Advancing Clinical Diagnostics: A Comparative Analysis of fMRI, EEG, and fNIRS for Accurate Neurological Disorder Assessment

Abstract

This article provides a comprehensive analysis of the evolving roles of functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS) in enhancing the diagnostic accuracy for neurological disorders. Tailored for researchers, scientists, and drug development professionals, it explores the foundational principles, methodological applications, and comparative strengths of these neuroimaging modalities. By examining integrative approaches, troubleshooting common challenges, and validating techniques through case studies, this review synthesizes current evidence to guide the selection and optimization of imaging tools. The article further discusses future directions, including the impact of machine learning and multimodal integration, on advancing personalized medicine and clinical trial methodologies in neurology.

Understanding the Neuroimaging Trio: Core Principles and Clinical Targets of fMRI, EEG, and fNIRS

Functional magnetic resonance imaging (fMRI) and functional near-infrared spectroscopy (fNIRS) have become cornerstone techniques in modern neuroscience for non-invasively mapping brain function. Both modalities rely on hemodynamic responses, specifically changes in blood oxygenation, to infer neural activity through the principle of neurovascular coupling [1] [2]. When neurons become active, local blood flow increases to deliver oxygenated blood, creating measurable changes in the relative concentrations of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) [2]. While fMRI and fNIRS share this common physiological basis, they differ fundamentally in their physical principles, technical capabilities, and clinical applications.

This guide provides a detailed comparison of fMRI and fNIRS, focusing on their respective abilities to measure blood oxygenation changes for functional brain mapping. We present experimental data comparing their performance, detailed methodologies for simultaneous acquisition protocols, and analysis of their complementary strengths and limitations within clinical and research contexts, particularly for neurological disorders.

Technical Foundations and Comparison

Fundamental Physical Principles

fMRI measures brain activity indirectly through the Blood Oxygenation Level Dependent (BOLD) contrast [1] [3]. The physical principles of fMRI are based on nuclear magnetic resonance, where hydrogen nuclei align in a strong magnetic field [1]. The BOLD signal originates from the magnetic susceptibility differences between hemoglobin species: deoxygenated blood is paramagnetic while oxygenated blood is diamagnetic [1]. Active brain regions exhibit increased blood flow that surpasses oxygen consumption, leading to a higher ratio of oxygenated to deoxygenated hemoglobin and increased MRI signal intensity [1].

fNIRS utilizes near-infrared light (650-950 nm) to measure changes in hemoglobin concentrations in cortical brain tissue [1] [2]. Biological tissues are relatively transparent to light in this spectrum, allowing light to penetrate and be absorbed by chromophores, primarily HbO and HbR [2]. Using modified Beer-Lambert law, fNIRS calculates concentration changes of these hemoglobin species based on measured light attenuation [2] [4]. Unlike fMRI which provides a combined oxygenation measure, fNIRS quantifies HbO and HbR separately, potentially offering more nuanced physiological information [2].

Technical Specifications Comparison

Table 1: Technical comparison between fMRI and fNIRS

Parameter	fMRI	fNIRS
Spatial Resolution	1-3 mm (excellent) [5]	1-3 cm (moderate) [5]
Temporal Resolution	1-2 seconds (limited by hemodynamic response) [5]	Millisecond to second range (excellent) [5]
Penetration Depth	Whole brain (including subcortical structures) [5]	Superficial cortex (2-3 cm from surface) [5] [2]
Measured Parameters	BOLD signal (combined oxygenation effect) [1] [3]	Separate HbO and HbR concentration changes [2]
Portability	No (fixed scanner environment) [1] [5]	Yes (bedside, naturalistic settings) [1] [5]
Tolerance to Motion	Low (requires head immobilization) [1]	Moderate (some tolerance for movement) [1]
Cost	High (equipment and maintenance) [1]	Relatively low [5] [4]
Population Flexibility	Limited (claustrophobia, metal implants) [1] [2]	High (infants, patients with implants) [1] [2]

Experimental Validation and Direct Comparison Studies

Spatial Correspondence Studies

Recent research has directly compared the spatial accuracy of fNIRS against the gold standard of fMRI. A 2024 study with 22 healthy adults performing motor and visual tasks during same-day fMRI and whole-head fNIRS scanning revealed promising correspondence [6].

Table 2: Spatial correspondence between fNIRS and fMRI in detecting task-related activity

Analysis Type	True Positive Rate (fNIRS overlap with fMRI)	Positive Predictive Value
Group Analysis	Up to 68%	51%
Within-Subject	Average of 47.25%	41.5%

The positive predictive value was lower for within-subject analyses, reflecting significant fNIRS activity in regions without corresponding fMRI activity, potentially due to physiological noise or different sensitivities to hemoglobin changes [6]. This study supports whole-head fNIRS as having promising clinical utility for functional assessment of superficial cortical regions [6].

Brain Fingerprinting Accuracy

A 2023 simultaneous fNIRS-fMRI study investigated "brain fingerprinting" - identifying individuals based on their unique functional connectivity patterns [7]. The research demonstrated that with proper preprocessing and sufficient data, fNIRS can achieve classification accuracy approaching that of fMRI.

Table 3: Brain fingerprinting classification accuracy with fNIRS vs. fMRI

Condition	fNIRS Classification Accuracy	fMRI Classification Accuracy
Optimal conditions	75% to 98% (depending on runs and regions)	99.9%
Impact factors	Number of runs, spatial coverage, optode positioning	Consistency of BOLD response across sessions

The accuracy of fNIRS-based identification was more impacted by the number of runs and spatial coverage than the choice of classification algorithm [7]. This highlights the critical importance of experimental design in fNIRS studies.

Methodologies for Combined fMRI-fNIRS Experiments

Simultaneous Acquisition Protocol

A validated protocol for simultaneous fMRI-fNIRS acquisition involves the following key steps [7]:

• Participant Preparation: Fit participants with an fNIRS cap compatible with MRI environments, ensuring all materials are non-metallic and MR-safe.

• Optode Placement: Position fNIRS sources and detectors on the head using a standard 10-20 system cap. Typical setups include 16 light sources (LEDs at 760 and 850 nm) and 32 detectors, creating 64 measurement channels with source-detector distances of 2.8-3.5 cm [7].

• Spatial Co-registration: Digitize optode locations using a magnetic motion tracking sensor (e.g., Fastrak, Polhemus) while the participant is in the scanner. Record five anatomical landmarks (Nz, Cz, Iz, A1, A2 in the 10-20 system) for precise co-registration with anatomical MRI [7].

• Data Acquisition: Collect six runs of simultaneous MRI and fNIRS data, each six minutes in duration, during resting-state or task conditions. Instruct participants to close eyes, stay relaxed, avoid movement, and not focus on specific thoughts [7].

• fMRI Parameters: Acquire T1-weighted structural images (TR = 7 ms, TE = 3.2 s) and five functional resting-state scans (180 volumes, TR = 2 s, TI = 900 ms) using a 3T scanner with a 32-channel head coil [7].

