Decoding Brain Networks: A Comparative Analysis of EEG and fNIRS for Structure-Function Relationships

Abstract

This article provides a comprehensive exploration of how Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) illuminate the relationship between brain structure and function, a cornerstone of modern neuroscience. Tailored for researchers, scientists, and drug development professionals, we dissect the foundational principles, methodological applications, and optimization strategies for using these modalities individually and in tandem. Covering topics from neurovascular coupling and multimodal integration to artifact correction and biomarker validation, this review synthesizes current evidence demonstrating that combined EEG-fNIRS approaches offer superior insights into brain network organization, with significant implications for accelerating CNS therapeutic development and improving diagnostic classification in clinical populations.

The Neural Blueprint: Uncovering Core Principles of Brain Connectivity with EEG and fNIRS

Understanding the relationship between the brain's structure and its function is a cornerstone of modern neuroscience. This endeavor relies heavily on non-invasive neuroimaging techniques, primarily Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS). These two modalities capture fundamentally different physiological processes: EEG measures the brain's direct, electrical activity, while fNIRS tracks the indirect, hemodynamic response associated with neural firing [1] [2]. This guide provides a detailed, objective comparison of EEG and fNIRS, framing their capabilities within the critical context of investigating structure-function relationships in the human brain. We summarize quantitative data, delineate experimental protocols, and outline essential research tools to inform researchers, scientists, and drug development professionals.

Core Principles and Comparative Profile

EEG and fNIRS are based on distinct biophysical phenomena. EEG records electrical potentials generated by the synchronized firing of cortical pyramidal neurons, offering a direct view of neural dynamics with millisecond temporal resolution [2]. In contrast, fNIRS is an optical technique that leverages the absorption properties of hemoglobin. It uses near-infrared light to measure concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) in the cortical vasculature, which are modulated by neural activity via neurovascular coupling [2] [3]. The relationship between these signals is governed by neurovascular coupling, the physiological process where active neurons trigger a localized increase in blood flow to meet metabolic demand [2].

The table below provides a side-by-side comparison of the technical characteristics of EEG and fNIRS.

Table 1: Technical comparison between EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity from cortical neurons	Hemodynamic response (HbO and HbR concentration)
Signal Source	Postsynaptic potentials in pyramidal cells [2]	Cerebral blood flow and oxygenation [3]
Temporal Resolution	High (milliseconds) [1] [2]	Low (seconds) [1] [2]
Spatial Resolution	Low (centimeter-level) [1]	Moderate (better than EEG, cortex-limited) [1]
Depth of Measurement	Cortical surface	Outer cortex (~1–2.5 cm deep) [1]
Key Rationale for Structure-Function Studies	Direct coupling to neural electrical activity	Indirect coupling via hemodynamic response
Best Use Cases	Fast cognitive tasks, ERPs, brain state monitoring [1]	Sustained cognitive states, ecological studies, clinical populations [1]

Investigating Structure-Function Relationships: Experimental Evidence

A pivotal 2024 study directly addressed how EEG and fNIRS capture the structure-function relationship differently. Using simultaneous EEG-fNIRS recordings and a graph signal processing framework, researchers characterized this coupling during rest and motor imagery tasks [4] [5].

Key Experimental Findings:

• Global Coupling: The study found that the structure-function coupling derived from fNIRS signals resembled the coupling observed in slower-frequency EEG oscillations at rest [4] [5].
• Local Heterogeneity: The relationship was not uniform across the brain. Both modalities showed greater structure-function coupling in the sensory cortex (unimodal areas) and increased decoupling in the association cortex (transmodal areas), following a known unimodal-to-transmodal organizational gradient [4] [5].
• Modality Discrepancies: Notably, discrepancies between EEG and fNIRS were observed, particularly within the frontoparietal network. Furthermore, cross-modal analysis revealed a lower correspondence between electrical and hemodynamic activity in the transmodal cortex, regardless of the brain state [4].

These findings highlight that the choice of modality can influence the observed structure-function relationship, with fNIRS providing a perspective that aligns with slower, hemodynamic processes.

Experimental Protocols for Multimodal Imaging

Combining EEG and fNIRS in a multimodal protocol leverages their complementary strengths. The following workflow outlines a standard methodology for a simultaneous recording experiment, such as one investigating motor imagery.

Diagram 1: Experimental workflow for simultaneous EEG-fNIRS

Detailed Methodology:

• Hardware Integration and Synchronization: A critical first step is the mechanical and electrical integration of the systems. This can involve:

  • Integrated Caps: Using a single headcap with pre-defined mounts for both EEG electrodes and fNIRS optodes [6] [3].
  • Synchronization: Employing a shared microcontroller or an external trigger (e.g., TTL pulses) to precisely time-lock the data acquisition from both modalities, ensuring the millisecond-precise EEG data and the slower fNIRS data can be accurately aligned [7] [6].

• Data Preprocessing: EEG and fNIRS require separate, modality-specific preprocessing pipelines before fusion [1] [2].

  • EEG Preprocessing: Typically involves filtering (e.g., 0.5-40 Hz for ERPs), artifact removal (e.g., for eye blinks, muscle activity), and often re-referencing.
  • fNIRS Preprocessing: Involves converting raw light intensity signals to optical density, then to HbO and HbR concentration changes using the Modified Beer-Lambert Law (MBLL). This is followed by bandpass filtering (e.g., 0.01-0.1 Hz to remove physiological noise) and motion artifact correction [4] [2].

• Data Fusion and Analysis: After preprocessing, the features from both modalities are fused for a comprehensive analysis. Common strategies include [2] [8]:

  • Feature-Level Fusion: Concatenating extracted features (e.g., EEG band powers and fNIRS HbO slopes) into a single feature vector for input into a classifier.
  • Decision-Level Fusion: Combining the classification outcomes or probabilistic outputs of separate classifiers trained on EEG and fNIRS data.

The Neurovascular Coupling Pathway

The connection between the signals measured by EEG and fNIRS is established through neurovascular coupling. The following diagram illustrates this biological pathway.

Diagram 2: Neurovascular coupling connects EEG and fNIRS signals

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful execution of EEG-fNIRS experiments requires specific hardware, software, and consumables. The following table details key components of a multimodal research setup.

Table 2: Essential research reagents and materials for simultaneous EEG-fNIRS studies

Item Name	Type/Model Examples	Critical Function in Research
Integrated EEG-fNIRS Cap	Custom 3D-printed helmet; Cryogenic thermoplastic sheet [6]	Ensures stable, co-registered placement of sensors, critical for spatial alignment and data quality.
EEG Amplifier & Electrodes	Active Ag/AgCl wet electrodes; Passive dry electrodes [7]	Measures voltage differences on the scalp. Active electrodes reduce noise, while wet electrodes provide better signal quality.
fNIRS Continuous-Wave (CW) System	Cortivision Photon Cap; NIRScout [7] [3]	Emits near-infrared light and detects attenuated light to calculate HbO and HbR changes. CW systems are popular for portability and cost.
Electrolyte Gel / Paste	Standard EEG electrolyte gels	Essential for wet EEG electrodes to facilitate ionic conduction and achieve low impedance at the scalp-electrode interface.
Synchronization Interface	Lab Streaming Layer (LSL); TTL Pulse Generator [7] [3]	Provides a shared time clock or trigger signals to synchronize data streams from independent EEG and fNIRS hardware.
Analysis Software Suites	MNE, Brainstorm, Homer2, NIRS-KIT [4] [2]	Provides toolboxes for modality-specific preprocessing, feature extraction, and multimodal data fusion.

EEG and fNIRS are not competing technologies but complementary pillars for probing the brain's structure-function relationship. EEG provides an unrivalled, direct window into the brain's millisecond-scale electrical dynamics, while fNIRS offers superior spatial localization of slower, metabolically coupled hemodynamic activity. The choice between them—or the decision to use them in tandem—depends entirely on the research question. For investigating rapid neural oscillations or evoked potentials, EEG is indispensable. For studying sustained cognitive load, emotional processing, or working with populations in naturalistic settings, fNIRS holds the advantage. Critically, as evidence shows, a multimodal approach can overcome the limitations of either method alone, providing a richer, more holistic picture of brain organization and paving the way for advanced applications in clinical diagnosis, therapeutic development, and cognitive neuroscience.

Neurovascular coupling (NVC) is the fundamental biological process that ensures rapid and precise matching between neuronal activity and regional cerebral blood flow. This mechanism delivers oxygen and nutrients to active brain regions, forming the physiological basis for functional brain imaging techniques like electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). Understanding the structure-function relationships in the brain requires a multimodal approach, as these techniques capture different yet complementary aspects of neural activity—EEG provides millisecond-scale temporal resolution of electrical potentials, while fNIRS maps hemodynamic changes with better spatial localization [9] [4]. This guide objectively compares how EEG and fNIRS research illuminates NVC, providing researchers and drug development professionals with experimental data, protocols, and analytical frameworks.

Defining the Neurovascular Unit and Coupling Mechanism

The neurovascular unit (NVU) is a functional complex comprising neurons, astrocytes, vascular endothelial cells, and pericytes that work in concert to maintain brain homeostasis [10]. The core function of this unit, NVC, describes the coordinated series of events where increased synaptic activity triggers a local hemodynamic response, typically characterized by an increase in cerebral blood flow (CBF) known as functional hyperemia [11] [12].

The widely accepted hypothesis is that activity-induced changes in neurons and astrocytes initiate the neurovascular response. Neurotransmitters and potassium ions (K+) released during synaptic activity lead to the generation of vasoactive agents that cause the dilation of arterioles and capillaries, thereby increasing blood flow to the active region [11]. A 2025 study identified that endothelial gap junction coupling, particularly in arterial segments, enables the long-range propagation of vasodilation signals through the vasculature, setting both the speed and spatial extent of this crucial response [12].

EEG vs. fNIRS: A Technical Comparison for NVC Research

EEG and fNIRS are non-invasive neuroimaging techniques that measure fundamentally different physiological signals, which are linked through the process of NVC.

Electroencephalography (EEG) records the brain's electrical activity from the scalp. The signals originate primarily from the postsynaptic potentials of cortical pyramidal neurons. Its greatest strength is its excellent temporal resolution (milliseconds), allowing it to capture rapid neural dynamics. However, its spatial resolution is limited due to the blurring effect of the skull and scalp on electrical signals [9] [13].

Functional Near-Infrared Spectroscopy (fNIRS) measures the hemodynamic response associated with neural activity. It uses near-infrared light to track changes in the concentration of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) in the microvasculature of the outer cortex. This is an indirect measure of neural activity, relying on the principle of NVC. During neural activation, a regional increase in blood flow typically leads to an increase in HbO and a decrease in HbR [11] [9]. fNIRS offers better spatial resolution than EEG for surface cortical areas but has slower temporal resolution (seconds) due to the delayed nature of the hemodynamic response [9].

Table 1: Technical Comparison of EEG and fNIRS for NVC Studies

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity of neurons (postsynaptic potentials)	Hemodynamic response (changes in HbO and HbR)
Signal Source	Direct neural electrical activity	Indirect, neurovascular-coupled blood oxygenation
Temporal Resolution	High (millisecond scale)	Low (second scale)
Spatial Resolution	Low (centimeter-level)	Moderate (better than EEG, cortical surface)
Depth of Measurement	Cortical surface	Outer cortex (~1–3 cm deep)
Key NVC Metrics	Event-Related Potentials (N1, P2), Band Power (Alpha, Beta)	HbO concentration, HbR concentration
Sensitivity to Motion	High (susceptible to artifacts)	Low (more tolerant)

Experimental Protocols for Investigating NVC

Multimodal experiments that concurrently record EEG and fNIRS are powerful for quantifying NVC, as they capture both the electrical ignition and the subsequent hemodynamic response within the same brain region and task.

Auditory Intensity-Dependent Amplitude Changes (IDAP) Paradigm

This protocol is designed to probe the integrity of NVC by presenting auditory stimuli of varying intensities.

• Stimuli: Pure tone bursts presented at multiple intensity levels (e.g., 70 dB, 80 dB, 90 dB). Each tone lasts for 50-500 ms [11].
• Design: Tones of different intensities are presented in a randomized order with a sufficient inter-stimulus interval (e.g., >10-15 seconds) to allow the hemodynamic response to return to baseline [11].
• EEG Measures: Analysis focuses on Event-Related Potential (ERP) components like N1 and P2, as well as the N1–P2 peak-to-peak amplitude. The expected result is a linear increase in the amplitude of these components with increasing sound intensity [11].
• fNIRS Measures: Changes in HbO and HbR concentrations over the auditory and prefrontal cortices. The expected result is an increase in HbO and a decrease in HbR with increasing sound intensity, demonstrating a coupled hemodynamic response [11].
• NVC Analysis: The relationship is quantified by performing Spearman correlation analysis between the EEG amplitudes (e.g., N1) and the fNIRS concentration changes (e.g., HbR) in the same anatomical region (e.g., left auditory cortex) [11].

Cognitive-Motor Dual-Task Paradigm

This protocol investigates how NVC is modulated during complex tasks that require divided attention, which is relevant for neurodegenerative diseases.

• Tasks: Participants perform a single motor task (e.g., grip force tracking), a single cognitive task (e.g., mental arithmetic or number detection), and a dual task combining both [14].
• Recording: Simultaneous EEG-fNIRS is collected throughout the experiment. fNIRS optodes are typically placed over the prefrontal cortex, while EEG electrodes cover motor and cognitive areas [14].
• Signal Processing: Task-Related Component Analysis (TRCA) is applied to both EEG and fNIRS signals to extract reproducible, task-related neural components from noise and artifacts [14].
• NVC Analysis: Neurovascular coupling strength is computed by calculating the correlation between the power of EEG rhythmic activities (in theta, alpha, and beta bands) and the fNIRS HbO signal from the same task period. A decrease in this correlation during the dual task compared to single tasks indicates impaired NVC due to cognitive-motor interference [14].

The following diagram illustrates the workflow of a typical multimodal NVC investigation.

Key Research Findings on Structure-Function Relationships

Simultaneous EEG-fNIRS studies have revealed critical insights into how the brain's structural connectivity supports its functional dynamics, as measured by electrical and hemodynamic signals.

• Regional Heterogeneity in Coupling: Research shows that the relationship between structural and functional connectivity is not uniform across the brain. There is greater structure-function coupling in the sensory cortex (unimodal areas) and increased decoupling in the association cortex (transmodal areas), following a unimodal-to-transmodal gradient. This gradient is observed in both EEG and fNIRS data, though with notable discrepancies in networks like the frontoparietal network [4] [5].

• Temporal Convergence at Slow Frequencies: The functional networks captured by fNIRS most closely resemble those captured by slower-frequency EEG oscillations (e.g., delta, theta bands) during the resting state. This highlights that fNIRS hemodynamic signals are more coupled to the slow, rhythmic neural activity than to high-frequency activity [4].

• Intensity-Dependent Activation: Studies using auditory stimuli have successfully demonstrated NVC by showing that increasing tone intensity leads to a corresponding increase in both ERP amplitudes (N1, P2) and hemodynamic responses (HbO increase) in the auditory cortex. Correlation analyses have confirmed a significant relationship between the left auditory cortex hemodynamic response and the N1 EEG amplitude [11].

Table 2: Representative Quantitative Findings from Multimodal NVC Studies

Experimental Paradigm	EEG Findings	fNIRS Findings	NVC Correlation Result
Auditory IDAP [11]	N1-P2 amplitude increased with tone intensity (77.9 dB to 89.5 dB).	HbO increased, HbR decreased with tone intensity in auditory cortex.	Spearman correlation: Left auditory cortex HbR with N1 amplitude.
Motor Imagery & Mental Arithmetic [13]	Specific patterns of band power desynchronization during tasks.	Specific patterns of HbO activation during tasks.	Feature-level fusion achieved high classification accuracy (96.74% for MI, 98.42% for MA).
Cognitive-Motor Dual-Task [14]	Changes in theta, alpha, and beta band power.	Altered HbO concentration in prefrontal areas.	NVC strength (EEG-fNIRS correlation) decreased in the dual task vs. single tasks.
"Where's Waldo?" Visual Task [15]	Reduction in alpha/low beta power across all electrodes (p < 0.001).	Increased HbO, decreased HbR in occipital cortex.	Cross-correlation showed hemodynamic changes were independent of systemic influences.

The following table details key materials and analytical tools essential for conducting rigorous EEG-fNIRS NVC research.

Table 3: Research Reagent Solutions for Multimodal NVC Studies

Item / Solution	Function / Application in NVC Research
High-Density fNIRS System (e.g., NIRSIT) [16]	Measures hemodynamic changes with multiple sources/detectors for improved spatial resolution over the prefrontal and auditory cortices.
EEG Recording System (e.g., 30-electrode setup) [4]	Captures electrical brain activity with high temporal resolution; often synchronized with fNIRS.
Transcranial Doppler (TCD) Ultrasound [15]	Provides an additional measure of macro-vascular blood flow velocity in large cerebral arteries during multimodal NVC assessment.
Task-Related Component Analysis (TRCA) [14]	A signal processing algorithm used to extract reproducible, task-related components from both EEG and fNIRS signals, enhancing SNR for NVC analysis.
Graph Signal Processing (GSP) Framework [4]	A mathematical tool for analyzing the relationship between the brain's structural connectome and its functional patterns (from EEG/fNIRS).
Joint Independent Component Analysis (jICA) [9] [13]	A data fusion technique used to identify linked, independent patterns of activity across simultaneously recorded EEG and fNIRS datasets.

The molecular signaling pathway underlying neurovascular coupling can be visualized as follows.

EEG and fNIRS are not competing technologies but complementary pillars in the study of neurovascular coupling. EEG delivers direct, millisecond-resolution insights into neuronal firing, while fNIRS provides a localized map of the consequent hemodynamic response. The fusion of these modalities in a multimodal framework offers the most comprehensive approach to investigating the brain's structure-function relationships, quantifying the integrity of the NVC bridge, and identifying its breakdown in neurological and psychiatric disorders. For drug development professionals, this integrated approach provides a robust platform for assessing the efficacy of novel therapeutics aimed at rescuing vascular function and cognitive deficits by targeting the neurovascular unit.

Understanding the intricate relationship between the brain's structure and its function is a central pursuit in modern neuroscience. Non-invasive neuroimaging technologies are the primary tools for this endeavor, allowing researchers to observe brain activity in real-time. Among these, Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) have emerged as two prominent modalities, each with distinct strengths and limitations rooted in the physiological signals they capture. EEG measures the electrical activity generated by populations of synchronously firing neurons, primarily pyramidal cells in the cerebral cortex [17]. In contrast, fNIRS is an optical technique that monitors hemodynamic responses, measuring changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) in the blood vessels of the cortex, providing an indirect marker of neural activity via neurovascular coupling [17] [18]. The choice between these techniques, or the decision to combine them, is fundamentally governed by a trade-off between their inherent spatial and temporal resolution. This guide provides a detailed, objective comparison of EEG and fNIRS to inform researchers and drug development professionals in selecting the optimal methodology for probing structure-function relationships within the human brain.

Technical Comparison: EEG vs. fNIRS

The core differences between EEG and fNIRS arise from their underlying biophysical principles. The following table provides a quantitative summary of their key technical specifications.

Table 1: Technical Specifications and Performance Comparison of EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity of neurons (postsynaptic potentials) [17]	Hemodynamic response (changes in HbO and HbR) [17]
Temporal Resolution	High (millisecond-scale) [17] [19]	Low (seconds-scale, limited by hemodynamic response) [17] [20]
Spatial Resolution	Low (centimeter-level) [17] [19]	Moderate (better than EEG; millimeter- to centimeter-level) [17] [19]
Depth of Measurement	Cortical surface [17]	Outer cortex (~1–2.5 cm deep) [17] [20]
Signal Source	Direct neural electrical activity [17]	Indirect hemodynamic correlate of neural activity [17]
Sensitivity to Motion Artifacts	High – susceptible to movement [17]	Low – more tolerant to subject movement [17] [20]
Portability	High (lightweight, wireless systems available) [17] [21]	High (wearable, mobile formats) [17] [20]
Best Use Cases	Fast cognitive tasks, Event-Related Potentials (ERPs), sleep research, seizure detection [17] [22]	Naturalistic studies, child development, motor rehabilitation, sustained cognitive states [17] [23]

Experimental Protocols for Unimodal and Multimodal Imaging

Protocol 1: Unimodal fNIRS-EEG Study on the Action Observation Network

A seminal study investigating motor execution, observation, and imagery provides a robust protocol for unimodal comparison or simultaneous multimodal recording [23].

• Objective: To elucidate the differences in neural activity across motor execution (ME), motor observation (MO), and motor imagery (MI) conditions and to identify the shared neural network (Action Observation Network, AON) using fused fNIRS-EEG data.
• Participants: 60 healthy adults were recruited. Final analysis included 21 participants after quality control to ensure >50% usable data from both modalities.
• Equipment and Reagents:
  • fNIRS System: A 24-channel continuous-wave system (Hitachi ETG-4100) measuring HbO and HbR concentrations at wavelengths of 695 nm and 830 nm with a 10 Hz sampling rate [23].
  • EEG System: A 128-electrode EEG system (Electrical Geodesics, Inc.) embedded within an elastic cap [23].
  • Integration Cap: The fNIRS probe was embedded within the same 128-electrode EEG cap. fNIRS optode positions were digitized using a 3D magnetic space digitizer (Fastrak, Polhemus) for precise anatomical co-registration [23].
• Procedure:
  • Setup: The participant is fitted with the integrated EEG-fNIRS cap. Optode and electrode positions are digitized.
  • Task Paradigm: The participant sits facing an experimenter. Three conditions are administered in a block design:
    • Motor Execution (ME): On an audio cue ("Your turn"), the participant uses their right hand to grasp, lift, and move a cup.
    • Motor Observation (MO): On an audio cue ("My turn"), the participant observes the experimenter perform the same cup-moving task.
    • Motor Imagery (MI): On an audio cue ("Imagine"), the participant mentally rehearses the action without physical movement.
  • Data Analysis:
    • Unimodal Analysis: fNIRS data is analyzed for HbO/HbR changes in channel space and projected to the cortical surface. EEG data is processed for event-related potentials and spectral power changes.
    • Multimodal Fusion: Structured Sparse Multiset Canonical Correlation Analysis (ssmCCA) is used to fuse fNIRS and EEG data, identifying brain regions where both electrical and hemodynamic activity are consistently detected [23].

Protocol 2: Joint EEG-fNIRS Source Reconstruction for Enhanced Resolution

To overcome the inherent limitations of each modality, advanced computational protocols for joint source reconstruction have been developed.

• Objective: To achieve spatiotemporal resolution superior to either modality alone by using fNIRS (as Diffuse Optical Tomography, DOT) to provide a spatial prior for EEG source reconstruction [19].
• Simulation Environment:
  • A head model is created using a pre-segmented brain atlas (e.g., ICBM152). A tetrahedral mesh is generated for scalp, skull, cerebrospinal fluid (CSF), and brain tissues [19].
  • Neuronal source activations are simulated on the cortical surface.
• Forward Modeling:
  • EEG Leadfield: The electric leadfield matrix is calculated for electrode placements (e.g., standard 10-20 system) using finite element methods [19].
  • DOT Light Model: The photon propagation model (light sensitivity) is calculated for the given optode placements on the scalp [19].
• Inverse Problem and Joint Reconstruction:
  • The algorithm utilizes a Restricted Maximum Likelihood (ReML) framework. The spatial information from the DOT reconstruction is used to constrain the highly ill-posed EEG inverse problem [19].
  • This approach has been shown in simulations to accurately resolve two neuronal sources separated by 2.3-3.3 cm and 50 ms, a feat not achievable with either modality in isolation [19].

Visualizing Signaling Pathways and Experimental Workflows

The Neurovascular Coupling Pathway

The following diagram illustrates the fundamental relationship between the signals measured by EEG and fNIRS, connected through the process of neurovascular coupling.

Diagram 1: From Neuronal Firing to Measurable Signals.

Multimodal fNIRS-EEG Experimental Workflow

This flowchart outlines the key steps in a simultaneous fNIRS-EEG experiment, from setup to data fusion.

Diagram 2: Simultaneous fNIRS-EEG Experimental Workflow.

The Scientist's Toolkit: Key Research Reagent Solutions

Successful execution of EEG, fNIRS, or combined studies requires specific hardware and software solutions. The following table details essential components.

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

Item	Function/Description	Example Use Case/Note
Integrated EEG-fNIRS Cap	An elastic cap with pre-defined placements for both EEG electrodes and fNIRS optodes, ensuring stable and co-registered positioning [24] [23].	Critical for simultaneous recordings. Can be based on standard 10-20 system layouts.
3D Magnetic Digitizer	Precisely records the 3D spatial coordinates of EEG electrodes and fNIRS optodes relative to anatomical landmarks (nasion, inion) [23].	Enables accurate co-registration of data with anatomical brain atlases for source localization.
Continuous-Wave fNIRS System	A common type of fNIRS device that emits continuous light at constant amplitude to measure changes in light attenuation related to HbO and HbR [23].	Suitable for most functional studies; other types include time-domain and frequency-domain systems for absolute quantification [18].
High-Impedance EEG Amplifiers	Amplifiers capable of handling very high contact impedances (>1 GOhm), often used with dry electrodes without skin preparation [21].	Enables quicker setup and improved comfort for mobile and long-duration studies.
Structured Sparse Multiset CCA (ssmCCA)	A advanced data fusion algorithm used to find correlated components across fNIRS and EEG datasets, identifying brain regions consistently active in both modalities [23].	Used to fuse hemodynamic and electrical data for a more complete picture of neural activity.
Joint Source Reconstruction Algorithm	Computational methods (e.g., based on ReML framework) that use fNIRS/DOT priors to constrain the EEG inverse problem, enhancing spatial resolution [19].	Key for overcoming the spatial resolution limit of standalone EEG.

EEG and fNIRS are not competing but complementary technologies in the quest to map cortical structure-function relationships. EEG remains the modality of choice for capturing the rapid dynamics of neural communication with millisecond precision, while fNIRS provides superior spatial localization of sustained cortical activity, especially in naturalistic settings. The fundamental trade-off between temporal and spatial resolution dictates the initial choice of technology. However, as evidenced by the experimental protocols and toolkits outlined herein, the integrated use of fNIRS-EEG is a powerful frontier in neuroscience. By leveraging multimodal fusion algorithms and joint source reconstruction, researchers can transcend the inherent limitations of each standalone method, achieving a spatiotemporal resolution that more closely mirrors the complex nature of brain activity itself. For drug development professionals and clinical researchers, this multimodal approach offers a more comprehensive framework for identifying biomarkers and evaluating therapeutic interventions.

The relationship between the brain's structural wiring and its dynamic functional activity is a cornerstone of modern neuroscience. Research demonstrates that this structure-function coupling is not uniform across the brain but follows a systematic unimodal-transmodal gradient. This article provides a comparative analysis of how Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) capture this heterogeneity. We synthesize experimental data showing that both modalities consistently reveal stronger coupling in unimodal sensory regions and progressive decoupling in transmodal association areas. However, key differences emerge in their sensitivity to temporal dynamics, neural mechanisms, and specific network properties. This guide objectively compares their performance, detailing methodologies and findings to inform tool selection for researchers and drug development professionals studying brain network organization.

The human brain is organized along a principal hierarchy, anchored by unimodal regions at one end and transmodal regions at the other. Unimodal cortex, including primary visual, auditory, and somatosensory areas, processes information from a single sensory modality and exhibits robust, predictable functional responses directly supported by underlying anatomical connections. Conversely, transmodal cortex, such as the default mode and frontoparietal networks, integrates information from multiple sensory streams and internal states to support high-order cognition; its functional patterns appear more decoupled from the static structural connectome [25] [4].

This unimodal-transmodal gradient is a fundamental principle of brain organization, reflected not only in function but also in microstructural properties like cytoarchitecture and laminar differentiation [26]. Investigating this gradient requires neuroimaging tools that can capture brain dynamics. EEG and fNIRS are two non-invasive techniques that measure complementary neural signals. EEG records electrical potentials from synchronized neuronal firing with millisecond temporal resolution, providing a direct measure of neural activity. fNIRS measures hemodynamic changes in blood oxygenation, an indirect marker of neural activity linked through neurovascular coupling, with superior spatial resolution to EEG and greater tolerance for movement [27] [4]. Understanding how these modalities map the same fundamental structure-function relationship is critical for advancing research and clinical applications.

Experimental Protocols for Assessing Structure-Function Coupling

A standardized experimental and analytical workflow is essential for comparing how EEG and fNIRS quantify structure-function coupling.

Data Acquisition and Preprocessing

Table 1: Core Data Acquisition Protocols

Step	EEG Protocol	fNIRS Protocol
Recording Setup	30+ electrodes placed via international 10-5 or 10-20 system.	36+ channels (sources & detectors) placed on scalp, often using 10-20 system for coregistration.
Signal Measured	Electrical potentials from cortical pyramidal neurons.	Changes in oxygenated (HbO) and deoxygenated (HbR) hemoglobin.
Sampling Rate	High (e.g., 1000 Hz, down-sampled to 200 Hz) [4].	Lower (e.g., 10-12.5 Hz) [4].
Preprocessing	Filtering, artifact removal (e.g., ocular, muscle), re-referencing.	Optical density conversion, bandpass filtering (e.g., 0.02-0.08 Hz for resting state), motion artifact correction [4].
Source Reconstruction	Signal is source-localized to cortical regions of interest (ROIs) using anatomical templates (e.g., Desikan-Killiany atlas) [4].	Hemodynamic signals are mapped to cortical ROIs using digitized optode positions and anatomical templates [4].

