biomarker validation follows a "fit-for-purpose" approach, meaning the level of evidence needed depends on the specific context of use and intended application [62]. The validation process addresses both analytical validation (assessing the performance characteristics of the measurement tool) and clinical validation (demonstrating that the biomarker accurately identifies or predicts the clinical outcome of interest) [62]. This tailored approach ensures rigorous assessment of each biomarker type according to its specific role in drug development or clinical decision-making.

• Analytical Validation: Includes assessment of accuracy, precision, analytical sensitivity, analytical specificity, reportable range, and reference range. The appropriate performance characteristics depend on the method of detection and the analyte of interest [62].
• Clinical Validation: Involves demonstrating the biomarker accurately identifies or predicts the clinical outcome through assessment of sensitivity and specificity, determination of positive and negative predictive values, and evaluation of performance in the intended population [62].

Different biomarker categories require distinct validation approaches. Diagnostic biomarkers prioritize sensitivity and/or specificity and require proof of accurate disease identification across diverse populations. Prognostic biomarkers require robust clinical data showing consistent correlation with disease outcomes, while predictive biomarkers prioritize sensitivity, specificity, and often causality with emphasis on a mechanistic link to treatment response [62]. Pharmacodynamic/response biomarkers require evidence of a direct relationship between drug action and biomarker changes, and safety biomarkers need to demonstrate consistent indication of potential adverse effects across different populations and drug classes [62].

Regulatory Pathways

The FDA provides several pathways for regulatory acceptance of biomarkers, with the optimal approach depending on the specific circumstances of development [62]:

• Early Engagement: Developers can engage with the FDA early in the drug development process to discuss biomarker validation plans via paths such as Critical Path Innovation Meetings (CPIM) or the pre-Investigational New Drug (IND) process [62].
• IND Process: Drug developers can pursue clinical validation and regulatory acceptance of biomarkers within the context of specific drug development programs through the IND application process [62].
• Biomarker Qualification Program (BQP): This program provides a structured framework for the development and regulatory acceptance of biomarkers for a specific COU, involving three stages: Letter of Intent, Qualification Plan, and Full Qualification Package [62]. Once qualified, a biomarker can be used by any drug developer without requiring FDA re-review, provided it is used within the specified COU [62].

Despite these pathways, the BQP has faced challenges with slow progress and lengthy review timelines. As of 2025, the FDA had only qualified eight biomarkers through the BQP, with most qualified prior to the 21st Century Cures Act's enactment in 2016 [61]. This has led sponsors to explore alternative pathways, such as "collaborative group interactions" for biomarker acceptance [61].

Table 1: Biomarker Categories and Examples in Drug Development

Biomarker Category	Primary Use in Drug Development	Examples
Susceptibility/Risk	Identify individuals with increased disease risk	BRCA1/BRCA2 mutations for breast/ovarian cancer [62]
Diagnostic	Identify presence or subtype of a disease	Hemoglobin A1c for diabetes mellitus [62]
Prognostic	Identify likelihood of disease recurrence or progression	Total kidney volume for autosomal dominant polycystic kidney disease [62]
Monitoring	Assess disease status or evidence of exposure to a medical product	HCV RNA viral load for Hepatitis C infection [62]
Predictive	Identify individuals more likely to respond to a specific treatment	EGFR mutation status in nonsmall cell lung cancer [62]
Pharmacodynamic/Response	Show biological response to a therapeutic intervention	HIV RNA (viral load) in HIV treatment [62]
Safety	Monitor for potential adverse effects during treatment	Serum creatinine for acute kidney injury [62]

Neuroimaging Biomarkers: EEG and fNIRS

Electroencephalography (EEG) Biomarkers

EEG is a common, widely used neuroimaging technique that records electrical activity in the brain with excellent temporal resolution at relatively high sampling rates [65]. As a non-invasive, cost-effective, and portable technology, EEG has emerged as a powerful tool for CNS drug discovery, particularly through the development of translational EEG biomarkers that bridge preclinical and clinical research [63].

In drug development, EEG biomarkers serve three critical functions [63]:

• Screen New Targets: EEG biomarkers can screen potential drug targets in model systems by assessing their effects on brain activity, enabling early identification of promising candidates.
• Confirm Target Engagement: In early-phase clinical trials, EEG biomarkers provide objective measurement of target engagement, delivering critical data on experimental drug efficacy.
• Track Functional Outcomes: EEG biomarkers track functional outcomes by assessing how drugs affect brain function and monitoring patient responses over time.

EEG biomarkers have demonstrated particular utility across multiple CNS disorders. In Alzheimer's disease, EEG measures such as theta coherence, alpha, and beta rhythms have proven effective in monitoring cognitive levels and treatment efficacy [66]. For neuropsychiatric disorders including depression, anxiety, and schizophrenia, EEG biomarkers can objectively evaluate drug candidate effects on mood and cognition [63]. In epilepsy, EEG has long been essential for diagnosis and management, with translational biomarkers enhancing understanding of seizure dynamics and supporting targeted therapy development [63].

Functional Near-Infrared Spectroscopy (fNIRS) Biomarkers

fNIRS is a non-invasive optical imaging technique that measures brain activity by assessing changes in cortical hemodynamic activity using near-infrared light [65]. The technique measures the blood oxygen level-dependent (BOLD) response of the brain, similar to fMRI, but with the advantages of being portable, cost-effective, and tolerant of movement [65]. fNIRS detects hemodynamic responses associated with cortical activation, measuring increases in total hemoglobin (THb) and oxygenated hemoglobin (OHb) levels while recording decreases in deoxygenated hemoglobin (DHb) concentrations in regions of interest [64].

The application of fNIRS has gained significant traction in studying neurodevelopmental disorders (NDDs), where it shows promise for assessing brain function through visually evoked hemodynamic responses (vHDR) [64]. In studies of X-linked NDDs, fNIRS has demonstrated potential as a biomarker for disease severity and treatment efficacy, with unique vHDR patterns identified across various neurological conditions [64]. The technique is particularly valuable for studying neurovascular coupling - the relationship between neural activity and subsequent changes in cerebral blood flow [64].

fNIRS has proven effective in detecting early neural deficits in high-risk populations, including pre-term newborns, and can track brain maturation even in challenging populations [64]. The reliability of fNIRS in capturing task-specific cortical activation has been validated across age groups, making it suitable for longitudinal assessment of neurodevelopmental trajectories and therapeutic interventions [64].

Table 2: Technical Comparison of EEG and fNIRS Biomarker Modalities

Characteristic	EEG	fNIRS
Measured Signal	Electrical activity from neuronal firing	Hemodynamic response (blood oxygenation)
Temporal Resolution	Excellent (milliseconds) [65]	Good (seconds)
Spatial Resolution	Limited (several centimeters) [65]	Better than EEG (~1-2 cm of activated area) [65]
Portability	High [65]	High [65]
Tolerance to Movement	Moderate	High [65]
Primary Applications in Drug Development	Target screening, engagement confirmation, functional outcome tracking [63]	Disease progression monitoring, treatment efficacy assessment, neurovascular coupling evaluation [64]
Key Biomarker Parameters	Theta coherence, alpha/beta rhythms, evoked potentials [66] [63]	Oxygenated hemoglobin (OHb), deoxygenated hemoglobin (DHb), total hemoglobin (THb) [64]

Multimodal Integration of EEG and fNIRS

Technical Foundations of Multimodal Integration

The integration of EEG and fNIRS represents a promising direction for brain activity decoding with high spatiotemporal resolution in naturalistic scenarios [43]. These modalities reflect distinct but closely related aspects of underlying neuronal activity, carrying complementary information about brain function [65]. While EEG directly measures the electrical activity of neurons with excellent temporal resolution, fNIRS measures the hemodynamic consequences of neural activity with better spatial localization [65]. Together, they enable exploration of neurovascular coupling - the fundamental relationship between electrical brain activity and subsequent cerebral blood flow changes [65].

The technical challenges in integrating EEG and fNIRS include coupling electrodes and optodes to the subject's head, achieving precise time synchronization between systems, and minimizing electrical crosstalk between modalities [65]. Recent advances have addressed these challenges through integrated systems that combine EEG electrodes and fNIRS optodes in co-located holders, enabling truly simultaneous measurement [65]. Additionally, robust artifact removal techniques and signal processing methods have been developed to handle the distinct noise profiles of each modality [43].

Data Fusion Approaches

Multimodal fusion of EEG and fNIRS data employs several computational strategies, each with distinct advantages and applications [43]:

• Data Concatenation: Combining feature vectors from both modalities into a single input for classification algorithms, maintaining the separate identity of each signal type.
• Model-Based Fusion: Developing computational models that incorporate the physiological relationship between electrical and hemodynamic activities, often based on neurovascular coupling principles.
• Decision-Level Fusion: Processing each modality separately and combining the results at the decision stage, allowing for modality-specific preprocessing and feature extraction.
• Source-Decomposition Techniques: Using unsupervised symmetric methods to identify latent variables that represent shared sources of variance across both modalities, potentially revealing complex neurovascular coupling processes [43].

The development of realistic synthetic datasets that simulate concurrent EEG and fNIRS responses to specific tasks has advanced method validation and comparison, addressing the scarcity of multimodal public datasets [43]. These synthetic datasets with known ground truth enable robust evaluation of fusion algorithm performance under controlled conditions.

Diagram 1: Multimodal EEG-fNIRS Data Fusion Workflow. This diagram illustrates the sequential processing steps from raw data acquisition through multimodal fusion to applications in drug development.

Experimental Protocols and Methodologies

Protocol Design for Visual Processing Assessment

The visual system serves as a powerful model for assessing overall brain function in both preclinical models and clinical studies [64]. Visual evoked potentials (VEPs) recorded via EEG have shown significant reductions in amplitude correlating with disease severity in X-linked neurodevelopmental disorders, suggesting VEP amplitude reflects overall neurological health and disorder progression [64]. Complementary fNIRS measurements of visually evoked hemodynamic responses (vHDR) provide insights into neurovascular coupling and energy metabolism disruptions in these disorders [64].

Established protocols for visual stimulation employ high-contrast black-and-white checkerboard patterns and pattern-reversal gratings, which have proven most effective for eliciting robust, measurable responses in both EEG and fNIRS [64]. Stimulation parameters including shape, contrast, chromaticity, and frequency can be optimized to enhance response reliability [64]. Higher stimulus frequencies produce linear increases in vHDR amplitude, while greater contrast elicits logarithmic changes in oxygenated and deoxygenated hemoglobin levels [64].

For fNIRS specifically, visual stimulation typically produces a distinct hemodynamic response pattern characterized by increased OHb levels and a smaller-magnitude decrease in DHb, peaking approximately 5 seconds after stimulus onset [64]. During sustained stimulation, OHb levels rise and remain elevated, while DHb initially decreases before returning to baseline and exhibiting a post-stimulus overshoot [64]. Multi-channel fNIRS studies have confirmed the regional specificity of these responses, with activation localized to the contralateral occipital cortex [64].

Protocol Implementation for Treatment Response Assessment

Comprehensive assessment of therapeutic interventions requires longitudinal measurement protocols that capture both immediate and long-term effects on brain function. For EEG, this includes quantification of spectral power across frequency bands (delta, theta, alpha, beta, gamma), functional connectivity measures, and event-related potentials in response to specific cognitive tasks [63]. For fNIRS, the focus is on hemynamic response magnitude, latency, spatial extent, and laterality patterns during task activation [64].

Protocols should be tailored to specific patient populations and clinical questions. For neurodegenerative diseases like Alzheimer's, protocols often incorporate cognitive tasks alongside sensory stimulation to engage higher-order processing networks affected by the disease [66]. For neurodevelopmental disorders, protocols must account for developmental trajectories and implement age-appropriate tasks with engaging stimuli to maintain participant attention [64]. Standardized protocols, engaging stimuli, and age-stratified analyses are crucial for enhancing clinical relevance and diagnostic accuracy [64].

Diagram 2: Experimental Protocol for Multimodal Visual Processing Assessment. This workflow outlines the key steps in acquiring and analyzing simultaneous EEG-fNIRS data during visual stimulation paradigms.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Tools for Multimodal EEG-fNIRS Biomarker Studies

Tool/Category	Specific Examples	Function/Application in Biomarker Research
Integrated EEG-fNIRS Systems	TMSi/Artinis integrated systems [65]	Simultaneous acquisition of electrical and hemodynamic data with synchronized timing and minimal crosstalk
EEG Electrodes	Gel, water, and dry electrodes [65]	Capture electrical brain activity with varying tradeoffs between signal quality and usability
fNIRS Optodes	Source-detector pairs with specific separation distances	Measure cortical hemodynamic activity through near-infrared light transmission and absorption
Visual Stimulation Equipment	Pattern generators, display systems, high-contrast stimuli [64]	Elicit robust, measurable neural responses in visual processing pathways
Data Processing Platforms	EEGLAB, NIRS-KIT, Homer2, Brainstorm	Preprocess, analyze, and visualize multimodal neuroimaging data
Multimodal Fusion Algorithms	Data concatenation, model-based fusion, decision-level fusion, source decomposition [43]	Integrate complementary information from EEG and fNIRS to enhance biomarker sensitivity
Artifact Removal Tools	ICA, PCA, motion correction algorithms [43]	Identify and remove non-neural signals from physiological noise and movement
Synthetic Validation Datasets	Realistic simulated HD-fNIRS-EEG data [43]	Validate fusion methods with known ground truth for algorithm development and comparison

The integration of multimodal neuroimaging approaches, particularly EEG and fNIRS, represents a transformative advancement in biomarker development for CNS drug discovery. These complementary technologies enable comprehensive assessment of both electrical neural activity and hemodynamic responses, providing a more complete picture of brain function and drug effects than either modality alone [63] [65] [64]. The development of standardized, fit-for-purpose biomarker validation frameworks ensures these tools meet regulatory standards while addressing critical drug development challenges [62].

As drug development evolves toward more targeted therapies and precision medicine approaches, biomarkers will play an increasingly vital role across therapeutic areas. While oncology has led biomarker adoption, promising applications are emerging in CNS disorders, with Alzheimer's disease showing particular traction [67]. Advances in artificial intelligence and machine learning further enhance biomarker utility by enabling integration of multimodal data streams and identification of complex patterns predictive of treatment response [68] [67].

Despite regulatory challenges and technical hurdles, the strategic implementation of validated biomarker strategies incorporating EEG, fNIRS, and other modalities holds tremendous potential to accelerate CNS therapeutic development, reduce late-stage failures, and ultimately deliver more effective treatments to patients with neurological and psychiatric disorders [63] [64].

Overcoming Technical Challenges: Artifact Removal, Signal Processing, and Data Quality Optimization

Multimodal neuroimaging, which combines electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), provides a comprehensive view of brain activity by capturing complementary signals: electrophysiological activity from EEG and hemodynamic responses from fNIRS [20] [12]. However, the fidelity of these signals is consistently challenged by various biological and technical contaminants, known as artifacts. Successful identification and removal of these artifacts is a critical prerequisite for accurate data interpretation in both basic neuroscience research and clinical drug development [69] [70]. Motion artifacts pose a particularly significant challenge in mobile or naturalistic study designs, where they can severely degrade signal quality and obscure underlying neural correlates [71]. This guide provides an in-depth technical overview of the primary artifact types in EEG and fNIRS, presents quantitative performance comparisons of correction methodologies, and outlines detailed experimental protocols for artifact management.

Artifacts in EEG and fNIRS can be broadly categorized into physiological artifacts, originating from the subject's body, and motion artifacts, resulting from subject or equipment movement.

Physiological Artifacts

• EEG Physiological Artifacts: The EEG signal is susceptible to contamination from several physiological sources [71]. Electrooculographic (EOG) artifacts are caused by eye movements and blinks, generating a strong electrical field that propagates to frontal EEG electrodes. Electromyographic (EMG) artifacts arise from the electrical activity of cranial, facial, and neck muscles, appearing as high-frequency, spiky activity in the EEG. Electrocardiographic (ECG) artifacts are the result of the electrical activity of the heart and can be observed as a periodic QRS complex, particularly in electrodes close to the neck.
• fNIRS Physiological Artifacts: fNIRS signals are contaminated by physiological noises originating from the systemic circulation [69]. These include cardiac oscillations (~1 Hz), respiratory cycles (~0.2-0.3 Hz), and Mayer waves in blood pressure (~0.1 Hz). These physiological noises often overlap with the frequency band of the hemodynamic response and must be separated using signal processing techniques.

Motion Artifacts

Motion artifacts are a major obstacle for both EEG and fNIRS, especially in mobile or long-duration recordings [71] [69].

• In EEG: Motion artifacts manifest in several ways. Vertical head movements during walking can cause baseline shifts and periodic oscillations. Muscle twitches in the scalp, neck, and jaw produce sharp transients that can mimic epileptic spikes. Sudden displacements of electrodes, such as during a heel strike in the gait cycle, result in large-amplitude, low-frequency bursts that can saturate the amplifier [71].
• In fNIRS: Motion artifacts primarily occur when the optode assembly shifts relative to the scalp. This movement alters the optical path length and the coupling between the optode and skin, leading to abrupt, high-amplitude shifts in the measured intensity signals for both 690 nm and 830 nm wavelengths [72] [69]. The severity of these artifacts is documented using simultaneously recorded accelerometer data [72].

Quantitative Comparison of Artifact Removal Techniques

The efficacy of artifact removal techniques is typically quantified using performance metrics such as the improvement in Signal-to-Noise Ratio (ΔSNR) and the percentage reduction in motion artifacts (η). The following tables summarize the performance of various state-of-the-art methods for EEG and fNIRS signals, as evaluated on a benchmark dataset [72] [69].

Table 1: Performance comparison of motion artifact removal techniques for single-channel EEG signals.

Method Category	Specific Technique	Average ΔSNR (dB)	Average η (%)	Key Characteristics
Single-Stage	WPD (db2 wavelet)	29.44	—	Wavelet Packet Decomposition [69]
Single-Stage	WPD (db1 wavelet)	—	53.48	Wavelet Packet Decomposition [69]
Two-Stage	WPD-CCA (db1 wavelet)	30.76	59.51	Combines WPD with Canonical Correlation Analysis [69]
Deep Learning	Motion-Net (CNN)	20.00 ± 4.47	86.00 ± 4.13	Subject-specific, uses Visibility Graph features [71]
Spatial Filtering	SPHARA (Improved)	—	—	Spatial harmonic analysis for dry EEG [70]

Table 2: Performance comparison of motion artifact removal techniques for single-channel fNIRS signals.