Data Processing Pipeline

fMRI Preprocessing:

• Process BOLD data with standard tools (e.g., SPM12, UF2C toolbox)
• Include normalization, motion artifact correction (framewise displacement and DVARS)
• Apply band-pass filtering (0.009-0.08 Hz) with stopband attenuation of 50 dB
• Perform regressions of white matter, cerebral spinal fluid, and global signal
• Exclude runs with less than 4 minutes without motion artifacts [7]

fNIRS Preprocessing:

• Prune channels with low signal-to-noise ratio (SNR < 8)
• Convert light intensity to optical density
• Correct motion artifacts using hybrid algorithms combining spline interpolation with wavelet decomposition
• Remove consistent bad channels across all participants to maintain consistent channel numbers [7]
• Apply global systemic physiology removal to reduce confounding physiological signals

Neurovascular Coupling and Signal Interpretation

The Hemodynamic Response Pathway

The fundamental connection between neural activity and measurable hemodynamic changes occurs through neurovascular coupling. This complex process involves multiple cell types and signaling pathways that ultimately translate neuronal activation into vascular responses [2].

Increased neuronal activity triggers glutamate release, activating astrocytes which release vasoactive factors including nitric oxide (NO) and prostaglandins [2]. These factors cause dilation of cortical arterioles, increasing regional cerebral blood flow (CBF) and cerebral blood volume (CBV) [2]. The resulting oxygen delivery typically exceeds local consumption, creating a surplus of oxygenated hemoglobin that forms the basis for both BOLD fMRI and fNIRS signals [2].

Temporal Dynamics of Hemodynamic Parameters

Studies comparing the latency differences between HbO and HbR responses during functional activation have revealed important insights into hemodynamic regulation. Research shows that the apparent latency between HbO and HbR changes (approximately 1.6±0.2 seconds in motor cortex) may be influenced by systemic confounds rather than representing fundamental physiological differences between cortical areas [8]. When systemic responses are minimized using specialized paradigms, these latencies disappear, suggesting simultaneous changes in both hemoglobin species [8].

Research Reagent Solutions and Essential Materials

Table 4: Essential research materials for combined fMRI-fNIRS studies

Item	Function/Purpose	Example Specifications
fNIRS System	Measures changes in hemoglobin concentrations	NIRScout (NIRx); 760 & 850 nm wavelengths; 7.8 Hz sampling rate [7]
MRI Scanner	Provides structural images and BOLD functional data	3T Philips Achieva with 32-channel head coil [7]
fNIRS Optodes	Light sources and detectors for signal acquisition	16 sources, 32 detectors forming 64 channels; 2.8-3.5 cm separation [7]
Digitization System	Records precise optode locations for co-registration	Fastrak (Polhemus) magnetic motion tracking sensor [7]
Analysis Software	Processes and analyzes neuroimaging data	SPM12, UF2C toolbox, HomER2, in-house Matlab scripts [7]
MRI-Compatible Cap	Holds fNIRS optodes in place during scanning	Standard 10-20 system cap with non-metallic components [7]

Clinical Applications in Neurological Disorders

The complementary strengths of fMRI and fNIRS make them valuable tools for researching and diagnosing neurological disorders. fMRI provides detailed spatial maps for localization of function, while fNIRS offers portable monitoring capabilities for naturalistic assessment and treatment evaluation.

In substance use disorders and behavioral addiction, neurofeedback training using fMRI, fNIRS, or EEG has shown promise as an adjunctive intervention [9]. These approaches enable patients to self-regulate brain activity patterns associated with craving, potentially improving treatment outcomes [9].

For neurological monitoring in critical care, fNIRS has become a standard tool for assessing cerebral oxygenation and autoregulation in patients with stroke and traumatic brain injury [2] [4]. Its portability allows for bedside monitoring that would be impossible with fMRI [2].

In cognitive and psychiatric disorders, fNIRS has been applied to conditions including Alzheimer's disease, schizophrenia, Parkinson's disease, and childhood disorders [4]. The tolerance of fNIRS for movement and its ability to function in naturalistic environments makes it particularly valuable for populations that cannot tolerate fMRI scanning [1] [4].

Limitations and Methodological Considerations

fNIRS Specific Limitations

A significant challenge in fNIRS, particularly for forehead measurements, is contamination by systemic physiological signals. Studies have demonstrated that during tasks such as verbal fluency, a substantial portion of the fNIRS signal may originate from skin blood flow changes rather than cerebral activity [10]. This confound can be addressed using short-distance channels (5 mm separation) to measure and subtract superficial contributions [10].

fNIRS also faces limitations in spatial resolution and depth penetration. The spatial resolution typically ranges from 1-3 cm, restricting precise functional localization, while penetration depth is limited to superficial cortical regions, making fNIRS unsuitable for investigating subcortical structures [5] [2].

fMRI Specific Limitations

fMRI is limited by its sensitivity to motion, requiring strict head immobilization that restricts the range of behaviors that can be studied [1] [5]. The scanner environment also imposes practical constraints including contraindications for individuals with metal implants, claustrophobia, and difficulties studying populations such as infants [1].

The relationship between neural activity and the BOLD signal is complex and influenced by multiple physiological variables including the efficiency of the hemodynamic response and unique properties of neural circuits being interrogated [1] [3]. This complexity inherently limits the interpretation of brain function possible with BOLD fMRI alone [1].

The integration of fMRI and fNIRS represents a powerful multimodal approach that leverages their complementary strengths. Future developments will likely focus on hardware innovations such as MRI-compatible fNIRS probes, standardized protocols for combined acquisition, and advanced data fusion techniques driven by machine learning [5]. These advances may help overcome current limitations, including fNIRS's depth limitation, potentially by combining it with other modalities to infer subcortical activities [5].

For researchers and clinicians, the choice between fMRI and fNIRS depends on specific application requirements. fMRI remains superior for precise spatial localization throughout the entire brain, while fNIRS offers advantages for temporal dynamics, portability, and studies requiring ecological validity. The combined use of both modalities provides a more comprehensive understanding of brain function, enhancing diagnostic and therapeutic strategies in neurological and psychiatric disorders.

Understanding the hemodynamic basis of both techniques is essential for proper interpretation of neuroimaging data and advancing our knowledge of brain function in health and disease. As both technologies continue to evolve, they will undoubtedly remain indispensable tools in the neuroscientist's toolkit for mapping brain function through hemodynamic responses.

Electroencephalography (EEG) occupies a unique position in neuroimaging due to its ability to capture neural oscillations with millisecond-level temporal resolution, directly reflecting the brain's synchronous electrical activity. This unparalleled temporal precision enables researchers to observe brain dynamics as they unfold in real time, from rapid cognitive processes to pathological neuronal discharges. While other neuroimaging modalities like fMRI provide superior spatial localization and fNIRS offers greater portability, EEG remains unmatched for studying the chronometry of brain function [11] [12]. The electrophysiological basis of EEG lies in its capacity to record postsynaptic potentials from pyramidal neurons, which when synchronized across large neuronal populations, generate oscillations detectable at the scalp [13]. These neural oscillations, organized into characteristic frequency bands, form the fundamental language of brain communication and provide critical insights into brain health and function that complement other neuroimaging techniques.