Quantifying Structure-Function Coupling

The relationship is typically quantified using a multilinear regression model or a graph signal processing (GSP) framework [25] [4].

• Structural Connectivity (SC): Derived from diffusion MRI (dMRI) data, representing the anatomical backbone of white matter fibers between brain regions.
• Functional Connectivity (FC): Derived from EEG or fNIRS time-series. It is calculated as the statistical dependence (e.g., correlation or phase synchronization) between the activity of different brain regions.
• Model Fitting: For a given brain region, its empirical functional connectivity profile (its connections to all other regions) is predicted using its structural connectivity profile, often supplemented with other predictors like Euclidean distance, shortest path length, and communicability [28].
• Coupling Metric: The goodness-of-fit (e.g., Pearson correlation ( R ) or R²) between the predicted and empirical functional profile is computed for each region. This value represents the regional structure-function coupling strength [25] [28].

The GSP framework employs a Structural-decoupling Index (SDI) to quantify the (dis)alignment between structural and functional networks for each brain region [4].

Comparative Data: EEG vs. fNIRS on the Unimodal-Transmodal Gradient

Both modalities consistently reveal the core gradient, but with nuanced differences critical for tool selection.

Table 2: Quantitative Comparison of EEG and fNIRS Findings

Feature	EEG (Electrical Networks)	fNIRS (Hemodynamic Networks)
Overall Coupling Strength	Generally shows stronger coupling in multiple frequency bands compared to hemodynamic measures [26].	Weaker overall coupling compared to neurophysiological measures like EEG/MEG [26].
Temporal Resolution	Millisecond scale, captures rapid neural synchronizations [27].	Seconds scale, limited by the slow hemodynamic response [27].
Spatial Patterning	Heterogeneous coupling, strongest in sensory cortex, decoupling in association cortex [4].	Heterogeneous coupling, strongest in sensory cortex, decoupling in association cortex [4].
Frequency Band Sensitivity	Coupling is strongest in slower (delta, theta) and intermediate (alpha, beta) bands [26].	fNIRS coupling resembles slower-frequency EEG coupling at rest [4].
Network-Specific Findings	Shows specific discrepancies, e.g., in the frontoparietal network [4].	Provides reliable mapping of unimodal areas (visual, somatomotor) and default mode network [4].
Response to Brain State	Coupling patterns vary dynamically with cognitive tasks and resting state [4] [28].	Coupling patterns are stable during sustained cognitive states; effective for resting-state and task-based studies [27] [4].

Key Experimental Findings

• Consistent Gradient Validation: A seminal study using simultaneous EEG-fNIRS recordings confirmed that both modalities exhibit a heterogeneous coupling pattern, with greater coupling in the sensory cortex and increased decoupling in the association cortex, faithfully following the unimodal-transmodal gradient [4].
• Divergence in High-Order Networks: Despite overall agreement, discrepancies are noted, particularly in the frontoparietal network, a key transmodal system. This suggests that electrical and hemodynamic functional networks, while related, are differentially sensitive to the complex neural computations in these regions [4].
• Temporal Dynamics: EEG's high temporal resolution reveals that structure-function coupling is not static but fluctuates moment-to-moment. These dynamics are regionally heterogeneous, with the most variable coupling found in intermediate regions of the hierarchy (e.g., salience network), while unimodal and transmodal cortices show more stable coupling [28]. fNIRS, by contrast, captures a more temporally smoothed picture of this relationship.

Visualizing the Workflow and Gradient

The following diagrams illustrate the core experimental workflow and the fundamental unimodal-transmodal gradient concept.

Workflow for Multimodal Structure-Function Analysis

The Unimodal-Transmodal Coupling Gradient

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Materials and Solutions for EEG/fNIRS Structure-Function Research

Item	Function & Rationale
High-Density EEG System	Measures electrical brain activity with high temporal resolution. Essential for capturing direct neural dynamics and oscillatory coupling across frequency bands.
fNIRS System (e.g., NIRSIT)	Measures hemodynamic responses. Ideal for studies requiring mobility, tolerance of movement, and localized cortical mapping of sustained states.
Diffusion MRI (dMRI) Data	Provides the structural connectome backbone. A prerequisite for defining structural connectivity matrices and calculating structure-function coupling.
Anatomical Brain Atlas (e.g., Desikan-Killiany)	Provides a common parcellation scheme to coregister EEG electrodes, fNIRS optodes, and dMRI tracts into standardized regions of interest (ROIs).
Graph Signal Processing (GSP) Toolbox	Offers mathematical framework (e.g., for calculating Structural-decoupling Index) to combine structure-function analyses and extract harmonic patterns from SC.
Synchronization Hardware/Software	Critical for multimodal studies. Enables temporal alignment of EEG and fNIRS data streams for direct comparison and data fusion.
Digitized Optode/Electrode Positions	Improves spatial accuracy. Using digitized positions rather than theoretical models enhances source reconstruction for both fNIRS and EEG.

The unimodal-transmodal gradient provides a powerful lens through which to interpret brain organization. The comparative data from EEG and fNIRS reveal a consistent story: the brain's structural architecture most directly supports function in sensory areas, while high-order association areas exhibit greater functional flexibility relative to their structural underpinnings [25] [4]. This decoupling in transmodal cortex may not be an intrinsic limitation but rather a reflection of specialized, high-frequency spatiotemporal propagation regimes that are not fully captured by standard models [25].

Performance Comparison and Selection Guide

The choice between EEG and fNIRS is not about which is superior, but which is optimal for the specific research question.

• Choose EEG when your research hinges on temporal precision. This includes studying rapid cognitive processes, event-related potentials (ERPs), brain-computer interfaces (BCIs), or the dynamic fluctuations of structure-function coupling across different neural oscillation bands [27] [28] [26].
• Choose fNIRS when your focus is on localized cortical activity during sustained cognitive, emotional, or motor states, especially in real-world environments or with populations prone to movement (e.g., children, clinical patients) [27] [16] [29]. fNIRS is particularly effective for mapping the hemodynamic correlates of the gradient in prefrontal and parietal association cortices.
• Adopt a Multimodal Approach when a comprehensive understanding is required. Combining EEG and fNIRS allows researchers to simultaneously capture the brain's millisecond electrical dynamics and their localized hemodynamic consequences, offering the most complete picture of structure-function relationships across the cortical hierarchy [4] [30].

In conclusion, both EEG and fNIRS objectively and reliably map the unimodal-transmodal gradient of structure-function coupling, validating it as a fundamental principle of brain organization. The decision to use one or the other—or both in tandem—should be guided by the specific neural signals of interest, the required spatiotemporal resolution, and the experimental context, ensuring that the chosen tool is precisely matched to the scientific inquiry at hand.

The quest to understand the relationship between the brain's structure and its function relies heavily on the tools available to measure neural activity. Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) have emerged as two prominent non-invasive neuroimaging techniques, each providing a unique window into brain organization. While both methods are used to study real-time brain function, they capture fundamentally different physiological processes and operate on distinct spatiotemporal scales [31]. EEG measures the brain's electrical activity through electrodes placed on the scalp, offering millisecond-scale temporal resolution to track rapid neural dynamics [31]. In contrast, fNIRS monitors hemodynamic responses by measuring changes in blood oxygenation using near-infrared light, providing better spatial localization of cortical activity [31]. This fundamental difference in measurement principles positions EEG as a tool for capturing global, temporally precise brain dynamics, while fNIRS offers more localized insights into region-specific brain function. Understanding these complementary strengths is crucial for advancing research in cognitive neuroscience and developing effective brain-computer interfaces (BCIs) [32] [13].

Technical Foundations: Measurement Principles and Physiological Bases

EEG: Electrical Activity of Neural Assemblies

EEG captures the electrical potentials generated by synchronized firing of cortical neurons, primarily pyramidal cells [31]. When these neurons fire in synchrony, their postsynaptic potentials summate to create electrical fields strong enough to be detected through the skull and scalp. These voltage fluctuations reflect the brain's electrical activity with exceptional temporal resolution—on the millisecond scale—allowing researchers to track the rapid dynamics of cognitive processes [31]. However, the electrical signals undergo significant dispersion and smearing as they pass through various tissues (brain membranes, cerebrospinal fluid, skull, and scalp), resulting in limited spatial resolution [31]. This "blurring" effect means that EEG is most sensitive to electrical activity from large populations of synchronously firing neurons located primarily on the brain's surface, making it less effective at detecting activity from deeper brain regions or precisely localizing neural sources [31] [19].

fNIRS: Hemodynamic Correlates of Neural Activity

fNIRS operates on an entirely different principle, measuring metabolic changes associated with neural activity rather than the electrical activity itself [31]. The technique leverages the relative transparency of biological tissue to near-infrared light (700-900 nm) and the differential absorption properties of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) [31] [32]. When neural activity increases in a specific brain region, it triggers a hemodynamic response that delivers oxygenated blood to that area. fNIRS devices measure changes in light absorption to quantify concentration changes in both HbO and HbR, providing an indirect measure of neural activity through neurovascular coupling [31]. This hemodynamic response unfolds more slowly than electrical activity, typically peaking 4-8 seconds after stimulus onset, which fundamentally limits fNIRS's temporal resolution [32]. However, because light scattering follows more predictable paths than electrical conduction through tissues, fNIRS offers superior spatial resolution for mapping activity in surface cortical areas, particularly the prefrontal cortex [31] [33].

Table 1: Fundamental Measurement Characteristics of EEG and fNIRS

Characteristic	EEG	fNIRS
What is Measured	Electrical potentials from synchronized neuronal firing	Changes in hemoglobin oxygenation concentrations
Signal Source	Postsynaptic potentials in cortical pyramidal neurons	Hemodynamic response via neurovascular coupling
Temporal Resolution	Milliseconds [31]	Seconds (typically 1-2 second lag) [32]
Spatial Resolution	Low (centimeter-level) due to signal dispersion [31]	Moderate (better than EEG) for cortical areas [31]
Depth Sensitivity	Cortical surface, limited deeper access [31]	Superficial cortex (1-2.5 cm depth) [31]

Spatial and Temporal Characteristics: Resolution Trade-offs

Temporal Dynamics: Milliseconds vs. Seconds

The temporal domain reveals the most striking trade-off between EEG and fNIRS. EEG's millisecond-scale resolution enables researchers to capture rapidly evolving neural processes with precision timing [31]. This exceptional temporal fidelity makes EEG ideal for studying event-related potentials (ERPs), sensory processing, and other fast cognitive operations where timing is critical. The electrical signals measured by EEG represent almost instantaneous neural communication, allowing direct observation of brain dynamics as they unfold in real time [31].

In contrast, fNIRS tracks the slower hemodynamic response that lags behind neural activity by 1-2 seconds, peaks at 4-8 seconds, and then gradually returns to baseline over several more seconds [32]. This fundamental limitation arises because fNIRS measures metabolic consequences of neural activity rather than the activity itself. While this slower time course is sufficient for studying sustained cognitive states, it prevents fNIRS from capturing rapid neural transitions or precise timing relationships between different brain events [31] [32].

Spatial Specificity: Localization Capabilities

While fNIRS outperforms EEG in spatial resolution, it's important to note that its capabilities remain limited to superficial cortical regions [33]. The technique can localize activity to specific gyri within the prefrontal, motor, and parietal cortices with greater precision than EEG [31]. High-density fNIRS configurations can achieve spatial resolution on the millimeter scale when used for source reconstruction [19]. However, consistent targeting of specific regions of interest remains challenging due to variations in cap placement and limited anatomical information without additional imaging [33].

EEG's spatial limitations are more pronounced, with resolution typically at the centimeter level [31]. The inverse problem—estimating the location of neural sources within the brain from electrical measurements on the scalp—is mathematically ill-posed, with multiple possible source configurations producing identical scalp potential distributions [19]. This fundamentally limits EEG's ability to precisely localize neural activity, particularly for deep or closely spaced sources [19].

Table 2: Spatiotemporal Resolution and Practical Considerations

Feature	EEG	fNIRS
Temporal Resolution	Milliseconds (ideal for rapid processes) [31]	Seconds (limited by hemodynamic delay) [31] [32]
Spatial Resolution	Low (cm-level), limited by skull conductivity [31] [19]	Moderate (mm-cm), limited to cortical surface [31] [33]
Depth Penetration	Cortical surface, sensitive to superficial sources [31]	Superficial cortex (1-2.5 cm) [31] [33]
Movement Tolerance	Low - highly susceptible to motion artifacts [31]	Moderate - more robust to movement [31]
Experimental Environment	Best in controlled lab settings [31]	Suitable for naturalistic, real-world settings [31]

Experimental Protocols and Methodological Approaches

Protocol 1: Working Memory Assessment Using N-back Tasks

The N-back task is a widely used protocol for studying working memory and cognitive load that can be implemented with both EEG and fNIRS [34]. In this task, participants monitor a series of stimuli and indicate when the current stimulus matches the one presented N trials back. The variable N (typically 0-back to 3-back) systematically manipulates working memory load.

fNIRS Implementation: Researchers typically place fNIRS optodes over the prefrontal cortex to measure hemodynamic changes during task performance. The primary measures are changes in HbO and HbR concentrations during high (2-back) versus low (1-back) memory load conditions [34]. A scale invariance analysis of the fNIRS time series can be performed by calculating the Hurst exponent (H), which quantifies long-range temporal dependencies in the signal. Studies have successfully demonstrated that H significantly differentiates between task and rest periods, with higher H values during 1-back compared to 2-back tasks, particularly in HbR signals [34].

EEG Implementation: EEG setups for N-back tasks typically use a full-cap electrode arrangement according to the international 10-20 system. The protocol focuses on event-related potentials (ERPs) like the P300 component, which emerges approximately 300ms after stimulus presentation and reflects working memory updating. EEG also permits analysis of oscillatory power in frequency bands such as theta (4-7 Hz) over frontal regions, which increases with working memory load.

Protocol 2: Motor Imagery and Mental Arithmetic Tasks

Motor imagery (MI) and mental arithmetic (MA) tasks are commonly used in brain-computer interface (BCI) research and can be implemented with both modalities [13].

Experimental Setup: For multimodal studies, participants perform either MI (imagining moving a body part without actual movement) or MA (silently solving arithmetic problems) tasks in randomized blocks interspersed with rest periods. A typical session includes 20-30 trials per condition with appropriate cueing and inter-trial intervals [13].

fNIRS Measurements: Optodes are positioned over the motor cortex for MI tasks (C3/C4 locations) and prefrontal cortex for MA tasks. The signal processing pipeline includes converting raw light intensity to optical density, then to HbO and HbR concentration changes using the modified Beer-Lambert law. Features typically include mean, slope, and variance of HbO/HbR during tasks [13].

EEG Measurements: Electrodes are placed over sensorimotor areas (C3, Cz, C4) for MI tasks, focusing on event-related desynchronization (ERD) in the mu (8-12 Hz) and beta (13-30 Hz) rhythms. For MA tasks, frontal electrodes capture theta (4-7 Hz) synchronization and alpha (8-12 Hz) desynchronization [13].

Protocol 3: Joint EEG-fNIRS Source Reconstruction

Advanced protocols combine simultaneous EEG and fNIRS measurements to overcome the limitations of each modality [19]. The integration leverages fNIRS's spatial precision to constrain EEG's source localization.

Implementation: Researchers use integrated EEG-fNIRS caps with co-registered electrodes and optodes. The forward models for both modalities are constructed using anatomical templates (e.g., ICBM152 brain atlas) or individual MRI data when available [19].

Fusion Algorithm: The joint reconstruction algorithm utilizes fNIRS reconstruction as a spatial prior for EEG source localization. This approach can accurately resolve neuronal sources separated by as little as 2.3-3.3 cm and 50 ms—a feat impossible with either modality alone [19]. Performance is enhanced by optimizing electrode and optode placement according to the locations of target neuronal sources.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Equipment and Software for EEG-fNIRS Research

Item	Function/Purpose	Example Applications
EEG Cap with Integrated fNIRS Optodes	Allows simultaneous measurement of electrical and hemodynamic activity	Multimodal brain imaging studies, neurovascular coupling research [24]
fNIRS Light Sources & Detectors	Emits near-infrared light and detects intensity changes to measure hemoglobin concentrations	Monitoring prefrontal cortex activity during cognitive tasks [31] [32]
EEG Amplifiers	Amplifies microvolt-level electrical signals from the scalp	Recording event-related potentials, oscillatory activity [31]
3D Digitizer	Records precise locations of EEG electrodes and fNIRS optodes	Co-registration with anatomical MRI, accurate source reconstruction [33]
Signal Synchronization Hardware	Temporally aligns EEG and fNIRS data streams	Multimodal experiments requiring precise timing relationships [24]
Motion Tracking System	Monitors and quantifies head movements	Artifact identification and correction, especially in movement-tolerant fNIRS [33]
Stimulus Presentation Software	Prescribes precise timing of experimental paradigms	Cognitive task delivery with millisecond precision [13]
Joint Source Reconstruction Algorithms	Fuses EEG and fNIRS data for enhanced spatiotemporal resolution	Resolving closely spaced neural sources with high timing precision [19]

Integrated Experimental Workflow: From Data Acquisition to Fusion

The complementary nature of EEG and fNIRS is best leveraged through integrated experimental designs that capitalize on their respective strengths. The following diagram illustrates a typical workflow for simultaneous EEG-fNIRS studies:

This workflow demonstrates how simultaneous data collection enables researchers to capture both the millisecond-scale electrical dynamics (via EEG) and the localized hemodynamic responses (via fNIRS), which can then be fused for enhanced spatiotemporal resolution of brain activity [24] [13] [19].

Multimodal Integration: Leveraging Complementary Strengths

The combination of EEG and fNIRS creates a powerful multimodal approach that overcomes the limitations of each individual technique [24] [13]. Three primary fusion strategies have emerged, each with distinct advantages:

Data-Level Fusion

Data-level fusion involves directly combining raw or minimally processed data from both modalities. This approach is computationally intensive but preserves the complete information content from both signals [13]. The major challenge lies in reconciling the different temporal resolutions and physical units of EEG and fNIRS signals.

Feature-Level Fusion

Feature-level fusion extracts relevant features from each modality separately before combining them into a joint feature vector [13]. For example, EEG power band features and fNIRS HbO/HbR concentration changes can be concatenated to create a multimodal feature set. This approach has demonstrated significant improvements in classification accuracy for BCI applications, with some studies reporting accuracy increases of up to 31.83% compared to single-modality approaches [13].

Decision-Level Fusion

Decision-level fusion involves processing each modality independently through separate classification pipelines, then combining the final decisions [13]. This approach offers flexibility as the modalities can be processed using their optimal algorithms before integration. Studies have shown decision-level fusion can improve mental stress detection rates by 7.76-10.57% compared to single modalities [13].

The following diagram illustrates how these fusion strategies capitalize on the complementary nature of EEG and fNIRS signals:

EEG and fNIRS provide fundamentally different yet complementary perspectives on brain organization, each with distinct strengths and limitations for probing structure-function relationships. EEG excels at capturing global neural dynamics with millisecond precision, making it ideal for studying rapid information processing, functional connectivity, and network-level interactions across the brain. In contrast, fNIRS offers more localized insights into region-specific cortical function with better spatial specificity, particularly valuable for investigating sustained cognitive states and prefrontal cortex functions in real-world settings [31].

The choice between these modalities should be guided by specific research questions, with EEG preferred for studies requiring high temporal resolution, and fNIRS better suited for investigations prioritizing spatial localization of cortical activity [31]. Importantly, these techniques are not mutually exclusive; integrated EEG-fNIRS approaches harness their complementary strengths to achieve enhanced spatiotemporal resolution of neural processes [24] [19]. As both technologies continue to advance, their synergistic application promises to further unravel the complex relationship between brain structure and function across multiple scales of organization.

From Theory to Practice: Methodological Frameworks and Cutting-Edge Applications

In the era of big data, scientific research and technological applications increasingly rely on insights gleaned from multiple, complementary data sources. Multimodal fusion is the process of integrating these diverse data types to produce more accurate, robust, and comprehensive information than could be obtained from any single source alone [35]. In fields ranging from neuroscience to drug discovery, the ability to effectively combine data from different sensors, instruments, or modalities has become crucial for advancing scientific understanding and improving practical outcomes. The fundamental premise of multimodal fusion rests on the concept that different data sources can provide complementary information, offering unique perspectives that, when combined, create a more complete picture of the system under investigation [36].

Multimodal fusion strategies are broadly categorized into three distinct levels based on the stage at which integration occurs: data-level, feature-level, and decision-level fusion. Each approach offers different advantages and involves specific methodological considerations. Data-level fusion (also called pixel-level or early fusion) operates directly on raw or minimally processed data. Feature-level fusion (sometimes called intermediate fusion) combines features extracted from each modality separately. Decision-level fusion (also known as late fusion) integrates the outputs from multiple classifiers or models, each trained on a different modality [37] [36]. The choice between these strategies depends on multiple factors, including data characteristics, computational resources, and the specific requirements of the application.

This guide provides a comprehensive comparison of these three fusion strategies, with particular emphasis on their application in neuroscience research comparing electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) for studying brain structure-function relationships. We present structured comparisons, experimental protocols, performance data, and practical implementation guidelines to assist researchers in selecting and applying the most appropriate fusion strategy for their specific research needs.

Comparative Analysis of Fusion Strategies

The table below provides a systematic comparison of the three primary multimodal fusion strategies, highlighting their key characteristics, advantages, and limitations.

Table 1: Comprehensive Comparison of Multimodal Fusion Strategies

Aspect	Data-Level Fusion	Feature-Level Fusion	Decision-Level Fusion
Fusion Stage	Raw data or input level	Feature representation level	Model output or decision level
Alternative Names	Early fusion, pixel-level fusion	Intermediate fusion, feature fusion	Late fusion, decision fusion
Data Requirements	Requires precise spatial-temporal alignment	Allows some feature alignment flexibility	No alignment needed; operates on decisions
Complexity	High (due to alignment needs)	Moderate (requires feature engineering)	Low (simpler implementation)
Information Preservation	High (retains all original information)	Moderate (depends on feature extraction)	Low (only final decisions are combined)
Noise Sensitivity	High (noise affects all subsequent processing)	Moderate (feature extraction can filter noise)	Low (robust to noisy inputs)
Modality Interdependence	Captures fine-grained interactions	Can learn cross-modal relationships	Treats modalities independently
Best-Suited Applications	Multisensor image fusion, aligned data sources	Cross-modal retrieval, pattern recognition	Ensemble methods, distributed systems

Data-level fusion, also called early fusion or pixel-level fusion, involves combining raw data from multiple sources before any significant feature extraction or processing has occurred [36]. This approach operates on the principle that the richest information content resides in the original data streams, and early integration preserves the full fidelity of this information. The primary challenge with data-level fusion is the requirement for precise spatial and temporal alignment between different data sources, which can be computationally demanding and sometimes practically difficult to achieve [38]. In neuroscience applications, this might involve precisely aligning EEG electrical activity measurements with fNIRS hemodynamic responses despite their different temporal resolutions and physiological origins.

Feature-level fusion, often termed intermediate fusion, represents a compromise between the rich information preservation of data-level fusion and the practical implementation advantages of decision-level fusion [37]. In this approach, each modality undergoes initial processing to extract relevant features, which are then combined into a unified representation before final model training or inference. This strategy allows for the learning of cross-modal relationships at an intermediate abstraction level, often capturing interactions that might be missed in other approaches. Feature-level fusion has demonstrated remarkable success in various applications; for instance, in schizophrenia detection using resting-state fMRI data, feature-level fusion achieved an impressive 98.57% accuracy, significantly outperforming single-modality approaches [39].

Decision-level fusion, also known as late fusion, operates by combining the final outputs or decisions from multiple models, each trained on a different modality [37]. This approach treats each data stream independently until the final decision stage, where results are aggregated through methods such as voting, weighted averaging, or meta-classification. The primary advantage of this strategy is its robustness to missing modalities and its flexibility in handling heterogeneous data types that may be difficult to align or process jointly. In medical applications, decision-level fusion has proven highly effective; for example, in schizophrenia detection, it achieved 97.85% accuracy by combining decisions from multiple fMRI feature classifiers [39]. Similarly, in drug discovery, late fusion strategies have been successfully employed to combine predictions from multiple molecular representations to enhance docking score predictions [40].

Fusion Strategies in EEG-fNIRS Research

The integration of EEG and fNIRS data provides a powerful approach for investigating brain structure-function relationships, as these modalities offer complementary information about neural activity. EEG captures electrical activity with millisecond temporal resolution, reflecting postsynaptic potentials of neuronal populations. In contrast, fNIRS measures hemodynamic responses through changes in blood oxygenation, providing better spatial localization but slower temporal response due to neurovascular coupling delays [4]. The multimodal fusion of these complementary signals enables researchers to overcome the limitations of each individual modality and gain more comprehensive insights into brain function.

In recent studies, researchers have employed graph signal processing to characterize both global and local structure-function coupling using source-reconstructed EEG and fNIRS signals during both resting state and task conditions [4]. Results demonstrate that fNIRS structure-function coupling resembles slower-frequency EEG coupling at rest, with variations across brain states and oscillations. Importantly, the relationship is heterogeneous across brain regions, with greater coupling in sensory cortex and increased decoupling in association cortex, following the unimodal to transmodal gradient [4] [5]. These findings highlight the importance of selecting appropriate fusion strategies that account for the distinct characteristics of EEG and fNIRS signals.

Table 2: Performance Comparison of Fusion Strategies in Different Applications

Application Domain	Fusion Strategy	Reported Performance	Comparative Baseline
Schizophrenia Detection (fMRI)	Feature-Level Fusion	98.57% accuracy, 99.71% sensitivity, 97.66% specificity [39]	Single-feature approaches: <90% accuracy
Schizophrenia Detection (fMRI)	Decision-Level Fusion	97.85% accuracy, 98.33% sensitivity, 96.83% specificity [39]	Single-feature approaches: <90% accuracy
Drug Discovery (Docking Scores)	Early Fusion (Data-Level)	Outperformed single-representation models and late fusion [40]	Late fusion and single-representation models
Land Cover Classification	Decision-Level Fusion	95% accuracy with data fusion vs. 85% with single data source [36]	Single data source: 85% accuracy
Crop Yield Prediction	Feature-Level Fusion	R² = 0.85 with data fusion vs. 0.65 with single data source [36]	Single data source: R² = 0.65

The workflow for implementing multimodal fusion in EEG-fNIRS studies typically involves several critical stages, as illustrated in the following diagram:

EEG-fNIRS Multimodal Fusion Workflow

This workflow demonstrates how the three fusion strategies can be implemented either independently or in combination for EEG-fNIRS integration. The precise alignment of temporal characteristics between EEG and fNIRS signals is particularly crucial for data-level and feature-level fusion approaches, given their different temporal dynamics and physiological origins.

Experimental Protocols and Methodologies

Experimental Design for EEG-fNIRS Fusion Studies

Implementing effective multimodal fusion requires careful experimental design and methodological rigor. For EEG-fNIRS studies examining structure-function relationships, the following protocol has been successfully employed [4]:

Subject Population and Data Acquisition:

• Recruit 18-30 healthy adult subjects following appropriate ethical guidelines
• Acquire simultaneous EEG and fNIRS recordings during both resting-state and task conditions (e.g., motor imagery tasks)
• For EEG: Use 30+ electrodes placed according to the international 10-5 system, sampled at 1000 Hz
• For fNIRS: Employ 36 channels (14 sources, 16 detectors) with inter-optode distance of 30mm, following the standardized 10-20 EEG system, sampled at 12.5 Hz
• Coregister EEG electrodes and fNIRS optodes by spatially aligning them to an anatomical MRI template using digitized positions relative to known scalp landmarks

Data Preprocessing Pipeline:

• Apply bandpass filtering appropriate to each modality (EEG: frequency bands of interest; fNIRS: 0.02-0.08 Hz for resting state)
• Implement artifact removal techniques (e.g., principal component analysis for physiological noise removal in fNIRS)
• Perform quality assessment metrics (e.g., scalp-coupled index for fNIRS signal quality)
• Exclude subjects with poor signal quality (>50% channels with SCI <0.7 for fNIRS)
• Segment data into appropriate epochs (e.g., 10-second task epochs)

Data Analysis and Fusion Implementation:

• Map structural and functional data onto common anatomical frameworks (e.g., Desikan-Killiany atlas)
• Employ mathematical frameworks like graph signal processing to quantify structure-function coupling
• Calculate metrics such as Structural-Decoupling Index (SDI) to quantify degree of structure-function dependency
• Implement cross-validation procedures appropriate to sample size
• Apply statistical tests accounting for multiple comparisons

Generalized Protocol for Multimodal Fusion

Beyond EEG-fNIRS specific applications, successful implementation of multimodal fusion typically follows this generalized protocol:

Data Preparation Phase:

• Data Collection: Acquire data from multiple modalities with appropriate synchronization
• Quality Control: Assess data quality and exclude poor-quality recordings
• Preprocessing: Apply modality-specific preprocessing (filtering, artifact removal, normalization)
• Alignment: Perform spatial, temporal, and semantic alignment across modalities

Fusion Implementation Phase:

• Strategy Selection: Choose appropriate fusion level based on data characteristics and research questions
• Feature Engineering: For feature-level fusion, extract relevant features from each modality
• Model Architecture: Design appropriate model architecture for the chosen fusion strategy
• Training/Validation: Implement cross-validation and hyperparameter tuning

Evaluation Phase:

• Performance Assessment: Evaluate using appropriate metrics (accuracy, sensitivity, specificity, etc.)
• Ablation Studies: Compare fused model performance against single-modality baselines
• Interpretation: Apply explainable AI techniques to interpret fusion model decisions
• Validation: Validate findings against established biological or physical principles

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of multimodal fusion requires both computational tools and experimental resources. The following table outlines key solutions and their functions in EEG-fNIRS fusion research.