Method Category	Specific Technique	Average ΔSNR (dB)	Average η (%)	Key Characteristics
Single-Stage	WPD (fk4 wavelet)	16.11	26.40	Wavelet Packet Decomposition [69]
Two-Stage	WPD-CCA (db1 wavelet)	16.55	—	Combines WPD with Canonical Correlation Analysis [69]
Two-Stage	WPD-CCA (fk8 wavelet)	—	41.40	Combines WPD with Canonical Correlation Analysis [69]

Detailed Experimental Protocols for Artifact Handling

Data Acquisition and Ground-Truth Establishment

A robust methodology for evaluating motion artifact removal techniques requires a dataset with a known ground truth. The following protocol, as utilized in the benchmark dataset, outlines this process [72]:

• Apparatus Setup: Two pairs of transducers (for fNIRS: one pair at 690 nm, one at 830 nm; for EEG: one pair of electrodes) are placed in close proximity on the prefrontal cortex. One transducer in each pair is intentionally disturbed to induce motion artifacts, while the other remains undisturbed to serve as a ground-truth (GT) signal. A 3-axis accelerometer is affixed to each transducer to objectively document motion.
• Signal Recording and Synchronization:
  • fNIRS Protocol: Record raw light intensity signals at a sampling frequency of approximately 25 Hz. Simultaneously record the 3-axis accelerometer signals at 200 Hz. A software-generated trigger signal, recorded by both systems, is used for synchronization. The trigger signal transitions to a low value during motion-artifact intervals and high during clean intervals.
  • EEG Protocol: Record raw EEG signals at a sampling frequency of 2048 Hz. Simultaneously record the 3-axis accelerometer signals at 200 Hz. Synchronize the two data streams using a trigger signal that marks the experiment's start and end.
• Preprocessing:
  • Synchronization: Align the physiological signals (EEG/fNIRS) with the accelerometer signals using the corresponding trigger signal transitions.
  • Resampling: Digitally resample signals to a common frequency for analysis (e.g., fNIRS resampled to 200 Hz, or accelerometers resampled to 2048 Hz to match their respective paired signals).
  • Baseline Correction: Apply baseline correction, such as by deducting a fitted polynomial, to improve the correlation between the motion-contaminated (MA) and ground-truth (GT) signals during clean intervals [71].

Implementation of the WPD-CCA Algorithm

The two-stage WPD-CCA method is a robust technique for correcting motion artifacts in single-channel data. The detailed workflow is as follows [69]:

• Signal Decomposition: Apply Wavelet Packet Decomposition (WPD) to the motion-artifact-contaminated single-channel signal. This decomposes the signal into a set of wavelet packet components (WPCs) across different frequency bands. Multiple wavelet families (e.g., Daubechies 'db1', Fejer-Korovkin 'fk4') can be evaluated for optimal performance.
• Component Reconstruction: Reconstruct the signal from each of the individual WPCs obtained in the previous stage. This creates a set of multiple channels, where each channel is the reconstructed signal from a single WPC.
• Artifact Removal with CCA: Apply Canonical Correlation Analysis (CCA) to the multichannel signal generated from the WPCs. CCA identifies and removes components that are highly correlated with the motion artifacts. The output of this stage is a cleaned signal.

Implementation of the Motion-Net Deep Learning Algorithm

Motion-Net is a subject-specific, CNN-based framework designed for motion artifact removal. The protocol for its application is [71]:

• Input Feature Extraction: For each subject, train and test the model separately on individual trials. The input to the model consists of the raw EEG signal. To enhance learning, especially with smaller datasets, supplementary features like Visibility Graph (VG) features can be incorporated. VG features convert the time-series signal into a graph structure, providing additional information about the signal's structural properties.
• Model Architecture: Employ a 1D U-Net convolutional neural network architecture. This network is designed for signal reconstruction, using an encoder-decoder structure with skip connections to preserve important signal details.
• Training and Validation: The model is trained in a supervised manner. The input is the motion-artifact-contaminated signal, and the target output is the corresponding ground-truth signal from the undisturbed transducer. Performance is validated using metrics such as artifact reduction percentage (η), SNR improvement, and Mean Absolute Error (MAE).

Visualization of Workflows and Signaling Pathways

Artifact Removal Workflow

WPD-CCA Processing Steps

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key materials and data resources for artifact research in multimodal neuroimaging.

Item Name	Function/Application	Specification Notes
Motion Artifact Contaminated fNIRS and EEG Data	Benchmark dataset for developing and validating artifact removal algorithms.	Publicly available on PhysioNet; contains simultaneous EEG/fNIRS, accelerometer, and trigger signals [72].
Dry EEG Cap with Integrated fNIRS Probes	Enables concurrent acquisition of EEG and fNIRS signals in a single helmet.	Custom designs using 3D printing or cryogenic thermoplastic sheets improve fit and reduce motion [12].
3-Axis Accelerometers	Objective quantification of transducer motion to identify artifact-contaminated intervals.	Affixed directly to EEG electrodes or fNIRS optodes; sampled at 200 Hz [72].
Software Trigger Signal Generator	Synchronizes independent recording systems (e.g., EEG, fNIRS, accelerometer).	Critical for aligning data streams during preprocessing for accurate artifact analysis [72].

The integration of Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) offers a powerful multimodal approach to studying brain function by combining EEG's millisecond-level temporal resolution with fNIRS's hemodynamic measures. However, the fidelity of the neural and hemodynamic information extracted from both modalities is critically dependent on the effective identification and removal of biological and motion artifacts. EEG signals are notoriously susceptible to contamination from physiological sources such as electrooculographic (EOG) signals from eye movements and electromyographic (EMG) signals from muscle activity [73]. Similarly, fNIRS signals are profoundly affected by motion artifacts and systemic physiological confounds like scalp blood flow, which can obscure the hemodynamic response of neural origin [10] [74]. Effective preprocessing is, therefore, not merely a preliminary step but a foundational one for ensuring the validity, reliability, and reproducibility of findings in multimodal EEG-fNIRS research [75] [10]. This guide provides an in-depth technical overview of advanced preprocessing techniques for mitigating these artifacts, framed within the context of a robust multimodal research paradigm.

Advanced Preprocessing for EEG

The goal of EEG preprocessing is to isolate brain-generated electrical activity from non-cerebral artifacts while preserving the underlying neural signals of interest.

Nature of EEG Artifacts

EEG artifacts can be broadly categorized into physiological and non-physiological types. Physiological artifacts, such as EOG and EMG, are particularly challenging because their frequency spectra often overlap with those of genuine neural signals. For instance, EMG artifacts from facial muscles can manifest as high-frequency noise, while EOG artifacts from eye blinks appear as large, low-frequency deflections [73]. The presence of these artifacts significantly reduces the quality of EEG recordings, posing challenges for accurate data analysis and impeding the development of EEG-related research and applications [73].

Technical Approaches for Artifact Removal

Traditional and Blind Source Separation Methods Conventional methods for artifact removal include regression, filtering, and Blind Source Separation (BSS) techniques like Independent Component Analysis (ICA). Regression methods use reference channels to estimate and subtract artifact components but their performance degrades significantly in the absence of a reference signal [73]. Filtering is effective for removing noise outside the frequency band of interest but is of limited use for EOG/EMG due to spectral overlap [73]. BSS methods, including ICA, project the data into a component space where artifact-related components can be identified and removed manually or via automated algorithms [75] [76]. A comparative study found that the specific choice of ICA algorithm (e.g., SOBI vs. Extended Infomax) has a relatively small effect on the overall cleaning procedure, whereas the initial segmentation of data and the re-referencing method are more critical steps [75].

Deep Learning-Based Methods Recently, deep learning (DL) has emerged as a powerful alternative for EEG artifact removal, overcoming limitations of traditional methods such as the need for manual intervention [73]. DL models can learn to map artifact-contaminated EEG to clean EEG in an end-to-end, automated fashion. For example, the CLEnet architecture integrates dual-scale Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, alongside an improved attention mechanism (EMA-1D), to extract both morphological and temporal features of EEG, thereby effectively separating EEG from artifacts [73]. CLEnet has demonstrated superior performance in removing mixed (EMG+EOG) artifacts, achieving a signal-to-noise ratio (SNR) of 11.498 dB and a correlation coefficient (CC) of 0.925, outperforming other mainstream models like 1D-ResCNN and NovelCNN [73].

Table 1: Performance Comparison of Advanced EEG Artifact Removal Algorithms on a Semi-Synthetic Dataset

Algorithm	Artifact Type	Signal-to-Noise Ratio (SNR)	Correlation Coefficient (CC)	RRMSE (Temporal)
CLEnet [73]	Mixed (EMG+EOG)	11.498 dB	0.925	0.300
1D-ResCNN [73]	Mixed (EMG+EOG)	Not Reported	Not Reported	Not Reported
NovelCNN [73]	Mixed (EMG+EOG)	Not Reported	Not Reported	Not Reported
DuoCL [73]	Mixed (EMG+EOG)	Not Reported	Not Reported	Not Reported

Practical Workflow and Re-Referencing A typical advanced EEG preprocessing pipeline involves filtering, epoching, artifact removal (e.g., via ICA or automated algorithms like Artifact Subspace Reconstruction - ASR), and re-referencing [76]. The choice of re-referencing method can significantly impact results. Common approaches include the Common Average Reference (CAR), robust CAR, and Reference Electrode Standardization Technique (REST). Studies have shown that CAR, REST, and RESIT yield similar topographical representations, while robust CAR can produce different event-related spectral perturbation patterns [75].

Application-Specific Considerations: Epilepsy Seizure Detection

In clinical applications like epilepsy seizure detection, advanced preprocessing can dramatically enhance performance. One study employed advanced artifact removal and introduced a novel Peak-to-Peak Amplitude Fluctuation (PPAF) metric to assess amplitude variability within event-related potential waveforms [76]. This approach, applied to data from epilepsy patients, identified the frontal and parietal regions (Cz, Pz, Fp2 electrodes) as primary contributors to seizures and achieved detection accuracies of up to 99% [76].

Advanced Preprocessing for fNIRS

fNIRS preprocessing aims to isolate the hemodynamic responses (changes in oxyhemoglobin - HbO and deoxyhemoglobin - HbR) related to neural activity from confounding noise, primarily motion artifacts and physiological oscillations.

Nature of fNIRS Artifacts

fNIRS is relatively robust to motion compared to fMRI, but motion artifacts remain a significant challenge. Motion can cause two main types of signal disruptions: spikes (rapid, large-amplitude changes) and baseline shifts [74]. Furthermore, systemic physiological confounds, such as cardiac pulsation, respiration, and blood pressure waves, introduce noise in the same frequency band as the task-evoked hemodynamic response (~0.01-0.1 Hz) [77]. A critical confound is scalp blood flow, which is non-neural in origin and can be a major source of false positives if not properly accounted for [10].

Technical Approaches for Artifact Removal

Standard Processing Pipeline A standard fNIRS preprocessing pipeline, as implemented in toolboxes like MNE-Python, involves several key stages [77]:

• Conversion to Optical Density: Raw light intensity is converted to optical density, which is more stable for subsequent processing.
• Quality Assessment: The Scalp Coupling Index (SCI) can be used to quantify the quality of the coupling between optodes and the scalp. Channels with poor coupling (e.g., SCI < 0.5) should be excluded.
• Conversion to Hemoglobin: The modified Beer-Lambert law is applied to convert optical density to relative concentrations of HbO and HbR.
• Filtering: A band-pass filter (e.g., 0.01 - 0.7 Hz) is applied to remove high-frequency noise (like heart rate) and slow drifts.

Handling Motion Artifacts and Physiological Confounds While band-pass filtering removes some noise, more sophisticated methods are often required. A major challenge in the field is the lack of standardization in how these methods are applied. The fNIRS Reproducibility Study Hub (FRESH) initiative, which involved 38 independent analysis teams, found that the primary sources of variability in fNIRS results were related to how poor-quality data were handled, how responses were modeled, and how statistical analyses were conducted [10]. This underscores the importance of transparent reporting of preprocessing choices. To address physiological confounds, short-separation channels are highly effective. These channels (typically with a source-detector distance of < 1 cm) are primarily sensitive to scalp blood flow and can be used as regressors to remove this noise from the standard channels (with longer separations) [43].

Motion Characterization and Correction Recent research has focused on better characterizing motion artifacts to improve correction algorithms. One study used computer vision (the SynergyNet deep neural network) to analyze video recordings of participants and extract ground-truth head movement data [74]. They found that repeated, upward, and downward movements particularly compromised signal quality, and that the susceptibility of different brain regions to motion artifacts varied with the type of movement [74]. This work lays the foundation for developing and validating more effective, data-driven motion correction algorithms.

Table 2: Impact of Specific Head Movements on fNIRS Signal Quality [74]

Movement Type	Axis	Impact on fNIRS Signal
Upward, Downward	Vertical	High impact; particularly compromises signal in occipital and pre-occipital regions.
Bend Left, Bend Right	Frontal	High impact; temporal regions most affected.
Turn Left, Turn Right	Sagittal	High impact; temporal regions most affected.
Repeated Rotations	Multiple	High impact; tends to compromise signal quality.

Multimodal EEG-fNIRS Fusion Considerations

The synergy of EEG and fNIRS lies in their complementary nature: EEG provides high temporal resolution, while fNIRS provides a more direct measure of hemodynamic activity with better spatial resolution than EEG alone [78] [43]. When preprocessing multimodal data, several key points must be considered. First, artifact removal should be performed separately for each modality before data fusion, as the nature and timing of artifacts differ. Second, temporal alignment of the two data streams is crucial, given their vastly different sampling rates (EEG often >250 Hz, fNIRS typically ~10 Hz). Fusion methods can occur at multiple levels, including data-level (concatenating features), model-level (using generative models), and decision-level (combining classifier outputs) [43]. The FRESH initiative highlights that despite analytical variability, clear hypotheses and standardized protocols can lead to strong consensus in results, advocating for the development of best practices in multimodal analysis [10].

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Essential Tools for Multimodal EEG-fNIRS Research

Tool Name	Type	Primary Function	Key Context
EEGLAB [79] [76]	Software Toolbox	Interactive EEG data processing & ICA	Standard environment for EEG preprocessing; includes artifact removal, time-frequency analysis.
MNE-Python [77]	Software Library	Python-based EEG & MEG data analysis	Comprehensive suite for EEG & fNIRS processing, including standardized pipelines.
CLEnet [73]	Deep Learning Model	Automated EEG artifact removal	End-to-end removal of EOG/EMG artifacts using CNN-LSTM architecture.
NIRScout System [38]	fNIRS Hardware	Data acquisition for fNIRS	Used in multimodal studies to collect high-quality fNIRS data concurrently with EEG.
Short-Separation Channels [43]	fNIRS Hardware/Technique	Scalp blood flow regression	Critical for removing systemic physiological noise from fNIRS signals.
EDF Browser [79]	Software Tool	EEG data visualization & inspection	Foundational tool for initial data examination and identifying potential anomalies.
Computer Vision (e.g., SynergyNet) [74]	Analysis Technique	Motion artifact characterization	Quantifies head movements from video to validate and improve motion correction.

Experimental Protocols & Workflows

Protocol for Validating a Novel EEG Artifact Removal Algorithm

A typical protocol for validating a new deep learning-based artifact removal method, such as CLEnet, involves the following steps [73]:

• Dataset Curation: Use a semi-synthetic dataset created by adding recorded EOG and EMG artifacts to clean EEG segments. This provides a ground truth for evaluation. A real, multi-channel EEG dataset containing unknown artifacts can also be used to test generalizability.
• Model Training: Train the neural network (e.g., CLEnet with its dual-branch CNN and LSTM) in a supervised manner using the artifact-contaminated EEG as input and the clean EEG as the target. The mean squared error (MSE) is a common loss function.
• Performance Evaluation: Quantify performance using metrics like Signal-to-Noise Ratio (SNR), average Correlation Coefficient (CC) between the cleaned and clean EEG, and Relative Root Mean Square Error in both temporal and frequency domains (RRMSEt, RRMSEf).
• Comparison and Ablation: Compare the model's performance against established mainstream models (e.g., 1D-ResCNN, NovelCNN). Conduct ablation studies (e.g., removing the EMA-1D attention module) to confirm the contribution of each network component.

Protocol for an fNIRS Motion Artifact Characterization Study

A protocol to characterize the relationship between specific head movements and motion artifacts, as in [74], is:

• Participant and Task Design: Recruit participants to perform a series of controlled, cued head movements along three rotational axes (pitch, roll, yaw) at varying speeds.
• Multimodal Data Acquisition: Record fNIRS signals simultaneously with video of the participant's head. Use a whole-head fNIRS cap for comprehensive coverage.
• Computer Vision Analysis: Process the video data frame-by-frame using a deep neural network (e.g., SynergyNet) to extract precise, ground-truth head orientation angles.
• Signal Analysis: Identify motion artifacts (spikes, baseline shifts) in the fNIRS data. Correlate the amplitude and timing of these artifacts with the kinematic metrics (amplitude, speed) derived from the computer vision data.
• Regional Susceptibility Analysis: Map the occurrence of artifacts to specific fNIRS channel locations to determine which brain regions are most susceptible to different types of movement.

Workflow Diagrams

Advanced EEG Preprocessing Workflow

Advanced fNIRS Preprocessing Workflow

Multimodal EEG-fNIRS Fusion Pathway

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) offers a powerful, multimodal approach to studying brain function by combining millisecond-temporal resolution with improved spatial localization of hemodynamic activity [2]. However, the analysis of these signals is significantly challenged by the presence of persistent artifacts that can obscure neural information and compromise data integrity. Artifacts in these modalities arise from diverse sources: EEG recordings are contaminated by ocular movements (EOG), muscle activity (EMG), cardiac rhythms, and environmental interference [80], while fNIRS signals are confounded by systemic physiological activities such as cardiac pulsations, respiration, blood pressure changes, and motion-induced hemodynamics in the scalp [2]. The problem is exacerbated in naturalistic experimental settings, where motion artifacts are more prevalent, and in clinical populations, where neurovascular coupling may be altered [81].