Technical Comparison of Neuroimaging Modalities

Hierarchy of Temporal and Spatial Resolution

EEG's primary advantage in temporal resolution comes with inherent limitations in spatial resolution, creating a technological trade-off that researchers must navigate based on their specific investigative needs. The table below provides a systematic comparison of key neuroimaging techniques across critical performance parameters:

Table 1: Technical comparison of major neuroimaging modalities in neurological disorders research

Technique	Temporal Resolution	Spatial Resolution	Invasiveness	Portability	Key Clinical Applications
EEG	Milliseconds (<100 ms) [13]	Low (scalp-level) [11]	Non-invasive [13]	High (wearable systems available) [13]	Epilepsy monitoring, sleep studies, consciousness assessment, cognitive event-related potentials [13] [14]
fMRI	Seconds (1-2 s hemodynamic response) [11]	High (mm-level) [11]	Non-invasive	Low (requires immobilization) [11]	Pre-surgical mapping, network connectivity analysis, structural abnormalities
fNIRS	Seconds (∼5 s delay) [12]	Moderate (cm-level) [11] [12]	Non-invasive	Moderate [11] [12]	Functional localization, pediatric neuroimaging, rehabilitation monitoring [11]
MEG	Milliseconds [11]	Moderate [11]	Non-invasive	Low	Source localization of epileptogenic zones, cognitive processing studies
PET	Minutes [11]	High [11]	Minimally invasive (radiotracer injection) [11]	Low	Amyloid/tau imaging in Alzheimer's, metabolic activity mapping [15]
ECoG	Milliseconds [11]	High (direct cortical surface) [11]	Invasive (surgical implantation) [11]	Low	Refractory epilepsy evaluation, pre-surgical mapping

Complementary Strengths in Clinical Diagnostics

The distinct biophysical principles underlying each technique create complementary diagnostic value, particularly evident in clinical settings:

• EEG-fNIRS Integration: Combined systems leverage EEG's millisecond temporal precision with fNIRS's improved spatial localization of hemodynamic responses, overcoming individual limitations to provide insights into both cortical electrical activity and metabolic hemodynamics [11] [12]. This bimodal approach demonstrates superior mental state classification accuracy compared to unimodal systems across various paradigms including motor imagery, cognitive workload assessment, and clinical diagnosis [12].

• Diagnostic Contextualization: While blood-based biomarkers like neurofilament light protein (NfL) and phosphorylated tau provide molecular specificity for neurodegenerative processes [15], and amyloid-PET offers definitive pathological confirmation in Alzheimer's disease [15], EEG captures the functional consequences of these pathologies through altered neural synchrony and connectivity [16] [14].

Figure 1: Electrophysiological pathway from neuronal activity to detectable EEG oscillations

Experimental Evidence: Diagnostic Performance Across Neurological Disorders

EEG Biomarkers in Alzheimer's Disease and MCI

EEG spectral power and functional connectivity metrics provide quantifiable biomarkers for neurodegenerative conditions, offering predictive value for disease progression:

Table 2: EEG biomarkers in Alzheimer's disease continuum and diagnostic performance

EEG Biomarker	MCI/AD Findings	Predictive Value for AD Conversion	Comparison to Healthy Controls
Delta Power (1-4 Hz)	Significantly increased in prefrontal, parietal, temporal, and central regions [16]	Associated with disease severity and progression [16] [14]	Marked elevation in AD continuum [16]
Theta Power (4-7 Hz)	Both increased and decreased patterns with fewer electrodes [16]	Mixed findings across studies [16]	Less consistent than delta/alpha changes [16]
Alpha Power (8-13 Hz)	Significantly decreased across entire brain, particularly frontal lobe [16]	Strong predictor with decreased posterior alpha power [16] [14]	Robust reduction in MCI/AD [16] [14]
Functional Connectivity	Altered dynamic FC in delta and theta bands [16]	Prefrontal-parietal network hyperconnectivity [16]	Network disintegration and hyperconnectivity patterns [16]
Event-Related Potentials	Prolonged P300 and N200 latencies [14]	Reliable predictors of conversion from MCI to AD [14]	Delayed cognitive processing speed [14]

Multimodal Diagnostic Protocols

Standardized experimental protocols enable consistent biomarker quantification across research sites:

• Resting-State EEG Acquisition: Participants remain seated in a relaxed state with eyes open for 5 minutes, minimizing ocular and muscle artifacts [16]. EEG signals are typically digitized at 250-1000 Hz sampling rate with appropriate anti-aliasing filters [13].

• Spectral Power Analysis: Power spectral density is computed using Fast Fourier Transform (FFT) or Welch's method, with absolute or relative power quantified within standard frequency bands (delta: 1-4 Hz, theta: 4-7 Hz, alpha: 8-13 Hz, beta: 13-30 Hz, gamma: 30-100 Hz) [16] [17].

• Functional Connectivity Assessment: Phase-based synchronization metrics (coherence, phase-locking value) or information-theoretic measures (mutual information) quantify neural network interactions [16] [18]. Directed transfer function (DTF) analysis enables effective connectivity mapping with directionality information [16].

• Dual-Modality EEG-fNIRS Protocol: Simultaneous acquisition systems with integrated caps provide temporally aligned electrophysiological and hemodynamic data, requiring careful co-registration of EEG electrodes and fNIRS optodes [11]. Customized helmets using 3D printing or thermoplastic materials improve probe-scalp contact stability [11].

Figure 2: Experimental workflow for unimodal and multimodal neuroimaging protocols

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential research materials for EEG and multimodal neuroimaging studies

Item	Function/Purpose	Technical Specifications
High-Density EEG Systems	Recording electrical brain activity with optimal spatial sampling	64-256 channels; Ag/AgCl electrodes; impedance <5 kΩ; sampling rate ≥500 Hz [13]
fNIRS Integration Modules	Simultaneous hemodynamic monitoring with EEG	Near-infrared lasers (690-830 nm); photodetectors; light source-driver synchronizer [11]
Multimodal Acquisition Caps	Secure positioning of EEG electrodes and fNIRS optodes	Flexible fabric with customizable holder positions; co-registration with 10-20 system [11]
Electrode Gel/Saline Solution	Ensuring optimal electrode-scalp conductivity	Electrolyte chloride compounds; minimal evaporation properties; non-irritating formulation
Signal Processing Software	Preprocessing, feature extraction, and data fusion	MATLAB Toolboxes (EEGLAB, FieldTrip); Python (MNE-Python); commercial packages (BrainVision) [13]
Artifact Removal Tools	Identifying and eliminating non-neural signals	Independent Component Analysis (ICA); regression methods; advanced machine learning filters [13] [17]
Validation Phantoms	Testing and validating combined system performance	Tissue-simulating materials with known optical and electrical properties [11]

The electrophysiological basis of EEG provides an essential window into brain dynamics with millisecond temporal precision that remains unmatched by other neuroimaging modalities. While techniques like fMRI and fNIRS offer complementary strengths in spatial localization and hemodynamic monitoring, EEG's capacity to directly capture neural oscillations positions it as a fundamental tool for understanding brain function and dysfunction. The growing evidence supporting EEG biomarkers in neurological disorders, particularly when integrated with complementary modalities in bimodal systems, demonstrates the significant potential of combined approaches to improve diagnostic accuracy, disease monitoring, and therapeutic development. For researchers and clinical professionals, this integrated neuroimaging strategy promises deeper insights into brain network dynamics and more precise diagnostic capabilities across the spectrum of neurological disorders.

Functional neuroimaging technologies are indispensable in modern neuroscience and clinical diagnostics, yet each modality presents a unique set of strengths and limitations. Functional Magnetic Resonance Imaging (fMRI) and Functional Near-Infrared Spectroscopy (fNIRS) both measure hemodynamic responses correlated with neural activity but differ profoundly in their spatiotemporal capabilities and practical deployment. fMRI provides high spatial resolution and deep brain access, making it the gold standard for precise localization of neural events. In contrast, fNIRS offers superior portability, tolerance to motion, and higher temporal sampling at the cost of more superficial coverage and lower spatial definition. This guide objectively compares their performance parameters, experimental validation, and clinical utility to inform researcher selection for specific neurological investigations.