Table 3: Essential Research Toolkit for Multimodal EEG-fNIRS Studies

Category	Item	Specification/Function
Hardware Equipment	EEG System	30+ electrodes, 1000Hz+ sampling rate, international 10-5 placement
	fNIRS System	30+ channels, 760nm & 850nm wavelengths, 10Hz+ sampling rate
	Structural MRI	High-resolution T1-weighted images for anatomical reference
Software & Analysis Tools	Brainstorm/MNE-Python	Open-source tools for EEG/MEG and fNIRS data processing
	Graph Signal Processing Toolbox	Mathematical framework for structure-function analysis
	Alignment Algorithms	Spatial coregistration of EEG electrodes and fNIRS optodes
Methodological Components	Desikan-Killiany Atlas	Standard brain parcellation for region-of-interest analysis
	Structural-Decoupling Index (SDI)	Quantifies degree of structure-function dependency per region
	Cross-Validation Framework	Ensures robust performance estimation, especially with limited samples
Data Quality Assessment	Scalp-Coupled Index (SCI)	fNIRS signal quality metric (threshold: >0.7)
	Global Variance in Temporal Derivative (GVTD)	Identifies motion artifacts in fNIRS data
	Modality Dropout	Training technique for robustness to missing modalities

Performance Metrics and Interpretation

The ultimate validation of any fusion strategy lies in its performance against established benchmarks and single-modality approaches. The diagram below illustrates the comparative performance of different fusion strategies across multiple applications, based on published research findings.

Comparative Performance of Fusion Strategies

When interpreting fusion results, several key considerations emerge from the research. First, the complementary nature of different modalities significantly influences fusion effectiveness. In EEG-fNIRS studies, the electrical activity captured by EEG and the hemodynamic responses measured by fNIRS provide distinct but related information about neural processes, creating ideal conditions for beneficial fusion [4]. Second, the temporal characteristics of each modality must be carefully considered. Research shows that fNIRS structure-function coupling resembles slower-frequency EEG coupling at rest, suggesting that effective fusion requires accounting for these temporal dynamics [5].

Third, regional variations in brain organization impact fusion outcomes. Studies consistently show heterogeneous structure-function relationships across brain regions, with stronger coupling in sensory areas and greater decoupling in association cortex, following the unimodal to transmodal gradient [4]. This regional specificity means that fusion approaches may need to be adapted for different brain systems or networks. Finally, task demands influence optimal fusion strategies. During motor imagery tasks, for instance, both EEG and fNIRS show specificity for the somatomotor network, suggesting that task context should guide fusion approach selection [4].

Multimodal fusion represents a powerful paradigm for advancing scientific research across multiple domains, particularly in neuroscience investigations of brain structure-function relationships using complementary techniques like EEG and fNIRS. Through this comparative analysis, we have delineated the distinct characteristics, implementation requirements, and performance profiles of data-level, feature-level, and decision-level fusion strategies.

The evidence indicates that feature-level fusion often provides an optimal balance between information preservation and practical implementation for many applications, achieving superior performance in tasks such as schizophrenia detection (98.57% accuracy) and crop yield prediction (R² = 0.85) [39] [36]. However, decision-level fusion offers compelling advantages in scenarios requiring robustness to missing data or noisy inputs, while data-level fusion remains valuable when precise alignment of high-quality data sources is feasible.

In EEG-fNIRS research specifically, effective fusion requires careful attention to the distinct temporal characteristics and physiological origins of each modality. The emerging finding of heterogeneous structure-function coupling across brain regions further underscores the need for sophisticated fusion approaches that account for this neurobiological complexity [4] [5]. As multimodal data becomes increasingly prevalent in scientific research, the strategic implementation of appropriate fusion techniques will continue to enhance our ability to extract meaningful insights from complex, complementary data sources, ultimately advancing both theoretical understanding and practical applications across diverse scientific domains.

Understanding the relationship between the brain's structural wiring and its dynamic functional activity is a fundamental pursuit in neuroscience. This structure-function relationship is crucial for uncovering the large-scale organizational principles of the human brain and has significant implications for understanding neural mechanisms underlying behavior, cognition, and various neurological disabilities [4]. The emergence of graph signal processing (GSP) provides a powerful mathematical framework to quantitatively explore this relationship by analyzing functional brain signals in the context of underlying structural connectivity networks [41] [42]. GSP extends traditional signal processing concepts to irregular domains represented by graphs, enabling researchers to investigate how functional signals—whether electrical or hemodynamic—are shaped by the brain's structural architecture [41].

Within this framework, comparative studies between different neuroimaging modalities have gained prominence, particularly comparisons between electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). These modalities capture complementary aspects of brain activity at different temporal scales, and GSP provides the mathematical tools to quantify their relationship to the common structural substrate [5] [4]. This article provides a comprehensive comparison of how GSP leverages EEG and fNIRS to quantify structure-function coupling, offering researchers evidence-based guidance for selecting appropriate methodologies for specific research questions.

Theoretical Foundations of Graph Signal Processing

Graph signal processing bridges spectral graph theory and conventional signal processing by defining fundamental operations for signals indexed by graphs [42]. In GSP, a brain network is represented as a graph (G = (V, A)), where (V = {0, 1, ..., n-1}) is a set of (n) nodes representing distinct brain regions, and (A \in \mathbb{R}^{n \times n}) is a weighted adjacency matrix with entries (a_{ij} \geq 0) representing the strength of structural connectivity between regions (i) and (j) [41].

Core Mathematical Framework

The GSP framework relies on several key mathematical constructs:

• Degree Matrix: A diagonal matrix (D \in \mathbb{R}+^{n \times n}) where (D{ii} = \sum{j=0}^{n-1} a{ij})
• Laplacian Matrix: Defined as (L = D - A), which is a real-valued positive semidefinite matrix
• Normalized Laplacian: Variants include (L{sn} = D^{-1/2}LD^{-1/2}) or (L{rw} = D^{-1}L) [41]

The eigenvector decomposition of the Laplacian matrix (L) enables the definition of the graph Fourier transform (GFT), which projects spatial brain signals onto frequency-ordered graph Fourier modes [42]. This transformation allows the analysis of brain signals in the graph spectral domain, analogous to how the classical Fourier transform analyzes temporal frequencies.

Key GSP Concepts for Brain Signal Analysis

GSP provides several analytical tools specifically valuable for neurophysiological signal analysis [42]:

• Graph Fourier Transform and Dimensionality Reduction: The GFT projects spatial signals onto graph eigenmodes, enabling dimensionality reduction by selecting specific graph frequencies.
• Analysis of Total Variation: Total variation measures how much a signal varies across the graph structure, providing insights into neural communication patterns.
• Graph Denoising: GSP can filter multivariate signals by applying graph filters that leverage the connectivity structure.
• Graph Learning: This approach optimizes the graph structure to reduce the total variation of signals.

Figure 1: GSP Workflow for Quantifying Structure-Function Coupling in Brain Networks

Experimental Approaches: EEG vs. fNIRS in GSP Studies

Multimodal Data Acquisition and Preprocessing

Recent research has employed simultaneous EEG-fNIRS recordings to investigate structure-function coupling across electrical and hemodynamic domains. A typical experimental setup includes:

• EEG Acquisition: 30 electrodes placed according to the international 10-5 system, recording at 1000 Hz sampling rate (often downsampled to 200 Hz) [4]
• fNIRS Acquisition: 36 channels (14 sources and 16 detectors) with an inter-optode distance of 30 mm, placed according to the 10-20 system, recording at 12.5 Hz sampling rate (often downsampled to 10 Hz) using two wavelengths (760 nm and 850 nm) [4]
• Experimental Paradigm: Typically includes resting-state sessions (1 minute) and task conditions (e.g., motor imagery tasks with 30 trials of 10-second each) [4]

Data preprocessing follows modality-specific pipelines. EEG preprocessing typically involves filtering, artifact removal, and source reconstruction. fNIRS preprocessing includes optical density transformation, signal quality assessment using scalp-coupled index (SCI), bandpass filtering (0.02-0.08 Hz for resting-state), motion artifact correction, and removal of systemic physiological effects using principal component analysis (PCA) [4].

Structural Connectivity Processing

The structural backbone for GSP analysis is typically derived from diffusion magnetic resonance imaging (dMRI) or diffusion tensor imaging (DTI) data, which visualize physical connections and reconstruct nerve fibers [4]. Alternatively, geometric-based connectivity models can generate adjacency matrices where connection strength between regions is proportional to the inverse square of the distance between them [41]. These structural networks are mapped onto standardized brain atlases (e.g., Desikan-Killiany atlas or HCP's multimodal parcellation) to define regions of interest for subsequent analysis [41] [4].

Key GSP Metrics for Structure-Function Coupling

The structural-decoupling index (SDI) has emerged as a primary metric for quantifying structure-function relationships in brain networks [4] [43]. SDI measures the degree of structure-function dependency for each brain region, with higher values indicating greater decoupling from the structural backbone. Additional GSP features include:

• Power Spectral Density: Analysis of signal energy distribution across graph frequencies
• Alignment and Liberality: Components analogous to low and high frequencies in classical Fourier analysis [41]
• Total Variation: Measuring signal smoothness over the graph structure [42]

Table 1: Core GSP Metrics for Structure-Function Coupling Analysis

Metric	Definition	Interpretation	Application in EEG/fNIRS
Structural-Decoupling Index (SDI)	Quantifies degree of structure-function dependency for each brain region	Higher values indicate greater decoupling from structural constraints; reveals regional heterogeneity	Applied to both EEG and fNIRS; shows modality-specific patterns [4] [43]
Graph Power Spectral Density	Distribution of signal energy across graph frequencies	Reveals how functional activity utilizes structural connectivity at different graph frequencies	Used for both modalities; identifies dominant graph frequency components [43]
Total Variation (TV)	Measures how much a signal varies across the graph structure	Lower TV indicates smoother signals on the graph; correlates with consciousness states [42]	Applied to EEG alpha-band signals; shows state-dependent variations [42]

Comparative Analysis: EEG vs. fNIRS in GSP Applications

Temporal Characteristics and Scale-Dependent Coupling

EEG and fNIRS capture brain activity at fundamentally different temporal scales, which significantly impacts their structure-function coupling profiles as revealed by GSP analysis:

• EEG measures direct electrical neural activity with high temporal resolution (milliseconds), capturing fast neural oscillations and transient synchronization/desynchronization patterns [4] [44]
• fNIRS measures hemodynamic responses through oxygenation changes with slower temporal resolution (seconds), reflecting the delayed metabolic consequences of neural activity [4]

GSP analysis reveals that fNIRS structure-function coupling resembles slower-frequency EEG coupling patterns during resting state, suggesting that hemodynamic networks align with low-frequency electrical oscillations in their relationship to structural connectivity [5] [4]. This relationship varies significantly across different brain states and oscillatory bands, highlighting the scale-dependent nature of structure-function coupling.

Regional Heterogeneity: Unimodal-Transmodal Gradient

Both EEG and fNIRS GSP analyses reveal consistent regional heterogeneity in structure-function coupling, following the unimodal to transmodal gradient of brain organization [4] [45]:

• Unimodal Regions (primary sensory and motor areas) show stronger structure-function coupling across both modalities, indicating tighter constraint by structural connectivity
• Transmodal Regions (association areas supporting higher-order cognition) exhibit greater decoupling (higher SDI), suggesting more flexible functional dynamics less constrained by anatomy

However, notable differences emerge between modalities in specific networks. The frontoparietal network, crucial for cognitive control and adaptive functions, shows particularly prominent discrepancies between EEG and fNIRS structure-function coupling patterns [4] [45].

Figure 2: Unimodal-Transmodal Gradient in Structure-Function Coupling

Task-Specific Modulation of Coupling Patterns

The structure-function relationship dynamically modulates between resting state and task conditions, with important differences between EEG and fNIRS:

• During motor imagery tasks, both modalities show specific activation of the somatomotor network, but with distinct coupling characteristics [4]
• Cross-band representations of neural activity reveal lower correspondence between electrical and hemodynamic activity in the transmodal cortex, regardless of brain state [4]
• Multilayer network approaches that integrate EEG and fNIRS data outperform unimodal analyses, offering a more comprehensive view of brain function, particularly during tasks [44]

Table 2: Performance Comparison of EEG vs. fNIRS in GSP Applications

Characteristic	EEG	fNIRS	Comparative Implications
Temporal Resolution	High (milliseconds) [44]	Lower (seconds) [4]	EEG captures rapid coupling changes; fNIRS reflects sustained processes
Spatial Resolution	Limited by source reconstruction	Better inherent spatial resolution [4]	fNIRS may provide more localized coupling estimates
Sensitivity to Neural Events	Direct electrical measurement	Indirect hemodynamic correlate	EEG captures neural synchrony; fNIRS reflects metabolic demands
Structure-Function Coupling at Rest	Frequency-dependent coupling patterns	Resembles slow-frequency EEG coupling [5]	Complementary aspects of brain organization
Task-Based Modulation	Precise timing of information transfer [44]	Specificity for sustained neural processes [44]	EEG excels for transient tasks; fNIRS for sustained activation
Frontoparietal Network Coupling	Distinct coupling patterns	Divergent from EEG patterns [4]	Modality-dependent network organization
Unimodal-Transmodal Gradient	Stronger coupling in unimodal areas [4]	Similar gradient but different magnitude	Consistent organizational principle across modalities
Practical Considerations	Sensitive to electrical artifacts	Less susceptible to movement artifacts [4]	fNIRS may be preferable in movement-prone populations

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Materials for GSP Studies in EEG/fNIRS Research

Item Category	Specific Examples	Function in GSP Research
Neuroimaging Hardware	30-channel EEG system (10-5 placement); 36-channel fNIRS system (14 sources, 16 detectors) [4]	Simultaneous acquisition of electrical and hemodynamic brain activity
Structural Imaging	Diffusion MRI/DTI sequences; Structural T1-weighted MRI	Reconstruction of structural connectivity matrices for GSP analysis
Software Platforms	Brainstorm; MNE-Python; MATLAB with Signal Processing Toolbox [41] [4]	Data preprocessing, graph construction, and implementation of GSP algorithms
Brain Atlases	Desikan-Killiany Atlas; HCP Multimodal Parcellation (MMP) [41] [4]	Standardized parcellation for defining graph nodes and aligning multi-modal data
GSP Analysis Tools	Custom scripts for Graph Fourier Transform; Structural-Decoupling Index calculation [4] [43]	Quantification of structure-function coupling metrics
Validation Frameworks	Multiple baseline models; Cross-validation approaches [42]	Verification of GSP method performance and specificity of results

Graph signal processing provides a powerful mathematical framework for quantifying structure-function relationships in brain networks, offering distinct yet complementary insights when applied to EEG and fNIRS data. The comparative analysis reveals that neither modality is superior; rather, they capture different aspects of brain organization operating at different temporal scales.

EEG excels in capturing rapid electrical neural synchronization tightly coupled to structural constraints in unimodal regions, while fNIRS reveals slower hemodynamic patterns that show stronger decoupling in transmodal areas. The emerging approach of multimodal integration using multilayer networks and simultaneous EEG-fNIRS recordings presents the most promising path forward, leveraging the complementary strengths of both modalities [44].

For researchers and drug development professionals, selection between EEG and fNIRS for GSP studies should be guided by specific research questions: EEG for investigations of fast neural dynamics and transient cognitive processes, fNIRS for studies of sustained metabolic demands and clinical applications requiring robustness to movement artifacts. Future methodological advances will likely focus on optimizing feature selection from the GSP framework—with power spectral density and structural decoupling index currently showing particular promise—and developing more sophisticated integration frameworks to fully leverage the complementary information provided by electrical and hemodynamic brain signals [43].

A fundamental pursuit in neuroscience is understanding the relationship between the brain's structural architecture and its dynamic functions. Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) are two non-invasive neuroimaging techniques that are pivotal in this endeavor, yet they capture distinctly different physiological phenomena [46]. EEG measures the brain's electrical activity, offering a direct window into neural firing with millisecond temporal resolution. In contrast, fNIRS monitors hemodynamic responses, specifically changes in blood oxygenation, providing superior spatial localization for cortical areas [46]. This case study objectively compares the performance of these two modalities, and their integrated use, in classifying Motor Imagery (MI) and Mental Arithmetic (MA) tasks—paradigms critical for brain-computer interfaces (BCIs) and cognitive neuroscience. The analysis is framed within the broader thesis of how electrical and hemodynamic signals contribute to our understanding of structure-function relationships in the human brain, highlighting that the choice of modality can fundamentally influence the observed neural correlates of cognitive processes [4].

Technical Comparison of EEG and fNIRS

Table 1: Fundamental Characteristics of EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity from postsynaptic potentials of cortical neurons [46]	Hemodynamic response: changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) [46]
Temporal Resolution	High (millisecond scale) [46]	Low (seconds-scale, limited by hemodynamic response delay) [46]
Spatial Resolution	Low (centimeter-level) due to signal dispersion through skull and scalp [46]	Moderate, superior to EEG for surface cortical areas [46]
Depth of Measurement	Primarily the cortical surface [46]	Outer cortex (approximately 1–2.5 cm deep) [46]
Key Strengths	Captures fast neural dynamics; ideal for event-related potentials (ERPs) [46]	Better motion tolerance; suitable for naturalistic, sustained tasks [46]
Key Limitations	Susceptible to motion artifacts; poor spatial localization [46]	Slow response; indirect measure of neural activity [46]

Performance Data in Task Classification

Classification of Motor Imagery (MI) Tasks

Motor imagery, the mental rehearsal of a movement without physical execution, activates brain regions such as the premotor cortex (PMC), supplementary motor area (SMA), and primary motor cortex (M1) [47]. The ability to accurately classify these imagined movements is a cornerstone of BCIs for rehabilitation.

Table 2: Performance in Motor Imagery Classification

Modality / Approach	Reported Performance (Accuracy)	Key Findings and Experimental Context
EEG-fNIRS Fusion (Deep Learning)	83.26% [48]	An end-to-end fusion model using evidence theory on the TU-Berlin-A dataset, representing a 3.78% improvement over state-of-the-art methods [48].
EEG-fNIRS Fusion (ECA-FusionNet)	Outperformed unimodal and existing fusion methods [49]	A hybrid network that performs feature-level and decision-level fusion on a public dataset, demonstrating improved adaptability and robustness [49].
EEG-only (Multi-Joint MI)	65.49% (highest, 2-class) [50]	Classification between hand and shoulder MI tasks using a deep learning method (ShallowConvNet) on a dataset of 8 MI tasks from 18 subjects [50].
fNIRS-only (MI vs. Other Modes)	N/A (Activation Analysis)	fNIRS measurements showed MI induced greater activity in the PMC and SMA compared to Action Observation (AO), and activation levels comparable to Motor Execution (ME) [47].

Classification and Processing of Mental Arithmetic (MA) Tasks

Mental arithmetic, a complex cognitive task, engages a network including the supramarginal gyrus (SMG), superior temporal gyrus (STG), and inferior frontal gyrus (IFG) [51]. The neurocognitive processing of arithmetic is influenced by both task complexity and individual math ability [51].

Table 3: Performance and Findings in Mental Arithmetic Studies

Aspect / Modality	Key Findings	Experimental Context
Temporal Dynamics (EEG vs. fNIRS)	Optimal signal segment lengths for state estimation differed between EEG and fNIRS [52]	A study with 10 participants performing an MA task evaluated classification accuracy of task vs. rest states using 30-second, 1-minute, and 2-minute signal segments [52].
Individual Differences (fNIRS-EEG)	Individuals with low math ability showed less activation in left SMG, STG, and IFG during complex arithmetic [51]	A combined fNIRS-EEG study on individuals with high and low math ability solving simple and complex multiplication and division problems [51].
Oscillatory EEG Activity	Theta and alpha desynchronization observed with increasing arithmetic complexity, but no interaction with math ability [51]	EEG analysis revealed general band power changes related to cognitive load during arithmetic tasks [51].

Experimental Protocols for Key Studies

Protocol: Multimodal MI Classification with Deep Learning

This protocol is based on the study that achieved 83.26% accuracy using EEG-fNIRS fusion [48].

• 1. Participants & Task: Participants perform a cue-guided MI task, such as imagining left-hand or right-hand movements.
• 2. Data Acquisition:
  • EEG: Recorded using a multi-electrode cap (e.g., 30 electrodes following the international 10-5 system) at a high sampling rate (e.g., 1000 Hz).
  • fNIRS: Recorded simultaneously using sources and detectors placed over the motor cortex (e.g., 36 channels), with wavelengths of 760 nm and 850 nm to measure HbO and HbR changes at a lower sampling rate (e.g., 10-12.5 Hz). Systems are synchronized via triggers [4] [50].
• 3. Signal Preprocessing:
  • EEG: Processing may include filtering, artifact removal (e.g., ocular, muscular), and re-referencing.
  • fNIRS: Raw light intensity is converted to optical density, then to HbO and HbR concentrations using the Modified Beer-Lambert Law. Processing includes bandpass filtering (e.g., 0.01-0.1 Hz) and motion artifact correction [4].
• 4. Feature Extraction:
  • EEG Features: Spatiotemporal features are extracted using deep learning modules like dual-scale temporal convolution and depthwise separable convolution. A hybrid attention mechanism can be used to highlight salient features [48].
  • fNIRS Features: Spatial convolution across channels explores activation differences between brain regions. Parallel temporal convolution combined with a Gated Recurrent Unit (GRU) captures the temporal dynamics of the hemodynamic response [48].
• 5. Decision Fusion & Classification: The extracted features from both modalities are used for initial classification. The decision outputs are then fused using a theory like Dempster-Shafer Theory (DST), which models uncertainty and combines evidence from both modalities to make a final classification [48].

Protocol: Structure-Function Coupling During Rest and MI

This protocol investigates the relationship between brain structure and function using simultaneous EEG-fNIRS [4] [5].

• 1. Participants & States: Data is collected from participants during two conditions: a resting state (e.g., 1 minute) and an MI task block (e.g., multiple trials of left/right-hand MI).
• 2. Multimodal Data Acquisition: Simultaneous EEG (e.g., 30 electrodes) and fNIRS (e.g., 36 channels) are recorded, with optodes and electrodes co-registered on the scalp using the international 10-20/10-5 system [4].
• 3. Data Preprocessing & Source Reconstruction: Both EEG and fNIRS data are preprocessed. The functional data is then source-reconstructed and mapped onto a common brain atlas (e.g., Desikan-Killiany atlas) to obtain regional time series.
• 4. Network Construction:
  • Structural Connectome: A group-consensus structural connectome derived from diffusion MRI (dMRI) data (e.g., from the ARCHI database) is used, representing the physical white matter connections between brain regions.
  • Functional Networks: Functional connectivity networks are computed from the EEG and fNIRS source time series for both rest and task states, representing statistical dependencies between regional signals.
• 5. Quantifying Structure-Function Coupling: Using the mathematical framework of Graph Signal Processing (GSP), a Structural-decoupling index (SDI) is computed for each brain region. This index quantifies the (dis)alignment between the functional patterns and the underlying structural substrate [4] [5]. This analysis reveals that fNIRS coupling resembles slower-frequency EEG coupling, and that coupling is stronger in sensory (unimodal) cortices than in association (transmodal) cortices [4].

Visualizing the Hybrid EEG-fNIRS Fusion Workflow

The following diagram illustrates the integrated experimental and analytical pipeline for a multimodal classification study, synthesizing the protocols above.

Diagram 1: Workflow for hybrid EEG-fNIRS task classification.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 4: Key Materials and Equipment for EEG-fNIRS Research

Item	Function / Description
Integrated EEG-fNIRS Cap	A specialized head cap with pre-defined openings and holders to accommodate both EEG electrodes and fNIRS optodes without interference, often based on the international 10-20/10-5 system [53] [50].
Amplifier & fNIRS System	Synchronized hardware for data acquisition. Examples include a g.HIamp amplifier for EEG and a continuous-wave fNIRS system like NirScan, synchronized via external triggers [53].
Conductive Gel & Abrasive Prep	For preparing the scalp and ensuring good electrical contact for EEG electrodes to maintain impedance below a threshold (e.g., 10 kΩ) [50].
Source-Detector Location Digitizer	A tool to record the precise 3D coordinates of EEG electrodes and fNIRS optodes on the scalp, which is crucial for accurate co-registration with anatomical brain templates [4].
Structural Connectome Database	A pre-existing dataset (e.g., the ARCHI database) providing diffusion MRI-derived structural connectivity matrices, used as a common reference for structure-function coupling studies [4].
Software Toolboxes (MNE, Brainstorm)	Open-source software packages used for comprehensive preprocessing, source reconstruction, and analysis of multimodal neuroimaging data [4].

Biomarkers are fundamentally reshaping the landscape of pharmaceutical research and development by providing objective indicators of biological processes, pathogenic states, or pharmacological responses to therapeutic interventions. These molecular, histological, radiographic, or physiological characteristics serve as essential tools throughout the drug development continuum, from initial discovery through clinical trials and regulatory approval. The U.S. Food and Drug Administration (FDA) categorizes biomarkers into several distinct types based on their specific application, including susceptibility/risk, diagnostic, monitoring, prognostic, predictive, and pharmacodynamic/response biomarkers, with some biomarkers fitting multiple categories depending on context of use [54].

The integration of biomarkers is particularly crucial in addressing one of the most significant challenges in pharmaceutical development: the high attrition rate of drug candidates during clinical trials. Biomarker-driven approaches help de-risk this process by enabling more informed target selection, better patient stratification, and earlier efficacy signals. Furthermore, the emergence of artificial intelligence (AI) is radically transforming biomarker analysis, uncovering hidden biological patterns in complex datasets that exceed human observational capacity [55]. AI-driven pathology tools and biomarker analysis provide deeper biological insights for clinical decision-making, particularly in complex disease areas like oncology where reliable biomarkers for personalized treatment remain a major challenge [55].

Table: Biomarker Categories and Applications in Drug Development

Biomarker Category	Primary Function	Example Application
Susceptibility/Risk	Identify individuals with increased disease risk	BRCA1/2 mutations for breast/ovarian cancer risk
Diagnostic	Detect or confirm presence of a disease	Hemoglobin A1c for diabetes mellitus
Prognostic	Identify likelihood of disease progression	Total kidney volume for autosomal dominant polycystic kidney disease
Monitoring	Assess disease status or response to treatment	HCV RNA viral load for Hepatitis C infection
Predictive	Identify responders to specific therapies	EGFR mutation status in non-small cell lung cancer
Pharmacodynamic/Response	Measure biological response to therapeutic intervention	HIV RNA viral load as surrogate endpoint in HIV trials

Structural and Functional Biomarkers in Neuroscience

The relationship between brain structure and function represents a fundamental concept in neuroscience, with significant implications for understanding neurological diseases and developing targeted therapies. Structure-function relationships provide crucial insights into neural mechanisms underlying behaviors, cognition, and disability, holding particular significance for early diagnosis and guiding therapeutic interventions [4]. Advanced neuroimaging technologies now enable high-throughput reconstruction of neural circuits across various spatiotemporal scales, offering unprecedented opportunities for biomarker discovery in neurological disorders.

Two complementary neuroimaging modalities—electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS)—provide distinct perspectives on brain function with implications for biomarker development. EEG measures electrical activity generated by neuronal firing with millisecond temporal resolution, making it ideal for capturing rapid neural dynamics [56]. In contrast, fNIRS measures hemodynamic responses through near-infrared light penetration, providing information about regional blood oxygenation changes that indirectly reflect neural activity with higher spatial resolution than EEG but slower temporal response due to neurovascular coupling delays [57] [56].

Recent research has revealed that the structure-function relationship in the brain is not uniform but follows a unimodal to transmodal gradient [4] [5]. Unimodal areas (sensory and motor regions) show stronger structure-function coupling, meaning functional activity aligns closely with underlying anatomical connections. Conversely, transmodal areas (association regions supporting higher-order cognition) demonstrate greater decoupling, allowing for more flexible integration of information across distributed networks [4] [5]. This organizational principle has profound implications for understanding neurological diseases, which often differentially affect specific regions along this gradient.