Data-driven approaches, particularly machine learning (ML) and deep learning (DL), have emerged as transformative solutions for distinguishing evoked neuronal activity from complex artifactual patterns. Unlike traditional model-based methods like the General Linear Model (GLM), which require precise a priori knowledge of stimulus timing and noise characteristics, data-driven methods can learn to separate signal from noise directly from the data itself. This capability is crucial for analyzing data from continuous brain imaging in naturalistic environments, assessing dynamic brain network activity, and advancing preprocessing for single-trial analysis in brain-computer interfaces (BCIs) [2]. This guide provides a comprehensive technical overview of modern, data-driven methodologies for artifact correction and feature extraction in simultaneous EEG-fNIRS research, framed within the context of multimodal neuroimaging.

Artifact Correction Methodologies

Statistical and Blind Source Separation Methods

Traditional and robust methods for artifact removal often rely on statistical feature extraction and blind source separation. These approaches are particularly valuable when ground-truth clean data is unavailable for training deep learning models.

A seminal approach for MRI-related artifact removal in EEG data utilizes Singular Value Decomposition (SVD) for gradient artifact removal and a hybrid of Independent Component Analysis (ICA) and SVD for pulse artifacts [82]. The methodology involves:

• Gradient Artifact Removal: Channel-wise filtering based on SVD is applied to remove highly repetitive, slice-acquisition-locked artifacts. The SVD-derived components representing gradient artifacts are identified and regressed out from the recorded data.
• Pulse Artifact Removal: The data is first decomposed into temporally independent components using ICA. A compact cluster of components with sustained high mutual information with the electrocardiogram (ECG) is selected automatically. After removing these components, the time courses of the remaining components are filtered by SVD to remove residual patterns phase-locked to cardiac markers [82].

This method's robustness has been validated on large datasets acquired during various behavioral tasks, sensory stimulations, and resting conditions [82]. The core principle is the extraction and selection of statistical features that characterize artifacts using reference signals like MRI triggers and ECG.

Deep Learning-Based Approaches

Deep learning models excel at learning complex, non-linear mappings from noisy to clean data, often surpassing the performance of traditional techniques.

Generative Adversarial Networks (GANs) have shown remarkable effectiveness in EEG denoising. The AnEEG model exemplifies this approach by integrating Long Short-Term Memory (LSTM) networks with a GAN architecture [80]. The generator, composed of a two-layered LSTM network, takes noisy EEG as input and produces a cleaned version. The discriminator, a one-dimensional convolutional neural network, judges the quality of the generated signal against ground-truth clean data. This adversarial training guides the generator to produce artifact-free EEG that preserves temporal dynamics and neural information [80].

For artifacts induced by Transcranial Electrical Stimulation (tES), a comparative benchmark of eleven ML methods revealed that the optimal model is highly dependent on the stimulation type [83]:

• For tDCS artifacts, a Complex CNN performed best.
• For the more complex tACS and tRNS artifacts, a multi-modular network (M4) based on State Space Models (SSMs) yielded superior results [83].

Evaluation of these models is typically conducted using semi-synthetic datasets, where clean EEG is artificially contaminated with known artifacts, allowing for rigorous calculation of metrics like Root Relative Mean Squared Error (RRMSE) and Correlation Coefficient (CC) against a ground truth [83].

Table 1: Performance Comparison of Deep Learning Models for Artifact Removal

Model	Architecture	Target Artifact	Key Performance Metrics
AnEEG [80]	LSTM-based GAN	General Biological & Environmental	Lower NMSE & RMSE, Higher CC & SNR vs. wavelet methods
Complex CNN [83]	Convolutional Neural Network	tDCS	Best RRMSE and Correlation for tDCS artifacts
M4 Network [83]	State Space Models (SSMs)	tACS, tRNS	Best RRMSE and Correlation for tACS/tRNS artifacts
GCTNet [80]	GAN with CNN & Transformer	Ocular & Muscular	11.15% RRMSE reduction, 9.81 SNR improvement

The following diagram illustrates a generalized workflow for a deep learning-based artifact removal pipeline, integrating elements from the aforementioned models:

Feature Extraction for Multimodal Fusion

Once artifacts are mitigated, the next critical step is extracting informative features from both EEG and fNIRS signals to leverage their complementary nature.

EEG Feature Engineering

EEG features capture different aspects of neural electrical activity and are typically extracted from specific frequency bands or event-related potentials (ERPs).

• Spectral Features: The power spectral density (PSD) within standard frequency bands (delta, theta, alpha, beta, gamma) is a fundamental feature. Studies on Internet Gaming Disorder (IGD) have successfully used the ratio of the area under the PSD curve and distinct power differences in frontal theta activity as discriminative features [84].
• Time-Domain Features: Simple statistical measures like mean, kurtosis, and skewness of the signal in sub-bands can be indicative of brain states. For instance, skewness has been found distinctive in EEG sub-bands for classifying IGD [84].
• Nonlinear Features: These are crucial for capturing the complex, dynamic nature of neural signals. Commonly used features include:
  • Higuchi Fractal Dimension
  • Lyapunov Exponent
  • Shannon Entropy
  • Lempel-Ziv Complexity
  • Entropy of Hilbert Transform [84]
• Event-Related Potential (ERP) Components: ERP analysis involves averaging EEG responses time-locked to stimuli. Key components like P100, N200, P300, and N450 provide insights into cognitive processes. Their amplitude and latency are critical features; for example, prolonged P300 latency and reduced amplitude are observed in IGD, indicating impaired attention [84].

fNIRS Feature Engineering

fNIRS features primarily describe the slow hemodynamic responses following neural activity.

• Hemodynamic Response Features: The most common features are changes in concentration of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR). These can be quantified as the mean amplitude during a task block, slope, or area under the curve [38] [84]. In motor imagery tasks, for example, HbO activation patterns in the sensorimotor cortex serve as valid features for decoding intention [38].
• Spatial Topography: The spatial pattern of hemodynamic activation across channels covering the prefrontal, motor, or other cortices provides valuable feature vectors for classification [38] [84].

Multimodal Fusion Strategies

Combining EEG and fNIRS features can yield more robust and accurate decoding than unimodal systems.

• Data-Level Fusion: This involves simply concatenating the raw or preprocessed feature vectors from both modalities into one large input vector for a classifier [2].
• Feature-Level Fusion: Features are extracted from each modality separately and then concatenated into a unified feature vector. This was demonstrated in an IGD study using the Stroop task, where combining EEG, ERP, and fNIRS features achieved a superior classification accuracy of 87.25% using Support Vector Machines, compared to 79.4% with EEG features alone [84].
• Decision-Level Fusion: Each modality is processed by its own classifier, and the final decision is made by combining the outputs (e.g., by averaging or voting) [2].
• Source-Decomposition and Symmetric Fusion: More advanced, data-driven techniques like joint ICA or canonical correlation analysis can be applied to uncover latent relationships between the modalities. These unsupervised, symmetric fusion methods are particularly promising for revealing complex neurovascular coupling processes, especially in naturalistic scenarios where stimulus timing is not available [2].

Table 2: Summary of Key Features for EEG and fNIRS Analysis

Modality	Feature Category	Specific Features	Application Example
EEG	Spectral	PSD, Power Ratios (e.g., Theta/Beta)	IGD Classification [84]
EEG	Temporal/Statistical	Mean, Skewness, Kurtosis	IGD Biomarker Investigation [84]
EEG	Nonlinear Complexity	Higuchi FD, Lyapunov Exponent, Entropy	Enhanced Biomarker Discovery [84]
EEG	Event-Related Potentials	P300/N200 Latency & Amplitude	Assessing Attention & Cognitive Control [84]
fNIRS	Hemodynamic	HbO/HbR Concentration Changes	Motor Imagery Decoding [38]
fNIRS	Spatial	Topographical Activation Maps	Prefrontal Cortex Assessment in IGD [84]

Experimental Protocols and Validation

Implementing and validating these methodologies requires rigorous experimental design. Below is a detailed protocol from a multimodal study investigating Internet Gaming Disorder, which exemplifies a comprehensive approach [84].

1. Participant Selection and Grouping:

• Cohort: 102 male university students (18-23 years).
• Groups: 51 with IGD and 51 healthy controls (HCs), matched for age and other demographics.
• Clinical Assessments: Use standardized scales for group confirmation and correlation with neural features.

2. Experimental Paradigm:

• Task: A classic Stroop task with congruent, incongruent, and neutral stimuli is presented.
• Procedure: Participants are instructed to respond to the color of the word while ignoring its meaning. The task is designed to elicit cognitive conflict and executive control processes.
• Simultaneous Recording: EEG and fNIRS data are acquired throughout the task execution.

3. Data Acquisition Specifications:

• EEG: Recorded using a multi-channel system (e.g., 32 channels). Data is sampled at a high frequency (e.g., 256 Hz or higher) with appropriate online filtering.
• fNIRS: A system with multiple sources and detectors is used to cover regions of interest (e.g., Prefrontal Cortex). Data is sampled at a lower rate (e.g., 10 Hz) suitable for hemodynamic signals.

4. Data Preprocessing and Analysis:

• EEG Processing:
  • Bandpass filtering (e.g., 0.5-45 Hz).
  • Artifact correction using one of the methods described in Section 2 (e.g., ICA, ASR, or a deep learning model).
  • Epoching around stimulus onset.
  • ERP computation for components like P300, N450.
  • Decomposition into frequency sub-bands and extraction of linear, nonlinear, and spectral features.
• fNIRS Processing:
  • Conversion of raw light intensity to optical density and then to HbO and HbR concentrations using the Modified Beer-Lambert Law.
  • Bandpass filtering to remove physiological noise (e.g., cardiac pulsations, very low-frequency drift).
  • Extraction of hemodynamic features (e.g., mean HbO during task blocks).

5. Machine Learning Classification:

• Feature Selection: Statistically significant features (e.g., p-value < 0.05 from t-tests) are selected from the pool of EEG and fNIRS features.
• Model Training: Classifiers like Support Vector Machines (SVM) are trained on the selected feature set.
• Validation: Model performance is evaluated using cross-validation, reporting metrics like accuracy, sensitivity, and specificity.

The logical flow of such an experiment, from data collection to insight, is captured in the following workflow:

The Scientist's Toolkit: Essential Research Reagents

The following table details key computational tools, datasets, and methodological "reagents" essential for research in this field.

Table 3: Essential Research Reagents for EEG-fNIRS Machine Learning Research

Item Name	Type	Function & Application	Example Use Case
Public Multimodal Datasets (e.g., [38] [81])	Data	Provides benchmark data for developing and testing algorithms; ensures reproducibility and comparison across studies.	HEFMI-ICH dataset [81] for stroke rehabilitation BCI development.
Semi-Synthetic Data	Data	Enables controlled, rigorous evaluation of artifact removal methods by mixing clean signals with known artifacts.	Benchmarking tES artifact removal models [83].
Independent Component Analysis (ICA)	Algorithm	Blind source separation for isolating and removing artifacts like eye blinks and cardiac signals from EEG.	Pulse artifact removal in simultaneous EEG-fMRI [82].
Generative Adversarial Network (GAN)	Model Architecture	Deep learning framework for learning to map noisy signals to clean versions, preserving neural information.	AnEEG model for general EEG denoising [80].
State Space Models (SSMs)	Model Architecture	Effective for modeling and removing complex, structured noise like tACS and tRNS artifacts from EEG.	M4 network for tES artifact removal [83].
Support Vector Machines (SVM)	Classifier	A robust, nonlinear classifier for BCI and clinical classification tasks using multimodal features.	Achieving 87.25% accuracy in IGD classification [84].
Image-Based Meta- & Mega-Analysis (IBMMA)	Software Tool	A unified R/Python framework for large-scale, multi-site neuroimaging analysis, handling missing data.	Analyzing large-n neuroimaging datasets across multiple cohorts [85] [86].

Addressing Neurovascular Uncoupling in Patient Populations

Neurovascular coupling (NVC) represents a fundamental physiological mechanism that ensures precise coordination between neuronal activity and cerebral blood flow (CBF), thereby maintaining brain homeostasis by delivering oxygen and nutrients to active brain regions [87] [88]. This process occurs through the integrated functioning of the neurovascular unit (NVU), a functional complex comprising neurons, vascular cells (endothelial cells, pericytes, and vascular smooth muscle cells), glial cells (astrocytes, microglia, and oligodendrocytes), and the extracellular matrix [87] [89]. The sophisticated dialogue between these cellular components enables rapid vascular responses to neuronal signaling, typically occurring within seconds of neural activation [89].

Neurovascular uncoupling describes the pathological disruption of this precise coordination, resulting in a mismatch between cerebral metabolic demand and blood supply [88]. This decoupling has emerged as a critical pathophysiological mechanism across a spectrum of neurological and psychiatric disorders, including Alzheimer's disease (AD), cerebral small vessel disease (CSVD), stroke, epilepsy, and major depressive disorder (MDD) [88] [90] [91]. The clinical significance of NVC dysfunction stems from its potential role as both an early biomarker of disease and a therapeutic target for intervention [89]. In encephalopathic conditions, NVC failure disrupts the critical equilibrium between CBF, oxygen delivery, and neuronal metabolic demands, ultimately driving neuronal energy crisis, excitotoxicity, and oxidative stress that accelerate disease progression [88].

Within the framework of multimodal neuroimaging with EEG and fNIRS research, understanding and addressing neurovascular uncoupling presents both challenges and opportunities. These complementary modalities provide unique windows into neurovascular function, with EEG capturing neuronal electrochemical activity with millisecond temporal resolution, while fNIRS measures hemodynamic responses with better spatial resolution than EEG [27] [16]. The integration of these technologies offers a powerful approach for investigating NVC dynamics in both healthy and diseased states, potentially enabling earlier detection and more precise monitoring of therapeutic interventions aimed at restoring neurovascular homeostasis.

Pathophysiological Mechanisms of Neurovascular Uncoupling

Cellular and Molecular Basis of NVC Dysfunction

At the cellular level, neurovascular uncoupling involves disruptions in the intricate signaling pathways that mediate communication between neurons, glia, and vascular cells. The potent vasoconstrictor endothelin-1 (ET-1) has been identified as a key mediator of NVC dysfunction across multiple encephalopathic conditions [88]. ET-1, produced by endothelial cells and astrocytes, binds to its receptors and triggers pathological vasoconstriction, oxidative stress, and neuroinflammatory cascades that collectively exacerbate cerebral hypoperfusion, metabolic dysregulation, and neuronal injury [88]. In Alzheimer's disease, amyloid-β (Aβ) oligomers exert direct neurotoxic effects on endothelial integrity and disrupt NVC, potentiating hypoperfusion and metabolic insufficiency [88]. Progressive NVC failure uncouples activity-dependent neurovascular signaling, thereby depriving neurons of hemodynamic-metabolic support and accelerating neurodegenerative cascades through feedforward mechanisms [88].

The cellular mechanisms underlying NVC involve both feedforward and feedback regulatory systems. Feedforward mechanisms are primarily mediated by non-metabolic signaling pathways, such as glutamate synaptic signaling, which directly initiate intracellular calcium-dependent cascades that produce vasoactive messengers to increase blood flow [88]. Feedback mechanisms involve metabolic factors like oxygen, glucose, and CO₂, where a drop in oxygen levels or an increase in CO₂ can trigger vasodilation to adjust blood flow [88]. When feedforward mechanisms function improperly, feedback mechanisms can theoretically compensate to ensure appropriate blood flow, though this compensatory capacity becomes compromised in disease states.

Table 1: Key Molecular Mediators in Neurovascular Coupling and their Dysregulation

Mediator	Physiological Role in NVC	Dysregulation in Disease
Endothelin-1 (ET-1)	Potent vasoconstrictor; fine-tunes vascular tone	Pathological vasoconstriction; oxidative stress; neuroinflammation [88]
Nitric Oxide (NO)	Vasodilation; mediated by neuronal NO synthase (nNOS)	nNOS inhibition reduces neurovascular response by ~64%; impaired vasodilation [89]
Prostaglandin E₂ (PGE₂)	Vasodilation via EP2/EP4 receptors; from pyramidal neurons	Dysregulated in cerebrovascular diseases; contributes to NVC dysfunction [89]
Amyloid-β (Aβ)	Not a physiological mediator; pathologically relevant	Oligomers impair endothelial function; disrupt NVC in Alzheimer's disease [88]

Research has revealed that different neuronal subtypes contribute distinctively to NVC mechanisms. Pyramidal neurons serve as "neurogenic hubs" within the context of NVC, while inhibitory GABA interneurons play a critical role primarily by modulating the output of these excitatory pyramidal cells [89]. The activity of neurons expressing neuronal nitric oxide synthase (nNOS) significantly influences alterations in both resting and reactive arterial dimensions, with pharmacological blockade of nNOS resulting in an average 64% reduction in neurovascular response across 11 studies [89]. These findings underscore the complexity of neuronal contributions to NVC and suggest multiple potential points of failure in disease states.

Disease-Specific Patterns of Neurovascular Uncoupling

Neurovascular uncoupling manifests with distinct patterns across different neurological and psychiatric conditions, reflecting the diverse pathophysiology and affected brain regions. In cerebral small vessel disease (CSVD), combined resting-state fMRI and arterial spin labeling studies have identified multiple regions with NVC dysfunction, particularly within the default mode network and subcortical nuclei [90]. The decreased NVC of the left superior frontal gyrus has been found to partially mediate the impact of white matter hyperintensities on delayed recall, suggesting this region may serve as a promising biomarker and therapeutic target for memory deficits in CSVD patients [90].

In major depressive disorder (MDD), recent research has revealed abnormal NVC patterns that vary with disease severity and sex [91]. First-episode drug-naïve MDD patients show reduced whole-brain NVC coupling, with spatial correlation analysis revealing significant reductions in severe and female MDD patients [91]. Temporal correlation analysis demonstrates that moderate MDD patients exhibit increased NVC coupling in the left insula, while severe MDD patients show reduced coupling in the left anterior cingulate cortex and increased coupling in the right superior occipital gyrus [91]. These findings highlight the potential of NVC metrics as sensitive indicators of disease state in neuropsychiatric disorders.