Fundamental Technical Principles and Measured Signals

Functional Magnetic Resonance Imaging (fMRI)

fMRI indirectly measures neural activity by detecting associated changes in blood flow and oxygenation. Its primary contrast mechanism is the Blood Oxygen Level Dependent (BOLD) signal, which reflects differences in magnetic susceptibility between oxygenated and deoxygenated hemoglobin [5]. When a brain region becomes active, a localized increase in blood flow occurs, leading to a greater concentration of oxygenated hemoglobin relative to oxygen consumption. This results in a measurable decrease in deoxygenated hemoglobin, a paramagnetic molecule that distorts the local magnetic field [19]. The BOLD signal is thus an indirect and complex proxy of neural activity, with changes typically lagging behind the underlying electrical events by 4–6 seconds [5].

Functional Near-Infrared Spectroscopy (fNIRS)

fNIRS is an optical neuroimaging technique that non-invasively measures hemodynamic changes in the cortex. It leverages the relative transparency of biological tissue (e.g., skin, skull, brain) to light in the near-infrared spectrum (650–950 nm), known as the "optical window" [4] [20]. fNIRS devices emit low-power near-infrared light into the scalp and detect the back-scattered light after it has passed through the tissue. The primary absorbing chromophores in this window are oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR), which have distinct absorption spectra [20]. By measuring the attenuation of light at multiple wavelengths, changes in HbO and HbR concentrations can be calculated using the modified Beer-Lambert law [4]. The resulting hemodynamic response is functionally analogous to the fMRI BOLD signal but provides separate, simultaneous measurements of oxy- and deoxy-hemoglobin dynamics [19] [21].

Quantitative Performance Comparison

The table below summarizes the core technical specifications and performance characteristics of fMRI and fNIRS, highlighting their inherent trade-offs.

Table 1: Technical and Performance Specifications of fMRI and fNIRS

Feature	fMRI	fNIRS
Spatial Resolution	High (Millimeter-level) ~1-3 mm [5]	Low (Centimeter-level) ~1-3 cm [5]
Temporal Resolution	Low (Seconds) Limited by hemodynamic response; sampling rate typically 0.33-2 Hz [5]	Moderate (Sub-second) Can achieve millisecond-level precision; typical sampling rate 2-10 Hz [20] [22]
Penetration Depth	Whole-brain Capable of imaging cortical and subcortical structures [5]	Superficial Cortex Limited to outer cortical layers (typically 1-3 cm) [5]
Measured Variables	BOLD signal (complex function of HbO, HbR, CBV) [19]	Direct concentration changes of HbO and HbR [4]
Portability & Cost	Low portability; high equipment and operational cost [4] [20]	High portability; relatively low-cost [4] [20]
Motion Tolerance	Highly sensitive to motion artifacts [4] [5]	Tolerant to moderate motion [20]
Experimental Environment	Restrictive scanner environment; loud noise [19]	Flexible; suitable for naturalistic, bedside, and real-world settings [4] [20]

Experimental Validation of Spatial and Temporal Correspondence

Spatial Correspondence Studies

Multimodal studies directly investigating the spatial overlap of activation detected by fMRI and fNIRS provide critical validation for the optical method.

• Motor and Visual Tasks: A 2024 study with 22 participants performing finger-tapping and visual checkerboard tasks found a 47.25% spatial overlap for within-subject analyses when comparing fNIRS-measured activation to the fMRI benchmark. This overlap increased to 68% in group-level analyses, demonstrating good spatial correspondence in superficial cortical regions [6].
• Motor Imagery and Execution: Research on asynchronous fMRI and fNIRS recordings during motor tasks confirmed that subject-specific fNIRS signals from motor regions could identify corresponding activation clusters in fMRI data. Significant peak activation was found overlapping the individually defined primary and premotor cortices [21].
• Factors Influencing Correspondence: The correlation between fNIRS and fMRI signals is influenced by the signal-to-noise ratio (SNR) of the fNIRS data and the distance from the scalp to the targeted brain region. Brain areas closer to the scalp typically show better correspondence [19].

Temporal Correspondence and Signal Characteristics

The temporal relationship between fNIRS chromophores and the BOLD signal is complex due to their different physiological underpinnings.

• Relationship to BOLD: The fMRI BOLD signal is most strongly correlated with changes in deoxygenated hemoglobin (HbR), as both are sensitive to the local concentration of this paramagnetic molecule [19] [21]. However, some studies report that the oxygenated hemoglobin (HbO) signal measured by fNIRS often has a higher SNR and can correlate more robustly with the BOLD response [19] [21].
• Comparative Temporal Fidelity: fNIRS provides a more direct measure of hemodynamic changes with a higher sampling rate, allowing it to better capture the temporal dynamics of the hemodynamic response function compared to the slower fMRI sampling rates [5].

Table 2: Experimental Evidence on fMRI-fNIRS Correspondence in Motor Tasks

Study Focus	Experimental Protocol	Key Findings on Correspondence
Spatial Overlap [6]	22 healthy adults; same-day fMRI & whole-head fNIRS during finger tapping and visual checkerboard tasks.	True Positive Rate (fNIRS vs fMRI): Up to 68% (group-level), 47.25% avg (within-subject). Positive Predictive Value: 51% (group-level), 41.5% (within-subject).
Spatial Interplay [21]	9 volunteers; asynchronous fMRI & fNIRS (NIRSport2, 54 channels) during motor imagery (MI) and motor action (MA).	fNIRS-based cortical signals (HbO, HbR, HbT) could identify corresponding fMRI activation clusters in primary (M1) and premotor (PMC) cortices. No significant difference between chromophores.
Temporal Correlation [19]	13 participants; simultaneous fMRI & NIRS during a battery of cognitive tasks (e.g., finger tapping, go/no-go).	NIRS signals were often highly correlated with fMRI, but correlation strength was dependent on fNIRS SNR and scalp-to-brain distance.

Clinical and Research Applications in Neurological Disorders

The complementary strengths of fMRI and fNIRS dictate their application across different clinical and research scenarios.

• fMRI in Clinical Research: fMRI is extensively used for mapping functional networks and identifying pathophysiological changes in deep brain structures across a wide range of disorders, including Alzheimer's disease, schizophrenia, and stroke [4] [5]. Its high spatial resolution is crucial for localizing epileptic foci and presurgical mapping of eloquent cortex.
• fNIRS in Clinical Monitoring and Rehabilitation: The portability and motion tolerance of fNIRS make it ideal for applications where fMRI is impractical.
  • Stroke Rehabilitation: fNIRS is used to monitor cortical activation changes in stroke patients before and after rehabilitation, such as detecting increased activation in the premotor cortex of the affected hemisphere following therapy [20].
  • Psychiatric Disorders: fNIRS has identified functional abnormalities, such as reduced brain activity and atypical functional connectivity within the prefrontal cortex, during verbal fluency tasks in individuals with schizophrenia [20].
  • Neurodevelopment: fNIRS is particularly valuable for studying object, face, and language processing in infants and children, including those with autism spectrum disorder, populations that are difficult to scan with fMRI due to motion restrictions [20].