Table: Technical Comparison of EEG and fNIRS for Biomarker Discovery

Parameter	EEG	fNIRS
Measured Signal	Electrical activity from neuronal firing	Hemodynamic response (oxygenated/deoxygenated hemoglobin)
Temporal Resolution	High (milliseconds)	Moderate (seconds)
Spatial Resolution	Limited (centimeters)	Moderate (1-3 cm)
Portability	High	High
Cost	Moderate	Moderate to High
Susceptibility to Artifacts	Sensitive to electrical noise and motion	Sensitive to scalp blood flow and motion
Primary Biomarker Applications	Epilepsy, sleep disorders, cognitive states	Cognitive impairment, motor recovery, neurodevelopment
Structure-Function Coupling	Stronger in unimodal regions	Resembles slow-frequency EEG coupling

Methodological Approaches in Biomarker Discovery

Emerging Technologies and Analytical Frameworks

The landscape of biomarker discovery is undergoing a technological renaissance, driven by breakthroughs in multi-omics, spatial biology, AI, and high-throughput analytics [58]. These approaches offer higher resolutions, faster speed, and enhanced translational relevance compared to traditional methods. Spatial biology techniques, including spatial transcriptomics and multiplex immunohistochemistry, have proven particularly valuable for characterizing the complex heterogeneity of tumors by revealing the spatial context of dozens of markers within intact tissue specimens [58]. This spatial information is crucial for biomarker identification, as the distribution of expression throughout a tumor often carries significant biological and clinical implications beyond mere presence or absence of a marker.

Artificial intelligence and machine learning now play transformative roles in biomarker analytics, capable of pinpointing subtle patterns in high-dimensional multi-omic and imaging datasets that conventional methods may miss [55] [58]. AI-driven analysis exceeds human observational capacity by integrating genomic, proteomic, transcriptomic, and histopathology data to reveal new relationships between biomarkers and disease pathways [55]. For instance, at DoMore Diagnostics, AI-based digital pathology has uncovered prognostic and predictive signals in standard histology slides that outperform established molecular and morphological markers for colorectal cancer [55].

Advanced model systems, including organoids and humanized systems, represent another significant advancement by better mimicking human biology and drug responses compared to conventional 2D or animal models [58]. Organoids recapitulate complex tissue architectures and functions, making them well-suited for functional biomarker screening, target validation, and exploration of resistance mechanisms. When integrated with multi-omic technologies, these advanced models enhance the robustness and predictive accuracy of biomarker studies, ultimately bridging the gap between bench research and clinical application [58].

Experimental Protocols for Structure-Function Biomarker Validation

The validation of biomarkers requires a fit-for-purpose approach where the level of evidence depends on the context of use (COU) and specific application [54]. For biomarkers intended to support regulatory decisions, the FDA's Biomarker Qualification Program (BQP) provides a structured framework spanning three stages: Letter of Intent, Qualification Plan, and Full Qualification Package [54].

Simultaneous EEG-fNIRS Recording Protocol:

• Participant Preparation: 18 healthy adult participants (28.5 ± 3.7 years) are prepared for simultaneous data collection [4].
• EEG Setup: 30 electrodes placed according to the international 10-5 system at a 1000 Hz sampling rate (down-sampled to 200 Hz) [4].
• fNIRS Configuration: 36 channels (14 sources and 16 detectors) with an inter-optode distance of 30 mm, placed following the standardized 10-20 EEG system at a 12.5 Hz sampling rate (down-sampled to 10 Hz) using two wavelengths (760 nm and 850 nm) [4].
• Experimental Paradigm: Recordings conducted during both 1-minute resting state sessions and 30 trials of 10-second left and right-hand motor imagery tasks [4].
• Signal Processing: EEG and fNIRS data are coregistered by spatially aligning them to an anatomical MRI template using digitized positions relative to known scalp landmarks [4].
• Data Analysis: Structural and functional data are mapped onto the Desikan-Killiany atlas to define regions of interest (ROIs). Structure-function coupling is quantified using graph signal processing (GSP) tools, particularly the structural-decoupling index (SDI) [4].

Multimodal Representation Learning Model (EFRM):

• Pre-training Stage: The model is trained on large-scale, publicly available multimodal datasets without labels using approximately 1250 hours of brain signal recordings from 918 participants [56].
• Feature Extraction: Modality-specific features are extracted based on a Masked Autoencoder (MAE) while shared domains between EEG and fNIRS are extracted through contrastive learning [56].
• Transfer Learning: The pre-trained model is optimized for specific downstream tasks using minimal labeled samples, working with EEG-only, fNIRS-only, or paired EEG-fNIRS inputs [56].
• Validation: Model performance is quantified using classification accuracy with few-shot learning scenarios, demonstrating competitive performance compared to state-of-the-art supervised learning models [56].

Comparative Analysis of EEG and fNIRS in Biomarker Applications

Structure-Function Relationship Mapping

Direct comparison of EEG and fNIRS in mapping structure-function relationships reveals distinct patterns with implications for biomarker development. Research using simultaneous EEG-fNIRS recordings demonstrates that fNIRS structure-function coupling resembles slower-frequency EEG coupling at rest, with significant variations across brain states and neural oscillations [4] [5]. This relationship is not uniform across the brain but exhibits substantial regional heterogeneity, with stronger coupling in sensory regions and greater decoupling in association cortices, following the unimodal to transmodal gradient [4].

During task conditions, such as motor imagery, both modalities capture brain state-dependent changes but with notable differences in temporal dynamics and spatial specificity. EEG provides millisecond-level resolution of electrical dynamics, while fNIRS captures the slower hemodynamic response that follows neural activation by 1-2 seconds due to neurovascular coupling [56]. Cross-band representations of neural activity show lower correspondence between electrical and hemodynamic activity in the transmodal cortex regardless of brain state, while demonstrating specificity for the somatomotor network during motor imagery tasks [4] [5].

Discrepancies between EEG and fNIRS are particularly evident in the frontoparietal network, a region critical for higher-order cognitive functions [4] [45]. These differences likely reflect the distinct physiological processes each modality captures—direct electrical activity versus indirect hemodynamic responses—and highlight the complementary value of multimodal approaches for comprehensive biomarker development.

Diagnostic Performance in Neurological Disorders

fNIRS has emerged as a particularly promising tool for early detection of mild cognitive impairment (MCI) and neurodegenerative diseases, especially in settings where advanced imaging is limited [57]. Systematic investigations reveal that NIRS effectively assesses cognitive function, identifying reduced prefrontal connectivity in MCI and subjective cognitive decline (SCD) [57]. Interestingly, while SCD patients maintain stronger brain network integrity, NIRS reveals decreased oxyhemoglobin levels in Alzheimer's disease patients' dorsolateral prefrontal cortex during cognitive tasks [57].

The combination of NIRS with graph analysis, cognitive tasks, and machine learning significantly enhances diagnostic accuracy for neurodegenerative conditions [57]. Furthermore, NIRS can differentiate between various neurodegenerative disorders; for instance, differences in Broca's area activation during language tasks help distinguish behavioral variant frontotemporal dementia from Alzheimer's disease, revealing unique cognitive profiles through NIRS [57].

EEG biomarkers, in contrast, have demonstrated particular utility in epilepsy management, sleep disorder diagnosis, and monitoring consciousness during anesthesia, leveraging its millisecond-level temporal resolution to capture transient neural events that may be missed by hemodynamic-based measures.

Table: Diagnostic Applications of fNIRS in Neurodegenerative Conditions

Condition	fNIRS Findings	Clinical Utility
Subjective Cognitive Decline (SCD)	Lower HbO2 levels than healthy individuals, but higher than MCI	Early risk stratification
Mild Cognitive Impairment (MCI)	Reduced prefrontal connectivity during cognitive tasks	Progression monitoring and early intervention
Alzheimer's Disease	Pronounced oxyhemoglobin reductions in dorsolateral prefrontal cortex	Differential diagnosis and treatment response assessment
Frontotemporal Dementia	Distinct Broca's area activation during language tasks	Differentiation from Alzheimer's disease
Parkinson's Disease	Altered prefrontal activation patterns during executive tasks	Cognitive symptom monitoring

Visualization of Workflows and Relationships

Multimodal Biomarker Discovery Workflow

Biomarker Discovery Workflow Integrating Multimodal Data

Neurovascular Coupling Mechanism

Neurovascular Coupling Linking EEG and fNIRS Signals

The Scientist's Toolkit: Essential Research Solutions

Table: Key Reagents and Technologies for Biomarker Research

Tool/Technology	Function	Application Context
Multi-omics Profiling Platforms	Comprehensive molecular characterization	Identification of novel biomarker signatures from genomic, epigenomic, and proteomic data
Spatial Biology Technologies	Preservation of tissue architecture while measuring multiple markers	Characterization of tumor microenvironment and cellular interactions
AI-Powered Analytics	Pattern recognition in high-dimensional data	Discovery of subtle biomarker patterns in complex datasets
Human Organoid Models	Recreation of human tissue architecture and function	Functional biomarker screening and target validation
Graph Signal Processing Tools	Quantification of structure-function relationships	Analysis of brain network organization and connectivity biomarkers
Masked Autoencoder Models	Self-supervised representation learning	Few-shot classification of brain signals with minimal labeled data
fNIRS Systems	Non-invasive hemodynamic monitoring	Cognitive assessment and neurodegeneration biomarker development
High-Density EEG Systems	Millisecond-resolution electrical brain activity mapping	Epilepsy biomarker identification and cognitive state monitoring

Biomarker discovery represents a critical frontier in advancing drug development and understanding target engagement, with significant evolution in methodological approaches and technological capabilities. The comparison between EEG and fNIRS modalities reveals complementary strengths for different biomarker applications—EEG provides unparalleled temporal resolution for capturing neural dynamics, while fNIRS offers practical advantages for clinical settings and specific sensitivity to hemodynamic changes in neurodegenerative conditions. The integration of these modalities through advanced computational approaches, including graph signal processing and multimodal representation learning, enables more comprehensive characterization of structure-function relationships in both healthy and diseased states. As biomarker science continues to evolve, the strategic combination of emerging technologies—from spatial biology and multi-omics to AI-powered analytics—will further transform our ability to develop validated, clinically meaningful biomarkers that accelerate therapeutic development and improve patient outcomes across neurological and psychiatric disorders.

A central challenge in modern neuroscience is linking the brain's structural architecture to its dynamic functional outputs. Structure-function relationships describe how the physical wiring of the brain's neural networks (structure) supports and constrains its patterns of activity and connectivity (function). Investigating this relationship is crucial for understanding the neural mechanisms underlying behavior, cognition, and neurological disability [4]. Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) provide complementary windows into these brain dynamics through distinct neurophysiological mechanisms. EEG captures the brain's electrical activity with millisecond temporal resolution, reflecting the direct effects of neural processing. In contrast, fNIRS measures hemodynamic responses coupled to neural activity via neurovascular coupling, providing better spatial resolution and resistance to motion artifacts [2] [59]. This comparative analysis examines how these modalities perform across three clinical domains, highlighting their unique advantages for specific translational applications.

Table 1: Fundamental Technical Comparison of EEG and fNIRS

Parameter	EEG	fNIRS
Temporal Resolution	Very High (milliseconds) [2]	Moderate (~0.1-10 Hz) [59] [60]
Spatial Resolution	Low [2] [61]	Moderate [2]
Measured Signal	Electrical activity from synchronized pyramidal neurons [2]	Hemodynamic changes (HbO, HbR) via neurovascular coupling [2] [59]
Penetration Depth	Cortical surface [59]	Superficial cortex (2-3 cm) [59] [60]
Key Strength	Excellent for tracking rapid neural dynamics	Robust against motion artifacts, suitable for natural environments
Primary Limitation	Poor spatial localization, sensitive to artifacts [2]	Limited to cortical surface, indirect neural measure [2] [59]

Methodological Foundations: Experimental Protocols and Analysis

Core Experimental Paradigms

Clinical translation of EEG and fNIRS relies on standardized experimental protocols that elicit reproducible neural signatures:

• Resting-State Paradigm: Participants remain awake while relaxing with eyes closed for a fixed duration (typically 5-10 minutes). This protocol assesses intrinsic brain network organization without task demands and is widely used in depression and neurodegenerative disease research [61] [62]. The methodology requires minimal participant compliance, making it suitable for cognitively impaired populations.

• Task-Based Activation Paradigms:

  • Verbal Fluency Test (VFT): Subjects generate words belonging to a specific category within a time limit. This protocol activates prefrontal regions and is extensively used in depression studies to probe executive dysfunction [63].
  • Motor Imagery Tasks: Participants imagine performing specific motor actions (e.g., hand movements) without actual execution. This is commonly employed in brain-computer interface applications and structure-function studies [4].
  • Drug Cue Reactivity: Individuals with substance use disorders are exposed to drug-related cues while brain responses are monitored. This paradigm targets craving-related circuitry in addiction research [16].

Signal Acquisition and Processing Workflows

The integrity of clinical findings depends on rigorous data acquisition and processing pipelines:

• EEG Processing Pipeline: Raw EEG signals undergo referencing, band-pass filtering (typically 0.5-70 Hz), and artifact removal (ocular, cardiac, muscle). Signals are then decomposed into frequency bands (delta: 1-4 Hz, theta: 4-8 Hz, alpha: 8-13 Hz, beta: 13-30 Hz, gamma: >30 Hz) for feature extraction [2] [61]. Brain network analysis employs graph theory metrics including clustering coefficient, characteristic path length, and local efficiency to quantify functional connectivity patterns [61].

• fNIRS Processing Pipeline: Raw optical signals are converted to oxygenated (HbO) and deoxygenated hemoglobin (HbR) concentrations using the Modified Beer-Lambert Law [2] [16]. Signal processing includes band-pass filtering (0.01-0.1 Hz) to remove physiological noise, motion artifact correction, and principal component analysis to separate cerebral from superficial signals [4] [16]. Hemodynamic features such as mean activation, laterality, and functional connectivity strength are then extracted for analysis.

Diagram 1: Multimodal Neuroimaging Processing Workflow

Clinical Application in Depression

Diagnostic Classification Performance

Depression research has demonstrated the complementary value of EEG and fNIRS for objective diagnosis:

• EEG Biomarkers: Abnormal patterns in brain network topology differentiate depressed patients from healthy controls. Specifically, increased clustering coefficient and local efficiency in delta and theta bands indicate altered functional integration in depression [61]. Spectral asymmetry features in frontal alpha and theta rhythms have shown diagnostic relevance with classification accuracies up to 88.33% using support vector machine classifiers [61].

• fNIRS Biomarkers: Reduced prefrontal cortex activation during cognitive tasks like the Verbal Fluency Test serves as a reliable diagnostic indicator. In adolescents with Major Depressive Disorder (MDD), age-dependent hemodynamic patterns emerge: younger patients (12-15 years) show increased right dorsolateral prefrontal cortex (DLPFC) activation with depression severity, while older adolescents (16-18 years) display the opposite pattern [63].

• Multimodal Superiority: Combined EEG-fNIRS approaches significantly outperform single-modality classification. One study achieved 81.8% accuracy using EEG features alone, which increased to 92.7% when incorporating fNIRS-derived biomarkers [61]. The fusion of electrical and hemodynamic information provides a more comprehensive characterization of the neural correlates of depression.

Table 2: Depression Classification Performance Across Modalities

Modality	Key Biomarkers	Classification Accuracy	Population
EEG Alone	- Theta/alpha band asymmetry [61]- Increased delta/theta local efficiency [61]	81.8% [61]	25 MDD patients, 30 healthy controls [61]
fNIRS Alone	- Prefrontal HbO changes during VFT [63]- Age-dependent DLPFC activation [63]	Not quantified in reviewed studies	30 adolescents with MDD [63]
EEG + fNIRS	- Combined electrical and hemodynamic features [61]	92.7% [61]	25 MDD patients, 30 healthy controls [61]

Structure-Function Insights in Depression

The structure-function relationship in depression reveals system-level disruptions:

• Network Decoupling: EEG studies demonstrate aberrant connectivity in default mode and frontoparietal networks, reflecting disrupted information integration [61].
• Prefrontal Dysfunction: fNIRS consistently identifies hypofrontality during executive tasks, indicating impaired top-down cognitive control [63].
• Neurovascular Uncoupling: Combined assessments reveal discordance between electrical oscillations and hemodynamic responses, particularly in transmodal association cortices, suggesting impaired neurovascular coupling as a potential disease mechanism [4].

Clinical Application in Addiction

Craving and Prefrontal Dysregulation

Addiction research leverages the portability of fNIRS and EEG to study cue-elicited craving:

• Orbitofrontal Cortex (OFC) Activation: fNIRS studies demonstrate that drug cue reactivity produces distinct OFC activation patterns across substance classes. Methamphetamine users show the highest OFC activation, followed by mixed-drug users, with heroin users displaying the lowest activation [16]. This graded response correlates with craving intensity and punishment tolerance.

• Prefrontal Functional Organization: Machine learning classification of fNIRS data differentiates substance users with high accuracy (LDA, SVM, and CNN algorithms have been successfully applied), providing objective assessment of drug-specific neural effects [16].

• EEG Correlates: Complementary EEG measures reveal altered frontal beta and gamma oscillations during craving states, reflecting incentive salience attribution to drug cues. However, motion artifacts during craving expression limit EEG data quality in non-restricted paradigms.

Table 3: Addiction Research Findings by Substance Type

Substance Type	Sample Characteristics	Key fNIRS Findings	Methodological Protocol
Methamphetamine	10 users [16]	Highest OFC activation [16]	Resting state and drug cue induction paradigm [16]
Heroin	10 users [16]	Lowest OFC activation [16]	Resting state and drug cue induction paradigm [16]
Mixed Drugs	10 users [16]	Intermediate OFC activation [16]	Resting state and drug cue induction paradigm [16]

Structure-Function Implications in Addiction

Addiction manifests as disrupted structure-function relationships in reward and control networks:

• OFC-Striatal Decoupling: Weakened structural connectivity between orbitofrontal and striatal regions correlates with cue-induced hemodynamic responses, reflecting impaired reward prediction signaling [16].
• Executive Control Deficits: EEG reveals slowed frontal theta oscillations during cognitive control tasks, indicating compromised top-down regulation of drug-seeking behavior.
• Neurovascular Uncoupling: The heightened OFC response to drug cues despite structural degradation suggests maladaptive neurovascular compensation mechanisms in addiction [16].

Clinical Application in Neurodegenerative Diseases

Diagnostic and Monitoring Applications

Neurodegenerative disease research benefits from the repeated-measurement capability of fNIRS and EEG:

• Alzheimer's Disease and Mild Cognitive Impairment: Resting-state fNIRS (rsfNIRS) identifies reduced prefrontal functional connectivity and hemodynamic complexity in Mild Cognitive Impairment (MCI) and Alzheimer's Disease (AD) [59] [62]. These alterations correlate with cognitive decline severity and may precede structural atrophy.

• Parkinson's Disease: Hybrid EEG-fNIRS modeling detects disrupted motor cortex connectivity and neurovascular coupling during movement planning and execution [59]. fNIRS also monitors therapeutic effects of interventions like transcranial direct current stimulation (tDCS) on cortical excitability [59].

• Frontotemporal Dementia: Disease-specific patterns of prefrontal hemodynamic asymmetry differentiate FTD from AD, with fNIRS providing complementary information to structural MRI for differential diagnosis [62].

Structure-Function Correlates in Neurodegeneration

Multimodal studies reveal characteristic structure-function disturbances:

• Hierarchical Vulnerability: Structure-function coupling follows a unimodal-to-transmodal gradient, with stronger coupling in sensory areas and greater decoupling in association cortices [4]. Neurodegeneration preferentially affects weakly coupled transmodal regions.
• Network Disintegration: EEG shows progressive slowing of dominant frequencies and disrupted functional connectivity topology that correlates with structural connectome degradation [4].
• Compensatory Reorganization: fNIRS identifies hyperactivation in preserved brain regions during early disease stages, indicating functional compensation for structural loss [59].

Table 4: Essential Materials and Analytical Tools for Multimodal Studies

Tool Category	Specific Examples	Function/Purpose
Acquisition Hardware	- Wireless EEG systems (e.g., 32-channel NeuSen W) [61]- High-density fNIRS devices (e.g., NIRSIT with 24 sources, 32 detectors) [16]	Simultaneous recording of electrical and hemodynamic activity
Analytical Software	- MNE toolbox [4]- Brainstorm software [4]- Custom MATLAB/Python scripts for graph theory analysis	Signal processing, source reconstruction, and network analysis
Feature Extraction Algorithms	- Modified Beer-Lambert Law (fNIRS) [2] [16]- Wavelet transform (EEG) [61]- Graph theory metrics (clustering coefficient, local efficiency) [61]	Conversion of raw signals to physiologically meaningful features
Classification Approaches	- Support Vector Machines (SVM) [61]- Linear Discriminant Analysis (LDA) [16]- Convolutional Neural Networks (CNN) [16]	Automated diagnosis and group classification
Experimental Paradigms	- Verbal Fluency Test [63]- Resting-state protocols [61] [62]- Motor imagery tasks [4]	Standardized cognitive activation for reproducible findings

EEG and fNIRS provide distinct yet complementary insights into brain structure-function relationships across clinical populations. Their comparative value varies by application:

• Diagnostic Specificity: fNIRS offers superior spatial localization for prefrontal dysfunction in depression and addiction, while EEG better captures network dynamics across distributed brain regions.

• Practical Implementation: fNIRS demonstrates advantages in motion-rich environments and with special populations, while EEG provides unparalleled temporal resolution for capturing rapid neural dynamics.

• Multimodal Superiority: Combined approaches consistently outperform single-modality assessments, with diagnostic accuracy improvements of approximately 10% demonstrated in depression classification [61].

• Structure-Function Insights: The relationship between electrical and hemodynamic functional networks varies across brain regions, with stronger coupling in unimodal sensory areas and greater decoupling in transmodal association cortices [4]. This hierarchical organization has important implications for understanding disease-specific vulnerability patterns.

Future clinical translation will benefit from standardized experimental protocols, improved multimodal integration algorithms, and the development of normative databases for structure-function relationships across the lifespan and disease continuum.

Navigating Technical Challenges: Optimization for Reliable and Reproducible Results

This guide provides an objective comparison of motion correction and signal quality assurance in electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), two pivotal non-invasive neuroimaging techniques. Ensuring data fidelity is paramount for investigating structure-function relationships in the brain, a core pursuit in modern neuroscience and drug development.

The quest to elucidate the relationship between the brain's structural wiring and its functional output is a central theme in neuroscience. Techniques like EEG and fNIRS are invaluable in this pursuit, offering complementary insights into brain dynamics. However, the signals they acquire are perpetually contaminated by motion artifacts, which can severely distort the underlying neural correlates and compromise the validity of any structure-function inference.

Motion artifacts manifest differently across modalities. In fNIRS, which measures hemodynamic changes by shining near-infrared light through the scalp and brain, motion primarily causes pronounced spike-like shifts or baseline drifts in the oxygenated (HbO) and deoxygenated hemoglobin (HbR) signals [64]. In EEG, which records electrical potentials on the scalp, motion often introduces high-amplitude, low-frequency drift and muscle activity noise [65]. The growing interest in multimodal EEG-fNIRS studies, especially for tasks like motor imagery, makes understanding and correcting these modality-specific artifacts not just a technical necessity but a foundational requirement for robust scientific discovery [4] [48] [44].

Comparative Analysis of Motion Artifacts and Correction Techniques

The susceptibility to motion artifacts and the efficacy of correction methods vary significantly between EEG and fNIRS. The table below summarizes the core characteristics and prevalent correction strategies for each modality.

Table 1: Comparative Analysis of Motion Artifacts and Correction in EEG and fNIRS

Aspect	EEG (Electrophysiological)	fNIRS (Hemodynamic)
Primary Signal	Electrical brain potentials (millisecond resolution) [65]	Hemodynamic changes in HbO and HbR (∼0.1-10 Hz resolution) [64]
Nature of Common Motion Artifacts	High-amplitude low-frequency drift, muscle activity (EMG) noise [65]	Sudden spikes, baseline shifts, and slow drifts [64] [66]
Common Software-Based Correction Methods	Filtering (e.g., high-pass), Blind Source Separation (e.g., ICA), regression techniques [65]	Moving Average (MA), Wavelet filtering, Spline interpolation, Principal Component Analysis (PCA), Correlation-Based Signal Improvement (CBSI) [64]
Typical Hardware/Setup Solutions	Electrode caps with conductive gel, active electrodes, accelerometers for motion tracking [65]	Secure cap design, foam pad holders, collodion-fixed fibers, accelerometers, short-separation channels [64]
Relative Robustness to Motion	Highly susceptible to electrical and motion noise [65]	More robust to electrical noise, but still significantly affected by motion [64] [66]
Key Consideration for Structure-Function	High temporal fidelity is crucial for mapping fast neural dynamics to structural networks.	Spatial specificity is key for localizing hemodynamic activity within structural regions [4].

fNIRS-Specific Motion Correction: Performance Data

A direct comparison of prevalent fNIRS motion correction techniques on real pediatric data (a population with high artifact prevalence) reveals performance differences. The following table summarizes findings from a study that evaluated these methods using five predefined metrics [64].

Table 2: Efficacy of Prevalent fNIRS Motion Artifact Correction Techniques [64]

Correction Method	Brief Description	Reported Efficacy / Key Finding
Moving Average (MA)	Replaces data points with the average of surrounding points in a defined window.	Among the best outcomes for improving signal quality in child data [64].
Wavelet Filtering	Uses wavelet transforms to identify and remove artifact components in the signal.	Among the best outcomes; particularly effective for spike-type artifacts [64].
Spline Interpolation	Identifies motion artifacts and interpolates over the corrupted segment using spline functions.	Less effective than MA and Wavelet methods for the studied pediatric data [64].
Principal Component Analysis (PCA)	Identifies and removes components of the signal correlated with motion.	Less effective than MA and Wavelet methods for the studied pediatric data [64].
Correlation-Based Signal Improvement (CBSI)	Utilizes the negative correlation between HbO and HbR to correct artifacts.	Less effective than MA and Wavelet methods for the studied pediatric data [64].
Trial/Block Rejection	Directly excludes data segments contaminated by motion artifacts.	Effective but often impractical with pediatric populations due to limited data yield [64].

Experimental Protocols for Motion Correction

To ensure reproducibility and provide a clear framework for implementation, this section outlines standardized protocols for key motion correction methodologies cited in this guide.

Protocol: Motion Artifact Correction in fNIRS Using Moving Average and Wavelet Methods

This protocol is adapted from studies evaluating correction techniques on real fNIRS data from children performing a language task [64].

• 1. Data Acquisition: Record fNIRS data at sampling rates of 10 Hz or higher, using standard wavelengths (e.g., 690 nm and 830 nm). Ensure optodes are securely fixed to the scalp using customized caps and foam holders to minimize motion.
• 2. Signal Conversion: Convert the raw intensity data into optical density (OD) units.
• 3. Motion Artifact Identification: Apply a motion artifact detection algorithm (e.g., hmrMotionArtifactByChannel in Homer2) to the OD data. Typical parameters for identifying spike artifacts (Type A) include a standard deviation threshold (tMotion) of 1.0 second.
• 4. Motion Correction Application:
  • For the Moving Average (MA) method: Implement an MA filter that slides a window across the signal, replacing the central point with the average of the points within the window. The window size is a critical parameter that must be optimized for the specific data.
  • For the Wavelet method: Apply a wavelet transformation (e.g., using the hmrMotionCorrectWavelet function in Homer2) to the OD data. This method identifies and thresholds coefficients in the wavelet domain that are characteristic of motion artifacts before reconstructing the signal.
• 5. Final Processing: Convert the corrected OD data into concentration changes of HbO and HbR using the Modified Beer-Lambert Law. Proceed with general linear model (GLM) or other statistical analyses for functional activation mapping.

Protocol: Multimodal EEG-fNIRS Integration for Network Analysis

This protocol details the steps for integrating simultaneously acquired EEG and fNIRS data to create a robust multilayer brain network, mitigating the limitations of either modality alone [4] [44].

• 1. Simultaneous Data Collection: Collect synchronized EEG and fNIRS data. For EEG, use a 30-electrode setup based on the international 10-5 system sampled at 1000 Hz. For fNIRS, use a system with multiple sources and detectors (e.g., 14x16) placed according to the 10-20 system, with an inter-optode distance of 30 mm, sampled at 12.5 Hz [4].
• 2. Coregistration: Coregister EEG electrode and fNIRS optode positions to a common anatomical template (e.g., MNI space) using digitized head landmarks to ensure spatial alignment.
• 3. Unimodal Preprocessing:
  • EEG: Apply bandpass filtering, independent component analysis (ICA) to remove ocular and muscle artifacts, and source reconstruction to estimate cortical activity.
  • fNIRS: Convert raw data to OD, apply bandpass filtering (e.g., 0.02-0.08 Hz for resting state), and use Principal Component Analysis (PCA) to remove components associated with systemic physiological noise (e.g., heart rate, blood pressure) [4].
• 4. Network Metric Calculation: Map both structural (from a standard atlas) and functional data onto the same parcellation scheme (e.g., Desikan-Killiany atlas). For each modality, calculate functional connectivity matrices (e.g., using Pearson correlation between regional time-series). Derive a group-consensus structural connectome from a public database (e.g., ARCHI) [4].
• 5. Multilayer Network Construction & Analysis: Integrate the EEG-derived and fNIRS-derived functional connectivity networks into a single multilayer network model. Analyze this model using graph theory tools (e.g., the NetworkX library in Python) to compute topology metrics like small-worldness, which captures the balance between network segregation and integration [44].