For Alzheimer's disease, neurovascular-metabolic dysregulation (MVD) represents one of the earliest changes in disease progression, occurring prior to hallmark pathological features such as amyloid-beta deposition [92]. This MVD persists across preclinical and clinical stages of the disease spectrum, suggesting its potential utility as an early diagnostic indicator and therapeutic target [92]. The progression of NVC dysfunction across different disorders follows distinctive trajectories that reflect underlying disease mechanisms and affected neural systems.

Table 2: Neurovascular Uncoupling Patterns Across Patient Populations

Disorder	Affected Brain Regions/Networks	Characteristic NVC Abnormalities
Alzheimer's Disease	Default Mode Network; Medial Temporal Lobe	Early neurovascular-metabolic dysregulation; impaired Aβ clearance; endothelial dysfunction [88] [92]
Cerebral Small Vessel Disease	Default Mode Network; Subcortical Nuclei; Superior Frontal Gyrus	NVC dysfunction mediates white matter hyperintensity impact on memory [90]
Major Depressive Disorder	Whole-brain; Anterior Cingulate Cortex; Insula; Superior Frontal Orbital Gyrus	Severity and sex-dependent patterns; reduced whole-brain coupling in severe cases [91]
Stroke/SAH	Ischemic Penumbra; Peri-infarct Regions	Inverse NVC; ischemic penumbral expansion; vasoconstrictive pathways [88] [89]

Assessment Methodologies: Multimodal Imaging of NVC

Integrated EEG-fNIRS for Neurovascular Assessment

The simultaneous acquisition of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) provides a powerful multimodal approach for investigating neurovascular coupling dynamics in health and disease [27] [16]. These techniques are technically complementary: EEG offers exceptional temporal resolution (milliseconds) but relatively low spatial resolution, whereas fNIRS achieves notable spatial resolution due to the exponential attenuation of incident light in tissues [27]. Neither technique exhibits atypical physical constraints, and both can be deployed in naturalistic settings beyond the laboratory environment, making them particularly well-suited for studying ecologically valid cognitive processes [27].

Two primary methods have been developed for integrating fNIRS and EEG signals. The first approach involves combining fNIRS and EEG data obtained separately using different systems (e.g., NIRScout and BrainAMP systems) and synchronized during acquisition and analysis via a host computer [27]. While relatively simple to implement, this method may not achieve the precision synchronization required for analysis of EEG data with microsecond time resolution. The second approach utilizes a unified processor to simultaneously process and acquire both EEG signals and fNIRS input and output, resulting in higher synchronization accuracy and streamlined analytical processes [27]. This method, though requiring more complex system design, represents the current gold standard for concurrent fNIRS and EEG detection.

Technical implementation of simultaneous fNIRS-EEG systems requires careful consideration of hardware integration. The joint-acquisition helmet design holds paramount importance, with current approaches typically integrating EEG electrodes and NIR probes on a shared substrate material or arranging them separately [27]. Some researchers have directly integrated NIR fiber optics into existing EEG electrode caps, though this approach presents challenges including uncontrollable variations in the distance between the NIR light source and detector across subjects with different head shapes, and limited effectiveness of elastic fabric in securing the NIR probe due to high stretchability [27]. To address these limitations, researchers have turned to 3D printing technology to craft customized joint-acquisition helmets tailored to experimental requirements, or utilized composite polymer cryogenic thermoplastic sheets that can be softened and shaped at around 60°C, retaining form stability upon cooling [27].

Analytical Approaches for Multimodal Data Fusion

The fusion of fNIRS and EEG data presents significant analytical challenges that have prompted the development of sophisticated computational methods. Structured sparse multiset Canonical Correlation Analysis (ssmCCA) has emerged as a powerful technique for fusing electrical and hemodynamic responses to pinpoint brain regions consistently detected by both modalities [15]. This method evaluates multivariate associations between two types of high-dimensional data using canonical vectors or matrices, effectively identifying shared neural substrates that may be missed when analyzing each modality separately [15].

Additional analytical frameworks have been developed to leverage the complementary nature of multimodal data. The EEG-informed fNIRS analysis framework investigates neuro-correlates between neuronal activity and cerebral hemodynamics by identifying specific EEG rhythmic modulations that improve fNIRS-based General Linear Model (GLM) analysis [16]. Furthermore, researchers have proposed an improved Normalized-ReliefF method to fuse and optimize multi-modal features from EEG and fNIRS, which has demonstrated effectiveness in improving classification accuracy for distinguishing between brain states, achieving up to 98.38% accuracy in distinguishing brain activity evoked by preferred music versus neutral music [16].

For clinical applications, machine learning approaches have shown considerable promise in analyzing complex multimodal data. The Machine Learning for Visualizing AD (ML4VisAD) framework represents a novel approach that generates color-coded visual images reflecting disease progression at different time points based on baseline multimodal measurements [93]. This method takes inputs including neuroimaging data (MRI, PET), neuropsychological test scores, cerebrospinal fluid biomarkers, and risk factors to produce visual renderings that augment diagnostic and prognostic capabilities in Alzheimer's disease [93].

Advanced MRI Techniques for NVC Assessment

Beyond EEG-fNIRS integration, advanced magnetic resonance imaging (MRI) techniques provide additional powerful approaches for evaluating neurovascular coupling in patient populations. The combination of resting-state functional MRI (rs-fMRI) with arterial spin labeling (ASL) has emerged as a particularly valuable method for investigating NVC dysfunctional patterns [90] [91]. This integrated approach enables the calculation of NVC coefficients through spatial and temporal correlations between the amplitude of low-frequency fluctuation (ALFF) derived from BOLD signals and cerebral blood flow measurements from ASL [91].

In research practice, spatial correlation analysis examines the coupling between ALFF and CBF maps across the whole brain, providing a global index of neurovascular integrity [91]. Temporal correlation analysis, in contrast, assesses regional NVC patterns by evaluating the synchronized fluctuations between neuronal activity and hemodynamic responses within specific brain regions [91]. These methods have revealed characteristic NVC dysfunction in conditions such as cerebral small vessel disease and major depressive disorder, with distinct patterns correlating with disease severity and specific cognitive deficits [90] [91].

Other MRI-based approaches include the use of blood-oxygen-level-dependent (BOLD) fMRI combined with diffusion tensor imaging (DTI) to investigate relationships between functional activation, cerebrovascular reactivity, and white matter integrity in conditions characterized by neurovascular uncoupling. These multimodal MRI protocols offer the advantage of whole-brain coverage with high spatial resolution, complementing the portability and ecological validity of EEG-fNIRS systems.

Experimental Protocols for NVC Investigation

Protocol 1: Simultaneous fNIRS-EEG for Motor Paradigms

The investigation of neurovascular coupling during motor execution, observation, and imagery provides a well-established paradigm for studying NVC mechanisms in both healthy and clinical populations. The following protocol details a comprehensive approach for simultaneous fNIRS-EEG recording during these conditions [15]:

Equipment and Setup:

• A 24-channel continuous-wave fNIRS system (e.g., Hitachi ETG-4100) measuring changes in oxygenated hemoglobin (HbO) and deoxyhemoglobin (HbR) concentration at two wavelengths (695 nm and 830 nm) with a sampling rate of 10 Hz.
• A 128-electrode EEG system (e.g., Electrical Geodesics, Inc.) embedded within an elastic cap.
• fNIRS probes positioned over sensorimotor and parietal cortices to index Action Observation Network (AON) hemodynamic activity.
• Custom integration of fNIRS optodes and EEG electrodes within the same cap, with optode digitization using a 3D-magnetic space digitizer (Fastrak, Polhemus) to account for positioning variances.

Experimental Paradigm: Participants sit face-to-face with an experimenter across a table. The protocol includes three conditions:

• Motor Execution (ME): A pre-recorded audio command ("Your turn") prompts the participant to grasp, lift, and move a cup approximately two feet toward themselves using their right hand.
• Motor Observation (MO): The pre-recorded audio ("my turn") prompts the participant to watch the experimenter lift and move the cup in an identical manner.
• Motor Imagery (MI): The audio command ("Imagine") prompts the participant to mentally rehearse the action without physical movement.

Each condition is presented in randomized blocks with rest periods between trials. The entire session includes practice trials (approximately 5 minutes) followed by data collection.

Data Processing Pipeline:

• Preprocessing: Separate preprocessing of fNIRS and EEG data following modality-specific pipelines.
• Feature Extraction: HbO/HbR concentration changes from fNIRS; time-frequency features from EEG.
• Data Fusion: Application of structured sparse multiset CCA (ssmCCA) to identify neural substrates consistently detected by both modalities.
• Statistical Analysis: Comparison of activation patterns across conditions with appropriate multiple comparison corrections.

This protocol has demonstrated robust activation of the Action Observation Network across all three conditions, with differentiated patterns between execution, observation, and imagery conditions [15].

Protocol 2: Multimodal MRI for NVC Assessment in CSVD

For investigating neurovascular uncoupling in cerebral small vessel disease, the following combined resting-state fMRI and arterial spin labeling protocol has been validated [90]:

Participant Selection:

• Inclusion of CSVD patients based on standardized criteria (e.g., presence of white matter hyperintensities, lacunes, cerebral microbleeds).
• Age-matched healthy control participants.
• Comprehensive neuropsychological assessment covering multiple cognitive domains.

Image Acquisition Parameters (3T MRI):

• T1-weighted images: 3D structural sequence for anatomical reference and tissue segmentation.
• Resting-state fMRI: TR = 2000 ms, TE = 30 ms, 43 slices, slice thickness = 3.2 mm, eyes-closed resting state for 6-8 minutes.
• Arterial Spin Labeling: 3D ASL sequence, TR = 4635 ms, TE = 10.5 ms, post-labeling delay = 1525 ms, 40 slices, slice thickness = 4 mm.
• T2-FLAIR: For white matter hyperintensity segmentation and CSVD burden quantification.

Data Processing and NVC Metric Calculation:

• Preprocessing: Standard preprocessing pipelines for each modality including motion correction, normalization, and smoothing.
• CBF Calculation: Processing of ASL data to generate quantitative cerebral blood flow maps.
• ALFF Calculation: Processing of rs-fMRI data to generate amplitude of low-frequency fluctuation maps.
• NVC Quantification: Calculation of spatial correlation between CBF and ALFF maps across the whole brain and within specific regions of interest using the Human Brain Atlas with 246 brain regions.
• Statistical Analysis: Between-group comparisons of NVC coefficients, correlation with CSVD burden markers, and mediation analysis to examine relationships between NVC dysfunction and cognitive performance.

This protocol has successfully identified NVC dysfunction in CSVD patients, particularly in regions of the default mode network and subcortical nuclei, with the decreased NVC of the left superior frontal gyrus mediating the impact of white matter hyperintensities on delayed recall [90].

Table 3: Key Research Reagent Solutions for NVC Investigation

Category	Specific Tools/Reagents	Function/Application
Neuroimaging Systems	24-channel fNIRS (Hitachi ETG-4100); 128-electrode EEG (Electrical Geodesics)	Simultaneous hemodynamic and electrical neural activity recording [15]
Integrated Headgear	3D-printed custom helmets; Cryogenic thermoplastic sheets; Flexible EEG caps with fNIRS integration	Secure multimodal sensor placement; Subject-specific customization [27]
Analysis Software/Packages	Structured Sparse Multiset CCA; RESTplus toolbox; SPM12; ExploreASL	Multimodal data fusion; fMRI/ASL processing; Statistical analysis [90] [15] [91]
Experimental Paradigms	Motor execution/observation/imagery; Resting-state protocols; Sensory stimulation tasks	Elicitation of standardized neural activation for NVC assessment [15]
Molecular Probes	Endothelin receptor antagonists (e.g., macitentan); NO pathway modulators	Pharmacological dissection of NVC mechanisms; Therapeutic interventions [88]

The investigation of neurovascular uncoupling in patient populations represents a critical frontier in clinical neuroscience, with implications for early diagnosis, disease monitoring, and therapeutic development across a spectrum of neurological and psychiatric disorders. The integration of multimodal neuroimaging approaches, particularly the simultaneous acquisition of EEG and fNIRS, provides powerful tools for deciphering the complex dynamics of neurovascular dysfunction in conditions such as Alzheimer's disease, cerebral small vessel disease, and major depressive disorder.

Future research directions in this field include the development of more sophisticated analytical frameworks for multimodal data fusion, the validation of NVC biomarkers in large-scale longitudinal studies, and the integration of molecular imaging with electrophysiological and hemodynamic measurements. Additionally, the translation of NVC assessment from controlled laboratory settings to more ecologically valid environments represents an important frontier for improving the real-world applicability of these techniques. As methodological advancements continue to enhance our ability to characterize neurovascular uncoupling across patient populations, these approaches hold significant promise for transforming diagnosis, treatment monitoring, and therapeutic development for a wide range of neurological and psychiatric conditions characterized by neurovascular dysfunction.

Optimizing Experimental Paradigms for Enhanced Signal Quality

In the evolving landscape of non-invasive brain monitoring, the combined use of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a paradigm shift toward comprehensive neurovascular assessment. This integration capitalizes on their complementary strengths: EEG provides millisecond-level temporal resolution of electrical neural activity, while fNIRS tracks hemodynamic responses with superior spatial localization and resistance to motion artifacts [43] [28]. The fusion of these modalities offers unprecedented insights into brain function, particularly for naturalistic research scenarios and clinical applications such as neurorehabilitation and drug development [43] [81].

However, the synergistic potential of EEG-fNIRS can only be fully realized through rigorous optimization of experimental paradigms that address their distinct signal properties and technical requirements. This technical guide examines current methodologies, identifies persistent challenges, and provides evidence-based protocols for enhancing signal quality in multimodal neuroimaging research.

Hardware Integration and Signal Acquisition Protocols

Integrated System Configurations

Successfully capturing synchronized neural signatures requires careful hardware selection and configuration. Commercial solutions now offer specialized integrated systems designed specifically for concurrent EEG-fNIRS acquisition:

Table 1: Commercial EEG-fNIRS Integrated Systems

System Name	EEG Channels	fNIRS Configuration	Key Features	Research Applications
g.Nautilus NIRx [94]	Up to 64	32 optode holders	Wireless, portable, real-time synchronization	Naturalistic experiments, cognitive workload monitoring
g.HIamp NIRx [94]	256	32 detectors/sources	High-density EEG, comprehensive spatial coverage	Brain-computer interfaces, neurorehabilitation
Brite & APEX Combination [95]	Configurable	Integrated optodes	Shielded cables to minimize crosstalk, multiple holder options	Laboratory studies with controlled movement

Addressing Crosstalk Challenges

The proximity of EEG electrodes and fNIRS optodes introduces electromagnetic crosstalk, where fNIRS driving currents create artifacts in EEG recordings [95]. Recent empirical investigations demonstrate this challenge can be effectively mitigated through:

• Active Shielding: Artinis systems incorporate shielded EEG cables that reduce susceptibility to electromagnetic interference from fNIRS optodes [95].
• Impedance Management: Maintaining EEG electrode impedances below 5 kΩ significantly reduces crosstalk vulnerability, even when electrodes and optodes share holders [95].
• Spectral Separation: Configuring fNIRS systems to sample at frequencies above 50 Hz prevents overlap with clinically relevant EEG bands (delta: 0.5-4 Hz, theta: 4-7 Hz, alpha: 8-12 Hz, beta: 13-30 Hz) [28] [95].

Experimental validation confirms that with proper configuration, EEG recordings show no observable peaks at fNIRS firing frequencies (17.4 Hz or 37 Hz in tested systems), confirming signal integrity in combined setups [95].

Figure 1: Hardware optimization pathway for minimizing EEG-fNIRS crosstalk

Experimental Design Considerations

Paradigm Structure and Timing Parameters

Motor imagery paradigms effectively demonstrate the temporal considerations in multimodal experiments. The hemodynamic response measured by fNIRS evolves over seconds, while EEG captures millisecond-scale electrical fluctuations [81] [15]. This necessitates carefully balanced task designs:

Table 2: Temporal Parameters in Motor Imagery Paradigms

Study Reference	Task Duration	Inter-Trial Interval	Number of Trials	Rationale
HEFMI-ICH [81]	10 seconds	15 seconds	30 left/right hand MI each	Allows full hemodynamic response development while minimizing fatigue
Multi-joint MI Dataset [38]	4 seconds	10-12 seconds (randomized)	320 total (40 per task)	Prevents fNIRS response overlap between trials; randomization reduces anticipation artifacts
Motor Execution/Observation/Imagery [15]	Self-paced	Variable	60 total (20 per condition)	Ecologically valid design for natural action processing

The HEFMI-ICH protocol exemplifies optimized timing: a 2-second visual cue presentation followed by a 10-second execution phase and 15-second inter-trial interval [81]. This structure accommodates the slow hemodynamic response while capturing event-related potentials in EEG.

Participant Preparation and Task Instruction

Standardized participant preparation significantly enhances signal quality and task compliance:

• Grip Strength Calibration: The HEFMI-ICH protocol incorporates dynamometer and stress ball exercises before motor imagery tasks to reinforce kinesthetic sensation and improve imagery vividness [81].
• Comprehensive Instruction: Some participants struggle with the abstract concept of motor imagery. Explicitly distinguishing between kinesthetic sensation (feeling the movement) and visual imagery (picturing the movement) improves paradigm validity [81].
• Artifact Mitigation Training: Instructing participants to suppress blinking during task periods and swallow only during rest periods minimizes ocular and muscle artifacts in EEG [38].

Figure 2: Optimized experimental timeline for simultaneous EEG-fNIRS acquisition

Signal Quality Validation and Artifact Management

Quality Metrics and Assessment Protocols

Establishing quantitative signal quality benchmarks ensures consistent data integrity across recording sessions:

• EEG Quality Standards: Electrode impedances should be maintained below 10 kΩ throughout experiments, with ideal performance achieved below 5 kΩ for integrated EEG-fNIRS setups [95] [38].
• fNIRS Optode Placement: Source-detector distances of 3 cm optimize the sensitivity to cerebral hemodynamics while minimizing scalp contribution [81].
• Synchronization Validation: Temporal alignment between EEG and fNIRS recording systems should be verified using simultaneous event markers transmitted to both systems [81].