Integrated and Multimodal Approaches

Given the spatiotemporal trade-offs, combining fMRI with fNIRS (and other modalities like EEG) offers a more comprehensive view of brain function [5].

• Synergistic Data Fusion: Integrated approaches use fMRI's high-resolution spatial maps to inform the source localization of fNIRS signals. Conversely, fNIRS can provide a higher temporal resolution sampling of the hemodynamic response within regions of interest identified by fMRI [5].
• Hyperscanning: fNIRS's portability enables "hyperscanning" paradigms, where multiple individuals' brain activities are recorded simultaneously during social interactions, a feat extremely challenging with fMRI [20] [5].
• Validation Paradigms: Simultaneous fMRI-fNIRS recordings serve as a critical method for validating the reliability and physiological basis of fNIRS signals against the established gold standard [5].

Essential Research Reagent Solutions

The table below lists key materials and tools essential for conducting experiments in this field.

Table 3: Key Research Reagents and Materials for fMRI and fNIRS Studies

Item	Function/Description	Example Use Case
fNIRS Optode Probe Set	A flexible cap or band holding light source emitters and detectors at specified distances (typically 3-4 cm).	Placed over the scalp region of interest (e.g., prefrontal or motor cortex) to measure cortical hemodynamics. [21]
MRI-Compatible fNIRS System	Specially designed fNIRS equipment with non-magnetic components to operate inside the MRI scanner without causing interference.	Enables simultaneous acquisition of fMRI and fNIRS data for direct temporal and spatial correlation. [19] [5]
Short-Distance Detectors (SDD)	fNIRS detectors placed close (~8 mm) to a source to preferentially sample signals from the scalp.	Used to measure and regress out systemic physiological noise (e.g., from scalp blood flow) from the cerebral fNIRS signal. [21]
Digitization System	A 3D stylus or camera system to record the precise spatial coordinates of fNIRS optodes/EEG electrodes on the scalp.	Crucial for coregistering fNIRS measurement channels with an individual's anatomical MRI scan for accurate spatial localization. [22]
Homer3 / BrainVoyager QX	Representative software packages for fNIRS (Homer3) and fMRI (BrainVoyager) data preprocessing and statistical analysis.	Standardized data processing pipelines (filtering, motion correction, GLM analysis) to ensure reproducibility. [21]

Signaling Pathways and Experimental Workflows

The Neurovascular Coupling Pathway

The following diagram illustrates the fundamental physiological process that underlies both fMRI and fNIRS signals, explaining their correlation despite different measurement techniques.

Diagram 1: Neurovascular Coupling Pathway

Multimodal Experimental Validation Workflow

This flowchart outlines a standard protocol for validating fNIRS measurements against the gold-standard fMRI, a common approach in methodological studies.

Diagram 2: Experimental Validation Workflow

The objective diagnosis and prognosis of major neurological disorders are being revolutionized by advanced neuroimaging techniques. Functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS) offer complementary windows into brain function, each with distinct strengths in clinical applications. This guide provides a comparative analysis of their performance against clinical gold standards and across key neurological domains, including disorders of consciousness (DoC), stroke, dementia, and movement disorders. Understanding the technical capabilities, validation status, and practical implementation of these tools is crucial for researchers and drug development professionals aiming to incorporate functional neuroimaging into biomarker development and therapeutic evaluation.

Comparative Performance of fMRI, EEG, and fNIRS

Table 1: Technical and Clinical Comparison of Neuroimaging Modalities

Feature	fMRI	EEG	fNIRS
Spatial Resolution	High (millimeters)	Low (centimeters)	Moderate (1-3 cm)
Temporal Resolution	Low (seconds)	High (milliseconds)	Moderate (seconds)
Portability	Low (fixed scanner)	High (mobile systems)	High (wearable systems)
Cost	Very High	Low to Moderate	Moderate
Tolerance to Motion	Low	Moderate	High
Primary Measured Signal	Blood-Oxygen-Level-Dependent (BOLD)	Electrical Potentials	Hemoglobin (HbO, HbR)
Key Clinical Strength	Biomarker validation, detailed localization	Rapid diagnosis, dynamic monitoring	Bedside, long-term monitoring

Table 2: Documented Performance Across Neurological Disorders

Disorder	Imaging Modality	Reported Performance	Clinical Context of Use
Dementia (MCI/AD)	fMRI (Complexity)	84% Accuracy, AUC 0.94 [23]	Classifying MCI from normal cognition
Dementia (MCI/AD)	Tau PET	Comparable Accuracy, AUC 0.92 [23]	Reference standard for tau pathology
Dementia (Multiple ND)	EEG (LDA Classifier)	Up to 100% Accuracy (AD vs. Control) [24]	Differentiating multiple neurological disorders
Stroke (Motor Outcome)	fNIRS (Connectivity Nomogram)	AUC 0.971 (Training), 0.804 (Validation) [25]	Predicting 3-month upper limb motor function
Disorders of Consciousness	fNIRS (Functional Connectivity)	76.92% Accuracy, AUC 0.818 [26]	Differentiating MCS from VS/UWS
Movement Disorders (PD Mortality)	EEG (LEAPD Algorithm)	100% Accuracy (LOOCV), 83% (Out-of-sample) [27]	Binary classification of 3-year mortality risk

Detailed Experimental Protocols and Methodologies

EEG for Multi-Class Neurological Disorder Classification

This protocol outlines the methodology for using EEG and machine learning to classify multiple neurological disorders, achieving high binary-classification accuracy [24].

• Data Acquisition: Resting-state EEG is recorded using a standard 19-electrode system (10-20 international placement) with a sampling rate of 500 Hz. The dataset includes participants from multiple groups: Alzheimer's disease (AD), mild cognitive impairment (MCI), schizophrenia, depression, and healthy controls (HC) [24].
• Preprocessing: Raw EEG signals are normalized to a 0-1 range. A 50 Hz notch filter is applied to remove mains interference, followed by a Butterworth bandpass filter (1-30 Hz) to isolate relevant neural frequencies [24].
• Feature Extraction & Selection: A broad set of features is extracted from the time, frequency, entropy, and complexity domains. The Least Absolute Shrinkage and Selection Operator (Lasso) algorithm is used for feature selection to identify the most discriminative EEG channels and features, reducing dimensionality and mitigating overfitting [24].
• Classification: Multiple machine learning classifiers (e.g., Linear Discriminant Analysis (LDA), Support Vector Machine (SVM)) are trained and evaluated. Performance is assessed for two-class (e.g., disease vs. disease, HC vs. disease), three-class, and four-class classification tasks using metrics such as accuracy [24].

fNIRS for Predicting Post-Stroke Motor Recovery

This protocol details the use of resting-state fNIRS functional connectivity to build a predictive model for upper limb motor recovery after ischemic stroke [25].