Visualization of Workflows and Signaling Pathways

The following diagrams, generated with Graphviz, illustrate the logical workflows for the key experimental and correction protocols described in this guide.

fNIRS Motion Correction Workflow

Diagram Title: fNIRS Motion Correction Pathways

Multimodal EEG-fNIRS Integration

Diagram Title: Multimodal EEG-fNIRS Analysis Pipeline

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful artifact correction and high-quality signal acquisition depend on both software algorithms and physical hardware. The following table details key materials and their functions in this domain.

Table 3: Essential Materials for EEG/fNIRS Motion Correction and Signal Assurance

Item Name / Category	Function / Purpose in Signal Quality Assurance
fNIRS Cap & Optode Holders	Custom-made caps with secure foam holders stabilize optodes, minimizing movement relative to the scalp and reducing motion artifact generation [64].
EEG Electrode Cap & Conductive Gel	Electrode caps placed according to the 10-20/10-10 system, used with conductive gel or paste, optimize the electrode-scalp interface and reduce electrical impedance, improving signal-to-noise ratio [65].
Accelerometers	Small sensors attached to the participant's head or the EEG/fNIRS cap provide independent, ground-truth measurement of head movement for use in regression-based motion correction algorithms [64].
Short-Separation Channels (fNIRS)	Special fNIRS channels with a very short source-detector distance (e.g., 8 mm) primarily measure systemic physiological noise from the scalp, which can be regressed out from the standard channels to improve signal quality [64].
Homer2 Software Package	A widely used, open-source MATLAB toolbox that provides standardized implementations of key fNIRS processing algorithms, including motion correction methods like wavelet filtering and moving average [64].
Python (MNE, NetworkX, etc.)	Open-source programming environment with libraries like MNE for EEG/fNIRS analysis and NetworkX for graph-theoretical network analysis, enabling flexible and reproducible multimodal pipeline development [4] [44].
Computer Vision Systems	Video recordings analyzed with deep neural networks (e.g., SynergyNet) provide detailed, contactless ground-truth head movement data (orientation, speed) to characterize and validate motion artifacts [66].

In cognitive neuroscience and drug development, establishing clear structure-function relationships is paramount. This requires neuroimaging techniques that can accurately map neuronal activity to specific cortical regions. Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) are two non-invasive, portable methods widely used for this purpose, yet they differ fundamentally in their spatial specificity. EEG measures the brain's electrical activity from the scalp surface, but these signals are blurred and attenuated as they pass through the skull and other tissues, leading to poor spatial resolution. In contrast, fNIRS measures hemodynamic changes—blood oxygenation and volume shifts associated with neural activity—offering better spatial localization but with a delayed temporal response [67]. This guide objectively compares the spatial performance of these modalities, detailing the protocols and technologies essential for ensuring accurate sensor placement to maximize data quality for research and clinical applications.

Technical Comparison: fNIRS vs. EEG Spatial Performance

The core difference in spatial specificity between EEG and fNIRS stems from their underlying biophysical principles. EEG records electrical potentials generated by synchronized neuronal firing. However, the skull disperses and smears these electrical signals, making it difficult to pinpoint their exact cortical origin; its spatial resolution is typically considered to be on the centimeter level [24] [67]. fNIRS, an optical method, uses near-infrared light to measure changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) concentrations. Light scattering in tissue is more localizable than electrical conduction, granting fNIRS moderate spatial resolution that is superior to EEG, though it is generally limited to the outer cortex (~1–2.5 cm deep) [67]. The following table summarizes their spatial characteristics.

Table 1: Spatial and Technical Characteristics of EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity of neurons [67]	Hemodynamic response (blood oxygenation levels) [67]
Spatial Resolution	Low (centimeter-level) [67]	Moderate (better than EEG, but limited to cortex) [67]
Temporal Resolution	High (milliseconds) [67]	Low (seconds) [67]
Depth of Measurement	Cortical surface [67]	Outer cortex (~1–2.5 cm deep) [67]
Key Spatial Limitation	Signal dispersion by skull and scalp [24] [67]	Limited penetration depth, sensitive to superficial layers [24]
Sensitivity to Motion Artifacts	High [67]	Low (More tolerant to subject movement) [67]

Methodologies for Accurate Placement and Co-Registration

Achieving spatial specificity requires meticulous placement and co-registration of sensors according to standardized anatomical reference systems.

Standardized Placement Systems

Both EEG and fNIRS predominantly use the international 10-20 system or its more detailed derivatives (10-10 and 10-5 systems) for positioning [67] [68]. This system uses proportional distances from cranial landmarks (nasion, inion, preauricular points) to define electrode/optode locations, ensuring consistent placement across subjects and sessions. This is crucial for reproducible data collection and for relating findings to underlying brain anatomy [68].

Advanced Co-Localized Probe Designs

A significant innovation for multimodal imaging is the development of integrated fNIRS-EEG probes. A leading design involves creating custom fNIRS optodes that attach directly to EEG electrodes. For instance, one study used 3D-printed optodes that mate with active wet EEG electrodes (e.g., BrainProducts LiveAmp), allowing the optode's light pipe to contact the scalp through the electrode's gel access hole at a minimal center-to-center distance of ~4.87 mm [69]. This co-localization preserves the integrity of both modalities without sacrificing modularity or portability. A breakaway attachment clip is integrated to control mechanical failure points, protecting the equipment during setup or accidental tugs [69].

Customized Headgear Solutions

Stable placement is key to signal quality. While standard elastic EEG caps are common, their high stretchability can lead to inconsistent optode-scalp contact pressure and variable source-detector distances, harming fNIRS data quality [24]. Advanced solutions include:

• 3D-Printed Helmets: Custom-fitted helmets designed using software like AtlasViewer and printed with flexible materials (e.g., NinjaFlex TPU) offer a secure, stable fit that maintains precise optode/electrode positioning across subjects [69] [24].
• Thermoplastic Sheets: Cost-effective, moldable material that softens at ~60°C to form a custom-fit helmet [24].

Table 2: Essential Research Reagents and Materials for fNIRS-EEG Placement

Item	Function	Example/Details
EEG Electrodes	Measures electrical potentials on the scalp.	Active wet electrodes (e.g., BrainVision LiveAmp) with a conductive gel application hole that also facilitates optode co-localization [69].
fNIRS Optodes	Emits and detects near-infrared light.	Custom 3D-printed shells (e.g., using SLS printing with Formlabs resin) housing sources (LEDs/lasers) and detectors [69].
Integrated Cap/Helmet	Holds all sensors in a stable, pre-defined arrangement.	3D-printed flexible helmets (e.g., NinjaFlex material) or modified elastic caps with dedicated fixtures for both EEG electrodes and fNIRS optodes [69] [24].
Conductive Gel	Ensures low-impedance electrical connection for EEG.	Standard electrolyte gel applied via the electrode's access hole [69].
3D Design & Atlas Software	Enables probe design and anatomical co-registration.	Software toolboxes (e.g., AtlasViewer) used to design high-density probe layouts and map them to anatomical atlases [69].

Experimental Protocols for Validating Placement and Specificity

Protocol: Validating a Co-Localized fNIRS-EEG Probe

This protocol is adapted from a study that designed and tested a custom co-localized optode-electrode system [69].

• Aim: To validate a novel co-localized fNIRS-EEG probe design by screening for inter-modal interference and measuring brain activation during a cognitive task.
• Probe Design:
  • The HD-fNIRS-EEG probe was designed in the AtlasViewer toolbox, targeting the prefrontal and temporal cortices [69].
  • The layout included 25 fNIRS sources (nine of which were co-localized with EEG electrodes), 58 fNIRS detectors, and 32 EEG electrodes (including the nine co-located positions) [69].
  • Source-detector distances were configured for high-density DOT: first-nearest neighbors at 8 mm, second-nearest at 19 mm, and third-nearest at 32.5 mm [69].
  • The final cap was 3D-printed using flexible NinjaFlex TPU [69].
• Interference Screening: A critical step was to verify that the fNIRS optodes did not introduce electrical noise into the EEG signals. Researchers performed a crosstalk analysis by examining the EEG spectral content in the presence of active fNIRS sources. The results showed no observable interference from the fNIRS sources in the EEG data [69].
• Functional Validation with a Stroop Task: 14 subjects performed a modified Stroop task, a classic cognitive conflict paradigm. The analysis of both fNIRS and EEG data revealed increased activation within the expected regions of interest, consistent with prior literature, thereby supporting the design's utility for capturing valid multimodal neural signals [69].

Protocol: Data Fusion for Enhanced Spatial Localization

Combining EEG and fNIRS data can improve the overall spatial accuracy of the measured brain activity.

• Aim: To leverage the complementary information from fNIRS and EEG to achieve a more precise localization of brain activity than possible with either modality alone.
• Synchronization: Precise temporal synchronization of fNIRS and EEG data streams is essential. This can be achieved using a unified processor or via external hardware triggers (e.g., TTL pulses) to align the datasets [24].
• Data Fusion Methods: Several computational approaches are used:
  • Data Concatenation: Combining preprocessed feature sets from both modalities into a single input vector for a classifier [68].
  • Source-Decomposition Techniques: Methods like joint Independent Component Analysis (jICA) can identify latent components that are shared across the electrophysiological (EEG) and hemodynamic (fNIRS) data domains, potentially revealing underlying neurovascular coupling processes [30].
  • Model-Based/Decision-Level Fusion: Integrating the outputs of separate classifiers for each modality (e.g., using a weighted average or a second-level classifier) to make a final decision [68].

The following diagram illustrates the workflow for setting up and validating a multimodal fNIRS-EEG experiment with co-localized sensors.

The quest for precise spatial specificity in non-invasive brain imaging is a fundamental challenge. While EEG provides a direct, millisecond-scale measure of neural electrical activity, its utility for pinpointing exact functional anatomy is limited by the skull's blurring effect. fNIRS offers a valuable complement with its superior spatial localization of hemodynamic changes in the cortical mantle, albeit with slower temporal dynamics.

The future of establishing robust structure-function relationships lies in the synergistic use of these modalities. Technological advancements in co-localized optode-electrode design and custom, stable headgear are critical for minimizing placement error and maximizing data quality [69] [24]. Furthermore, sophisticated data fusion algorithms are unlocking the potential to integrate these complementary signals, yielding a more holistic and spatially accurate picture of brain function than either method can provide alone [30] [68]. For researchers and drug development professionals, a thorough understanding of these placement principles and technologies is essential for designing rigorous, reproducible studies that can reliably link brain structure to function.

Real-Time Processing Constraints for Neurofeedback and BCI Applications

Brain-Computer Interfaces (BCIs) and Neurofeedback (NFB) represent transformative technologies that rely on the real-time monitoring and interpretation of brain activity to either control external devices or enable self-regulation of neural functions [70]. These systems create direct communication pathways between the brain and computers, bypassing conventional neuromuscular channels [71]. The effectiveness of both BCIs and NFB is critically dependent on their ability to process complex neural signals under strict temporal constraints, making real-time processing capabilities a fundamental determinant of system performance and clinical utility [70] [72].

The core challenge lies in the fact that neural signals are inherently noisy, susceptible to artifacts, and computationally demanding to process, particularly for modalities that provide high spatial resolution. Real-time processing requires establishing measurement baselines, simultaneously correcting artifacts across multiple channels, and in some cases, performing computationally intensive image reconstruction—all while maintaining low latency to ensure effective user interaction and learning [70]. These constraints manifest differently across neuroimaging modalities, creating distinct trade-offs between temporal resolution, spatial accuracy, and practical implementation that researchers must navigate when designing BCI and NFB systems.

Neuroimaging Modalities for BCI and Neurofeedback

Electroencephalography (EEG): The Temporal Specialist

Electroencephalography (EEG) measures electrical activity generated by the synchronized firing of neurons through electrodes placed on the scalp. It represents the most established and widely used modality for non-invasive BCI and NFB applications [73] [72]. EEG's principal strength lies in its excellent temporal resolution (millisecond precision), enabling it to capture rapid neural dynamics essential for real-time applications [70]. This high temporal fidelity allows EEG-based systems to detect transient cognitive events and generate responsive feedback without perceptible delays.

However, EEG faces significant spatial limitations. The electrical signals are blurred by the skull and other tissues between the brain and electrodes, resulting in relatively low spatial resolution [70]. Additionally, EEG signals are highly susceptible to various artifacts including muscle activity, eye blinks, and environmental interference, necessitating sophisticated preprocessing pipelines to extract meaningful neural information [73] [71]. A typical EEG-based BCI system consists of five sequential components: brain activity measurement, preprocessing, feature extraction, classification, and translation into commands [73].

Functional Near-Infrared Spectroscopy (fNIRS): The Spatial Compromise

Functional near-infrared spectroscopy (fNIRS) measures hemodynamic responses in the brain by detecting changes in near-infrared light absorption by hemoglobin, providing an indirect measure of neural activity through neurovascular coupling [70] [74]. This optical imaging technique tracks concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) in cortical regions, which are closely linked to local neuronal activity [70].

FNIRS offers several advantages for real-time applications, including higher spatial resolution than EEG, reduced sensitivity to motion artifacts, and greater tolerance for environmental noise [70] [74]. The technique is relatively portable, low-cost, and practical for use in more naturalistic settings compared to fMRI [74]. These characteristics make fNIRS particularly suitable for applications requiring some spatial precision without sacrificing practical implementation, such as motor rehabilitation and assistive technology for mobility support [70].

The primary limitation of fNIRS is its temporal resolution, which is constrained by the relatively slow hemodynamic response (typically several seconds) [74]. While this is sufficient for many NFB applications where users learn to modulate brain states over extended periods, it presents challenges for BCI applications requiring rapid command execution. Additionally, fNIRS signals are contaminated by systemic physiological noise from extracerebral tissues (e.g., scalp blood flow), requiring specialized correction methods to isolate cerebral activity [75].

Table 1: Comparative Analysis of EEG and fNIRS for Real-Time BCI/NFB Applications

Parameter	EEG	fNIRS
Temporal Resolution	Millisecond range [70]	Seconds (limited by hemodynamic response) [74]
Spatial Resolution	Low (skull blurs electrical signals) [70]	Moderate to high (can resolve cortical regions) [70] [74]
Artifact Vulnerability	High sensitivity to motion, EMG, EOG artifacts [73] [71]	Moderate sensitivity; mainly to systemic physiological fluctuations [70] [75]
Portability & Cost	Highly portable, low to moderate cost [72]	Portable, low cost compared to fMRI [70] [74]
Primary Real-Time Challenge	Rapid artifact rejection and feature extraction [73]	Systemic activity correction and slower inherent response [75]
Ideal Application Scope	Rapid communication, motor imagery classification [73]	Neurofeedback training, motor rehabilitation [70] [76]

Structural-Functional Relationships in EEG vs. fNIRS

Understanding the relationship between structural and functional brain networks is crucial for advancing BCI and NFB systems, as this coupling fundamentally constrains how neural signals can be interpreted and leveraged for real-time applications. Recent comparative studies using simultaneous EEG and fNIRS recordings have revealed distinctive patterns in how each modality captures structure-function relationships across different brain states [5].

Research examining both resting state and motor imagery tasks has demonstrated that fNIRS structure-function coupling resembles slower-frequency EEG coupling during rest, with notable variations across different brain states and neural oscillations [5]. At the local level, this relationship exhibits significant heterogeneity, with stronger coupling observed in sensory regions and increasing decoupling along the unimodal to transmodal gradient, particularly in association cortices [5].

These differences have profound implications for BCI and NFB design. EEG-based systems primarily capture synchronized electrical activity from large neuronal populations, making them ideal for tracking rapid state changes and network dynamics. In contrast, fNIRS measures the hemodynamic consequences of neural activity, providing better spatial specificity for targeting specific cortical regions in neurofeedback protocols [5] [74]. Discrepancies between EEG and fNIRS are particularly pronounced in the frontoparietal network, suggesting modality-specific biases in how functional processes are represented [5].

Experimental Approaches and Methodologies

EEG Experimental Protocol for Attention Monitoring

The NeuroPilot study exemplifies a sophisticated EEG approach to real-time attention monitoring in educational contexts [77]. This system utilizes a consumer-grade EEG headband (FocusCalm) with electrodes positioned at Fp1, Fp2, and Fpz according to the international 10-10 system, targeting prefrontal regions crucial for sustained attention [77].

Data Collection & Validation: Participants watched educational videos while EEG data was collected. A novel intra-video questionnaire assessment provided ground-truth labels for attentive versus non-attentive states, addressing a critical validation gap in previous research [77].

Signal Processing Pipeline:

• Segmentation: Raw signals were divided using sliding windows
• Filtering: A Butterworth bandpass filter was applied to isolate relevant frequencies
• Artifact Removal: High-amplitude artifacts and EOG contaminants (e.g., eye blinks) were systematically removed
• Feature Extraction: Multiple domains were explored including time, frequency, wavelet, and statistical features
• Feature Selection: Recursive feature elimination (RFE) with support vector machines (SVM) identified optimal feature subsets [77]

Classification & Validation: The final classifier achieved 88.77% accuracy in distinguishing attention states using leave-one-subject-out (LOSO) cross-validation, demonstrating robust generalization across individuals [77].

fNIRS Experimental Protocol for Working Memory Enhancement

A randomized sham-controlled study investigated fNIRS-informed neurofeedback for modulating working memory performance by targeting the dorsolateral prefrontal cortex (DLPFC), a region critically involved in working memory processes [76].

Experimental Design: Sixty-two healthy male participants were randomized into either active neurofeedback or sham control groups. The protocol included:

• fNIRS-NF Training: Four runs of real-time fNIRS neurofeedback targeting left DLPFC activation
• Transfer Assessment: A subsequent working memory challenge during fMRI to evaluate transfer effects in the absence of feedback [76]

Real-Time Processing Pipeline:

• Signal Acquisition: Hemodynamic responses (HbO/HbR) measured from DLPFC using NIRScout system
• Baseline Establishment: Initial resting-state period to calibrate measurements
• Motion Artifact Correction: Sliding window approach with denoising autoencoder (DAE) model
• Feature Translation: Converted hemodynamic changes to visual feedback signals [70] [76]

Key Findings: The active neurofeedback group demonstrated significantly increased DLPFC activation over training runs and during the subsequent working memory challenge without feedback. Exploratory analyses revealed a negative correlation between DLPFC activity and reaction times, suggesting improved neural efficiency [76].

Real-Time Processing Pipelines and Computational Constraints

EEG Processing Architecture

The EEG processing pipeline must operate under strict temporal constraints to maintain system responsiveness. The primary computational burden occurs during the artifact removal and feature extraction stages, where techniques like Independent Component Analysis (ICA) and wavelet transforms demand significant processing resources [73] [77]. For real-time applications, systems often employ simplified artifact detection algorithms and focus on a limited set of clinically validated features to reduce latency.

fNIRS Processing Architecture

FNIRS processing introduces unique computational challenges, particularly for high-density systems and Diffuse Optical Tomography (DOT). The conversion from optical density to hemoglobin concentrations requires solving matrix inversion problems that become increasingly complex with channel count [70]. For DOT systems generating 3D images, the image reconstruction process involves computing and inverting a Jacobian matrix that describes the correlation between tissue absorption changes and detected optical intensity—a traditionally time-consuming operation [70]. Advanced approaches address this through pre-calculated inverse Jacobian matrices and deep learning-based denoising autoencoders (DAE) trained on extensive HD-DOT datasets to enable real-time processing of approximately 750 channels simultaneously [70].

Table 2: Systemic Artifact Correction Methods in fNIRS

Method Type	Specific Technique	Performance Summary	Computational Demand
Without SDCs	Simple Spatial Filter	Improves signal quality but tends to overcorrect [75]	Low
Without SDCs	Advanced Spatial Filter	Better than no correction; limited by lack of reference [75]	Moderate
With SDCs	Regression (Single SDC)	Good improvement; limited by regional specificity [75]	Low
With SDCs	GLM (Multiple SDCs)	Superior performance; best for spatial specificity [75]	High
With SDCs	GLM (A Priori SDCs)	Balanced approach; efficient use of resources [75]	Moderate

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Research Tools for BCI and Neurofeedback Studies

Tool Category	Specific Examples	Function & Application
EEG Hardware	FocusCalm Headband, Emotiv EPOC+, OpenBCI Ultracortex [77]	Acquire electrical brain activity; consumer-grade to research-grade options
fNIRS Systems	NIRScout, NIRx Systems [75] [76]	Measure hemodynamic responses through near-infrared spectroscopy
Short-Distance Channels (SDCs)	NIRx Short-Distance Detectors [75]	Reference channels for systemic artifact correction in fNIRS
Signal Processing Libraries	MNE-Python, EEGLAB, BrainStorm [72]	Preprocess, filter, and analyze neural data
Classification Algorithms	Support Vector Machines (SVM), Deep Q-Learning [77] [73]	Translate neural features into device commands or cognitive states
Experimental Platforms	PsychoPy, Presentation, NeuroFeedback Suite [77] [72]	Design and implement stimulus presentation and feedback protocols
Quality Metrics	Mean Squared Error (MSE), Correlation Coefficient, Spatial Specificity [70] [75]	Quantify signal improvement and system performance

The real-time processing constraints for neurofeedback and BCI applications present researchers with fundamental trade-offs that must be carefully balanced against specific application requirements. EEG offers superior temporal resolution essential for rapid communication systems but struggles with spatial precision and artifact vulnerability. FNIRS provides better spatial localization and noise resilience but is constrained by the inherent latency of hemodynamic responses.

The choice between modalities ultimately depends on the specific neural targets and performance requirements of the intended application. For tasks requiring millisecond timing such as reactive BCI systems, EEG remains the modality of choice. For applications targeting specific cortical regions where timing is less critical, such as many neurofeedback protocols, fNIRS offers distinct advantages. Emerging approaches that combine multiple modalities (hybrid BCIs) or leverage advanced signal processing techniques like deep learning are progressively overcoming these historical limitations, creating new possibilities for both basic research and clinical translation.

Future progress will likely involve increased standardization of processing pipelines, validation of real-time artifact correction methods, and the development of more sophisticated computational approaches that can extract meaningful neural information within the stringent latency requirements of effective brain-computer interaction.

The integration of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) represents a paradigm shift in neuroimaging, enabling researchers to investigate brain function through complementary lenses of electrophysiology and hemodynamics. This multimodal approach is particularly powerful for exploring the fundamental structure-function relationships in the human brain, as it concurrently captures the millisecond-scale electrical neural oscillations and the slower, metabolically-supported blood oxygenation changes that underlie cognitive processes [24] [4]. While EEG provides exquisite temporal resolution for tracking rapid neural dynamics, fNIRS offers superior spatial localization and resistance to motion artifacts, creating a synergistic pairing that overcomes the limitations of either modality used in isolation [78]. The emerging field of structure-function research has particularly benefited from this dual-modality approach, revealing that functional networks captured by EEG and fNIRS show varying degrees of coupling with the underlying structural connectome, following a unimodal-to-transmodal gradient across brain regions [4] [5].

However, the full potential of simultaneous EEG-fNIRS recordings can only be realized through rigorously standardized preprocessing pipelines that account for the unique characteristics and artifacts inherent in each modality. The development of such pipelines presents distinctive challenges, including the need for precise temporal synchronization, differentiated artifact removal strategies, and sophisticated fusion approaches that respect the different physiological origins and temporal scales of EEG and fNIRS signals [24] [79]. This guide systematically compares best practices in EEG-fNIRS data preprocessing, providing researchers with experimentally-validated methodologies to enhance data quality and maximize the analytical value of multimodal recordings for probing the complex interplay between brain structure and function.

Table 1: Fundamental Characteristics of EEG and fNIRS Modalities

Characteristic	EEG	fNIRS
Measured Signal	Electrical potential from synchronized neuronal firing	Hemodynamic response via light absorption by hemoglobin
Temporal Resolution	Millisecond level (~1-1000 Hz)	~2-10 Hz typically
Spatial Resolution	~2-3 cm	~1-2 cm
Primary Artifacts	Ocular, muscle, cardiac, line noise	Motion, physiological (cardiac, respiratory), scalp blood flow
Depth Sensitivity	Cortical surface	Cortical surface (up to 2-3 cm depth)
Relationship to Neural Activity	Direct measure of electrical activity	Indirect measure via neurovascular coupling

Experimental Protocols for Multimodal Data Acquisition

The foundation of any successful EEG-fNIRS study lies in a meticulously planned experimental protocol that ensures high-quality, temporally aligned data from both modalities. Recent studies have established standardized approaches for simultaneous recordings, with particular attention to hardware integration, temporal synchronization, and participant preparation.

Hardware Integration and Setup

Successful multimodal recordings require integrated systems that minimize interference between modalities. The most effective approaches utilize either a unified acquisition system or precisely synchronized separate systems. In a practical implementation, researchers have successfully combined a g.HIamp EEG amplifier with a continuous-wave multifunctional fNIRS system (NirScan), using a custom-designed hybrid cap that incorporated 32 EEG electrodes, 32 optical sources, and 30 photodetectors arranged to achieve 90 fNIRS measurement channels through source-detector pairing [53]. The integration of EEG electrodes and fNIRS optodes follows specific spatial configurations, typically placing EEG electrodes between fNIRS optodes since EEG electrodes are smaller and easier to position [78]. Critical to this setup is maintaining proper source-detector separation distances (typically 3 cm for fNIRS) [53] and ensuring the cap material is dark to prevent ambient light from contaminating fNIRS signals [78].

Temporal Synchronization Methods

Precise temporal alignment between EEG and fNIRS data streams is paramount for meaningful multimodal analysis. Two primary synchronization approaches have emerged: (1) using a unified processor to simultaneously process and acquire both EEG and fNIRS signals, which achieves high-precision synchronization but requires more complex system design; and (2) combining fNIRS and EEG data obtained from separate systems (e.g., NIRScout and BrainAMP) synchronized during acquisition and analysis via a host computer, which is simpler to implement but may lack the precision needed for microsecond-level EEG analysis [24]. Modern implementations commonly use event markers transmitted from experimental presentation software (e.g., E-Prime) to simultaneously trigger both recording systems [53]. The EEG amplifier typically acts as the "master device" due to its higher sampling frequency, with data streamed in real-time to a computer for synchronized storage and analysis [78].

Participant Preparation and Experimental Design

Participant preparation for simultaneous EEG-fNIRS recordings typically requires approximately 10 minutes with modern active electrode systems [78]. Key considerations include proper cap sizing to ensure consistent optode-scalp contact pressure, which is crucial for fNIRS signal quality [24]. For motor imagery paradigms, studies have implemented grip strength calibration procedures using dynamometers and stress balls to enhance participants' kinesthetic awareness and improve motor imagery vividness [53]. Experimental designs should carefully consider the different temporal characteristics of EEG and fNIRS responses, with adequate trial durations to capture the slower hemodynamic response (typically 10+ seconds for fNIRS) while maintaining sufficient trials for EEG analysis [53]. Blocked designs are often preferable for fNIRS, while both blocked and event-related designs work well for EEG.

Comparative Analysis of Preprocessing Pipelines

EEG-Specific Preprocessing Workflows

EEG preprocessing requires specialized approaches to address its unique artifact profiles while preserving neural signals of interest. The following workflow outlines a standardized pipeline validated across multiple multimodal studies:

Diagram 1: EEG preprocessing pipeline workflow

Bandpass filtering represents a critical first step, with typical settings between 0.5-45 Hz to remove slow drifts and high-frequency noise while preserving neural oscillations of interest [4]. For studies focusing on event-related potentials (ERPs), researchers commonly apply more restrictive bandpass filters (e.g., 0.1-30 Hz) [80]. Subsequent steps include bad channel detection and removal using metrics like abnormal variance or correlation patterns, followed by re-referencing to a common average reference (CAR) or mastoid references. Artifact removal constitutes the most complex stage, with independent component analysis (ICA) emerging as the gold standard for identifying and removing ocular, cardiac, and muscular artifacts [4] [80]. Finally, data is epoched around experimental events, with baseline correction applied to remove pre-stimulus offsets.

fNIRS-Specific Preprocessing Workflows

fNIRS preprocessing addresses fundamentally different artifacts stemming from physiological processes and motion, with particular attention to preserving hemodynamic response patterns:

Diagram 2: fNIRS preprocessing pipeline workflow

Initial quality assessment using metrics like the scalp-coupling index (SCI) identifies poor-quality channels, with studies typically rejecting channels with SCI values below 0.7 [4]. Conversion to optical density followed by application of the Modified Beer-Lambert Law transforms raw light intensity measurements into concentration changes of oxygenated (HbO) and deoxygenated hemoglobin (HbR) [79]. Bandpass filtering with cutoffs between 0.01-0.1 Hz removes physiological noise from cardiac (~1 Hz), respiratory (~0.3 Hz), and very low-frequency drifts [4]. Motion artifact correction represents a critical step, with wavelet-based methods, principal component analysis (PCA), and targeted PCA (tPCA) effectively identifying and removing motion-induced signal distortions [4] [79]. Finally, additional physiological noise from systemic fluctuations can be addressed through PCA-based approaches or inclusion in general linear models (GLM) [79].