The Artinis validation protocol provides a methodology for quantifying crosstalk: recording EEG with fNIRS systems alternating between "on" and "off" states, then computing power spectral density to identify interference peaks [95].

Artifact Removal Techniques

Multimodal recordings require tailored artifact removal strategies:

• EEG Artifact Handling: Robust artifact removal techniques for EEG are well-established, including regression methods, independent component analysis, and advanced filtering [43].
• fNIRS Motion Correction: Motion artifacts in fNIRS remain challenging, with limited adoption of confounder correction beyond filtering or basic motion removal [43].
• Underutilized Methods: Short-separation measurements and other auxiliary signals for fNIRS artifact correction remain underutilized despite their potential benefits [43].

Data Fusion and Analytical Approaches

Multimodal Integration Strategies

Data-driven fusion methods extract synergistic information from combined EEG-fNIRS recordings:

• Concatenation Approaches: Simple concatenation of features from both modalities remains prevalent, offering straightforward implementation but limited interaction modeling [43].
• Source Decomposition: Techniques like structured sparse multiset Canonical Correlation Analysis (ssmCCA) identify latent variables that maximize correlation between electrical and hemodynamic data modalities [15].
• Decision-Level Fusion: Separate classification of EEG and fNIRS signals with subsequent integration of outcomes provides robustness to modality-specific artifacts [43].

In motor execution, observation, and imagery studies, ssmCCA fusion consistently identified activation in the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus across conditions—regions associated with the Action Observation Network that were not consistently detected by unimodal analyses [15].

Classification Performance in Hybrid Systems

Empirical evidence demonstrates the tangible benefits of multimodal integration:

• Accuracy Improvements: Hybrid EEG-fNIRS systems demonstrate 5%-10% improved classification accuracy compared to unimodal systems in normal subjects [81].
• Population-Specific Considerations: The performance advantage of hybrid systems may differ in clinical populations such as intracerebral hemorrhage patients, where neurovascular uncoupling alters signal dynamics [81].
• Complementary Strength Utilization: fNIRS provides superior spatial localization of motor cortex activation (5-10 mm resolution), while EEG captures the precise timing of event-related desynchronization [81] [96].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Critical Components for EEG-fNIRS Experimental Setup

Component Category	Specific Examples	Function	Implementation Considerations
Acquisition Systems	g.Nautilus NIRx, g.HIamp NIRx, NIRScout, Brite & APEX	Simultaneous recording of electrical and hemodynamic activity	Wireless systems preferred for naturalistic paradigms; channel count should match spatial resolution requirements
Headgear & Electrodes	g.GAMMAcap, combined optode-electrode holders, sintered Ag/AgCl electrodes	Secure sensor placement with consistent scalp contact	Combined holders reduce crosstalk; active electrodes preferred for high-impedance situations
Conductive Materials	NeuroPrep gel, Ten20 paste	Optimize electrode-scalp interface conductivity	Gel reduces impedance but requires cleaning; paste offers longer stability
Synchronization Tools	E-Prime 3.0, custom triggers	Temporal alignment of multimodal data	Event markers should be simultaneously transmitted to all recording systems
Data Validation Tools	BIDS validator, power spectral analysis, synthetic datasets	Verify data quality and format compliance	BIDS format enables data sharing and reproducibility

Optimizing experimental paradigms for simultaneous EEG-fNIRS recording requires meticulous attention to hardware integration, temporal paradigm design, artifact management, and analytical fusion techniques. The protocols and methodologies outlined provide a framework for enhancing signal quality in multimodal neuroimaging research. As the field evolves, increased adoption of standardized validation procedures, public datasets with shared formatting conventions like BIDS [97], and advanced source decomposition methods will further strengthen the validity and reproducibility of findings. When implemented systematically, these optimized approaches unlock the full potential of multimodal neuroimaging to illuminate complex brain dynamics in both fundamental research and clinical applications.

Benchmarking Performance and Clinical Validation: fNIRS-EEG Against Established Modalities

Functional near-infrared spectroscopy (fNIRS) has emerged as a portable and cost-efficient neuroimaging technology that measures cortical brain activity through hemodynamic responses. Unlike functional magnetic resonance imaging (fMRI), the gold standard for functional neuroimaging, fNIRS provides superior tolerance to motion artifacts and can be deployed in more naturalistic settings [98]. However, establishing the validity of fNIRS-derived signals through direct comparison with fMRI remains a critical step for legitimizing its application in both basic neuroscience and clinical practice. This technical guide examines the spatial and temporal correspondence between fNIRS and fMRI hemodynamic responses, detailing the experimental protocols and analytical frameworks that enable effective multimodal integration.

The neurophysiological basis for comparing these modalities lies in their shared dependence on neurovascular coupling. fMRI detects changes in blood oxygenation level-dependent (BOLD) contrast, which is sensitive to variations in deoxygenated hemoglobin [98]. fNIRS directly measures concentration changes in both oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) through differential absorption of near-infrared light [99]. While these signals originate from the same underlying hemodynamic processes, their relationship is complex due to differences in sensitivity, physiological confounds, and measurement physics.

Neurovascular Coupling and Signal Relationships

Physiological Basis of Hemodynamic Responses

The relationship between fNIRS and fMRI signals stems from their shared foundation in the hemodynamic response to neural activity. The balloon model provides a theoretical framework for understanding the interplay between cerebral blood flow (CBF), cerebral blood volume (CBV), and the cerebral metabolic rate of oxygen consumption (CMRO₂) [98]. During neural activation, localized increases in CBF typically exceed oxygen metabolic demands, resulting in an initial dip in HbR followed by a substantial overshoot in HbO concentration. The fMRI BOLD signal primarily reflects changes in HbR concentration, though not exclusively, as it is also influenced by CBV and CBF changes [98].

Figure 1: Neurovascular coupling pathway showing the relationship between neural activity and the hemodynamic signals measured by fNIRS and fMRI.

Signal Characteristics Across Modalities

The temporal and spatial characteristics of fNIRS and fMRI differ substantially. fNIRS provides higher temporal resolution (typically 1-10 Hz) compared to fMRI (typically 0.3-1 Hz), allowing for more detailed capture of the hemodynamic response shape [27]. However, fNIRS suffers from limited spatial resolution and penetration depth, primarily sampling cortical surfaces, whereas fMRI provides whole-brain coverage with superior spatial localization [98]. The concordance between fNIRS and fMRI varies across brain regions and depends on physiological factors, including superficial scalp blood flow that can contaminate fNIRS signals.

Quantitative Comparison of fNIRS and fMRI Signals

Spatial Correspondence Evidence

Multiple studies have systematically investigated the spatial overlap between fNIRS and fMRI activation patterns during controlled tasks. Research using motor execution and imagery paradigms has demonstrated significant spatial correspondence in primary motor cortices (M1) and premotor cortices (PMC) [98]. In one multimodal assessment, subject-specific fNIRS cortical signals were used to identify corresponding activation clusters in fMRI data, with significant peak activation found overlapping individually-defined primary and premotor cortices for all chromophores (HbO, HbR, and total hemoglobin) [98].

Table 1: Spatial Correspondence Between fNIRS and fMRI Across Brain Regions

Brain Region	Task Paradigm	Spatial Correlation	Highest Correlating Chromophore	Reference
Primary Motor Cortex	Motor Execution	High	HbO, HbR (equivalent)	[98]
Premotor Cortex	Motor Imagery	High	HbO, HbR (equivalent)	[98]
Prefrontal Cortex	Working Memory	Moderate-High	HbO	[100]
Parietal Cortex	Action Observation	Moderate	HbO	[15]

Notably, no statistically significant differences have been observed in multimodal spatial correspondence between HbO, HbR, and total hemoglobin for both motor execution and imagery tasks [98]. This suggests the possibility of translating neuronal information from fMRI into fNIRS setups using both oxy- and deoxyhemoglobin data, with important implications for translating well-established fMRI paradigms to fNIRS applications in cognitive and clinical neuroscience.

Temporal Correspondence Evidence

Temporal correlations between fNIRS and fMRI signals vary considerably across studies and experimental conditions. A review of concurrent fMRI-fNIRS studies concluded that temporal correlation with the BOLD contrast was most consistently observed with HbR changes, though not exclusively dependent on this chromophore [98]. Some studies have reported higher temporal correspondence with total hemoglobin, while others found similar correlation levels between HbO and HbR [98].

Table 2: Temporal Correlation Ranges Between fNIRS and fMRI Signals

fNIRS Chromophore	Correlation Range with fMRI BOLD	Factors Influencing Correlation	Typical Lag Observations
HbO (Oxyhemoglobin)	0 to 0.8	Signal-to-noise ratio, region specificity	Variable, generally minimal
HbR (Deoxyhemoglobin)	-0.76 to 0	Physiological noise, scalp influence	More consistent than HbO
HbT (Total Hemoglobin)	Moderate-High	CBV contribution	Less studied

More recent studies have reported high levels of temporal correlation between fMRI BOLD signal and both HbO (r = 0.65) and HbR (r = -0.76), while others have reported mean correlations as low as |r| ~ 0.2 [98]. This variability highlights the impact of analytical approaches, experimental design, and signal processing pipelines on observed correlations.

Methodological Frameworks for Multimodal Validation

Experimental Design Considerations

Effective validation of fNIRS against fMRI requires carefully controlled experimental paradigms that accommodate the constraints of both modalities. Block designs are commonly employed, with motor tasks being particularly prevalent due to their robust and well-characterized hemodynamic responses [98] [99]. For example, a typical motor paradigm might include alternating 30-second blocks of activity (e.g., finger tapping) and rest, repeated multiple times [99].

Advanced paradigms have incorporated more complex conditions such as motor execution, motor observation, and motor imagery within the same study design [15]. These approaches allow researchers to probe different aspects of the action observation network while comparing hemodynamic responses across modalities. When designing such experiments, consideration must be given to the differential sensitivity of each modality to motion artifacts, with fNIRS generally offering greater tolerance to movement [15].

Data Acquisition Protocols

Simultaneous fNIRS-fMRI acquisition provides the most direct method for comparison, though technical challenges related to electromagnetic interference and hardware compatibility must be addressed [99]. Asynchronous acquisition with careful maintenance of identical experimental parameters presents a viable alternative [98].

fMRI Acquisition Parameters: Typical parameters for motor task studies include: 3T scanner, echo-planar imaging sequence with TR/TE = 1500-3000/30-35 ms, in-plane resolution 3×3 mm, slice thickness 3-4 mm [98] [99]. Coverage is typically focused on regions of interest while maintaining whole-brain acquisition for comprehensive analysis.

fNIRS Acquisition Parameters: Standard setups include 16-24 sources and 15-24 detectors with wavelengths of 760 nm and 850 nm, sampling at 5.08-10.25 Hz [98] [101]. Optode placement follows the international 10-20 system, with source-detector distances of 30 mm for cerebral measurements and 8 mm for short-distance channels to correct for superficial confounds [98].

Figure 2: Multimodal fNIRS-fMRI experimental workflow showing the stages from data acquisition through validation analysis.

Analytical Approaches for Multimodal Fusion

Several analytical frameworks have been developed to quantify the relationship between fNIRS and fMRI signals:

Joint Independent Component Analysis (jICA): This approach identifies linked components between fNIRS temporal patterns and fMRI spatial maps, enabling inferences about associations between modality-specific sources [99]. The jICA model can be represented as:

[ \begin{bmatrix} X{fNIRS} \ X{fMRI} \end{bmatrix} = A \begin{bmatrix} S{fNIRS} \ S{fMRI} \end{bmatrix} ]

where X represents the observed data, A is the mixing matrix, and S contains the independent sources [99].

Spatiotemporal Fusion: This method generates dynamic "snapshots" of brain activity by combining fMRI spatial components weighted by their joint fNIRS time courses [99]. The resulting visualizations enable examination of the dynamic interplay between fNIRS and fMRI measurements across both space and time.

General Linear Model (GLM) Approaches: These models use subject-specific fNIRS cortical signals as predictors for fMRI data, testing the ability of fNIRS to identify corresponding brain regions in previously acquired fMRI data [98]. This approach has successfully identified group-level activation in fMRI data modeled from corresponding fNIRS measurements.

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Tools for fNIRS-fMRI Validation Studies

Tool Category	Specific Examples	Function/Purpose	Technical Notes
fNIRS Hardware	NIRSport2 (NIRx), NIRScout (NIRx), ETG-4100 (Hitachi)	Measures HbO/HbR concentration changes	16-24 sources, 15-24 detectors, 760/850 nm wavelengths [98] [101]
fMRI Systems	3T Siemens Magnetom, 3T ISOL MRI	Acquires BOLD contrast images	EPI sequence, TR/TE ~1500-3000/30-35 ms [98] [99]
Analysis Software	BrainVoyager QX, Homer3, SPM, NIRS-SPM	Preprocessing and statistical analysis	Homer3 for fNIRS preprocessing; BrainVoyager for fMRI analysis [98]
Multimodal Integration Tools	jICA Toolbox, NIRS-fMRI Fusion Pipeline	Data fusion and joint analysis	Identifies spatiotemporal correspondences [99]
Experimental Control	PsychoPy, Presentation, E-Prime	Paradigm presentation and synchronization	Precise timing critical for temporal correlation [102]
Head Localization	3D Digitizers (Polhemus Fastrak)	Optode/electrode co-registration	Ensures accurate spatial mapping [15]

Applications and Clinical Translation

The validation of fNIRS against fMRI has enabled the translation of neuroimaging applications to more naturalistic settings and clinical populations. In stroke rehabilitation, fNIRS has been used to monitor cortical activation patterns during digital therapeutic exercises, with activation in ipsilesional primary motor cortex correlating with motor recovery outcomes [101]. In cognitive neuroscience, fNIRS has been applied to study working memory deficits in mild cognitive impairment, with reduced activation in dorsolateral prefrontal cortex during 1-back tasks distinguishing patients from healthy controls [100].

The combination of fNIRS with EEG provides additional complementary information, with EEG offering millisecond temporal resolution to complement fNIRS hemodynamic measures [15] [27]. Structured sparse multiset Canonical Correlation Analysis (ssmCCA) has been used to fuse fNIRS and EEG data, consistently identifying activation in the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus during motor execution, observation, and imagery [15].

The convergence of evidence from multiple methodological approaches demonstrates a robust spatial and temporal correspondence between fNIRS and fMRI hemodynamic responses. While variations exist across brain regions, experimental paradigms, and analytical techniques, the overall pattern supports the validity of fNIRS as a reliable measure of task-related brain activity. The continued refinement of multimodal integration frameworks, including joint ICA and spatiotemporal fusion algorithms, promises to further enhance the quantitative relationship between these complementary modalities. As fNIRS technology becomes increasingly accessible and portable, its validated correspondence with the fMRI gold standard positions it as a powerful tool for extending neuroimaging beyond traditional laboratory settings into clinical environments and real-world applications.

Motor Imagery (MI), the mental rehearsal of a motor act without any physical movement, has emerged as a critical paradigm in neuroscience and neurorehabilitation. As a component of the shared Action Observation Network (AON), MI activates brain regions similar to those involved in actual motor execution, including the premotor cortex, supplementary motor area, and parietal regions [15]. The decoding of MI signals using electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) provides a window into neural processing across different populations, offering unique insights for both basic research and clinical applications. This technical guide examines the fundamental differences in MI decoding between healthy individuals and patient populations, with particular emphasis on methodological considerations for multimodal neuroimaging research.

Multimodal approaches that integrate EEG and fNIRS are particularly valuable for MI research, as they combine EEG's excellent temporal resolution with fNIRS's superior spatial localization capabilities [20] [15]. This integration allows researchers to capture both the rapid electrophysiological changes and the underlying hemodynamic responses associated with motor imagery processes. The fusion of these complementary signals through advanced analytical techniques like structured sparse multiset Canonical Correlation Analysis (ssmCCA) enables more precise identification of activated brain regions and provides a more comprehensive understanding of neural mechanisms underlying MI in diverse populations [15].

Neural Correlates of Motor Imagery Across Populations

Core Brain Networks in Motor Imagery

The neural substrates of motor imagery involve a distributed network of brain regions that facilitate the planning, preparation, and mental simulation of movements. Research using simultaneous fNIRS-EEG recordings has consistently identified activation in the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus during MI tasks across populations [15]. These regions form part of the Action Observation Network (AON), which is recruited during motor execution, observation, and imagery tasks. The premotor area (PMA) and primary motor cortex (MA) show significant positive correlation during hand MI in both healthy individuals and stroke patients, though the strength and pattern of connectivity differ between populations [103].

The PMA plays a particularly crucial role in movement preparation and planning, serving as a crucial component of motor control [103]. During MI, the PMA demonstrates greater power spectral density values in EEG signals compared to the primary motor cortex in both healthy and stroke populations [103]. This differential activation highlights the hierarchical organization of motor processing, where higher-order motor areas take precedence during imagined movements without actual execution.

Population-Specific Activation Patterns

Table 1: Comparative Neural Activation During Motor Imagery in Different Populations

Population	Key Activated Regions	Connectivity Patterns	Special Considerations
Healthy Adults	Bilateral central, right frontal, and parietal regions (EEG); Angular gyrus, supramarginal gyrus, superior/inferior parietal lobes (fNIRS) [15]	Normal functional connectivity between PMA and MA [103]	Stable activation patterns across sessions; minimal network reorganization
Stroke Patients	Enhanced ipsilateral activation; Abnormal connectivity in left PMA area (lead 18) [103]	Increased functional connectivity correlation during MI, especially left-handed MI [103]	Connectivity more enhanced during synergy movements than isolated movements [103]
Chronic Pain Patients	Reduced cortical areas on affected side; Altered premotor cortex activity [104]	Less distinguishable activation patterns due to cortical reorganization [104]	Pain-induced changes in motor cortex excitability

Methodological Approaches in Motor Imagery Research

Experimental Paradigms and Protocols

Motor imagery studies employ standardized experimental paradigms to elicit comparable neural responses across participants and sessions. A typical MI experiment consists of multiple blocks with randomized trials of different imagined movements. Each trial generally follows a structured timeline: a brief instruction period (1-2 seconds) indicating which movement to imagine, followed by the MI period itself (typically 4 seconds), and ending with a rest interval (2-4 seconds) [105] [106]. Participants are instructed to mentally rehearse the specified movement during the imagery period without executing any physical motion.