• Participants & Setting: Ischemic stroke patients are recruited during rehabilitation. A 5-minute resting-state fNIRS scan is performed with a multi-channel system (e.g., 106 leads) covering key motor and cognitive regions like the dorsolateral prefrontal cortex (DLPFC) and primary somatosensory motor cortex (PSMC) [25].
• Data Preprocessing: The raw fNIRS signal is converted to oxygenated (HbO), deoxygenated (HbR), and total hemoglobin (HbT) concentrations using the modified Beer-Lambert law. Motion artifacts are detected and corrected (e.g., with spline interpolation). A band-pass filter (0.01-0.1 Hz) is applied to remove physiological noise [25].
• Feature Engineering: Functional connectivity is calculated by determining the correlation between the time series of all possible pairs of channels or regions of interest (ROIs). The number of significant connectivity "edges" between specific brain regions (e.g., affected DLPFC to temporal lobe) is extracted as features [25].
• Model Building & Validation: Least Absolute Shrinkage and Selection Operator (LASSO) regression selects the most predictive connectivity features from a large initial pool. These features are used to build a logistic regression model, which can be presented as a clinical nomogram. The model is validated in a separate cohort, assessing discrimination via Area Under the Curve (AUC) and clinical utility via Decision Curve Analysis (DCA) [25].

fMRI Complexity for Classifying Cognitive Impairment

This protocol uses brain entropy mapping from resting-state fMRI as a non-invasive alternative to tau PET for classifying mild cognitive impairment (MCI) and Alzheimer's disease (AD) [23].

• Data Source: Using data from large, publicly available cohorts like the Alzheimer's Disease Neuroimaging Initiative (ADNI), researchers acquire resting-state fMRI scans from cognitively normal (CN) and cognitively impaired (MCI/AD) individuals [23].
• Complexity Feature Extraction: Sample entropy and multiscale entropy measures are computed from the fMRI time series to quantify the brain's signal complexity. This results in 3D brain entropy maps for each subject [23].
• Deep Learning Classification: A 3D Convolutional Neural Network (CNN) is trained on these entropy maps to distinguish between diagnostic groups. The model is trained and evaluated using robust methods like 5-fold cross-validation [23].
• Validation & Comparison: The classifier's performance is tested on an independent external validation cohort. Its classification accuracy, F1 score, and Area Under the Curve (AUC) are directly compared to a model based on tau PET imaging, the current reference standard for tau pathology [23].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Reagents and Solutions for Neuroimaging Research

Item	Primary Function	Example Application
High-Density EEG System	Records electrical brain activity with high temporal resolution.	Classification of neurological disorders; mortality prediction in Parkinson's disease [24] [27].
fNIRS System (e.g., 63-channel)	Measures cortical hemodynamics (HbO/HbR) non-invasively.	Bedside assessment of DoC; predicting motor recovery post-stroke [25] [26].
3T MRI Scanner	Acquires high-resolution structural and functional (BOLD) images.	Validation of fluid biomarkers; mapping disease progression in Alzheimer's [28] [23].
Tau PET Radiotracers (e.g., [18F]MK6240)	Binds to and quantifies tau neurofibrillary tangles in vivo.	Reference standard for Alzheimer's pathology staging and biomarker validation [28].
LASSO Regression Algorithm	Performs feature selection and regularization to prevent overfitting.	Identifying most predictive EEG or fNIRS features from high-dimensional datasets [24] [25].
GREENBEAN Checklist	Guidelines for reporting EEG biomarker validation studies.	Ensuring transparent and reproducible study design and reporting [29].

Integrated Discussion and Clinical Pathway

The presented data demonstrates a paradigm where neuroimaging modalities are selected based on the specific clinical question, context, and required balance between spatial/temporal resolution and practicality.

• fMRI excels as a reference tool for validating novel fluid biomarkers and providing detailed topographic mapping of pathology, as seen in the updated Alzheimer's diagnostic criteria that incorporate tau PET [28]. Its high spatial resolution is invaluable for delineating network-level dysfunction, though its cost and lack of portability are limitations.
• EEG offers an unparalleled combination of high temporal resolution and practicality for classifying a broad spectrum of neurological disorders and predicting functional outcomes, such as mortality in Parkinson's disease [24] [27]. Its main constraints are lower spatial resolution and sensitivity to artifacts.
• fNIRS has emerged as a powerful bedside tool for patient populations that are difficult to transport or scan with fMRI. Its utility in differentially diagnosing DoC and predicting long-term motor recovery after stroke highlights its growing role in personalized prognosis and treatment planning in rehabilitation settings [30] [25] [26].

The convergence of these technologies with standardized reporting guidelines [29] and advanced machine learning is creating a new era of objective, biomarker-driven neurology. Future developments will likely focus on the integration of multi-modal data (e.g., EEG-fNIRS) to provide a more comprehensive picture of brain function and accelerate therapeutic development for complex neurological conditions.

From Theory to Practice: Methodological Implementations and Diagnostic Applications

Functional magnetic resonance imaging (fMRI) has become a cornerstone of non-invasive brain research, providing invaluable insights into the neural mechanisms underlying neurological disorders. Two primary methodological approaches—resting-state fMRI (rs-fMRI) and task-based fMRI—enable researchers to investigate brain network organization and dysfunction. In the clinical research domains of stroke and Alzheimer's disease (AD), these protocols offer distinct advantages and face unique challenges concerning diagnostic accuracy, prognostic value, and practicality for specific patient populations [31] [32] [33].

Resting-state fMRI measures spontaneous, low-frequency fluctuations in the blood-oxygen-level-dependent (BOLD) signal while participants lie motionless in the scanner without performing any structured task. This approach identifies functionally connected brain regions that form intrinsic networks, such as the default mode network (DMN), which is crucial for understanding AD pathology [32] [34]. Conversely, task-based fMRI detects BOLD signal changes during the performance of specific cognitive, motor, or sensory paradigms, allowing researchers to map brain activation patterns associated with particular functions [35]. The complementary nature of these approaches provides a more comprehensive understanding of brain function in neurological conditions.

This comparison guide objectively evaluates the performance of resting-state and task-based fMRI protocols within the context of stroke and Alzheimer's disease research. We examine experimental data, detailed methodologies, and practical considerations to inform researchers and drug development professionals about optimal protocol selection for specific clinical research objectives.

Comparative Performance in Neurological Disorders

Table 1: Protocol Performance Comparison in Stroke Research

Performance Metric	Resting-State fMRI	Task-Based fMRI
Patient Applicability	Suitable for patients with severe impairment or inability to follow commands [31]	Limited to patients with sufficient cognitive/motor capacity for task performance [31]
Key Biomarkers in Stroke	Interhemispheric M1 connectivity; Ipsilesional M1 connectivity with contralesional thalamus/SMA [31]	Cortical activation shifts; Recruitment of additional regions; Performance-dependent activation [31]
Temporal Resolution	Limited by hemodynamic response (4-6s lag) [5]	Limited by hemodynamic response (4-6s lag) [5]
Spatial Resolution	High (millimeter level) [5]	High (millimeter level) [5]
Prognostic Value	Functional connectivity of ipsilesional M1 at onset correlates with 6-month motor recovery [31]	Variable due to performance confounds; Dependent on task selection [31]
Longitudinal Stability	High consistency across sessions in stable patients [31]	Variable due to performance improvement during recovery [31]
Data Interpretation	Based on correlation metrics of spontaneous activity [33]	Compares activation during task vs. baseline conditions [35]