Table 2: Quantitative Performance Comparison of Preprocessing Methods

Preprocessing Step	Method	Performance Metrics	Computational Cost
EEG Artifact Removal	Independent Component Analysis (ICA)	~90-95% artifact reduction [4]	High
	Regression-based Methods	~80-85% artifact reduction	Medium
fNIRS Motion Correction	Wavelet-Based	~85-90% improvement in SNR [4]	Medium
	tPCA	~80-88% improvement in SNR [79]	Low-Medium
fNIRS Physiological Noise Removal	PCA	~70-80% systemic noise reduction [4]	Low
	Bandpass Filtering	~60-70% noise reduction	Very Low
Multimodal Classification	Feature-Level Fusion	5-10% accuracy improvement over unimodal [53]	Medium
	MDNF Model [81]	Superior to state-of-the-art methods	High

Integrated Multimodal Fusion Approaches

Data Integration Strategies

The ultimate value of simultaneous EEG-fNIRS recordings lies in the strategic integration of both modalities to leverage their complementary strengths. Three primary fusion approaches have emerged, each with distinct advantages for exploring structure-function relationships:

• Hardware-level integration: Creating physically integrated systems with precise co-registration of EEG electrodes and fNIRS optodes. This approach minimizes temporal synchronization errors and facilitates precise spatial alignment, which is crucial for investigating regional variations in structure-function coupling [24] [78].

• Feature-level fusion: Extracting discriminative features from each modality separately, then combining them for classification or analysis. This approach has demonstrated 5-10% improvement in classification accuracy compared to unimodal systems in motor imagery tasks [53] [81]. For example, one innovative method transforms EEG data into 2D spectrogram images using short-time Fourier transform, applies transfer learning for feature extraction, then integrates these with fNIRS-derived spectral entropy features [81].

• Decision-level fusion: Processing each modality through separate classification pipelines, then combining the decisions at the output stage. This approach preserves modality-specific processing optimizations while still benefiting from multimodal information.

Advanced Analytical Frameworks

For structure-function relationship studies, graph signal processing (GSP) has emerged as a powerful mathematical framework for quantifying how functional patterns from EEG and fNIRS align with the underlying structural connectome [4] [5]. The structural-decoupling index (SDI) specifically quantifies the degree of structure-function dependency for each brain region, revealing heterogeneous coupling across the cortex [4]. Studies applying this framework have discovered that fNIRS structure-function coupling resembles slower-frequency EEG coupling at rest, with variations across brain states and oscillations [5]. Additionally, cross-band representations of neural activity have revealed lower correspondence between electrical and hemodynamic activity in the transmodal cortex, irrespective of brain state, while showing specificity for the somatomotor network during motor imagery tasks [4].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Equipment and Solutions for Simultaneous EEG-fNIRS Research

Item	Function/Purpose	Example Specifications
Hybrid EEG-fNIRS Cap	Secure placement of electrodes/optodes	32 EEG channels, 32 sources, 30 detectors [53]
EEG Amplifier	Signal acquisition and amplification	g.HIamp, g.Nautilus, 256+ Hz sampling [78] [53]
fNIRS System	Hemodynamic response measurement	Continuous-wave, 760/850 nm wavelengths [4] [53]
Electrode Gel	Ensure electrical conductivity	Standard EEG electrolyte gel
Light-Blocking Cap Material	Prevent ambient light contamination	Dark-colored, opaque fabric [78]
3D Digitizer	Precise localization of sensors	Electromagnetic or optical tracking systems
Synchronization Interface	Temporal alignment of modalities	Hardware triggers, TTL pulses [53]
Participant Preparation Kit	Scalp preparation and sensor placement	Abrasive gel, measuring tape, markers

The development of robust preprocessing pipelines for simultaneous EEG-fNIRS recordings represents an essential step toward unlocking the full potential of multimodal neuroimaging. By implementing the standardized methodologies and best practices outlined in this guide, researchers can significantly enhance data quality and analytical validity for probing the complex interplay between brain structure and function. The comparative analysis presented here demonstrates that while EEG and fNIRS require modality-specific preprocessing approaches, their strategic integration through advanced fusion methods yields substantial benefits, including improved classification accuracy and more comprehensive characterization of neural processes across temporal and spatial domains.

As the field advances, key challenges remain in further optimizing real-time processing capabilities, standardizing pipeline parameters across diverse experimental paradigms, and developing more sophisticated integration frameworks that respect the neurobiological relationships between electrical and hemodynamic signals. The ongoing development of publicly available multimodal datasets [82] [53] and consensus guidelines [79] will continue to drive innovation in this rapidly evolving field, ultimately enabling more precise and comprehensive investigations of the human brain in health and disease.

In neuroscience research and pharmaceutical development, establishing a direct link between measured signals and genuine underlying brain activity is paramount. Cross-modal validation—the practice of using multiple, complementary neuroimaging techniques to measure the same neural events—has emerged as a powerful solution to this challenge. By simultaneously capturing the brain's electrical activity (via electroencephalography, EEG) and its hemodynamic response (via functional near-infrared spectroscopy, fNIRS), researchers can obtain a more complete and verifiable picture of brain function [24] [23]. This approach is particularly valuable for investigating the structure-function relationship in the brain, a central theme in understanding both healthy cognition and neurological disorders [4] [5].

The fundamental premise is that while EEG and fNIRS measure different physiological processes, they are both driven by the same underlying neural activity. EEG records postsynaptic electrical potentials from cortical neurons with millisecond precision, providing an direct measure of neural firing. In contrast, fNIRS measures changes in oxygenated and deoxygenated hemoglobin concentration in the blood, an indirect marker of neural activity that unfolds over seconds but offers better spatial localization for cortical areas [83] [23]. When these signals correlate in time and space during a cognitive or motor task, it provides strong evidence that both are tracking the same genuine neural event, thereby reducing the likelihood that observations are due to artifact, noise, or non-neural physiological processes [24].

Technical Comparison of EEG and fNIRS Modalities

Fundamental Principles and Measured Phenomena

Electroencephalography (EEG) measures the brain's electrical activity through electrodes placed on the scalp. It detects voltage fluctuations resulting from ionic current flows within the neurons of the brain, particularly the postsynaptic potentials of pyramidal cells. These electrical events occur on a millisecond timescale, making EEG ideal for tracking the rapid dynamics of brain function, such as those involved in sensory processing, motor planning, and cognitive tasks [83]. However, the electrical signals are blurred as they pass through the skull and scalp, resulting in limited spatial resolution typically at the centimeter level.

Functional Near-Infrared Spectroscopy (fNIRS) employs near-infrared light to measure hemodynamic changes in the brain. Light sources and detectors placed on the scalp measure the absorption of light at different wavelengths, which varies with the changing concentrations of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR). These hemodynamic changes are linked to neural activity through neurovascular coupling—the mechanism by which active brain regions receive increased blood flow [4] [33]. While the hemodynamic response is relatively slow (peaking 4-6 seconds after neural activity), fNIRS provides better spatial localization of cortical activity than EEG, though it is generally limited to the outer layers of the cortex (1-2.5 cm depth) [83].

Table 1: Core Technical Specification Comparison between EEG and fNIRS

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity of neurons	Hemodynamic response (blood oxygenation levels)
Signal Source	Postsynaptic potentials in cortical neurons	Changes in oxygenated and deoxygenated hemoglobin
Temporal Resolution	High (milliseconds)	Low (seconds)
Spatial Resolution	Low (centimeter-level)	Moderate (better than EEG, but limited to cortex)
Depth of Measurement	Cortical surface	Outer cortex (~1–2.5 cm deep)
Sensitivity to Motion	High – susceptible to movement artifacts	Low – more tolerant to subject movement
Best Use Cases	Fast cognitive tasks, ERP studies, sleep research	Naturalistic studies, child development, motor rehab

Complementary Strengts in Structure-Function Research

The technical differences between EEG and fNIRS make them particularly complementary for probing structure-function relationships in the brain. Research comparing these modalities has revealed that fNIRS-based functional connectivity more closely resembles the structure-function coupling observed with low-frequency EEG oscillations at rest [4]. Furthermore, studies show that this relationship is not uniform across the brain; both modalities detect stronger structure-function coupling in unimodal sensory regions (e.g., visual, auditory, sensorimotor areas) compared to transmodal association areas (e.g., frontoparietal network) [4] [5]. This regional heterogeneity aligns with the known hierarchical organization of the cortex and provides a multimodal framework for understanding how brain structure constrains and guides functional networks.

Experimental Designs for Cross-Modal Validation

Simultaneous Recording Protocols

Implementing a robust cross-modal validation study requires careful integration of both recording systems. The most common approach involves embedding fNIRS optodes within a standard EEG electrode cap, allowing both systems to reference the same scalp landmarks, typically using the international 10-20 or 10-5 systems [24] [50]. For example, in a study investigating the Action Observation Network, researchers used a 128-electrode EEG cap with a 24-channel fNIRS system embedded within it, placing optodes over sensorimotor and parietal cortices to target regions of interest [23].

Synchronization between systems is critical and can be achieved through several methods. Some researchers use separate EEG and fNIRS systems synchronized via external triggers (TTL pulses) or shared clock systems [83]. Others employ a unified processor that simultaneously acquires and processes both EEG and fNIRS signals, resulting in higher synchronization precision [24]. The choice often depends on whether the research prioritizes simplicity or temporal precision.

Table 2: Performance Metrics in Motor Imagery Decoding Using Different Modalities

Modality/Approach	Reported Classification Accuracy	Experimental Context	Key Brain Regions Identified
EEG Alone	65.49% (highest between hand and shoulder tasks)	8 MI tasks from 4 joint types [50]	Bilateral central, right frontal, and parietal regions [23]
fNIRS Alone	Not explicitly quantified in results	8 MI tasks from 4 joint types [50]	Left angular gyrus, right supramarginal gyrus, right superior/inferior parietal lobes [23]
Multimodal EEG-fNIRS	"Stronger performance" than standalone methods [50]	ME, MO, and MI conditions [23]	Left inferior parietal lobe, superior marginal gyrus, post-central gyrus [23]

Data Processing and Artifact Removal

A critical challenge in cross-modal validation is addressing motion artifacts, which affect both modalities differently. fNIRS signals are particularly vulnerable to motion-induced perturbations, while EEG is susceptible to muscle and movement artifacts [64]. Several software-based correction methods have been developed and evaluated, particularly for pediatric populations where motion is more prevalent. A comparative study of six prevalent motion artifact correction techniques for fNIRS identified Moving Average (MA) and Wavelet methods as particularly effective [64].

For EEG, standard preprocessing typically includes filtering (e.g., 0.5-100 Hz bandpass filter), artifact rejection (e.g., for eye movements or muscle activity), and often re-referencing [50]. fNIRS processing pipelines generally involve converting raw intensity signals to optical density, then to hemoglobin concentration changes using the modified Beer-Lambert law, followed by filtering to remove physiological noise (e.g., cardiac and respiratory cycles) and motion artifacts [4] [50].

Analytical Approaches for Data Fusion and Validation

Multimodal Fusion Techniques

Once the data from both modalities are preprocessed, several analytical approaches can be employed to fuse the signals and validate their neural origin:

• Structured Sparse Multiset Canonical Correlation Analysis (ssmCCA): This method identifies shared sources of variance across multiple datasets (modalities). In one study, ssmCCA was used to fuse fNIRS and EEG data during motor execution, observation, and imagery tasks, consistently revealing activation in the left inferior parietal lobe, superior marginal gyrus, and post-central gyrus across all conditions—regions associated with the Action Observation Network [23].

• Joint Independent Component Analysis (jICA): This technique separates mixed signals into statistically independent components across modalities, potentially revealing shared underlying neural sources [83].

• Graph Signal Processing (GSP): This mathematical framework extracts harmonic basis functions from structural connectivity data, which then serves to obtain a graph-spectral representation of functional data. Researchers have used GSP to quantify structure-function coupling for both EEG and fNIRS, enabling direct comparison of how electrical and hemodynamic networks relate to their common structural substrate [4].

The following diagram illustrates a generalized workflow for cross-modal validation, from data acquisition to fused interpretation:

Cross-Modal Validation Workflow: This diagram outlines the sequential process from simultaneous data acquisition to the final interpretation of validated neural signals.

Validation Metrics and Interpretation

The success of cross-modal validation is typically assessed through several metrics:

• Temporal Correlation: Assessing whether EEG events (e.g., event-related potentials) are temporally coupled with fNIRS hemodynamic responses (e.g., HbO increases) within expected neurovascular coupling timeframes [23].

• Spatial Co-localization: Determining whether electrical and hemodynamic activations occur in anatomically consistent regions, often requiring co-registration with structural MRI data [4].

• Task Performance Correlation: Examining whether the magnitude of neural responses in both modalities correlates with behavioral performance or clinical measures [84].

• Classification Improvement: In brain-computer interface applications, assessing whether combined EEG-fNIRS features improve classification accuracy over single-modality approaches [50].

Essential Research Reagents and Equipment

Successful implementation of cross-modal validation studies requires specific hardware, software, and methodological components. The following table details key solutions and their functions in simultaneous EEG-fNIRS research:

Table 3: Research Reagent Solutions for Cross-Modal Validation Studies

Solution Category	Specific Examples/Models	Function in Cross-Modal Validation
Integrated Acquisition Systems	NIRScout with BrainAMP [24], Hitachi ETG-4100 with EEG cap [23]	Enables simultaneous recording with precise temporal synchronization
Specialized Headgear	3D-printed custom helmets [24], Composite polymer cryogenic thermoplastic sheets [24]	Ensures stable optode and electrode placement; improves reproducibility
Motion Correction Algorithms	Moving Average (MA) [64], Wavelet methods [64]	Removes motion artifacts that could confound neural signals
Data Fusion Toolboxes	Structured Sparse Multiset CCA [23], Graph Signal Processing tools [4]	Identifies shared neural sources across electrical and hemodynamic data
Co-registration Software	3D magnetic space digitizers (Fastrak, Polhemus) [23]	Aligns EEG electrodes and fNIRS optodes with anatomical landmarks

Cross-modal validation using simultaneous EEG and fNIRS represents a powerful methodological approach for ensuring that measured signals reflect genuine brain activity rather than artifact or noise. The convergence of electrical and hemodynamic signals provides stronger evidence for neural activation than either modality could provide alone, particularly when investigating the complex relationship between brain structure and function. This approach is especially valuable in clinical populations where motion artifacts may be more prevalent, in pharmacological studies where drug effects on neurovascular coupling must be accounted for, and in naturalistic research settings where traditional neuroimaging methods are impractical.

As both hardware integration and analytical techniques continue to advance, cross-modal validation is likely to become increasingly standard practice in neuroscience research and drug development. The development of more sophisticated fusion algorithms, improved motion correction techniques, and standardized reporting practices will further enhance the reliability and interpretability of multimodal brain imaging studies [24] [33]. Ultimately, by leveraging the complementary strengths of EEG and fNIRS, researchers can make more confident inferences about brain function, leading to more robust biomarkers and therapeutic targets for neurological and psychiatric disorders.

Head-to-Head Validation: Assessing the Complementary Value of EEG and fNIRS

The study of brain networks has evolved from examining single layers of connectivity to embracing multilayer network frameworks, which provide a more comprehensive representation of the brain's complex organizational architecture. Multilayer networks capture the inherent multidimensionality of neural systems, where different types of connections—structural, functional, electrical, and hemodynamic—coexist and interact within the same biological substrate. This framework enables researchers to move beyond traditional single-layer analyses and investigate how different network layers influence one another, creating emergent properties that cannot be understood by examining each layer in isolation [85].

In the context of neuroimaging, the multilayer approach is particularly valuable for integrating data from complementary modalities like electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). These techniques capture distinct yet interrelated aspects of brain activity: EEG measures electrical potentials generated by neuronal firing with millisecond temporal resolution, while fNIRS detects hemodynamic responses through near-infrared light absorption with superior spatial resolution to EEG. The mathematical foundation of multilayer networks allows researchers to quantify the complementarity between these modalities by analyzing how network properties in one layer relate to properties in another layer [4] [86].

The concept of "complementarity" in this context extends beyond simple correlation to encompass how the information from one modality fills gaps in the other, how interactions between layers give rise to system-level behaviors, and how these relationships change across different brain states. By applying formal multilayer network models, researchers can precisely quantify these relationships, leading to deeper insights into the structure-function principles that govern brain organization [85] [87].

Theoretical Foundations of Multilayer Network Analysis

Mathematical Frameworks for Multilayer Networks

Multilayer network analysis provides a formal mathematical foundation for representing and analyzing complex systems with multiple types of relationships. In neuroscience applications, this typically involves defining networks where each layer corresponds to a different modality (EEG, fNIRS), frequency band, or cognitive state. The fundamental mathematical representation is a quadruple M = (N, L, V, E), where N represents the set of nodes (brain regions), L represents the set of layers (modalities or conditions), V ⊆ N × L indicates node presence across layers, and E contains the edges (connections) between nodes across layers [87].

A critical advancement in this field is the property matrix representation, which enables systematic comparison across layers. This matrix organizes network properties such that columns represent network structures (nodes, edges, triangles) and rows represent contexts (layers), with each cell containing the value of an observational function mapping each structure-context pair to a numerical value [87]. This representation facilitates the computation of layer similarity through three primary approaches: (1) comparing aggregations of layer property vectors using statistical summaries, (2) comparing distributions of values across layers, and (3) comparing corresponding properties for each structure across layers [87].

Quantifying Complementarity Between Layers

Complementarity between EEG and fNIRS layers can be quantified through several mathematical approaches that capture different aspects of their relationship. The structural-decoupling index (SDI) is one such measure, derived from graph signal processing frameworks, which quantifies the degree of structure-function dependency for each brain region [4]. This index allows researchers to determine how tightly the functional patterns observed in each modality are constrained by the underlying structural connectivity.

Another important approach involves measuring layer similarity through correlation-based methods, mutual information, or distance metrics applied to property vectors [87]. These measures help determine whether nodes that are highly connected in one layer (EEG functional networks) tend to be highly connected in another layer (fNIRS functional networks), and whether this relationship is uniform across the brain or exhibits regional variations.

Inter-layer degree correlation measures assess whether nodes that have high degree (number of connections) in one layer also have high degree in another layer, revealing whether hub regions are consistent across modalities [87]. Additionally, multiplexity quantifies the tendency for connections to exist between the same nodes across multiple layers, indicating consistent functional partnerships across different neurophysiological domains [85].

Table 1: Key Mathematical Measures for Quantifying Complementarity

Measure	Definition	Interpretation in EEG-fNIRS Context
Structural-Decoupling Index (SDI)	Quantifies deviation between functional activity and structural constraints	High values indicate decoupling where functional activity is less constrained by structure
Layer Similarity	Correlation between property vectors of different layers	High values indicate similar network organization across modalities
Inter-layer Degree Correlation	Correlation between node degrees across layers	High values indicate consistent hub regions across EEG and fNIRS
Multiplexity	Proportion of edges present across multiple layers	High values indicate consistent functional connections across modalities

Experimental Protocols for Multilayer Network Analysis

Simultaneous EEG-fNIRS Data Acquisition

The foundation of robust multilayer network analysis lies in the careful acquisition of simultaneous EEG and fNIRS data. The experimental protocol must ensure precise temporal synchronization between modalities and accurate spatial coregistration of recording elements to brain anatomy. In a representative study examining structure-function relationships, researchers collected data from 18 healthy adult participants (28.5 ± 3.7 years) during both resting-state conditions and task performance (motor imagery) [4].

For EEG recording, the protocol utilized 30 electrodes arranged according to the international 10-5 system, with data acquired at a sampling rate of 1000 Hz (subsequently downsampled to 200 Hz). The fNIRS recording employed 36 channels (14 sources and 16 detectors) with an inter-optode distance of 30 mm, following the standardized 10-20 EEG system for positioning, with data acquired at a sampling rate of 12.5 Hz (downsampled to 10 Hz). The fNIRS system utilized two wavelengths (760 nm and 850 nm) to measure changes in oxyhemoglobin and deoxyhemoglobin concentrations [4].

Critical to the protocol is the precise coregistration of EEG electrodes and fNIRS optodes to anatomical landmarks, enabling accurate mapping of recorded signals to brain regions. This is typically achieved using digitized positions relative to known scalp landmarks (nasion, inion, preauricular points) followed by spatial normalization to a standardized brain template [4]. The experimental design should include both resting-state recordings (typically 5-10 minutes with eyes closed or open) and task-based conditions relevant to the research questions, such as motor imagery, cognitive tasks, or sensory stimulation.

Data Preprocessing and Quality Control

Robust preprocessing pipelines are essential for ensuring data quality and minimizing artifacts that could confound multilayer network analysis. For fNIRS data, the standard preprocessing pipeline includes: (1) conversion of raw intensity signals to optical density values; (2) assessment of signal quality using metrics like the scalp-coupled index (SCI), with exclusion of channels demonstrating SCI < 0.7; (3) bandpass filtering (typically 0.02-0.08 Hz for resting-state studies) using finite impulse response (FIR) filters; (4) identification and rejection of motion-contaminated segments using metrics like the global variance in temporal derivative (GVTD); and (5) removal of systemic physiological artifacts using principal component analysis (PCA) [4].

For EEG data, standard preprocessing includes: (1) filtering to remove line noise and artifacts; (2) identification and removal of bad channels; (3) segmentation into epochs; (4) independent component analysis (ICA) to remove ocular, cardiac, and muscle artifacts; and (5) source reconstruction to estimate activity in specific brain regions [4]. The source reconstruction step is particularly important for spatial alignment with fNIRS data and structural connectivity information.

Quality control metrics should be established a priori, with clear exclusion criteria. In the representative study, participants with more than 50% of fNIRS channels displaying poor signal quality (SCI < 0.7) were excluded from analysis [4]. For task-based analyses, epochs are typically defined relative to stimulus onset (e.g., 0-10 seconds for motor imagery tasks) and averaged across trials for each participant.

Diagram 1: Experimental workflow for multilayer EEG-fNIRS network analysis

Comparative Analysis of EEG and fNIRS in Structure-Function Relationships

Global Structure-Function Coupling Across Modalities

The relationship between structural connectivity (typically derived from diffusion MRI) and functional connectivity differs substantially between EEG and fNIRS, reflecting their sensitivity to distinct physiological processes operating at different temporal and spatial scales. Research using simultaneous EEG-fNIRS recordings has revealed that fNIRS-based structure-function coupling closely resembles slow-frequency EEG coupling (specifically in the delta and theta bands) during resting state [4] [5]. This alignment suggests that fNIRS captures neural processes associated with slower oscillatory activity, which may be more tightly constrained by the underlying structural architecture.

The graph signal processing (GSP) framework provides a mathematical foundation for quantifying these relationships by decomposing functional signals into patterns that are strongly or weakly supported by the structural connectome [4]. When applied to EEG and fNIRS data, this approach reveals that the alignment between functional activity and structural constraints is not uniform across frequency bands or hemodynamic responses. Specifically, faster EEG oscillations (gamma band) show greater decoupling from structural constraints, suggesting more dynamic, state-dependent functional reorganization that is less constrained by the underlying white matter architecture [4].

Table 2: Global Structure-Function Coupling by Modality and Frequency

Modality/Frequency Band	Structure-Function Coupling Strength	Temporal Characteristics	Spatial Specificity
fNIRS (hemodynamic)	Moderate to Strong	Slow (0.02-0.1 Hz)	High regional variation
EEG Delta (1-4 Hz)	Strong	Very Slow	Consistent with fNIRS
EEG Theta (4-8 Hz)	Moderate to Strong	Slow	Similar to fNIRS pattern
EEG Alpha (8-13 Hz)	Moderate	Medium	Posterior dominance
EEG Beta (13-30 Hz)	Weak to Moderate	Fast	Task-dependent
EEG Gamma (30-80 Hz)	Weak	Very Fast	Highly dynamic

Regional Heterogeneity in Structure-Function Relationships

The coupling between structural and functional networks exhibits substantial regional variation across the brain, following a systematic pattern along the unimodal-to-transmodal hierarchy. Both EEG and fNIRS show stronger structure-function coupling in unimodal regions (primary sensory and motor areas) compared to transmodal regions (association areas involved in higher-order cognition) [4] [5]. This gradient reflects fundamental principles of brain organization, with unimodal regions exhibiting more stable functional patterns tightly constrained by anatomy, while transmodal regions display more flexible functional configurations that can deviate from structural constraints.

However, important modality-specific differences emerge in specific networks. Most notably, the frontoparietal network, critical for cognitive control and adaptive functioning, shows significant discrepancies between EEG and fNIRS measures of structure-function coupling [4] [5]. This suggests that the electrophysiological and hemodynamic correlates of higher-order cognition may be differentially constrained by structural connectivity, with important implications for interpreting each modality in isolation.

During task performance, both modalities show increased structure-function coupling in task-relevant regions. For example, during motor imagery tasks, the somatomotor network demonstrates enhanced alignment between structural and functional patterns in both EEG and fNIRS [4]. However, the specific patterns of reorganization differ between modalities, with EEG showing more rapid dynamic adjustments and fNIRS capturing slower, more sustained changes that may reflect metabolic support for neural activity.

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of multilayer network analysis requires careful selection of equipment, software, and analytical tools. The following table summarizes key components of the research toolkit for EEG-fNIRS multilayer studies:

Table 3: Essential Research Toolkit for EEG-fNIRS Multilayer Network Studies

Category	Specific Tool/Equipment	Specifications	Function in Research Pipeline
EEG System	30+ electrode cap	10-5 placement system	Electrical potential measurement
fNIRS System	30+ channels	760nm & 850nm wavelengths	Hemodynamic response measurement
Structural Imaging	Diffusion MRI	60+ directions, b=1000+ s/mm²	White matter connectivity mapping
Coregistration	Digitization system	3D spatial coordinate capture	Anatomical alignment of sensors
Preprocessing	MNE-Python, Brainstorm	Signal processing pipelines	Artifact removal & quality control
Network Construction	Brain Connectivity Toolbox	Graph theory metrics	Network metric computation
Multilayer Analysis	Multilayer Python libraries	Customizable similarity measures	Cross-layer coupling quantification
Statistical Analysis	R, Python with specialized packages	Mixed-effects models	Group-level inference

Methodological Considerations and Best Practices

Addressing Technical Challenges in Multilayer Analysis

The integration of EEG and fNIRS within a multilayer network framework presents several technical challenges that require careful methodological consideration. Temporal misalignment between modalities stems from their fundamentally different physiological origins: EEG captures direct neuronal activity with millisecond precision, while fNIRS measures hemodynamic responses with a delay of 4-6 seconds due to neurovascular coupling [86] [88]. This can be addressed through temporal filtering approaches that align the frequency content of both signals or by incorporating hemodynamic response function modeling when analyzing event-related responses.

Spatial resolution disparities between EEG (which requires inverse modeling to source-localize activity) and fNIRS (which has better spatial specificity but limited depth penetration) must be reconciled through careful coregistration procedures [4]. The use of standardized anatomical atlases (such as the Desikan-Killiany atlas) provides a common coordinate system for comparing network properties across modalities [4]. Additionally, the property matrix framework offers a flexible approach for handling different types of network features across layers, allowing researchers to compare everything from simple degree distributions to more complex topological patterns [87].

Analytical Recommendations for Robust Quantification

To ensure robust quantification of complementarity, researchers should adopt several best practices. First, multiple similarity measures should be employed to capture different aspects of layer relationships, as no single metric fully characterizes the complex interplay between EEG and fNIRS networks [87]. Correlation-based measures, mutual information, and graph-based similarity indices each provide complementary insights into how the modalities relate to one another.

Second, null model testing is essential to establish whether observed layer similarities exceed what would be expected by chance. This involves comparing empirical similarity values against distributions generated from appropriate null models that preserve certain network properties while randomizing others [87]. Finally, multilevel validation approaches that examine consistency across different parcellation schemes, thresholding methods, and preprocessing parameters help ensure that findings are not methodological artifacts.

The application of these multilayer network approaches to EEG and fNIRS data has revealed that these modalities provide complementary rather than redundant information about brain function. Their relationship is not static but varies systematically across brain regions, frequency bands, and cognitive states, reflecting the multifaceted nature of neural information processing [4] [5]. By formally quantifying these complementarities, researchers can develop more comprehensive models of brain organization that integrate electrical and hemodynamic perspectives into a unified framework.

Diagram 2: Multilayer framework for quantifying EEG-fNIRS complementarity

The quest to decode human brain activity relies on two fundamental pillars: electrical dynamics and hemodynamic responses. Electroencephalography (EEG) captures millisecond-scale neuro-electrical activity with exquisite temporal precision but limited spatial resolution, while functional near-infrared spectroscopy (fNIRS) tracks slower blood oxygenation changes with better spatial localization but lower temporal fidelity. This structural-functional dichotomy presents both a challenge and an opportunity for brain-computer interface (BCI) research and clinical neurodiagnostics. Multimodal integration seeks to harmonize these complementary signals, potentially offering a more complete picture of brain activity than either modality can provide alone.