For upper limb MI, common tasks include left or right hand grasping movements [105] [106], while lower limb paradigms often involve foot movements or leg flexion/extension [104]. The specific instructions and cueing methods vary across studies, with some employing visual cues (videos or images of the movement) [106] and others using auditory instructions [15]. The number of trials per session typically ranges from 40-100, distributed across multiple blocks with rest periods to maintain participant attention and minimize fatigue [105] [104].

Data Acquisition and Preprocessing

Table 2: Technical Specifications for Multimodal MI Data Acquisition

Parameter	EEG Specifications	fNIRS Specifications
Equipment Type	64-channel wireless systems (e.g., Neuracle) [105] or portable saline-based devices [106]	Continuous-wave systems (e.g., Hitachi ETG-4100) [15]
Electrode/Optode Placement	International 10-20 or 10-10 system [105] [106]	Bilateral coverage of sensorimotor and parietal cortices [15]
Key Regions of Interest	Premotor area (FC5, FC3, FC1, FC2, FC4, FC6); Primary motor cortex (C5, C3, C1, C2, C4, C6) [103]	Primary motor cortex, premotor cortex, parietal lobule [15]
Sampling Rate	500 Hz [106] or 100 Hz after down-sampling [103]	10 Hz [15]
Filtering	Bandpass filtering (8-30 Hz) for event-related desynchronization analysis [103]	Not typically specified in studies
Additional Measures	Independent component analysis for artifact removal [103]	Digitization of optode positions for spatial registration [15]

Comparative Case Studies: Healthy vs. Patient Populations

Stroke Population Case Study

A 2024 study with 23 stroke inpatients and 21 healthy controls revealed significant differences in functional connectivity during motor imagery tasks [103]. The research demonstrated that stroke patients exhibited higher functional connectivity correlation during MI compared to healthy individuals, particularly during left-handed motor imagery. The regions with abnormal functional connectivity were localized to the 18th lead in the left premotor area (PMA), suggesting this area may serve as a target for non-invasive neuromodulation therapies [103].

For acute stroke patients (within 30 days post-stroke), a dataset of 50 patients performing hand-grip MI tasks showed the feasibility of classifying left vs. right hand imagery with decoding accuracy of 72.21% using specialized algorithms (TWFB + DGFMDM) [106]. This demonstrates that despite potential cortical damage, residual MI capabilities remain that can be leveraged for rehabilitation. The study employed 29 EEG recording electrodes with 2 EOG electrodes, focusing on sensorimotor rhythms in the alpha (8-15 Hz) and beta (15-30 Hz) frequency bands that reflect distinct brain activities [106].

Chronic Pain Population Case Study

Research with 30 knee osteoarthritis pain patients performing lower-limb MI tasks revealed unique challenges in decoding MI signals from this population [104]. Chronic pain induces cortical reorganization that alters brain function and affects the brain's ability to process motor intentions. Patients showed reduced cortical activation areas during MI on the affected side, along with changes in premotor cortex activity [104].

Traditional decoding algorithms achieved modest accuracies of 51.43-76.21% in classifying left vs. right leg imagery in this population, while a novel Optimal Time-Frequency Window in Riemannian Geometric Distance Classification Algorithm (OTFWRGD) reached significantly higher average accuracy of 86.41% [104]. This highlights both the challenges of MI decoding in pain populations and the potential of specialized algorithms to overcome these limitations. The study collected data over five independent sessions with 100 trials per session, generating 15,000 total trials for analysis [104].

Healthy Population Benchmarking

Large-scale datasets from healthy populations provide important benchmarks for MI decoding performance. A comprehensive MI dataset from 62 healthy participants across three recording sessions demonstrated average classification accuracy of 85.32% for two-class tasks (left vs. right hand-grasping) using EEGNet, and 76.90% for three-class tasks (adding foot-hooking) using DeepConvNet [105]. This dataset, collected during the 2019 World Robot Conference Contest, highlights the high-quality signals achievable with naive BCI users under controlled conditions.

The stability of MI patterns across multiple sessions in healthy individuals contrasts with the more variable patterns observed in patient populations. Healthy participants typically show progressive improvement in MI ability after multiple sessions, while patient populations may exhibit either compensatory reorganization or deficit-related variations depending on their specific condition [105].

Analytical Approaches and Decoding Techniques

Signal Processing and Feature Extraction

The processing of MI signals involves multiple stages to extract meaningful features from raw EEG and fNIRS data. For EEG signals, preprocessing typically includes filtering (8-30 Hz range to capture alpha and beta bands associated with sensorimotor rhythms), down-sampling to 100 Hz, and independent component analysis to remove artifacts and improve signal quality [103]. Power Spectral Density (PSD) values are then calculated using transforms like the modified S-Transform (MST) to analyze signal power variations across frequencies [103].

Event-Related Desynchronization/Synchronization (ERD/ERS) patterns serve as key features for MI decoding, particularly in the alpha (8-15 Hz) and beta (15-30 Hz) frequency bands [104]. These patterns represent decreases or increases in oscillatory power relative to baseline during motor imagery. The Common Spatial Pattern (CSP) technique is widely used for feature extraction, providing optimal spatial filters for discriminating between different MI tasks [106] [104]. For multimodal integration, techniques like structured sparse multiset Canonical Correlation Analysis (ssmCCA) fuse fNIRS and EEG data to identify brain regions consistently detected by both modalities [15].

Classification Algorithms and Performance

Table 3: Motor Imagery Decoding Performance Across Populations and Algorithms

Population	Task	Algorithm	Performance	Notes
Healthy [105]	Left vs. right hand grasping (2-class)	EEGNet	85.32% accuracy	Average across 62 subjects, 3 sessions
Healthy [105]	Left/right hand grasping + foot hooking (3-class)	DeepConvNet	76.90% accuracy	Average across subjects and sessions
Acute Stroke [106]	Left vs. right hand imagery	TWFB + DGFMDM	72.21% accuracy	50 patients, specialized algorithm
Knee Pain [104]	Left vs. right leg flexion/extension	Traditional Algorithms	51.43-76.21% accuracy	30 patients, varying by algorithm
Knee Pain [104]	Left vs. right leg flexion/extension	OTFWRGD	86.41% accuracy	Novel specialized algorithm

The Scientist's Toolkit: Essential Research Reagents

Table 4: Key Research Tools and Solutions for Motor Imagery Studies

Tool Category	Specific Examples	Function/Purpose
Neuroimaging Hardware	64-channel wireless EEG systems (Neuracle) [105]; Portable saline-based EEG devices (ZhenTec NT1) [106]; Continuous-wave fNIRS systems (Hitachi ETG-4100) [15]	Signal acquisition with specific spatial/temporal resolution characteristics
Experimental Paradigm Software	Custom MATLAB/Python scripts; Presentation software for visual cues; BCI2000 platform [104]	Precisely timed stimulus presentation and trial structure implementation
Signal Processing Tools	EEGLAB toolbox for MATLAB [106]; Modified S-Transform for PSD calculation [103]; Independent Component Analysis algorithms	Data preprocessing, artifact removal, and initial feature extraction
Analysis Algorithms	Common Spatial Patterns (CSP) [106] [104]; Structured Sparse Multiset CCA (ssmCCA) [15]; EEGNet, DeepConvNet [105]	Feature optimization, multimodal data fusion, and classification
Validation Metrics	Classification accuracy; Event-Related (De)Synchronization indices; Functional connectivity measures [103]	Quantification of decoding performance and neural activation patterns

Discussion and Future Directions

The comparative analysis of motor imagery decoding across healthy and patient populations reveals both significant challenges and promising opportunities for neurorehabilitation. The consistent finding of altered functional connectivity in patient populations—whether increased as in stroke [103] or reduced/distributed as in chronic pain [104]—suggests that successful decoding algorithms must accommodate population-specific neural reorganization. The enhanced functional connectivity observed in stroke patients during MI, particularly in the premotor areas, may represent compensatory mechanisms that could be leveraged for more effective rehabilitation protocols.

Future research directions should focus on the development of adaptive decoding algorithms that can accommodate the unique neural signatures of different patient populations. The success of specialized algorithms like OTFWRGD for knee pain patients [104] and TWFB + DGFMDM for stroke patients [106] demonstrates the limitations of one-size-fits-all approaches to MI decoding. Additionally, larger standardized datasets spanning multiple patient populations and recording sessions will be essential for advancing the field. The collection of data across multiple independent sessions, as demonstrated in several recent studies [105] [104], provides valuable insights into the stability of MI patterns over time and the potential for longitudinal tracking of rehabilitation progress.

The integration of multimodal neuroimaging approaches, particularly simultaneous EEG-fNIRS recording, represents a promising avenue for improving decoding accuracy across populations [20] [15]. By combining complementary information from electrical and hemodynamic signals, researchers can obtain a more comprehensive picture of neural activity during motor imagery. This approach may be particularly valuable for patient populations where traditional unimodal signals may be compromised by pathological cortical changes or medications. As these technologies become more accessible and analytical techniques more sophisticated, personalized MI-based rehabilitation protocols based on population-specific and even individual-specific neural signatures may become feasible, potentially revolutionizing neurorehabilitation for diverse patient populations.

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a promising direction in multimodal neuroimaging, combining EEG's millisecond-scale temporal resolution with fNIRS's superior spatial localization of hemodynamic activity [2]. This complementary relationship enables a more comprehensive decoding of brain activity, which is crucial for applications in brain-computer interfaces (BCIs), clinical neurology, and neurorehabilitation [2] [3]. However, the high-dimensional nature of combined EEG and fNIRS data introduces significant computational challenges, including the "curse of dimensionality," increased model complexity, longer training times, and reduced generalization capability [107] [108]. This technical guide explores how hybrid AI-driven systems, employing sophisticated feature selection and fusion strategies, can optimize feature selection to achieve substantial improvements in classification accuracy within multimodal neuroimaging research.

Theoretical Foundations: Why Hybrid Systems Improve Accuracy

The Role of Feature Selection in High-Dimensional Data

Feature selection (FS) is critical for datasets with multiple variables, as it helps eliminate irrelevant elements, thereby improving classification accuracy [107]. In the context of multimodal EEG-fNIRS, FS achieves four primary objectives:

• Reduces Model Complexity: By minimizing the number of parameters, FS creates simpler, more interpretable models [107] [108].
• Decreases Training Time: Fewer features accelerate the model training process [108].
• Enhances Generalization: By reducing overfitting, FS improves model performance on unseen data [107].
• Avoids the Curse of Dimensionality: FS mitigates the performance degradation that occurs in high-dimensional space [107] [108].

Neurovascular Coupling as a Basis for Multimodal Fusion

EEG and fNIRS are linked via neurovascular coupling (NVC), which relates transient neural activity to subsequent hemodynamic changes [2]. While EEG captures synchronous neuro-electrical activity, fNIRS measures slow changes in cerebral blood flow. This biological relationship provides the foundation for information fusion, as the modalities are sensitive to complementary aspects of brain activity and different types of physiology and measurement artifacts [2].

Experimental Protocols and Hybrid Methodologies

Hybrid Feature Selection Algorithms

Recent research has introduced several hybrid algorithms for identifying significant features in high-dimensional datasets. The performance of these algorithms has been systematically compared across multiple datasets [107] [108].

Table 1: Hybrid Feature Selection Algorithms for Neuroimaging Data

Algorithm	Full Name	Key Mechanism	Advantages
TMGWO	Two-phase Mutation Grey Wolf Optimization	Incorporates a two-phase mutation strategy to enhance exploration-exploitation balance [108]	Achieved superior results in both feature selection and classification accuracy [107]
ISSA	Improved Salp Swarm Algorithm	Utilizes adaptive inertia weights, elite salps, and local search techniques [108]	Significantly boosts convergence accuracy [108]
BBPSO	Binary Black Particle Swarm Optimization	Velocity-free mechanism that preserves global search efficiency [108]	Offers simplicity and improved computational performance [108]

Multimodal Fusion Strategies for EEG-fNIRS

The integration of EEG and fNIRS signals can be implemented through various architectural approaches:

Table 2: Data Fusion Strategies for EEG-fNIRS Hybrid Systems

Fusion Strategy	Implementation	Advantages	Limitations
Data Concatenation	Combining raw or pre-processed signals from both modalities into a single feature vector [2]	Simple to implement; preserves all original information	Can result in high-dimensional data; ignores modality-specific characteristics
Model-Based Fusion	Using structured algorithms to model the relationship between EEG and fNIRS signals [2]	Can capture complex, non-linear relationships between modalities	Requires precise knowledge of neurovascular coupling parameters
Decision-Level Fusion	Combining outputs from separate EEG and fNIRS classifiers [2]	Leverages modality-specific expertise; more robust to single-modality failure	May miss important cross-modal interactions
Source Decomposition	Using unsupervised symmetric techniques to extract latent components [2]	Reveals complex neurovascular coupling processes; does not require stimulus timing	Underutilized in current research; requires sophisticated implementation

Experimental Protocol: Multimodal Neurofeedback for Motor Imagery

A representative experimental protocol for evaluating hybrid EEG-fNIRS systems involves motor imagery (MI) tasks with neurofeedback (NF). The following workflow illustrates a standardized approach [3]:

Diagram 1: Neurofeedback Experimental Workflow

This protocol enables direct comparison of unimodal versus multimodal approaches, with the hypothesis that presenting participants with visual NF based on both EEG and fNIRS signals will result in more specific task-related brain activity in the sensorimotor cortices [3].

Performance Metrics and Evaluation Framework

Key Classification Metrics for Imbalanced Neuroimaging Data

Evaluating hybrid systems requires moving beyond simple accuracy to more nuanced metrics that account for class imbalance common in neuroimaging data [109] [110].

Table 3: Classification Metrics for Evaluating Hybrid System Performance

Metric	Formula	Interpretation	Use Case
Accuracy	(TP+TN)/(TP+TN+FP+FN)	Overall correctness of the model	Useful for balanced datasets; coarse-grained measure of quality [109]
Precision	TP/(TP+FP)	Proportion of positive predictions that are correct	When false positives are costly; ensures positive predictions are reliable [109]
Recall (TPR)	TP/(TP+FN)	Proportion of actual positives correctly identified	When false negatives are critical; e.g., disease detection [109]
F1 Score	2TP/(2TP+FP+FN)	Harmonic mean of precision and recall	Balanced measure for imbalanced datasets [109]

Quantitative Performance of Hybrid Approaches

Experimental results demonstrate the significant impact of hybrid feature selection on classification performance:

Table 4: Comparative Performance of Hybrid FS Algorithms on Medical Datasets

Dataset	Algorithm	Accuracy	Precision	Recall	Key Findings
Breast Cancer Wisconsin	TMGWO-SVM	96.0%	Not Reported	Not Reported	Outperformed TabNet (94.7%) and FS-BERT (95.3%) using only 4 features [108]
Diabetes Early Diagnosis	TMGWO-KNN + SMOTE	98.85%	Not Reported	Not Reported	Provided greater accuracy with less computation time than using all features [108]
General High-Dimensional Data	BBPSO with adaptive chaotic jump	Improved vs. Baselines	Not Reported	Not Reported	Better discriminative feature selection and classification performance [108]
Feature Selection Challenges	CHPSODE (PSO + Differential Evolution)	Reliable Solutions	Not Reported	Not Reported	Balanced exploration and exploitation for realistic FS solutions [108]

Implementation Framework for Hybrid Systems

Architectural Framework for Performance Optimization

The following diagram illustrates the integrated workflow for achieving classification accuracy improvements in hybrid EEG-fNIRS systems:

Diagram 2: Hybrid System Optimization Framework

The Researcher's Toolkit: Essential Materials and Solutions

Implementation of hybrid EEG-fNIRS systems requires specific hardware, software, and analytical components.

Table 5: Essential Research Toolkit for Hybrid EEG-fNIRS Systems

Component	Specification	Function/Purpose
Wearable EEG System	High-density, portable systems with multiple electrodes [2]	Captures electrical brain activity with millisecond resolution
Wearable fNIRS/HD-DOT	Fiberless continuous wave systems with source-detector pairs [2]	Measures hemodynamic changes in cortical regions
Integrated EEG-fNIRS Cap	Custom caps with integrated sensors positioned over sensorimotor cortices [3]	Enables simultaneous multimodal data acquisition
Real-time Signal Processing	Custom software for NF calculation and visualization [3]	Processes signals and provides real-time feedback
Hybrid FS Algorithms	TMGWO, ISSA, BBPSO implementations [107] [108]	Identifies most relevant features from high-dimensional data
Classification Algorithms	SVM, Random Forest, MLP, KNN, Logistic Regression [107]	Provides benchmarked classification performance
Validation Metrics	Accuracy, Precision, Recall, F1 Score calculations [109] [110]	Quantifies system performance and improvement

Hybrid systems that combine multimodal neuroimaging data with sophisticated feature selection algorithms represent a significant advancement in brain activity decoding. The integration of EEG and fNIRS, coupled with hybrid AI-driven approaches like TMGWO, ISSA, and BBPSO, demonstrates measurable improvements in classification accuracy while reducing computational complexity. These improvements are particularly evident in applications requiring robust performance despite high-dimensional data and potential class imbalances. As research in this field evolves, the continued refinement of fusion strategies and feature selection methodologies will further enhance the accuracy and practical utility of hybrid neuroimaging systems in both clinical and research settings.

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) is forging a new path in clinical neuroimaging, significantly enhancing the accuracy of disorder identification and monitoring. This multimodal approach synergistically combines EEG's millisecond-level temporal resolution with fNIRS's superior spatial localization of hemodynamic responses, providing a more complete picture of brain function than either modality alone [111]. The clinical utility of this hybrid methodology is being demonstrated across a growing spectrum of conditions, including Internet Gaming Disorder (IGD), stroke, and intracerebral hemorrhage, where it supports improved machine learning classification accuracy and offers a robust platform for neurofeedback interventions in motor rehabilitation [112] [84] [81]. This technical guide details the experimental protocols, data fusion techniques, and reagent solutions that underpin this advanced diagnostic paradigm, providing a framework for its application in clinical research and therapeutic development.