Table 2: Protocol Performance Comparison in Alzheimer's Disease Research

Performance Metric	Resting-State fMRI	Task-Based fMRI
Patient Applicability	Suitable across disease stages; Minimal performance demands [32]	Challenging in moderate-severe AD due to cognitive demands [32]
Key Biomarkers in AD	DMN connectivity; Hippocampal-PMC connectivity [32] [34]	Task-induced deactivations; Encoding/retrieval activation patterns [32] [34]
Effect Size for Risk Detection	Effect size of 3.35 for distinguishing AD risk groups [32]	Effect size of 1.39 for distinguishing AD risk groups [32]
Sensitivity to Early Pathology	Detects connectivity changes in preclinical stages [34]	Variable sensitivity; dependent on task engagement [34]
Relationship to Pathology	Hypoconnectivity within PMC related to Aβ accumulation [34]	Suppression of deactivation related to amyloid burden [32]
APOE4 Modulation	Connectivity changes differ by APOE4 status [34]	Encoding-associated deactivations differ by APOE4 status [32]

Experimental Protocols and Methodologies

Resting-State fMRI Protocols

Stroke Motor Recovery Protocol A longitudinal resting-state fMRI study investigating motor recovery after stroke implemented the following methodology [31]:

• Participants: 12 first-ever ischemic stroke patients with moderate to severe motor deficits and 11 age-matched healthy controls
• Scanning Schedule: Four sessions over 6 months (within 2 weeks post-stroke, 1 month, 3 months, and 6 months)
• Image Acquisition: 3T Philips scanner, T2*-weighted gradient echo EPI sequence (TR=3000ms, TE=35ms, 35 slices, slice thickness=4mm, matrix size=128×128, FOV=220×220mm)
• Resting-State Parameters: 100 whole-brain volumes collected over 5 minutes; participants instructed to keep eyes closed, remain motionless, and not fall asleep
• Data Preprocessing: Spatial realignment, normalization to MNI space, band-pass filtering (0.01-0.08 Hz), spatial smoothing (8mm FWHM Gaussian kernel)
• Functional Connectivity Analysis: Seed-based correlation using ipsilesional primary motor cortex (M1) as reference region; correlation maps converted to z-scores using Fisher's transformation
• Clinical Correlation: Functional connectivity measures correlated with Fugl-Meyer Assessment (FMA) scores at each time point

Alzheimer's Disease Default Mode Network Protocol A study comparing individuals at high and low risk for AD employed this resting-state protocol [32]:

• Participants: 17 cognitively normal individuals with family history of AD and at least one APOE4 allele compared to 12 low-risk controls
• Image Acquisition: Resting-state BOLD fMRI during eyes-closed rest
• DMN Analysis: Identified regions demonstrating synchronous low-frequency fluctuations characteristic of the default mode network
• Statistical Analysis: Compared DMN connectivity strength between risk groups, calculating effect sizes for group discrimination

Task-Based fMRI Protocols

Stroke Executive Function Protocol A functional near-infrared spectroscopy (fNIRS) study examining post-stroke executive dysfunction implemented this task-based protocol [35]:

• Participants: 20 post-stroke executive dysfunction patients and 20 healthy controls
• Tasks: Stroop task (inhibitory control) and 1-back task (working memory)
• Data Acquisition: Measured oxygenated hemoglobin concentration signals from bilateral inferior parietal lobule, pre-motor area, dorsolateral prefrontal cortex, Broca's area, and frontopolar cortex
• Analysis: Compared task-based functional connectivity and cortical activation between groups during task performance
• Clinical Correlation: Analyzed correlations between Montreal Cognitive Assessment (MoCA) scores, task performance, and cortical activation patterns

Alzheimer's Episodic Memory Protocol A longitudinal study investigating functional connectivity changes during memory tasks used this protocol [34]:

• Participants: 152 cognitively unimpaired older adults from the PREVENT-AD cohort
• Tasks: Object-location episodic memory encoding and retrieval
• Encoding Task: Participants viewed 48 objects on left or right side of screen, indicating side via button press
• Retrieval Task: 20 minutes after encoding, participants viewed 48 old and 48 new objects, providing forced-choice retrieval responses (familiar, remembered left, remembered right, or new)
• Image Acquisition: fMRI during rest, encoding, and retrieval conditions
• Analysis: Examined functional connectivity within medial temporal lobe, within posteromedial cortex, and between MTL-PMC across task conditions
• Longitudinal Assessment: Repeated assessments over up to four years with correlation to subsequent amyloid- and tau-PET burden

Experimental Workflows and Signaling Pathways

Figure 1: fMRI Experimental Workflow Comparison for Network Analysis

Figure 2: Neural Signaling Pathways in fMRI Protocols

The Scientist's Toolkit: Essential Research Materials

Table 3: Key Research Reagents and Solutions for fMRI Network Analysis

Tool/Category	Specific Examples	Function/Purpose
Data Acquisition Hardware	3T MRI Scanner (e.g., Philips Achieva, GE Discovery)	High-field magnetic resonance imaging for BOLD signal detection [31] [36]
Analysis Software Platforms	SPM8, FSL, AFNI, DPABI, DPARSF	fMRI data preprocessing, statistical analysis, and visualization [31] [36] [37]
Brain Parcellation Atlases	AAL3, Brainnetome Atlas, Shen 268 ROI Atlas	Standardized brain region definition for network node construction [38] [34] [37]
Functional Connectivity Tools	Pearson Correlation, MVPA, SNBG Method	Quantifying temporal dependencies between brain regions [38] [37]
Clinical Assessment Tools	Fugl-Meyer Assessment (FMA), MoCA, MMSE	Standardized clinical evaluation for correlation with imaging biomarkers [31] [35]
Task Paradigm Software	E-Prime, Presentation, PsychoPy	Precisely controlled stimulus delivery and response recording [35] [34]
Motion Correction Tools	Realignment algorithms, Friston 24-parameter model	Minimizing confounding effects of head motion [36] [37]

Integrated Discussion

The comparative analysis of resting-state and task-based fMRI protocols reveals a complementary relationship rather than a competitive one in neurological disorders research. For stroke populations, resting-state fMRI offers particular advantages in acute and severe cases where patient compliance with task demands is challenging [31]. The ability to detect interhemispheric connectivity changes, particularly between primary motor cortices, provides valuable prognostic information that correlates with long-term motor recovery [31] [33]. In contrast, task-based protocols reveal specific patterns of cortical reorganization during recovery, including both compensatory recruitment and maladaptive activation [35].

In Alzheimer's disease research, resting-state fMRI demonstrates superior effect sizes for distinguishing at-risk populations (effect size 3.35 for resting-state vs. 1.39 for task-based) [32]. The detection of DMN alterations, particularly reduced connectivity in the posteromedial cortex, provides an early biomarker of pathology that precedes clinical symptoms [32] [34]. Task-based approaches, however, reveal important dynamic processes such as failure to suppress DMN activity during memory encoding, which correlates with amyloid burden and memory performance [32] [34].

The emerging paradigm in clinical neuroscience recognizes that combined multi-modal approaches, integrating both resting-state and task-based fMRI with complementary techniques like fNIRS and EEG, provide the most comprehensive assessment of brain network integrity [11] [35] [5]. This integrated approach is particularly valuable for drug development, where sensitive biomarkers are needed to detect subtle treatment effects in clinical trials.

For researchers designing neuroimaging studies in stroke or Alzheimer's disease, protocol selection should be guided by specific research questions, patient characteristics, and clinical objectives. Resting-state protocols offer practical advantages for severely impaired populations and provide robust network-level biomarkers, while task-based approaches enable detailed investigation of specific cognitive and motor functions through carefully designed behavioral paradigms.