The theoretical foundation for multimodal fusion rests on neurovascular coupling—the established physiological relationship between neuronal electrical activity and subsequent hemodynamic changes in cerebral blood flow [30]. While EEG directly measures synchronous electrical activity from pyramidal neurons, fNIRS detects the hemodynamic consequences of this activity through neurovascular coupling mechanisms. This biological interdependence suggests that combined analysis should provide more robust brain state classification than unimodal approaches. This review synthesizes recent experimental evidence quantifying the performance benefits of multimodal EEG-fNIRS setups across various cognitive and motor tasks, with particular emphasis on classification accuracy benchmarks that directly inform translational applications in neurotechnology and pharmaceutical development.

Quantitative Performance Comparison: Single-Modal vs. Multimodal Classification

Recent studies consistently demonstrate that multimodal EEG-fNIRS systems achieve significantly higher classification accuracy compared to single-modal approaches across diverse experimental paradigms. The performance advantage stems from the complementary nature of electrical and hemodynamic signals, which provides more discriminative features for machine learning algorithms.

Table 1: Classification Accuracy Benchmarks for Motor Imagery Tasks

Modality	Classification Approach	Accuracy (%)	Dataset/Paradigm	Citation
EEG-only	Common Spatial Patterns + SVM	59.81 ± 0.97	Hand motion/imagery	[89]
EEG-only	Common Spatial Patterns + LDA	69.00 ± 11.42	Hand motion/imagery	[89]
fNIRS-only	Statistical features + Classifier	~57.00	Left vs. right hand MI	[90]
EEG-only	Conventional processing	~65.00	Left vs. right hand MI	[90]
EEG-fNIRS (multimodal)	Early fusion Y-shaped network	76.21	Left vs. right hand MI	[90]
EEG-fNIRS (multimodal)	Multi-domain features + progressive learning	96.74	Motor imagery tasks	[13]
EEG-fNIRS (multimodal)	Transfer learning framework	74.87	ICH patients MI	[91]

Table 2: Classification Performance Across Cognitive States

Modality	Classification Approach	Accuracy (%)	Task Type	Citation
EEG-fNIRS (multimodal)	Multi-domain features + progressive learning	98.42	Mental arithmetic	[13]
EEG-fNIRS (multimodal)	Cross-modal attention (MBC-ATT)	Superior to conventional	N-back & WG tasks	[92]
EEG-fNIRS (multimodal)	p-th order polynomial fusion	90.19	Mental arithmetic	[13]
EEG-fNIRS (multimodal)	Transfer learning	82.30-87.24	Public benchmark datasets	[91]

The performance advantage of multimodal systems is particularly pronounced in clinical populations. A recent transfer learning study demonstrated that when trained with optimally selected normal templates, a multimodal EEG-fNIRS classifier achieved 74.87% mean accuracy on intracerebral hemorrhage (ICH) patients during motor imagery tasks [91]. This represents a significant advancement for clinical applications where neurophysiological heterogeneity typically impedes cross-subject generalization.

Experimental Protocols and Methodologies

Data Acquisition and Preprocessing Standards

Multimodal experiments require careful temporal synchronization and artifact management. Typical protocols involve:

• Simultaneous recording using integrated EEG-fNIRS systems with embedded optodes in EEG caps to ensure spatial correspondence [23]
• EEG preprocessing: Band-pass filtering (8-25 Hz) for μ-band and low-β band preservation, common average referencing, and artifact removal [90]
• fNIRS preprocessing: Conversion of optical density to oxygenated (HbO) and deoxygenated hemoglobin (HbR) concentrations, motion artifact correction, and band-pass filtering (0.01-0.2 Hz) to isolate hemodynamic responses [90] [23]
• Temporal alignment using shared trigger signals and downsampling to common sampling rates (typically 128 Hz) to facilitate fusion operations

Fusion Strategies and Computational Architectures

The timing and method of information integration critically impact classification performance:

Early Fusion

Early fusion integrates raw or minimally processed signals before feature extraction. A Y-shaped neural network with separate EEG and fNIRS processing branches that merge before classification has demonstrated significantly higher performance (P < 0.05) compared to middle and late-stage fusion in motor imagery tasks [90]. The early fusion approach preserves complementary low-level information but requires careful handling of differing temporal resolutions and signal characteristics.

Middle Fusion

Hybrid approaches leverage cross-modal attention mechanisms to dynamically weight feature importance. The MBC-ATT framework employs independent convolutional branches for each modality with a modality-guided attention mechanism that selectively emphasizes task-relevant signals [92]. Similarly, HA-FuseNet integrates multi-scale dense connectivity with hybrid attention mechanisms to fuse complementary local and global spatio-temporal features [93].

Late Fusion

Decision-level fusion combines classifier outputs from separately processed modalities. The multi-level progressive learning framework extracts multi-domain features from each modality, applies feature selection via atomic search optimization, and fuses decisions through progressive machine learning, achieving remarkable accuracy (96.74% for motor imagery, 98.42% for mental arithmetic) [13].

The Scientist's Toolkit: Essential Research Reagents and Solutions

Table 3: Essential Research Materials for EEG-fNIRS Experiments

Category	Specific Solution	Function/Application	Representative Examples
Acquisition Systems	Integrated EEG-fNIRS	Simultaneous multimodal recording	Hitachi ETG-4100 with EEG integration [23]
	Wearable fiberless fNIRS	Mobile brain imaging in naturalistic settings	Portable continuous-wave systems [30]
Analysis Algorithms	Common Spatial Patterns (CSP)	EEG feature extraction for motor imagery	Dimensionality reduction for SVM/LDA [89]
	Structured sparse multiset CCA	Multimodal data fusion	Identifying concordant activation regions [23]
	Cross-modal attention	Dynamic feature weighting	MBC-ATT framework [92]
Classification Frameworks	Transfer learning	Cross-subject generalization	Wasserstein metric-driven domain selection [91]
	Multi-level progressive learning	Feature and decision fusion	Atomic search optimization [13]
Experimental Paradigms	Motor imagery (MI)	Sensorimotor cortex engagement	Left/right hand grasping imagery [90] [23]
	N-back tasks	Working memory assessment	0-back, 2-back, 3-back conditions [92]
	Mental arithmetic (MA)	Prefrontal cortex activation	Silent arithmetic operations [13]

Signaling Pathways and Experimental Workflows

The experimental workflow for benchmarking single-modal versus multimodal classification follows a structured pipeline from data acquisition to performance validation:

The neurophysiological basis for multimodal integration stems from the neurovascular coupling pathway:

The empirical evidence consistently demonstrates that multimodal EEG-fNIRS systems achieve statistically superior classification accuracy compared to single-modal approaches across diverse cognitive and motor tasks. The performance advantage ranges from 7-21% absolute improvement in motor imagery tasks to near-perfect classification (>96%) in mental arithmetic paradigms when sophisticated fusion strategies are employed.

The complementary nature of EEG's millisecond temporal resolution and fNIRS's superior spatial localization creates a synergistic effect that enhances BCI robustness, particularly in real-world applications where signal quality is compromised by artifacts and individual neurophysiological variability. For clinical applications in neurorehabilitation and pharmaceutical development, multimodal approaches offer enhanced capability to track both rapid electrophysiological changes and slower hemodynamic adaptations, providing a more comprehensive assessment of intervention effects.

Future research directions should focus on optimizing fusion architectures for specific clinical applications, developing standardized protocols for cross-study comparisons, and addressing the computational complexity challenges associated with real-time multimodal processing. As the field advances, multimodal EEG-fNIRS is poised to become the gold standard for robust brain state classification in both research and clinical settings.

Understanding the dynamic relationship between the brain's structural wiring and its functional activity is a fundamental pursuit in neuroscience. This structure–function relationship is not static; it varies significantly across different brain states, primarily categorized as task-based and resting-state conditions. The resting state, characterized by wakeful relaxation without explicit cognitive tasks, reveals the brain's intrinsic, spontaneous activity patterns. In contrast, task-based states involve specific, goal-directed cognitive operations that evoke distinct neural responses. The efficacy of neuroimaging modalities in capturing these state-dependent changes is crucial for both basic research and clinical applications [4] [94].

Two prominent non-invasive techniques for probing brain function are Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS). EEG measures the brain's electrical activity with millisecond temporal resolution, directly reflecting neuronal firing. fNIRS, on the other hand, measures hemodynamic changes—specifically, concentrations of oxygenated and deoxygenated hemoglobin—serving as an indirect proxy for neural activity through the mechanism of neurovascular coupling [4]. This fundamental difference in what they measure—electrical versus hemodynamic activity—suggests that EEG and fNIRS may have differential sensitivities to resting-state and task-based brain states. This article objectively compares the performance of EEG and fNIRS in these two conditions, framing the analysis within the broader context of structure–function relationships in the human brain.

Fundamental Principles and Physiological Bases

The physiological signals captured by EEG and fNIRS originate from distinct yet interconnected processes. EEG recordings capture post-synaptic potentials from pyramidal neurons, providing a direct, high-temporal-resolution view of neural electrical activity. These signals are often analyzed in specific frequency bands (e.g., delta, theta, alpha, beta, gamma), each associated with different cognitive states and functions [95] [96].

fNIRS relies on the principle of neurovascular coupling, the tight relationship between local neural activity and subsequent changes in cerebral blood flow and oxygenation. When a brain region becomes active, it triggers a hemodynamic response, increasing local blood flow and oxygen delivery. This leads to a rise in oxygenated hemoglobin (HbO) and a more subtle decrease in deoxygenated hemoglobin (HbR). fNIRS uses near-infrared light to measure these concentration changes, offering better spatial resolution than EEG but slower temporal resolution due to the sluggish nature of the hemodynamic response [97] [98].

The relationship between these electrical and hemodynamic signals is complex. Research using simultaneous EEG-fNIRS has shown that the fNIRS global signal, particularly in very low frequencies (<0.05 Hz), is negatively correlated with EEG measures of vigilance, indicating that lower arousal states are associated with larger hemodynamic fluctuations [97]. Furthermore, studies on structure–function coupling reveal that fNIRS-derived functional networks resemble those from slower-frequency EEG bands at rest, suggesting a link between the hemodynamic signal and specific electrical oscillation patterns [4].

Comparative Sensitivity in Resting-State Paradigms

The resting state provides a window into the brain's intrinsic functional architecture, often quantified through measures of functional connectivity (FC)—the temporal correlation between spatially remote neurophysiological events.

Insights from fNIRS

Resting-state fNIRS has proven highly effective in mapping the developing brain connectome. It reliably identifies small-world topology and hemispheric asymmetry in infants and children, with scans as short as 2.5 minutes yielding stable network metrics [94]. This demonstrates fNIRS's robustness for characterizing the brain's intrinsic organizational principles without task demands. The global signal in resting-state fNIRS is not merely noise; it contains valuable neurophysiological information. Studies have shown its amplitude is modulated by EEG vigilance, with lower vigilance states (e.g., drowsiness) associated with higher amplitude fNIRS global signals [97]. This relationship is particularly prominent in the very low-frequency range (<0.05 Hz), underscoring the importance of considering arousal states in resting-state fNIRS analysis.

Insights from EEG

Resting-state EEG (rs-EEG) is a powerful tool for identifying trait-like neural biomarkers. In predicting vulnerability to Major Depressive Disorder (MDD), rs-EEG has demonstrated exceptional classification performance, with one study reporting 98.06% accuracy using a 1D-Convolutional Neural Network model [95]. This suggests that rs-EEG captures stable, clinically relevant neural signatures that are highly sensitive to individual differences. Features such as the Higuchi fractal dimension, phase lag index, correlation, and coherence were identified as particularly important for this prediction, highlighting the utility of both linear and non-linear EEG metrics in characterizing the resting brain [95].

Table 1: Key Resting-State Studies and Findings

Modality	Study Focus	Key Finding	Performance/Accuracy
EEG [95]	Predicting vulnerability to depression	rs-EEG superior to task-based EEG for classification	98.06% accuracy with 1D-CNN
fNIRS [94]	Brain connectome development	Stable network topology in children from short scans	Stable FC after 1 min; accurate network metrics after 2.5 min
EEG-fNIRS [97]	Global signal & vigilance	Negative correlation between EEG vigilance & fNIRS global signal amplitude	Prominent in frequency <0.05 Hz

Comparative Sensitivity in Task-Based Paradigms

Task-based paradigms engage specific cognitive processes, allowing researchers to investigate the brain's functional response to controlled stimuli and cognitive demands.

Insights from fNIRS

fNIRS exhibits high sensitivity to cognitive state and load. In a working memory n-back task, fNIRS detected linear increases in prefrontal cortex activation corresponding to increasing cognitive load [98]. Furthermore, task-based functional connectivity measured with fNIRS also scaled with working memory load, with fronto-parietal and interhemispheric dorsolateral prefrontal cortex connectivity strengthening under higher loads [98]. fNIRS also effectively captures pathological changes in task-based responses. In post-stroke executive dysfunction (PSED) patients, fNIRS revealed reduced cortical activation in executive regions during Stroop and 1-back tasks, alongside complex patterns of functional connectivity reorganization, including both compensatory hyperconnectivity and maladaptive hypoconnectivity [99].

Insights from EEG

Task-based EEG excels at capturing the rapid neural dynamics underlying cognition. In a working memory study, task-based EEG functional connectivity slightly outperformed resting-state EEG in predicting individual working memory performance [96]. The peak correlation between predicted and observed scores reached r = 0.5, with alpha and beta band connectivity being the strongest predictors. In the context of depression, task-based EEG during a Sustained Attention to Response Task (SART) was used to probe "sticky" thoughts related to rumination. While its classification accuracy (91.42% in the delta band) was high, it was lower than that achieved by rs-EEG for the same clinical group [95].

Table 2: Key Task-Based Studies and Findings

Modality	Study/Context	Key Finding	Performance/Sensitivity
fNIRS [98]	Working Memory (n-back)	Activation & functional connectivity scale with cognitive load	Linear sensitivity to load in PFC
fNIRS [99]	Post-Stroke Executive Dysfunction	Hypoactivation in executive regions & altered FC	Identified compensatory & maladaptive connectivity
EEG [96]	Working Memory Prediction	Task-based FC superior to resting-state FC for prediction	Peak correlation with behavior: r = 0.5
EEG [95]	Vulnerability to Depression	Task-based (SART) classification accuracy	91.42% accuracy with LSTM

Structure–Function Relationships Across Modalities and States

The relationship between the brain's anatomical structure (white matter tracts) and its functional patterns differs between EEG and fNIRS, and this coupling is modulated by brain state.

Research using graph signal processing has shown that the structure–function relationship is not uniform across the brain. Both EEG and fNIRS show stronger coupling in unimodal sensory regions (e.g., visual, somatomotor) and greater decoupling in transmodal association cortices (e.g., frontoparietal network), following a unimodal-to-transmodal hierarchy [4] [5]. However, discrepancies exist between the modalities. fNIRS-derived structure–function coupling closely resembles that of slower-frequency EEG oscillations at rest [4]. During a motor imagery task, the cross-modal correspondence between electrical and hemodynamic activity was specific to the somatomotor network, indicating that the relationship between EEG and fNIRS is both state- and region-dependent [4]. This suggests that the functional information captured by EEG and fNIRS converges in primary sensory areas but diverges in higher-order associative regions, which are more influenced by internal models and contextual information.

Experimental Protocols and Methodologies

To ensure reproducibility and validate the findings discussed, this section outlines standard experimental protocols for key studies cited.

• Subject Preparation: Apply high-density, whole-head EEG caps and fNIRS optodes. Ensure proper impedance for EEG and light coupling for fNIRS.
• Data Acquisition:
  • Resting-State Recording: Conduct two 10-minute sessions (e.g., eyes-open and eyes-closed) in a supine, sitting, or standing position. Randomize session order.
  • EEG Parameters: Standard sampling rate (e.g., 1000 Hz), recording vigilance.
  • fNIRS Parameters: Standard wavelengths (e.g., 760 nm & 850 nm), calculate HbO and HbR concentration changes.
• Data Preprocessing:
  • fNIRS: Bandpass filter signals (e.g., 0.009–0.08 Hz for very low-frequency analysis). Calculate global signal as the average across all channels.
  • EEG: Compute vigilance metrics (e.g., power spectral ratios).
• Analysis: Correlate fNIRS global signal amplitude with EEG vigilance measures temporally and across sessions.

• Task Design: Use a block-design letter n-back task. Conditions include 0-back, 1-back, 2-back, and 3-back blocks interspersed with baseline (e.g., fixation) periods.
• fNIRS Setup: Place optodes over frontal and parietal cortices covering dorsolateral PFC (dlPFC), ventrolateral PFC (vlPFC), and inferior parietal lobule (IPL).
• Data Acquisition: Record continuous HbO and HbR during task performance.
• Data Analysis:
  • Activation: General Linear Model (GLM) to contrast task vs. baseline for each load condition. Test for linear trends across load.
  • Functional Connectivity: Calculate correlation matrices between channels' time-series for each load condition. Compare connectivity strength (e.g., fronto-parietal) across load levels.

• Experimental Sessions:
  • Resting-State: 5-10 minutes of eyes-closed EEG recording.
  • Task-Based: Auditory working memory task (e.g., n-back) during EEG recording.
• EEG Recording: High-density EEG system (e.g., 30+ electrodes) following the 10-5 system.
• Data Preprocessing:
  • Filter data and remove artifacts.
  • Parcellate data using a standard atlas.
• Feature Extraction: Calculate functional connectivity matrices (e.g., using phase-based or amplitude-based metrics) for standard frequency bands (delta, theta, alpha, beta, gamma) separately for rest and task.
• Modeling: Use Connectome-Based Predictive Modeling (CPM) with cross-validation to build models that predict behavioral scores from rest and task connectivity features. Compare model performance.

Signaling Pathways and Experimental Workflows

The following diagrams illustrate the core physiological pathway and a generalized experimental workflow for multimodal studies, as discussed in the reviewed literature.

Diagram 1: The neurovascular coupling pathway linking EEG and fNIRS signals. Neural activity, measurable by EEG, triggers neurotransmitter release and increases metabolic demand. This process, often mediated by astrocytes, leads to neurovascular coupling—the dilation of blood vessels and increased local blood flow. This hemodynamic response is measured by fNIRS as changes in hemoglobin concentration [4] [98].

Diagram 2: A generalized workflow for comparing brain states and modalities. Data is acquired simultaneously via EEG and fNIRS during both resting-state and task-based conditions. After modality-specific preprocessing, key features like spectral power (EEG) and activation maps (fNIRS) are extracted. These features are then used for a comparative analysis to determine the differential sensitivity of each modality to each brain state [97] [4] [96].

The Scientist's Toolkit: Essential Research Reagents and Materials

The following table details key materials and tools essential for conducting rigorous EEG and fNIRS research, as derived from the experimental protocols in the cited literature.

Table 3: Essential Research Toolkit for EEG and fNIRS Studies

Item Name	Function/Application	Key Characteristics
High-Density EEG System [97] [96]	Recording electrical brain activity with high temporal resolution.	30+ electrodes arranged in 10-5 or 10-20 system; compatible with fNIRS integration.
Continuous-Wave fNIRS System [97] [98]	Measuring hemodynamic changes (HbO/HbR) in the cortex.	Sources & detectors at 760nm & 850nm; optode distance ~30mm for cortical penetration.
fNIRS Short-Distance Detectors [97]	Separating superficial (scalp) from cerebral hemodynamic signals.	Placed close (~8mm) to sources to measure systemic confounds for signal correction.
Graph Theoretical Analysis Software [94]	Quantifying network topology (e.g., small-worldness, efficiency).	Computes metrics from functional connectivity matrices (e.g., using FC-NIRS toolbox).
Connectome-Based Predictive Modeling (CPM) [96]	Building models to predict behavior from brain connectivity.	Machine learning pipeline using functional connectivity features from rest or task.
Standardized Brain Atlases [4] [96]	Parcellating the brain into regions of interest (ROIs) for analysis.	Anatomical framework (e.g., Desikan-Killiany) for consistent ROI definition across subjects.

The collective evidence indicates that the choice between resting-state and task-based paradigms, and between EEG and fNIRS, is not a matter of superiority but of complementarity, dictated by the specific research question.

• For Trait-like Phenotypes and Clinical Classification: Resting-state paradigms show significant promise. The high accuracy of rs-EEG in identifying vulnerability to depression [95] and the reliability of resting-state fNIRS in characterizing the developmental connectome [94] suggest that the intrinsic, spontaneous brain state is a powerful biomarker for stable individual differences and neurological disorders.
• For State-like Cognitive Processes and Load Dependence: Task-based paradigms are more sensitive. The ability of fNIRS to track linear increases in activation and connectivity with working memory load [98] and the slight advantage of task-based EEG in predicting working memory performance [96] demonstrate that engaging the brain with specific cognitive challenges is essential for unpacking the neural mechanisms of cognition.
• For a Comprehensive View of Brain Organization: A multimodal approach is optimal. The structure–function relationship is best understood by combining the direct neural temporal dynamics of EEG with the hemodynamic spatial information from fNIRS, as their coupling varies across brain regions and states [4]. Furthermore, the finding that the fNIRS global signal is modulated by EEG-measured vigilance [97] crucially shows that "resting-state" is not a uniform baseline and must be interpreted with vigilance in mind.

In conclusion, both resting-state and task-based conditions offer unique and valuable insights into brain function. EEG excels in capturing rapid neural dynamics and trait biomarkers, while fNIRS is highly sensitive to hemodynamic changes induced by cognitive load and is robust for developmental connectome mapping. The integration of both modalities and both paradigms within the framework of the brain's structural architecture provides the most fertile ground for future discoveries in cognitive neuroscience and the development of objective biomarkers for neuropsychiatric drug development.

Understanding the relationship between the brain's structural wiring and its dynamic functional activity is a fundamental pursuit in neuroscience. The frontoparietal network (FPN), critical for high-level cognitive functions like working memory and cognitive control, and the somatomotor network (SMN), essential for sensory and motor processing, serve as prime models for this exploration. Contemporary neuroimaging techniques, particularly electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), offer distinct lenses through which to study these networks. EEG measures electrical neural activity with millisecond temporal resolution, directly reflecting the brain's fast electrophysiological processes. In contrast, fNIRS detects hemodynamic changes in blood oxygenation, an indirect correlate of neural activity that provides better spatial resolution but evolves over seconds [5] [4]. This comparison guide objectively examines the divergent findings generated by these two modalities when investigating the FPN and SMN, framing the analysis within the broader thesis of structure-function relationships in the human brain.

Methodological Comparison: EEG vs. fNIRS

Core Technical Principles

Table 1: Fundamental Characteristics of EEG and fNIRS

Feature	Electroencephalography (EEG)	Functional Near-Infrared Spectroscopy (fNIRS)
Measured Signal	Electrical potential from synchronized neuronal firing	Hemodynamic response (oxygenated & deoxygenated hemoglobin)
Temporal Resolution	Millisecond range (direct neural activity)	~1-10 seconds (indirect vascular response) [44]
Spatial Resolution	Limited (several centimeters)	Better than EEG (centimeter-scale) [23]
Portability/Tolerance	High (wearable caps, movement-tolerant)	Very High (portable, highly movement-tolerant) [100] [101]
Primary Strength	Timing of neural information transfer	Localization of sustained neural processes; ecological validity [44]

Experimental Protocols for Network Analysis

Typical protocols for investigating the FPN and SMN involve a combination of resting-state recordings and task-based paradigms.

• Resting-State Recordings: Participants are instructed to remain awake and relaxed with their eyes open or closed for a period of 3-10 minutes, during which spontaneous brain activity is recorded. This state reveals the brain's intrinsic functional architecture [100] [102].
• Task-Based Paradigms:
  • Motor Execution/Observation/Imagery (for SMN): Participants may physically perform a motor action (e.g., grasping a cup), observe an experimenter performing the action, or mentally rehearse the action without movement. These conditions probe different levels of the sensorimotor system [23].
  • Inhibitory Control Tasks (for FPN): Tasks like the Go/No-Go paradigm require participants to execute a response for frequent "Go" stimuli but inhibit that response for rare "No-Go" stimuli, heavily engaging the FPN's inhibitory control functions [101].
  • Complex Cognitive Tasks (for FPN): Semi-ecological tasks like the Space Fortress video game, which involves multiple demanding sub-tasks (e.g., ship control, target destruction, memory challenges), are used to engage higher-order cognitive functions of the FPN in a more naturalistic context [100] [103].

Divergent Findings in Frontoparietal and Somatomotor Networks

Quantitative Comparison of Network Findings

Table 2: Divergent Findings in Frontoparietal (FPN) and Somatomotor Networks (SMN) with EEG and fNIRS

Network	Experimental Condition	EEG Findings	fNIRS Findings	Interpretation of Divergence
Frontoparietal Network (FPN)	Structure-Function Coupling	Shows significant discrepancies with fNIRS, especially in cross-band representations [5] [45].	Lower correspondence with electrical activity in the transmodal cortex [5] [4].	FPN is a transmodal association area; EEG and fNIRS capture different physiological processes here.
Frontoparietal Network (FPN)	Inhibitory Control Task (Go/No-Go)	Not Available in Search Results	ADHD children showed decreased occurrence of strong, task-related FPN connectivity states [101].	fNIRS reveals atypical dynamic network recruitment to accommodate task demands.
Frontoparietal Network (FPN)	Resting-State / Complex Task	Not Available in Search Results	Intrinsic resting-state FPN connectivity predicts performance in a complex video game [100] [103].	fNIRS-derived FPN connectivity is a potential biomarker for real-world performance prediction.
Somatomotor Network (SMN)	Motor Imagery Task	Shows specificity for the SMN during the task [5] [45].	Shows specificity for the SMN during the task; activation in parietal regions (e.g., angular gyrus) [23].	Good convergence in this unimodal sensory-motor region.
Somatomotor Network (SMN)	Motor Execution, Observation, Imagery	Activation in bilateral central, right frontal, and parietal regions [23].	Activation in left angular gyrus, right supramarginal gyrus, and superior/inferior parietal lobes [23].	Different activated regions suggest each modality detects distinct facets of a shared network.

The Unimodal-Transmodal Gradient as an Explanatory Framework

A key concept for interpreting these divergent findings is the unimodal-transmodal gradient of cortical organization [5] [4]. This framework posits that the structure-function relationship is not uniform across the brain.

• Unimodal Cortex (e.g., Somatomotor Network): Regions dedicated to a single sensory or motor modality show stronger structure-function coupling. Both EEG and fNIRS findings tend to converge in these areas, such as the SMN, because they process immediate, externally-driven information that is more directly constrained by the underlying structural wiring [5] [45].
• Transmodal Cortex (e.g., Frontoparietal Network): Higher-order association areas that integrate information from multiple modalities show weaker structure-function coupling and greater decoupling. This is where the most significant discrepancies between EEG and fNIRS emerge. The FPN supports complex, abstract cognition that is less tied to instantaneous structural constraints and more influenced by dynamic, internal processes. The physiological signals captured by EEG (direct electrical activity) and fNIRS (indirect hemodynamic response) are therefore less correspondent in these regions, irrespective of the brain state [5] [4].

Visualizing Experimental and Analytical Workflows

Multimodal Experimental Protocol for Network Analysis

(Diagram Title: Multimodal Network Analysis Workflow)

Conceptual Model of Divergent Findings

(Diagram Title: Unimodal-Transmodal Gradient Explains Divergence)

The Scientist's Toolkit: Essential Research Reagents & Solutions

Table 3: Key Materials and Analytical Tools for Multimodal Network Research

Tool / Solution	Category	Function / Application	Example Use Case
Simultaneous EEG-fNIRS System	Hardware	Enables co-registration of electrical and hemodynamic brain activity from the same task epoch.	Core setup for directly comparing FPN/SMN activity across modalities [4] [23].
Graph Signal Processing (GSP)	Analytical Software	Mathematical framework for analyzing signals on graphs (networks). Extracts harmonic basis from structural connectomes.	Quantifying structure-function coupling and calculating the Structural-Decoupling Index (SDI) [5] [45].
Structured Sparse Multiset CCA (ssmCCA)	Analytical Software	A data fusion method to find correlated components across multiple datasets (e.g., EEG and fNIRS).	Identifying brain regions consistently activated across both modalities, validating findings [23].
High-Density EEG Cap (10-5 system)	Hardware	Provides dense spatial sampling of scalp electrical potentials.	Improving source reconstruction accuracy for locating FPN and SMN activity [4].
3D Magnetic Digitizer	Hardware	Precisely records the 3D locations of EEG electrodes and fNIRS optodes relative to head landmarks.	Crucial for coregistering multimodal data with an anatomical atlas for accurate source analysis [23].
fNIRS Optode Arrays	Hardware	Probes containing light sources and detectors placed on the scalp to measure hemodynamics.	Designed to cover cortical regions of interest (e.g., prefrontal-to-parietal for FPN) [100] [101].
Desikan-Killiany Atlas	Software/Template	A parcellation scheme that divides the cerebral cortex into anatomical regions of interest (ROIs).	Provides a common anatomical framework for mapping and comparing structural and functional data [4].
Dynamic Functional Connectivity Analysis	Analytical Software	Methods to study how connectivity between brain regions shifts over time during a task or at rest.	Revealing atypical recruitment of network states (e.g., in ADHD) that static analysis misses [101].