Quantitative Performance of Multimodal EEG-fNIRS

The enhanced diagnostic capability of combined EEG-fNIRS is quantitatively demonstrated through its superior performance in machine learning classification tasks, a critical step towards objective biomarker development.

Table 1: Classification Accuracy for Internet Gaming Disorder (IGD) Identification Using Machine Learning [84]

Modality	Classifier	Accuracy	Key Differentiating Features
EEG alone	Support Vector Machines	79.4%	Linear/nonlinear EEG features (e.g., skewness, power ratios in sub-bands); Prolonged latency and lower amplitude of ERP components (P100, N200, P300, N450).
EEG + fNIRS	Support Vector Machines	87.25%	Combined EEG features with fNIRS hemodynamic activation (e.g., lower oxygenation in prefrontal cortex during Stroop task).

Table 2: Clinical Applications and Associated Neural Targets

Clinical Disorder	Research Context	Primary Neural Targets/Tasks	Key Findings
Internet Gaming Disorder (IGD)	Identification & Classification [84]	Prefrontal Cortex (PFC) during Stroop Task	IGD group showed lower hemodynamic activity and distinct ERP components, indicating impaired executive control and cognitive processing.
Post-Stroke / Intracerebral Hemorrhage (ICH)	Motor Rehabilitation & Neurofeedback [112] [81]	Sensorimotor Cortices during Motor Imagery (MI)	Hybrid systems aim to stimulate neuroplasticity; Multimodal datasets (e.g., HEFMI-ICH) facilitate the development of precision rehabilitation systems.
Eating Disorders	Neurofeedback Outcomes [112]	Not Specified	EEG- and fNIRS-based neurofeedback produced similar outcomes, suggesting convergent validity for therapeutic applications.

Detailed Experimental Protocols

Standardized experimental protocols are essential for generating reliable and reproducible data in multimodal EEG-fNIRS studies. The following methodologies are cited from key literature.

Protocol 1: Evaluating Multimodal Neurofeedback for Motor Imagery

This protocol is designed to assess the benefits of EEG-fNIRS neurofeedback (NF) for motor imagery (MI), with applications in post-stroke motor rehabilitation [112] [3].

• Objective: To compare the efficacy of unimodal (EEG-only, fNIRS-only) versus multimodal (EEG-fNIRS) NF in modulating sensorimotor cortex activity during left-hand motor imagery.
• Participants: Thirty right-handed healthy participants (as a model for later clinical application).
• Equipment:
  • Cap-Integrated System: A custom EasyCap integrating a 32-channel EEG system (ActiCHamp, Brain Products GmbH) and a continuous-wave fNIRS system (NIRScout XP, NIRx) with 16 detectors, 16 LED sources, and 8 short channels.
  • Sensor Placement: EEG electrodes and fNIRS optodes are positioned over the sensorimotor cortices according to the international 10-10 system.
• NF Task & Design:
  • A block design is used, requiring an initial calibration step to parameterize the NF score calculation.
  • Participants undergo three NF conditions in a randomized order: EEG-only, fNIRS-only, and simultaneous EEG-fNIRS.
  • During each trial, participants perform kinesthetic left-hand MI (e.g., imagining a grasping movement) for a defined period.
  • Real-Time Feedback: A visual representation (e.g., a ball on a one-dimensional gauge) moves upwards in real-time based on the computed NF score, which reflects the activity level in the right primary motor cortex.
• Data Analysis: The NF score is computed from features of the right primary motor cortex activity. The association between the NF score, the neuroimaging modality, and the MI strategy is analyzed to determine the most effective condition.

Protocol 2: Machine Learning Identification of Internet Gaming Disorder

This protocol employs a classic cognitive task within a multimodal framework to identify biomarkers and classify individuals with IGD [84].

• Objective: To develop a machine learning-based diagnostic model for IGD in male university students using multimodal EEG-fNIRS data.
• Participants: 102 male university students (51 with IGD, 51 healthy controls), aged 18-23.
• Equipment: Simultaneous EEG and fNIRS recording systems.
• Paradigm: A Stroop task with congruent, incongruent, and neutral stimuli is presented to participants while recording neural data.
• Feature Extraction:
  • EEG/ERP: Statistical, complexity, and frequency-domain features from EEG sub-bands; amplitude and latency of ERP components (P100, N200, P300, N450).
  • fNIRS: Oxygenation changes (HbO/HbR) in the prefrontal cortex.
• Machine Learning Pipeline:
  • Feature selection based on statistical significance (t-test).
  • Classification using well-known ML algorithms (e.g., Support Vector Machines).
  • Performance comparison between unimodal (EEG) and multimodal (EEG+fNIRS) feature sets.

Protocol 3: Data Collection for Intracerebral Hemorrhage (ICH) Rehabilitation

This protocol outlines the procedure for creating a hybrid BCI dataset for motor imagery in an ICH population, addressing a critical gap in neurorehabilitation research [81].

• Objective: To acquire a synchronized hybrid EEG-fNIRS dataset from both healthy subjects and patients with ICH during standardized left- and right-hand motor imagery tasks.
• Participants: 17 normal subjects and 20 patients with ICH.
• Equipment:
  • A custom hybrid cap with 32 EEG electrodes and a high-density fNIRS array (32 sources, 30 detectors yielding 90 channels).
  • Synchronized data acquisition using a g.HIamp amplifier (EEG) and a NirScan system (fNIRS), triggered by E-Prime 3.0.
• Motor Imagery Paradigm:
  • Preparation: A grip strength calibration procedure using a dynamometer and stress ball is introduced to enhance the kinesthetic vividness of MI.
  • Trial Structure:
    • Visual Cue (2 s): A directional arrow indicates left or right-hand MI.
    • Execution (10 s): Participants imagine a grasping movement with the cued hand at 1 Hz.
    • Rest (15 s): Blank screen for baseline recovery.
  • Each session contains 30 trials (15 left/right), with at least two sessions per participant.
• Data Output: The dataset includes raw data, preprocessed trial-ready data, and clinical characteristics (e.g., Fugl-Meyer Assessment for Upper Extremities).

Workflow and Signaling Pathway Visualizations

Multimodal Neuroimaging Experimental Workflow

Neurovascular Coupling and Signal Pathway

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Equipment and Software for Multimodal EEG-fNIRS Research

Item Name	Category	Function / Application	Exemplar Models / Methods (from search results)
Hybrid EEG-fNIRS Cap	Hardware	Integrates electrodes and optodes for simultaneous data acquisition; ensures proper sensor placement over cortical regions.	Custom EasyCap with 32 EEG channels & fNIRS sources/detectors [112]; Custom Model M cap with 32 EEG, 32 fNIRS sources, 30 detectors [81].
EEG Amplifier System	Hardware	Records electrical brain activity with high temporal resolution.	ActiCHamp (Brain Products GmbH) [112]; g.HIamp amplifier (g.tec) [81].
fNIRS System	Hardware	Measures hemodynamic responses by detecting changes in near-infrared light absorption.	NIRScout XP (NIRx) [112]; NirScan (Danyang Huichuang) [81].
Stimulus Presentation Software	Software	Prescribes and delivers experimental paradigms, and sends synchronization triggers to recording hardware.	E-Prime 3.0 (Psychology Software Tools) [81].
Data Synchronization Interface	Hardware/Software	Temporally aligns EEG and fNIRS data streams for precise multimodal analysis.	TTL pulses, parallel ports, shared clock systems, or software triggers [111].
Signal Processing & ML Toolboxes	Software	Provides pipelines for preprocessing, feature extraction, data fusion, and machine learning classification.	Joint ICA (jICA), Canonical Correlation Analysis (CCA), Support Vector Machines [84] [111].
Grip Force Calibration Tools	Hardware	Enhances the kinesthetic vividness of motor imagery in rehabilitation protocols.	Dynamometer, Stress Ball [81].
Clinical Assessment Scales	Method	Quantifies clinical status and motor function for correlation with neural data.	Fugl-Meyer Assessment for Upper Extremities (FMA-UE), Modified Barthel Index (MBI) [81].

The quest to understand complex brain functions has driven the development of increasingly sophisticated neuroimaging technologies. Within this landscape, electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) have emerged as prominent non-invasive modalities, each with distinct strengths and limitations. EEG measures the brain's electrical activity with millisecond temporal resolution, providing exquisite detail about the timing of neural processes but suffering from limited spatial resolution and susceptibility to artifacts [37]. Conversely, fNIRS measures hemodynamic responses associated with neural activity through near-infrared light, offering better spatial localization and robustness to motion artifacts but capturing slower signal changes due to neurovascular coupling [2].

While unimodal approaches using either EEG or fNIRS alone have yielded significant insights, they inherently struggle to provide comprehensive spatiotemporal characterization of brain activity. This limitation has prompted growing interest in multimodal integration, combining these complementary signals to overcome their individual constraints. The fusion of EEG and fNIRS creates a synergistic relationship where their combined value exceeds the sum of their parts, enabling researchers to capture both rapid neural dynamics and their underlying vascular correlates [27] [113].

This technical guide provides a comprehensive analysis of multimodal EEG-fNIRS approaches relative to unimodal methodologies. We examine quantitative performance comparisons, detail experimental protocols from seminal studies, visualize core concepts and architectures, and equip researchers with practical tools for implementing these advanced neuroimaging techniques within the broader context of multimodal neuroimaging research.

Performance Comparison: Multimodal vs. Unimodal Approaches

Empirical evidence consistently demonstrates that multimodal EEG-fNIRS systems achieve superior performance across various applications compared to unimodal configurations. The complementary nature of electrical and hemodynamic signals enables more robust brain state classification, enhanced artifact rejection, and improved spatial and temporal resolution.

Table 1: Quantitative Performance Comparison Across Applications

Application Domain	Unimodal EEG Performance	Unimodal fNIRS Performance	Multimodal EEG-fNIRS Performance	Key Improvement Metrics
Motor Imagery Classification	76-80% accuracy [114]	75-78% accuracy [114]	83.26% accuracy [114]	+3.78% vs. state-of-the-art [114]
Cognitive State Decoding	Not Reported	Not Reported	Significantly outperforms conventional approaches [37]	Enhanced cross-modal synergy and dynamic dependency modeling [37]
Stroke Rehabilitation Outcome Prediction	Higher prediction error	Higher prediction error	20.36% reduction in prediction error [115]	Improved neurovascular coupling assessment [115]
Few-Shot Brain Signal Classification	Limited with minimal labeled data	Limited with minimal labeled data	Competitive with state-of-the-art supervised models [49]	Effective shared domain learning with minimal labels [49]

The performance advantages of multimodal integration stem from several fundamental mechanisms. First, the temporal complementarity allows researchers to capture both immediate neural responses (via EEG) and subsequent hemodynamic changes (via fNIRS), providing a more complete picture of brain activity across different timescales [2]. Second, spatial enhancement occurs when fNIRS's better localization capabilities help resolve EEG's spatial ambiguities, while EEG's extensive sensor coverage can guide fNIRS placement to regions of interest [27]. Third, artifact discrimination improves significantly since artifacts affecting one modality often have distinct signatures, enabling more effective noise removal through cross-validation [2].

Methodological Approaches and Fusion Strategies

Multimodal EEG-fNIRS integration employs diverse methodological frameworks ranging from traditional machine learning to advanced deep learning architectures. The fusion strategy critically influences the effectiveness of integration, with each approach offering distinct advantages for specific research contexts.

Deep Learning Architectures with Attention Mechanisms

The MBC-ATT (Multi-Branch Convolutional Neural Network with Attention) framework represents a sophisticated approach to multimodal fusion. This architecture employs independent branch structures to process EEG and fNIRS signals separately, leveraging modality-specific characteristics before integration [37]. A cross-modal attention mechanism dynamically weights the importance of features from each modality, strengthening the model's ability to focus on task-relevant signals while suppressing noise [37]. Experimental validation on n-back and word generation datasets demonstrated superior classification performance compared to conventional approaches, particularly for complex cognitive tasks where cross-modality correlations are essential [37].

Another innovative approach utilizes Dirichlet distribution parameter estimation combined with Dempster-Shafer Theory (DST) to model uncertainty in decision fusion. This method quantifies decision outputs from both modalities, followed by a two-layer reasoning process that fuses evidence from basic belief assignment methods [114]. When applied to motor imagery classification, this approach achieved 83.26% accuracy, representing a 3.78% improvement over state-of-the-art methods [114].

Representation Learning for Limited Data Scenarios

The EFRM (EEG-fNIRS Representation-learning Model) addresses the challenge of limited labeled data through a novel two-stage approach. The pre-training stage learns both modality-specific and shared representations across EEG and fNIRS using a Masked Autoencoder (MAE) and contrastive learning [49]. The model is trained on large-scale unlabeled datasets (1250 hours from 918 participants) before the transfer learning stage adapts the pre-trained model to specific downstream tasks with minimal labeled samples [49].

This approach uniquely enables adaptation to EEG-only, fNIRS-only, or paired EEG-fNIRS scenarios, addressing the practical challenge of acquiring paired multimodal datasets [49]. Quantitative evaluations demonstrate competitive performance with state-of-the-art supervised approaches while requiring significantly fewer labeled samples, making it particularly valuable for clinical applications where labeled data is scarce [49].

Multilayer Brain Network Analysis

For clinical applications such as stroke rehabilitation, EEG-fNIRS multilayer brain network analysis has proven effective in revealing functional neural reorganization. This method involves reconstructing and aligning EEG-fNIRS signals in a unified cortical source space, then quantifying neurovascular coupling strength through subject-specific estimation of the hemodynamic response function [115]. These measures are used to construct a multilayer brain network combining unimodal intra-layer edges with bimodal inter-layer edges [115].

In stroke recovery studies, this approach showed significant improvement in neurovascular coupling levels and multiplex clustering coefficients in patients receiving repetitive transcranial magnetic stimulation (rTMS) with motor training compared to sham groups [115]. These neural changes significantly correlated with motor function improvements (R=0.600 and 0.618), demonstrating the clinical relevance of multimodal biomarkers [115].

Multimodal Fusion Architecture

Experimental Protocols and Methodologies

Implementing effective multimodal EEG-fNIRS research requires careful experimental design, appropriate equipment selection, and rigorous data processing pipelines. This section details standardized protocols from validated studies.

Cognitive Task Paradigms

The n-back working memory task has been extensively used in multimodal studies. In a standardized protocol [37], each task block consists of:

• 2-second instruction display indicating task type (0-back, 2-back, or 3-back)
• 40-second task period where random one-digit numbers are displayed every 2 seconds (0.5s display followed by 1.5s fixation cross)
• 20-second rest period with fixation cross display
• 180 total trials per participant (20 trials × 3 series × 3 sessions) to ensure data adequacy and reliability

For motor imagery studies [81], a typical trial structure includes:

• Visual cue presentation (2s): Yellow directional arrow (left/right) indicating required MI
• Execution phase (10s): Central yellow fixation cross display with auditory cue (200ms beep)
• Inter-trial interval (15s): Blank screen for rest
• 30 trials per session (15 left/right each) with at least two consecutive sessions

Data Acquisition and Hardware Configuration

Simultaneous EEG-fNIRS recording requires synchronized systems with compatible sampling rates. A standard configuration includes [81]:

• EEG System: g.HIamp amplifier with 32 electrodes at 256 Hz sampling rate
• fNIRS System: Continuous-wave multifunctional system (e.g., NirScan) at 11 Hz sampling rate
• Hybrid Cap: Custom-designed integrating 32 EEG electrodes, 32 optical sources, and 30 photodetectors
• Synchronization: Event markers from stimulus presentation software (e.g., E-Prime 3.0) simultaneously triggering both systems

Proper optode placement follows the international 10-20 system, with fNIRS source-detector pairs positioned at ~3cm separation distances to ensure adequate penetration depth while maintaining signal quality [81]. For motor imagery tasks, coverage should focus on sensorimotor cortices, while cognitive tasks may require broader prefrontal and parietal coverage [3].

Table 2: Research Reagent Solutions for Multimodal EEG-fNIRS

Component Category	Specific Examples	Function and Purpose
Acquisition Hardware	g.HIamp amplifier, NirScan fNIRS system [81]	Simultaneous signal recording with temporal synchronization
Stimulus Presentation	E-Prime 3.0, MATLAB with Psychtoolbox [81]	Precise experimental control and event marker generation
Hybrid Caps/Helmets	Custom-designed caps with integrated EEG electrodes and fNIRS optodes [81] [27]	Ensures proper co-registration and stable sensor placement
Data Processing Tools	EEGLAB, NIRS-KIT, Homer2, SPM [2]	Preprocessing, artifact removal, and feature extraction
Fusion Frameworks	MBC-ATT, EFRM, Custom deep learning architectures [37] [49]	Implements cross-modal attention, representation learning, and decision fusion

Signal Processing and Analysis Pipelines

A robust preprocessing pipeline is essential for quality multimodal data. For EEG signals [2]:

• Filtering: Bandpass filtering (0.5-45 Hz) to remove drift and high-frequency noise
• Artifact Removal: Blind source separation (ICA) for ocular and muscle artifacts
• Re-referencing: Common average or mastoid reference
• Epoching: Segmenting data relative to task events

For fNIRS signals [2]:

• Optical Density Conversion: Raw intensity to optical density
• Motion Artifact Correction: Spline interpolation or wavelet-based methods
• Hemodynamic Response Conversion: Modified Beer-Lambert law to oxygenated/deoxygenated hemoglobin concentrations
• Bandpass Filtering: 0.01-0.2 Hz to remove physiological noise (cardiac, respiratory)

Multimodal fusion can then be implemented at various levels:

• Early Fusion: Concatenating preprocessed features before classification
• Late Fusion: Combining decisions or high-level features from separate classifiers
• Intermediate Fusion: Cross-modal attention or shared representation learning

Experimental Workflow

Advantages and Limitations of Multimodal Integration

Strengths of Multimodal EEG-fNIRS

The synergistic combination of EEG and fNIRS offers several compelling advantages over unimodal approaches:

• Enhanced Classification Accuracy: Multimodal systems consistently achieve 5-10% higher classification accuracy in brain-computer interface applications compared to unimodal systems [81]. This improvement stems from the complementary nature of electrical and hemodynamic information, providing more discriminative features for pattern recognition.

• Improved Neurovascular Coupling Assessment: Simultaneous measurement enables direct investigation of relationships between neural electrical activity and subsequent hemodynamic responses [115] [49]. This is particularly valuable for understanding pathological conditions where neurovascular uncoupling may occur, such as stroke or neurodegenerative diseases.

• Superior Artifact Rejection: The distinct artifact profiles of each modality enable more effective artifact identification and removal through cross-validation [2]. Motion artifacts affecting fNIRS differently from EEG allow for development of robust correction algorithms leveraging both signal types.

• Increased Information Content: Multimodal systems capture a more comprehensive picture of brain activity across different temporal and spatial scales [27] [3]. This expanded information dimensionality enhances the detection sensitivity for subtle neural phenomena that might be missed by single modalities.

Limitations and Technical Challenges

Despite these advantages, multimodal EEG-fNIRS implementation faces several significant challenges:

• Technical Complexity: Integrating hardware systems with different physical requirements presents substantial engineering challenges [27]. Creating comfortable headgear that maintains proper optode and electrode contact without signal interference requires careful design and customization.

• Synchronization Precision: Achieving millisecond-level temporal alignment between systems is non-trivial, especially when using separate acquisition units [27]. Even minor synchronization errors can compromise the analysis of neurovascular coupling dynamics.

• Data Processing Burden: Multimodal datasets dramatically increase computational requirements and analytical complexity [2]. Developing integrated processing pipelines that handle both data types efficiently remains challenging, particularly for real-time applications.

• Interpretation Challenges: Correlating and interpreting complementary but distinct physiological signals requires sophisticated analytical frameworks and deeper theoretical understanding of neurovascular physiology [2] [49]. The field still lacks standardized approaches for quantifying cross-modal relationships.

Multimodal EEG-fNIRS integration represents a significant advancement in neuroimaging methodology, offering substantial benefits over unimodal approaches through enhanced spatiotemporal resolution, improved classification accuracy, and more comprehensive brain activity characterization. The complementary nature of electrical and hemodynamic signals creates a synergistic relationship that enables researchers to address fundamental questions about brain function that cannot be adequately explored with single modalities.

While technical challenges remain in hardware integration, signal processing, and data interpretation, recent advances in fusion algorithms, deep learning architectures, and experimental protocols are rapidly addressing these limitations. The development of standardized frameworks like MBC-ATT and EFRM provides researchers with powerful tools for implementing multimodal approaches across diverse applications from basic cognitive neuroscience to clinical rehabilitation.

As the field progresses, future efforts should focus on standardizing acquisition protocols, validating multimodal biomarkers across patient populations, and developing more accessible analytical tools. With these advancements, multimodal EEG-fNIRS is poised to become an increasingly essential methodology in neuroscience research and clinical practice, ultimately enhancing our understanding of brain function and dysfunction.

Conclusion

Multimodal fNIRS-EEG neuroimaging represents a transformative approach that harnesses the complementary strengths of both modalities to overcome the limitations of unimodal systems. The integration of EEG's unparalleled temporal resolution with fNIRS's improved spatial localization and motion tolerance creates a powerful platform for studying brain function in naturalistic settings, particularly valuable for clinical populations and real-world applications. Current research demonstrates significant advances in data fusion methodologies, artifact handling, and clinical validation across diverse domains from stroke rehabilitation to drug development. Future directions should focus on hardware innovation for enhanced compatibility, standardized data processing pipelines, development of robust unsupervised fusion algorithms, and expansion of clinical applications through large-scale validation studies. As artificial intelligence and machine learning continue to evolve, fNIRS-EEG systems are poised to become indispensable tools for personalized medicine, closed-loop therapeutic systems, and accelerated CNS drug development, ultimately advancing our fundamental understanding of brain function and dysfunction.

References

• 1. Functional Near-Infrared Spectroscopy and Its Clinical ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7364176/]
• 2. Multimodal fNIRS–EEG sensor fusion: Review of data-driven ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12592382/]
• 3. Evaluating the effects of multimodal EEG-fNIRS ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331177]
• 4. Introduction - Electroencephalography (EEG) - NCBI - NIH [https://www.ncbi.nlm.nih.gov/books/NBK390346/]
• 5. EEG: Basic Principles and Overview [https://www.openanesthesia.org/keywords/eeg-basic-principles-and-overview/]
• 6. Basic EEG Electrophysiology [https://www.learningeeg.com/basic-eeg-electrophysiology]
• 7. Functional near Infrared Spectroscopy - an overview [https://www.sciencedirect.com/topics/materials-science/functional-near-infrared-spectroscopy]
• 8. Functional near-infrared spectroscopy [https://en.wikipedia.org/wiki/Functional_near-infrared_spectroscopy]
• 9. High-density multidistance fNIRS enhances detection of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12412631/]
• 10. fNIRS reproducibility varies with data quality, analysis ... [https://www.nature.com/articles/s42003-025-08412-1]
• 11. Enhanced spatiotemporal resolution imaging of neuronal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7778454/]
• 12. A Cross-Disciplinary Review of the fNIRS-EEG Dual ... [https://www.mdpi.com/2076-3425/14/10/1022]
• 13. Fusion of fNIRS and EEG: a step further in brain research [https://www.bitbrain.com/blog/fusion-fnirs-eeg-brain-activity-research]
• 14. The Advantages of Combining fNIRS and EEG for Brain ... [https://www.artinis.com/blogpost-all/advantages-of-combining-fnirs-and-eeg]
• 15. Simultaneous multimodal fNIRS-EEG recordings reveal ... [https://www.nature.com/articles/s41598-023-31609-5]
• 16. Multi-Modal Integration of EEG-fNIRS for Characterization ... [https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2022.823435/full]
• 17. Simultaneous EEG-fNIRS Data Classification Through ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11379445/]
• 18. Simultaneous EEG and fNIRS recordings for semantic ... [https://www.nature.com/articles/s41597-025-04967-0]
• 19. A multimodal EEG-fNIRS-TCD investigation [https://www.sciencedirect.com/science/article/pii/S1053811924004075]
• 20. A Cross-Disciplinary Review of the fNIRS-EEG Dual ... [https://pubmed.ncbi.nlm.nih.gov/39452034]
• 21. Neurovascular coupling during auditory stimulation [https://pmc.ncbi.nlm.nih.gov/articles/PMC10517045/]
• 22. An EEG-fNIRS neurovascular coupling analysis method to ... [https://www.sciencedirect.com/science/article/abs/pii/S001048252300433X]
• 23. Neurovascular Coupling by Functional Near Infra-Red ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2020.00042/full]
• 24. Structured sparse multiset canonical correlation analysis of... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9046278/]
• 25. An EEG-fNIRS neurovascular coupling analysis method to ... [https://pubmed.ncbi.nlm.nih.gov/37196454/]
• 26. Low frequency oscillations reflect neurovascular coupling ... [https://www.nature.com/articles/s41598-024-61819-4]
• 27. A Cross-Disciplinary Review of the fNIRS-EEG Dual-Modality ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11506513/]
• 28. Current implications of EEG and fNIRS as functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11629311/]
• 29. Bimodal EEG-fNIRS in Neuroergonomics. Current ... [https://www.frontiersin.org/journals/neuroergonomics/articles/10.3389/fnrgo.2022.934234/full]
• 30. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOor73NjScGcy0JSdsOrWmBUZNSH0z2M-UNoPu7JadKNYlWOt_Lnx]
• 31. Applications and advances of combined fMRI-fNIRs ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1542075/full]
• 32. Current State and Future Prospects of EEG and fNIRS in ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2019.00172/full]
• 33. E-FNet: A EEG-fNIRS dual-stream model for Brain– ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809424010012]
• 34. Neural correlates of motor imagery and execution in real- ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2024.1412307/full]
• 35. Neural correlates of motor imagery and execution in real ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11224467/]
• 36. Neural correlates of fine motor grasping skills - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC10784192/]
• 37. Multimodal MBC-ATT: cross-modality attentional fusion of ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2025.1660532/full]
• 38. A multi-modal dataset of electroencephalography and ... [https://www.nature.com/articles/s41597-025-05286-0]
• 39. Current Trends in Hyperscanning research [https://mbraintrain.com/trends-in-hyperscanning-research/]
• 40. Hardware — Blog-content | fNIRS Systems | NIRS Devices [https://nirx.net/blog-content/category/Hardware]
• 41. BrainBody Connect 2025 | Call for Proposal [https://www.biopac.com/brainbody-connect-proposal/]
• 42. Investigating the interaction between EEG and fNIRS [https://www.sciencedirect.com/science/article/pii/S1877750324002096]
• 43. Multimodal fNIRS-EEG sensor fusion: Review of data- ... [https://pubmed.ncbi.nlm.nih.gov/41211102/]
• 44. Data Fusion Based on Subspace Decomposition for Distributed State Estimation in Multi-Hop Networks - PMC [https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6339053/]
• 45. Multi-source heterogeneous data fusion using vector concatenation for quality fluctuation analysis - ScienceDirect [https://www.sciencedirect.com/science/article/pii/S2666764925000529]
• 46. Multi-modal data fusion using source separation: Two effective models based on ICA and IVA and their properties - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC4624202/]
• 47. Protocol for enhancing visualization clarity for categorical ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11103473/]
• 48. Advancing brain-computer interfaces through fNIRS and ... [https://www.frontiersin.org/research-topics/71646/advancing-brain-computer-interfaces-through-fnirs-and-eeg-integration]
• 49. EFRM: A Multimodal EEG–fNIRS Representation-learning ... [https://www.sciencedirect.com/science/article/abs/pii/S0010482525016464]
• 50. Simultaneous EEG-fNIRS study of visual cognitive processing [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0325017]
• 51. Your Data Visualization Color Guide: 7 Best Practices | Sigma [https://www.sigmacomputing.com/blog/7-best-practices-for-using-color-in-data-visualizations]
• 52. Over the rainbow: Guidelines for meaningful use of colour ... [https://www.sciencedirect.com/science/article/pii/S1053811921009010]
• 53. Data Visualization Colors: Best Practices & Palettes (2025) [https://letdataspeak.com/mastering-color-in-data-visualizations/]
• 54. Application and future directions of brain-computer ... [https://www.sciencedirect.com/science/article/pii/S2589238X2500083X]
• 55. Brain–machine interface - Latest research and news [https://www.nature.com/subjects/brain-machine-interface]
• 56. Clinical Applications of Brain-Computer Interfaces: Current ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2862632/]
• 57. Brain-Computer Interface Clinical Trials [https://www.hopkinsmedicine.org/neurology-neurosurgery/clinical-trials/brain-computer-interface]
• 58. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOopbp_R3gpWzVMFOYquZA_BO5PBZ1ylNV9Oo0YA6VVkjF5-tlHBd]
• 59. A brain implant that could rival Neuralink's enters clinical ... [https://www.nature.com/articles/d41586-025-03849-0]
• 60. Study of promising speech-enabling interface offers hope for ... [https://med.stanford.edu/news/all-news/2025/08/brain-computer-interface.html]
• 61. AgencyIQ – Sluggish biomarker program offers FDA ... [https://friendsofcancerresearch.org/news/agencyiq-sluggish-biomarker-program-offers-fda-opportunity-for-reform/]
• 62. Use of Biomarkers in Drug Development for Regulatory ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12489174/]
• 63. EEG Biomarkers: Transforming CNS Drug Discovery [https://www.neuroelectrics.com/blog/eeg-biomarkers-transforming-cns-drug-discovery]
• 64. fNIRS as a biomarker for X-linked neurodevelopmental ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1560786/full]
• 65. Multimodal Measurement: Integration of EEG & fNIRS - TMSi [https://www.tmsi.artinis.com/blog/eeg-fnirs-integration]
• 66. The effect of EEG and fNIRS in the digital assessment and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10703187/]
• 67. How AI & Biomarkers Are Reshaping the Next Frontier in ... [https://www.precisionformedicine.com/blog/how-ai-biomarkers-are-reshaping-the-next-frontier-in-drug-development]
• 68. Top 5 Drug Discovery Trends 2025 Driving Breakthroughs [https://www.pelagobio.com/cetsa-drug-discovery-resources/blog/drug-discovery-trends-2025/]
• 69. Motion Artifacts Correction from Single-Channel EEG and ... [https://www.mdpi.com/1424-8220/22/9/3169]
• 70. Combination of spatial and temporal de-noising and artifact ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1576954/full]
• 71. Motion Artifact Removal from EEG Signals Using the ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809425013886]
• 72. Motion Artifact Contaminated fNIRS and EEG Data v1.0.0 [https://physionet.org/content/motion-artifact/]
• 73. A novel EEG artifact removal algorithm based on an ... [https://www.nature.com/articles/s41598-025-98653-1]
• 74. Characterizing motion artifacts in functional near-infrared ... [https://www.sciencedirect.com/science/article/pii/S1746809425007670]
• 75. Selecting methods for a modular EEG pre-processing ... [https://www.sciencedirect.com/science/article/pii/S1746809423012636]
• 76. Enhancing Epilepsy Seizure Detection Through Advanced ... [https://www.mdpi.com/2075-4418/14/22/2525]
• 77. Preprocessing functional near-infrared spectroscopy ... [https://mne.tools/stable/auto_tutorials/preprocessing/70_fnirs_processing.html]
• 78. Applications and advances of combined fMRI-fNIRs ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11958174/]
• 79. Top 10 EEG Data Analysis Tools for Effective Brain Research [https://www.neuroelectrics.com/blog/top-10-eeg-data-analysis-tools-for-effective-brain-research]
• 80. AnEEG: leveraging deep learning for effective artifact ... [https://www.nature.com/articles/s41598-024-75091-z]
• 81. HEFMI-ICH: a hybrid EEG-fNIRS motor imagery dataset for ... [https://www.nature.com/articles/s41597-025-06100-7]
• 82. Statistical Feature Extraction for Artifact Removal from ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3254729/]
• 83. A deep learning approach to artifact removal in ... [https://www.sciencedirect.com/science/article/pii/S030645222500990X]
• 84. An efficient multimodal electroencephalography and ... [https://www.sciencedirect.com/science/article/abs/pii/S0952197625017749]
• 85. Image-Based Meta- and Mega-Analysis (IBMMA): A Unified ... [https://www.sciencedirect.com/science/article/pii/S1053811925005579]
• 86. Image-Based Meta- and Mega-Analysis (IBMMA): A ... - PubMed [https://pubmed.ncbi.nlm.nih.gov/40667083/]
• 87. Neurovascular Coupling: Scientometric Analysis of 30 Years ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12606022/]
• 88. Neurovascular coupling dysfunction in encephalopathy [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1522485/full]
• 89. From Mechanisms to Medicine: Neurovascular Coupling in ... [https://www.mdpi.com/2073-4409/14/1/16]
• 90. A combined resting-state fMRI and arterial spin labeling study [https://www.sciencedirect.com/science/article/pii/S0531556524001207]
• 91. Neurovascular coupling abnormalities in first-episode drug ... [https://bmcpsychiatry.biomedcentral.com/articles/10.1186/s12888-025-07503-x]
• 92. Neuro-Vascular-Metabolic Dysregulation ... [https://www.preprints.org/manuscript/202511.1035]
• 93. A Unique Color-Coded Visualization System with Multimodal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10204620/]
• 94. Introducing our new products and events in [https://www.gtec.at/2025/01/14/introducing-new-products-and-events-in-2025/]
• 95. Multimodal fNIRS - EEG measurements: Ensuring high- quality ... EEG [https://www.artinis.com/blogpost-all/ensuring-high-quality-eeg-recordings-in-integrated-fnirs-eeg]
• 96. Exploring Imagined Movement for Brain–Computer Interface ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12468679/]
• 97. The Brain Imaging Data Structure: BIDS [https://bids.neuroimaging.io/index.html]
• 98. Multimodal assessment of the spatial correspondence ... [https://www.nature.com/articles/s41598-023-29123-9]
• 99. Fusion of fNIRS and fMRI data: identifying when and where ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3797402/]
• 100. An fNIRS-based investigation of cerebral hemodynamic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12094942/]
• 101. Use of cortical hemodynamic responses in digital therapeutics ... [https://jneuroengrehab.biomedcentral.com/articles/10.1186/s12984-024-01404-y]
• 102. Validating the reproducibility of a low-cost single-channel ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1351341/full]
• 103. Variation in brain connectivity during motor imagery and motor ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10869475/]
• 104. Enhancing classification of a large lower-limb motor ... [https://www.nature.com/articles/s41597-025-05767-2]
• 105. A multi-day and high-quality EEG dataset for motor imagery ... [https://www.nature.com/articles/s41597-025-04826-y]
• 106. An EEG motor imagery dataset for brain computer interface in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10811218/]
• 107. Optimizing high dimensional data classification with a ... [https://pubmed.ncbi.nlm.nih.gov/41062545/]
• 108. Optimizing high dimensional data classification with a ... [https://www.nature.com/articles/s41598-025-08699-4]
• 109. Classification: Accuracy, recall, precision, and related metrics [https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall]
• 110. Accuracy vs. precision vs. recall in machine learning [https://www.evidentlyai.com/classification-metrics/accuracy-precision-recall]
• 111. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOooVG9F8aDdaI2uTX5pX0eOY_5cS9Zg7zwbW5xpWUh9-qIV_Dl8z]
• 112. Evaluating the effects of multimodal EEG-fNIRS ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12425317/]
• 113. Functional Near-Infrared Spectroscopy and... | SpringerLink [https://link.springer.com/chapter/10.1007/978-3-642-02812-0_50]
• 114. EEG–fNIRS signal integration for motor imagery ... [https://www.sciencedirect.com/science/article/pii/S2772528625000299]
• 115. EEG-fNIRS multilayer brain network analysis revealed ... [https://pubmed.ncbi.nlm.nih.gov/40531644/]