The integration of multimodal neuroimaging data represents a paradigm shift in clinical diagnostics and therapeutic development. By combining electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), researchers can now access complementary information about both the brain's electrical activity and its underlying hemodynamic responses. This comparative guide examines the technical capabilities, experimental protocols, and clinical applications of these modalities, highlighting how their fusion provides a more complete picture of brain structure-function relationships than either approach alone. We present quantitative data demonstrating the superior classification accuracy of hybrid systems, detailed methodologies for concurrent data acquisition, and visualizations of the underlying neurophysiological processes. For researchers and drug development professionals, this integration offers a powerful translational tool for screening therapeutic targets, confirming engagement, and tracking functional outcomes in central nervous system disorders.

Understanding the relationship between brain structure and function remains a fundamental challenge in neuroscience with profound implications for clinical diagnostics and therapeutic development [4]. While individual neuroimaging techniques have provided valuable insights, each modality possesses inherent limitations that constrain its diagnostic capability. Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) have emerged as particularly complementary techniques for probing brain function [24]. EEG measures the electrical activity generated by synchronized neuronal firing with millisecond temporal resolution, directly capturing neural dynamics but with limited spatial precision [104] [105]. In contrast, fNIRS measures hemodynamic responses by detecting changes in oxygenated and deoxygenated hemoglobin concentrations, providing better spatial localization of cortical activity but constrained by the slower temporal course of blood flow changes [106] [104].

The integration of these modalities is grounded in the physiological principle of neurovascular coupling—the intimate relationship between neural electrical activity and subsequent hemodynamic responses that deliver oxygen and nutrients to active brain regions [105]. This coupling forms the theoretical basis for combining EEG and fNIRS, enabling researchers to capture both the "when" of neural processing (via EEG) and the "where" of cortical activation (via fNIRS) within a single experimental framework [24] [107]. The translational advantage of this multimodal approach lies in its ability to provide a more comprehensive assessment of brain function, which is particularly valuable for understanding complex neurological and psychiatric conditions where both electrical and metabolic processes may be disrupted.

Technical Comparison: EEG and fNIRS Capabilities

Table 1: Fundamental characteristics of EEG and fNIRS neuroimaging techniques

Feature	EEG (Electroencephalography)	fNIRS (Functional Near-Infrared Spectroscopy)
What It Measures	Electrical activity from synchronized neuronal firing	Hemodynamic responses (oxygenated and deoxygenated hemoglobin)
Signal Source	Postsynaptic potentials in cortical pyramidal neurons	Changes in hemoglobin concentration due to neurovascular coupling
Temporal Resolution	High (milliseconds)	Low (seconds)
Spatial Resolution	Low (centimeter-level)	Moderate (better than EEG, but limited to cortex)
Depth of Measurement	Cortical surface	Outer cortex (1-2.5 cm deep)
Key Advantages	Direct neural activity measurement, excellent temporal resolution, portable	Better spatial resolution than EEG, tolerant to movement artifacts, portable
Primary Limitations	Susceptible to motion artifacts, poor spatial resolution, sensitive to electrical noise	Limited penetration depth, indirect measure of neural activity, lower temporal resolution
Best Use Cases	Fast cognitive tasks, ERP studies, seizure detection, sleep research	Naturalistic studies, child development, motor rehabilitation, sustained cognitive states

The complementary nature of EEG and fNIRS arises from their fundamental measurement principles. EEG records electrical potentials generated primarily by synchronized postsynaptic activity in cortical pyramidal neurons, which must be sufficiently aligned and coordinated to generate signals detectable at the scalp surface [105]. This electrical activity can be divided into various frequency bands (theta: 4-7 Hz, alpha: 8-14 Hz, beta: 15-25 Hz, gamma: >25 Hz) that reflect different brain states and cognitive processes [105]. However, the electrical signals are blurred and attenuated as they pass through the meninges, skull, and scalp, resulting in limited spatial resolution.

In contrast, fNIRS utilizes near-infrared light (650-925 nm) that can penetrate biological tissues to measure changes in hemoglobin concentrations in cortical blood vessels [106] [105]. Based on the modified Beer-Lambert law, fNIRS systems compute concentration changes of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) by measuring light attenuation at multiple wavelengths [105]. The spatial resolution of fNIRS is superior to EEG because light scattering, while substantial, is more predictable and localized than the spread of electrical potentials through resistive tissues [104]. However, fNIRS signals are indirect measures of neural activity, relying on neurovascular coupling, and thus have inherent latency of 2-6 seconds following neural activation [104].

Experimental Evidence: Quantifying the Multimodal Advantage

Table 2: Classification performance comparison across modalities in various tasks

Study Task	EEG Alone Accuracy	fNIRS Alone Accuracy	Fused EEG-fNIRS Accuracy	Improvement with Fusion	Citation
Mental Arithmetic	~88%	~88%	98.42%	+10.42%	[13]
Motor Imagery	~91%	~89%	96.74%	+5.74%	[13]
Mental Stress Assessment	83.1%	84.4%	91.0%	+7.9%	[108]
Amyotrophic Lateral Sclerosis	72.5%	70.8%	85.2%	+14.4%	[108]
Driver Drowsiness	81.5%	79.2%	87.0%	+5.5%	[108]

Empirical evidence demonstrates that combining EEG and fNIRS consistently outperforms single-modality approaches across diverse applications. The performance improvements shown in Table 2 highlight the synergistic value of integrating electrical and hemodynamic information. For instance, in mental arithmetic and motor imagery tasks, a multimodal framework utilizing multi-domain features and multi-level progressive learning achieved exceptional classification accuracies of 98.42% and 96.74% respectively, significantly outperforming either modality alone [13].

The structure-function relationship also varies meaningfully between these modalities. Research shows that fNIRS-based structure-function coupling resembles slower-frequency EEG coupling at rest, with notable variations across brain states and oscillations [4] [5]. Regionally, this relationship is heterogeneous, with stronger coupling in sensory cortex and greater decoupling in association cortex, following a unimodal to transmodal gradient [4]. These differences underscore how EEG and fNIRS capture complementary aspects of brain organization, together providing a more complete picture of brain function than either modality alone.

Different fusion strategies have been developed to optimize the integration of EEG and fNIRS data. Feature-level fusion involves concatenating or transforming features from both modalities before classification, while decision-level fusion combines the outputs of separate classifiers [13] [108]. Advanced feature selection methods, such as mutual information-based approaches that optimize complementarity, redundancy, and relevance between multimodal features, have demonstrated considerable improvements in hybrid classification performance compared to individual modalities and conventional classification without feature selection [108].

Methodological Protocols: Implementing Multimodal Imaging

Experimental Setup and Hardware Integration

Implementing concurrent EEG-fNIRS measurements requires careful consideration of hardware integration and experimental design. Three primary approaches exist for combining these modalities:

• Sequential separate systems: EEG and fNIRS data are acquired using separate systems synchronized during acquisition and analysis [24]. While simpler to implement, this approach may lack the precision required for microsecond-level EEG analysis.

• Unified processor systems: A single processor simultaneously processes and acquires both EEG signals and fNIRS input/output, achieving precise synchronization and streamlining analysis [24]. This method, though more complex, is widely used for concurrent fNIRS and EEG detection.

• Integrated headgear: EEG electrodes and fNIRS optodes are integrated into a single acquisition helmet [24]. Design approaches include shared substrate materials, separate arrangement of components, or direct integration of NIR fiber optics into existing EEG electrode caps.

Customized solutions using 3D printing or cryogenic thermoplastic sheets have been developed to create joint-acquisition helmets tailored to experimental requirements and individual head sizes, though these approaches involve higher costs or potential comfort issues [24].

Signal Acquisition Parameters

EEG Acquisition:

• Electrode placement: International 10-5 or 10-20 systems
• Sampling rate: 200-1000 Hz (typically downsampled to 200 Hz for analysis)
• Reference electrode placement: Typically vertex or mastoid
• Impedance: Maintained below 5-10 kΩ for optimal signal quality

fNIRS Acquisition:

• Optode placement: International 10-20 system with source-detector separation of 2.5-3 cm for adults
• Wavelengths: Typically 760 nm and 850 nm to distinguish HbO and HbR
• Sampling rate: 10-12.5 Hz (typically downsampled to 10 Hz for analysis)
• Signal quality assessment: Scalp-coupling index (SCI) >0.7 for acceptable signal quality [4]

Data Preprocessing Pipelines

EEG Preprocessing:

• Filtering: Bandpass filter (e.g., 0.5-40 Hz) to remove drift and high-frequency noise
• Bad channel identification and interpolation
• Ocular and cardiac artifact removal using ICA or regression approaches
• Epoching relative to task events
• Baseline correction and artifact rejection

fNIRS Preprocessing:

• Conversion of raw light intensity to optical density
• Signal quality assessment using SCI metric [4]
• Bandpass filtering (e.g., 0.02-0.08 Hz for resting state) to remove physiological noise
• Motion artifact correction using wavelet filtering or robust regression
• Conversion to hemoglobin concentration changes using modified Beer-Lambert law
• Identification and rejection of time segments with excessive head movement using global variance in temporal derivative (GVTD) metric [4]

Multimodal Data Fusion Approaches

Three primary methodological categories have been established for concurrent fNIRS-EEG data analysis [105]:

• EEG-informed fNIRS analyses: Utilizing EEG-derived features to inform or constrain the analysis of fNIRS data, particularly valuable for exploiting the superior temporal resolution of EEG.

• fNIRS-informed EEG analyses: Employing hemodynamic information to guide EEG source localization or interpretation, leveraging the better spatial resolution of fNIRS.

• Parallel fNIRS-EEG analyses: Analyzing both modalities separately and integrating results at the feature or decision level, often using techniques like joint Independent Component Analysis (jICA), canonical correlation analysis (CCA), or machine learning approaches that combine feature sets from both modalities [104].

Diagram 1: Neurovascular coupling and multimodal integration basis. EEG directly measures neural electrical activity with high temporal resolution, while fNIRS measures the hemodynamic response resulting from neurovascular coupling with better spatial resolution. Combining both provides complementary information that enhances classification accuracy in clinical applications.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential equipment and computational tools for multimodal EEG-fNIRS research

Category	Item	Specification/Function	Representative Examples
Hardware Components	EEG System	Amplifiers, electrodes, caps for electrical signal acquisition	30-electrode systems (10-5 placement), active/passive electrodes
	fNIRS System	Light sources, detectors, optodes for hemodynamic measurement	Continuous-wave systems, dual-wavelength (760/850 nm)
	Integrated Caps	Combined holder systems for simultaneous measurement	Custom 3D-printed helmets, thermoplastic sheets, modified EEG caps
	Synchronization Interface	Hardware/software for temporal alignment of modalities	TTL pulses, shared clock systems, parallel port triggers
Software & Analytical Tools	Preprocessing Tools	Artifact removal, filtering, quality assessment	EEGLAB, MNE-Python, Brainstorm, Homer2, NIRS-KIT
	Fusion Algorithms	Multimodal feature integration and analysis	Joint ICA, CCA, mutual information-based feature selection
	Classification Libraries	Machine learning for pattern recognition	Scikit-learn, TensorFlow, PyTorch, custom MATLAB toolboxes
Experimental Materials	Conductive Gels	EEG electrode-scalp interface	Electrolyte gels, pastes (for wet electrodes)
	Optode Positioning	Scalp contact and light transmission	Fiber optic holders, spring-loaded optodes
	Anatomical Co-registration	Spatial localization of signals	3D digitizers, MRI template alignment, photogrammetry

Successful implementation of multimodal EEG-fNIRS studies requires careful selection of compatible hardware and software components. Integrated systems should ensure precise temporal synchronization, either through hardware triggers or shared clock systems, with sampling rates appropriately matched to capture both fast electrical dynamics (EEG: typically 200-1000 Hz) and slower hemodynamic changes (fNIRS: typically 10-12.5 Hz) [24] [4].

Customized headgear solutions are particularly important for optimizing signal quality in simultaneous recordings. Research indicates that integrating EEG electrodes and fNIRS optodes on a shared substrate material or arranging them separately with precise co-registration enables accurate spatial localization of brain regions probed by both modalities [24]. While commercially available elastic caps offer convenience, they may result in uncontrollable variations in optode placement and contact pressure across subjects. Customized solutions using 3D printing or thermoplastic materials provide better individualization but at higher cost or potential comfort trade-offs [24].

Diagram 2: Experimental workflow for multimodal EEG-fNIRS studies. The process begins with careful experimental design and subject preparation, followed by synchronized data acquisition with quality monitoring. Parallel preprocessing pipelines for each modality lead to feature extraction and multimodal fusion, ultimately supporting clinical interpretation for diagnostics and therapeutic development.

Clinical Applications and Translational Impact

The integration of EEG and fNIRS has demonstrated particular value across multiple clinical domains where both electrical and hemodynamic information provides complementary diagnostic insights:

Epilepsy and Seizure Detection: Multimodal monitoring improves seizure focus localization by combining EEG's precise temporal resolution ofictal events with fNIRS's spatial localization of hemodynamic changes [24] [106]. The simultaneous assessment helps distinguish epileptiform activity from other artifacts and enhances understanding of seizure propagation through neurovascular networks.

Psychiatric and Neurodevelopmental Disorders: For conditions like attention-deficit hyperactivity disorder (ADHD) and schizophrenia, combined EEG-fNIRS approaches can identify both functional connectivity abnormalities (via EEG) and metabolic alterations (via fNIRS) that may not be detectable with single modalities [24] [109]. This is particularly valuable for evaluating treatment efficacy and providing precise diagnostic options beyond behavioral observations.

Neurological Rehabilitation: In motor imagery-based rehabilitation for stroke and paralysis, hybrid BCI systems leveraging both EEG and fNIRS have shown superior performance in detecting patient intent and tracking recovery progress [13]. The combination allows for more robust control signals by leveraging both immediate electrical signatures and sustained hemodynamic responses.

Drug Development and Target Engagement: EEG biomarkers are increasingly used in CNS therapeutic development to screen new targets, confirm target engagement, and track functional outcomes [109]. The addition of fNIRS provides complementary information about metabolic effects and neurovascular coupling integrity, offering a more comprehensive assessment of drug effects on brain function.

Consciousness Monitoring: During anesthesia and in critical care settings, simultaneous EEG-fNIRS monitoring provides information about both electrical brain states and cerebral oxygenation, helping to prevent neurological complications and optimize sedation levels [24] [106].

The integration of EEG and fNIRS represents a significant advancement in clinical neuroimaging, offering a translational advantage that transcends the capabilities of either modality alone. By capturing complementary electrical and hemodynamic aspects of brain function, this multimodal approach provides researchers and clinicians with a more comprehensive tool for understanding structure-function relationships in both healthy and diseased brains. The experimental evidence clearly demonstrates that fused EEG-fNIRS data consistently outperforms single-modality approaches in classification accuracy across diverse tasks and clinical populations.

For drug development professionals, this integration offers particularly promising applications in target screening, engagement confirmation, and outcome tracking. The ability to simultaneously monitor electrical brain activity and metabolic demands provides a more complete picture of therapeutic effects on neural function. As hardware integration becomes more sophisticated and analysis techniques more refined, simultaneous EEG-fNIRS is poised to become an increasingly valuable tool in the clinical diagnostics and therapeutic development arsenal, ultimately accelerating the pace of CNS therapeutic development and improving patient outcomes across a spectrum of neurological and psychiatric disorders.

Conclusion

The integration of EEG and fNIRS provides a powerful, multifaceted lens through which to view the brain's structure-function relationships, offering a more complete picture than either modality alone. The key takeaway is their profound complementarity: EEG delivers millisecond precision for electrical dynamics, while fNIRS provides superior spatial localization of hemodynamic changes. This synergy, especially when analyzed with advanced frameworks like graph signal processing and multilayer networks, leads to more robust biomarkers, higher classification accuracy in BCI applications, and richer insights into both healthy and diseased brain states. For future biomedical and clinical research, the path forward involves standardizing multimodal protocols, developing more sophisticated real-time fusion algorithms, and leveraging these combined tools to bridge the critical gap between preclinical models and human clinical trials, ultimately accelerating the development of novel CNS therapeutics.

References

• 1. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOopueiQL08VbW4t1cNxxJdhLtc6WkoPwfcO0BzcA-KQLxtsq49hN]
• 2. Concurrent fNIRS and EEG for Brain Function Investigation [https://pmc.ncbi.nlm.nih.gov/articles/PMC9371171/]
• 3. Fusion of fNIRS and EEG: a step further in brain research [https://www.bitbrain.com/blog/fusion-fnirs-eeg-brain-activity-research]
• 4. Comparing structure–function relationships in brain ... [https://www.nature.com/articles/s41598-024-79817-x]
• 5. Comparing structure-function relationships in brain ... [https://pubmed.ncbi.nlm.nih.gov/39578593/]
• 6. A Cross-Disciplinary Review of the fNIRS-EEG Dual ... [https://www.mdpi.com/2076-3425/14/10/1022]
• 7. Wearable, Integrated EEG–fNIRS Technologies: A Review [https://pmc.ncbi.nlm.nih.gov/articles/PMC8469799/]
• 8. Improved classification performance of EEG - fNIRS multimodal... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9388144/]
• 9. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOopd3eLOT4xQ_VWYNYusBPqLPjXKWH7qXbDS455S6Y2mSCqJYw9Y]
• 10. Neurovascular Coupling: Scientometric Analysis of 30 Years ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12606022/]
• 11. Neurovascular coupling during auditory stimulation: event-related potentials and fNIRS hemodynamic - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10517045/]
• 12. Brain endothelial gap junction coupling enables rapid ... [https://www.sciencedirect.com/science/article/pii/S0092867425007329]
• 13. Improved classification performance of EEG-fNIRS ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2022.973959/full]
• 14. An EEG-fNIRS neurovascular coupling analysis method to investigate cognitive-motor interference - ScienceDirect [https://www.sciencedirect.com/science/article/abs/pii/S001048252300433X]
• 15. Quantifying neurovascular coupling through a concurrent assessment of arterial, capillary, and neuronal activation in humans: A multimodal EEG-fNIRS-TCD investigation - ScienceDirect [https://www.sciencedirect.com/science/article/pii/S1053811924004075]
• 16. Prefrontal fNIRS-based clinical data analysis of brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8620253/]
• 17. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOopyiSfmvqLSZ4-LrRoTekWW4NXmgtvYcRhuBMqRSA_n71sH4aXQ]
• 18. Advances in Functional Near-Infrared Spectroscopy [https://pubmed.ncbi.nlm.nih.gov/40796273/]
• 19. Enhanced spatiotemporal resolution imaging of neuronal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7778454/]
• 20. Applications and advances of combined fMRI-fNIRs ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1542075/full]
• 21. Why Brain Wearables Are Making Traditional EEG Labs ... [https://globalrph.com/2025/10/why-brain-wearables-are-making-traditional-eeg-labs-obsolete-in-2025/]
• 22. Disrupted temporal structure of the M/EEG meta-states ... [https://pubmed.ncbi.nlm.nih.gov/41161361/]
• 23. Simultaneous multimodal fNIRS-EEG recordings reveal ... [https://www.nature.com/articles/s41598-023-31609-5]
• 24. A Cross-Disciplinary Review of the fNIRS-EEG Dual-Modality ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11506513/]
• 25. Enhanced brain structure-function tethering in transmodal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10598018/]
• 26. Spatially heterogeneous structure-function coupling in ... [https://www.sciencedirect.com/science/article/pii/S1053811923004275]
• 27. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOoptJ8TWdirBRZGCphRGPy6sEVkVtucC4nT5g7K2nD3nVgL5zCOE]
• 28. Time-resolved structure-function coupling in brain networks [https://www.nature.com/articles/s42003-022-03466-x]
• 29. Reproducibility of fNIRS within Subject for Visual and Motor ... [https://www.sciencedirect.com/science/article/pii/S1053811925004951]
• 30. Multimodal fNIRS–EEG sensor fusion: Review of data-driven ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12592382/]
• 31. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOooQAl5q-MtEfLsWz70Y-SJVbLaqhFxz0SHZek4_NKXs0avdcV4i]
• 32. What we can and cannot (yet) do with functional near ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4033094/]
• 33. Optimizing spatial specificity and signal quality in fNIRS [https://www.frontiersin.org/journals/neuroergonomics/articles/10.3389/fnrgo.2024.1286586/full]
• 34. Scale invariance in fNIRS as a measurement of cognitive ... [https://www.sciencedirect.com/science/article/abs/pii/S0010945222001551]
• 35. Decision fusion in healthcare and medicine: a narrative review [https://pmc.ncbi.nlm.nih.gov/articles/PMC8800206/]
• 36. Unlocking Environmental Insights [https://www.numberanalytics.com/blog/ultimate-guide-data-fusion-environmental-remote-sensing]
• 37. Multimodal Models and Fusion - A Complete Guide [https://medium.com/@raj.pulapakura/multimodal-models-and-fusion-a-complete-guide-225ca91f6861]
• 38. Quantifying the sensitivity of SAR and optical images three- ... [https://www.sciencedirect.com/science/article/pii/S156984322200070X]
• 39. Feature and decision-level fusion for schizophrenia ... [https://pubmed.ncbi.nlm.nih.gov/35609033/]
• 40. Predicted and Explained: Transforming drug discovery with ... [https://www.sciencedirect.com/science/article/abs/pii/S0010482525004962]
• 41. Graph Signal Processing and Brain Signal Analysis [https://www.mathworks.com/help/signal/ug/graph-signal-processing-and-brain-signal-analysis.html]
• 42. Understanding Concepts in Graph Signal Processing for ... [https://arxiv.org/html/2312.03371v2]
• 43. Large-scale comparative analysis reveals top graph signal ... [https://pubmed.ncbi.nlm.nih.gov/40909572/]
• 44. Investigating the interaction between EEG and fNIRS [https://www.sciencedirect.com/science/article/pii/S1877750324002096]
• 45. Comparing structure–function relationships in brain ... [https://sano.science/research/comparing-structure-function-relationships-in-brain-networks-using-eeg-and-fnirs/]
• 46. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOoqVXwg_771C-wMRDl_X-qY0yL94UHy_AXqOewyFiPxm7lz4ju5P]
• 47. Comparison of Brain Activation Between Different Modes ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11710888/]
• 48. EEG–fNIRS signal integration for motor imagery ... [https://www.sciencedirect.com/science/article/pii/S2772528625000299]
• 49. ECA-FusionNet: a hybrid EEG-fNIRS signals network for MI ... [https://pubmed.ncbi.nlm.nih.gov/39874664/]
• 50. A multi-modal dataset of electroencephalography and ... [https://www.nature.com/articles/s41597-025-05286-0]
• 51. Individual Differences in Math Ability Determine ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2019.00227/full]
• 52. Temporal Comparison Between NIRS and EEG Signals ... [https://pubmed.ncbi.nlm.nih.gov/27526147/]
• 53. HEFMI-ICH: a hybrid EEG-fNIRS motor imagery dataset for ... [https://www.nature.com/articles/s41597-025-06100-7]
• 54. Use of Biomarkers in Drug Development for Regulatory ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12489174/]
• 55. AI and the future of biomarker analysis in early R&D [https://www.drugtargetreview.com/article/190428/ai-and-the-future-of-biomarker-analysis-in-early-rd/]
• 56. EFRM: A Multimodal EEG–fNIRS Representation-learning ... [https://www.sciencedirect.com/science/article/abs/pii/S0010482525016464]
• 57. Diagnostic potential of near‐infrared spectroscopy in mild ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12505193/]
• 58. Emerging Technologies in Biomarker Discovery You ... [https://blog.crownbio.com/emerging-technologies-in-biomarker-discovery-you-should-know-about]
• 59. Functional Near-Infrared Spectroscopy in neurodegenerative ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11479976/]
• 60. Simultaneous functional near-infrared spectroscopy and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5566595/]
• 61. Automatic depression diagnosis through hybrid EEG and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10483285/]
• 62. Resting-state functional near-infrared spectroscopy in ... [https://www.sciencedirect.com/science/article/pii/S2213158225000038]
• 63. A Functional Near-infrared Spectroscopy Study [https://www.cpn.or.kr/journal/view.html?uid=1563&vmd=Full]
• 64. Comparison of motion correction techniques applied to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9900395/]
• 65. Exploring Imagined Movement for Brain–Computer Interface ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12468679/]
• 66. Characterizing motion artifacts in functional near-infrared ... [https://www.sciencedirect.com/science/article/pii/S1746809425007670]
• 67. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOopVFQE89nN2EdCYfnFTwnkRmWWWttJZUaRkMkOC7q1ViGzqbqlN]
• 68. Feature Extraction and Classification Methods for Hybrid ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2018.00246/full]
• 69. Co-localized optode-electrode design for multimodal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11978466/]
• 70. A Deep-Learning Empowered, Real-Time Processing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12270413/]
• 71. Brain–computer interface: trend, challenges, and threats [https://braininformatics.springeropen.com/articles/10.1186/s40708-023-00199-3]
• 72. Brain–computer interfaces for EEG neurofeedback [https://www.sciencedirect.com/science/article/abs/pii/S016787601300247X]
• 73. Current Status, Challenges, and Possible Solutions of EEG ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7283463/]
• 74. The Potential of Functional Near-Infrared Spectroscopy ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.00594/full]
• 75. Performance comparison of systemic activity correction in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9555616/]
• 76. A brief real-time fNIRS-informed neurofeedback training of ... [https://www.sciencedirect.com/science/article/abs/pii/S0278584624000368]
• 77. NeuroPilot: A Real-time Brain-Computer Interface system ... [https://arxiv.org/html/2510.20958v2]
• 78. A Practical Guide for Combined fNIRS & EEG Recordings [https://www.gtec.at/2021/04/12/combined-fnirs-eeg-recordings/?srsltid=AfmBOoorllyKwxPJUUC5J_dI4jglDmbC7VWxAUv28W-CVd9xnvWeiin5]
• 79. Best practices for fNIRS publications - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC7793571/]
• 80. Simultaneous EEG-fNIRS study of visual cognitive processing [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0325017]
• 81. Simultaneous EEG-fNIRS Data Classification Through ... [https://pubmed.ncbi.nlm.nih.gov/39247844/]
• 82. Simultaneous EEG and fNIRS recordings for semantic ... [https://www.nature.com/articles/s41597-025-04967-0]
• 83. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOoorsGBuhAkDwfedmcI7xXz487ccX1rYfyWR3QvFBs2zdLFvErA-]
• 84. Cross-modal fusion of brain imaging and clinical data for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12633903/]
• 85. A model of the formation of multilayer networks [https://www.sciencedirect.com/science/article/abs/pii/S002205312300114X]
• 86. EEG vs MRI vs fMRI vs fNIRS [https://neuronlab.cn/%E7%9F%A5%E8%AF%86%E7%A7%91%E6%99%AE/eeg-vs-mri-vs-fmri-vs-fnirs.html]
• 87. Quantifying layer similarity in multiplex networks [https://pmc.ncbi.nlm.nih.gov/articles/PMC6124071/]
• 88. fNIRS的跨学科视角翻译 [https://blog.csdn.net/cc1609130201/article/details/145208054]
• 89. Improvement of classification accuracy of functional near- ... [https://www.sciencedirect.com/science/article/pii/S2667102625000245]
• 90. Early-stage fusion of EEG and fNIRS improves ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.1062889/full]
• 91. A Novel Transfer Learning‐Based Hybrid EEG‐fNIRS Brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12631896/]
• 92. Multimodal MBC-ATT: cross-modality attentional fusion of ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2025.1660532/full]
• 93. A study of motor imagery EEG classification based on ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12307489/]
• 94. Applications of Resting-State fNIRS in the Developing Brain [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.00476/full]
• 95. Comparing resting state and task-based EEG using ... [https://www.nature.com/articles/s41598-023-34298-2]
• 96. Direct Comparison of EEG Resting State and Task ... [https://pubmed.ncbi.nlm.nih.gov/40317131/]
• 97. Amplitude of fNIRS Resting-State Global Signal Is Related to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7744746/]
• 98. Sensitivity of fNIRS to cognitive state and load - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3930096/]
• 99. Resting-state connectivity and task-based cortical ... [https://www.sciencedirect.com/science/article/pii/S2666956025000042]
• 100. Functional Connectivity Within the Fronto-Parietal Network ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10790952/]
• 101. An Insight Into Task-Based Dynamic Connectivity Through ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2020.00003/full]
• 102. Study on resting-state functional connectivity ... [https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2025.1567526/full]
• 103. Functional Connectivity Within the Fronto-Parietal Network ... [https://www.frontiersin.org/journals/neuroergonomics/articles/10.3389/fnrgo.2021.718176/full]
• 104. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOop38BZHsVOmHl2C_Vr1zlhqnOs79RiqHkWo3uBcNy35XDvvM9a0]
• 105. Concurrent fNIRS and EEG for Brain Function Investigation [https://www.mdpi.com/1424-8220/22/15/5865]
• 106. Functional Near-Infrared Spectroscopy and Its Clinical ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.00724/full]
• 107. The Advantages of Combining fNIRS and EEG for Brain ... [https://www.artinis.com/blogpost-all/advantages-of-combining-fnirs-and-eeg]
• 108. Multimodal fusion of EEG-fNIRS: a mutual information-based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7984774/]
• 109. EEG Biomarkers: Transforming CNS Drug Discovery [https://www.neuroelectrics.com/blog/eeg-biomarkers-transforming-cns-drug-discovery]