Quantitative electroencephalography (qEEG) represents a modern evolution of traditional EEG, involving the computational processing of digital EEG signals to extract objective metrics that characterize brain activity. By applying sophisticated mathematical algorithms, qEEG enables precise quantification of neural dynamics that are difficult to assess through visual inspection alone [39]. In clinical neurology and drug development, two qEEG biomarkers have demonstrated particular significance for prognostic applications: Power Spectral Density (PSD) and the Brain Symmetry Index (BSI). These biomarkers provide non-invasive, real-time windows into brain function, offering valuable insights for outcome prediction across a spectrum of neurological conditions including stroke, disorders of consciousness, and neurodegenerative diseases [40] [39].

The growing importance of qEEG biomarkers stems from their ability to address critical limitations in conventional neuroimaging. While techniques like fMRI provide exceptional spatial resolution, they lack the temporal precision to capture rapid neural dynamics and are often impractical for continuous monitoring [41] [5]. qEEG bridges this gap by offering millisecond-level temporal resolution at a fraction of the cost, making it particularly suitable for longitudinal tracking of disease progression and therapeutic response [40]. Furthermore, as the healthcare landscape increasingly emphasizes precision medicine, objective electrophysiological biomarkers like PSD and BSI provide quantifiable endpoints for clinical trials and individualized treatment planning.

Analytical Foundations of Key qEEG Biomarkers

Power Spectral Density (PSD) Analysis

Power Spectral Density analysis applies computational techniques, typically Fast Fourier Transform (FFT), to decompose the complex EEG signal from the time domain into its constituent frequency components. This process quantifies the oscillatory power distributed across conventional frequency bands: delta (δ, 1-4 Hz), theta (θ, 4-8 Hz), alpha (α, 8-13 Hz), beta (β, 13-30 Hz), and gamma (γ, >30 Hz) [40]. The resulting metrics can be expressed as absolute power (the actual magnitude within a specific band) or relative power (the proportion of power in a particular band relative to the total power spectrum) [40].

In pathological states, characteristic shifts in spectral power emerge. The phenomenon of "EEG slowing," marked by increased power in lower frequencies (delta, theta) and decreased power in higher frequencies (alpha, beta), represents a robust indicator of neurological dysfunction [40] [42]. For instance, in Alzheimer's disease and mild cognitive impairment (MCI), researchers have consistently observed increased theta and delta power alongside decreased alpha and beta power [40]. These alterations likely reflect disruptions in thalamocortical oscillatory coordination and diminished synaptic efficiency due to underlying neurodegeneration [40].

Brain Symmetry Index (BSI)

The Brain Symmetry Index quantifies interhemispheric asymmetry in brain electrical activity by calculating the mean absolute difference in power between homologous electrode pairs across the two hemispheres [43] [39]. Originally developed for monitoring cerebral ischemia during carotid surgery, BSI has evolved into a valuable prognostic tool for unilateral brain injuries, particularly stroke [39]. The formula for BSI, as described by van Putten and Tavy (2004), represents "the mean of the absolute value of the difference in mean hemispheric power in a frequency range from 1 to 25 Hz" [39].

BSI values range from 0 to 1, where 0 indicates perfect symmetry between hemispheres and 1 indicates maximal asymmetry [43]. Clinically, elevated BSI values reflect significant interhemispheric imbalances due to focal pathology. Research has demonstrated that BSI can be calculated in specific frequency ranges (BSIslow for 1-7 Hz and BSIfast for 7-25 Hz) and for specific electrode pairs, each with potential differential diagnostic and prognostic implications [43].

Complementary Biomarkers and Advanced Analyses

Beyond PSD and BSI, the qEEG arsenal includes other valuable analytical approaches:

Functional Connectivity measures the temporal correlation or coherence between signals from different brain regions, providing insights into network integrity [44] [40]. In disorders of consciousness, stronger global functional connectivity has been associated with better long-term outcomes [44].

Non-linear Measures including entropy analyses quantify the complexity and irregularity of EEG signals [40]. Reduced entropy values, indicating decreased signal complexity, have been observed in Alzheimer's disease and MCI, potentially serving as early diagnostic markers [40].

Global Field Power (GFP) measures the spatial standard deviation of voltage values across all electrodes, representing the instantaneous strength and synchronization of global brain activity [40].

Table 1: Key qEEG Biomarkers and Their Clinical Significance

Biomarker	Analytical Principle	Physiological Correlation	Prognostic Value
Power Spectral Density	Quantifies power distribution across frequency bands	Synaptic activity efficiency, thalamocortical coordination	EEG slowing (↑δ/θ, ↓α/β) predicts poorer outcomes [40] [42]
Brain Symmetry Index	Measures interhemispheric power asymmetry	Unilateral cortical dysfunction, ischemic damage	Higher values predict mortality in hemispheric stroke [43] [39]
Functional Connectivity	Assesses coherence between brain regions	Network integrity, information sharing	Stronger connectivity associates with better recovery in DoC [44]
Entropy Measures	Quantifies signal complexity/irregularity	Neural complexity, network efficiency	Reduced values in AD/MCI may serve as early marker [40]

Experimental Protocols and Methodologies

Standardized qEEG Data Acquisition

Implementing qEEG biomarkers in research and clinical trials requires meticulous attention to acquisition protocols. The technical foundation begins with proper EEG recording using Ag/AgCl electrodes positioned according to the international 10-20 system, with impedance typically maintained below 5-10 kΩ [43] [44]. The sampling rate should exceed 250 Hz (often 500 Hz or higher) to adequately capture neural dynamics, with appropriate bandpass filtering (e.g., 0.5-70 Hz) to remove non-physiological artifacts [43].

For resting-state qEEG analyses, researchers typically collect at least 20-30 minutes of continuous EEG data in a wakeful, relaxed state with eyes closed [44]. Artifact management represents a critical step, employing both automated algorithms and manual review to identify and remove segments contaminated by muscle activity, eye movements, or environmental interference [43]. In clinical populations with limited cooperation, such as severe stroke or disorders of consciousness, shorter recording periods (e.g., 10 minutes of clean data) may be acceptable if data quality is maintained [42].

Quantitative Processing Pipelines

Following data acquisition, standardized processing pipelines ensure reproducible biomarker extraction:

Preprocessing involves importing raw EEG data, applying re-referencing (often to average reference), downsampling if appropriate, and bandpass filtering (typically 0.5-30 Hz for most clinical applications) [43] [42]. Contemporary approaches frequently employ Independent Component Analysis (ICA) to identify and remove stereotypical artifacts such as blinks, eye movements, and cardiac interference [42].

Spectral Analysis utilizes Fast Fourier Transform (FFT) or similar algorithms to convert preprocessed time-domain signals into frequency-domain power distributions. For PSD calculation, EEG data is typically segmented into epochs (e.g., 2-4 seconds) with 50% overlap, and the Welch method is applied to reduce variance in power estimates [43].

BSI Calculation involves computing the absolute power difference between homologous electrode pairs (e.g., C3-C4, O1-O2) across specified frequency ranges, then averaging these differences across all electrode pairs to derive a global asymmetry index [43]. Research indicates that frequency-specific BSI (e.g., BSIfast for 7-25 Hz) may have particular prognostic value in certain conditions [43].

Diagram 1: qEEG Biomarker Processing Workflow. This workflow outlines the standardized pipeline from raw EEG acquisition to clinically actionable biomarkers, highlighting critical preprocessing steps that ensure data quality.

Prognostic Performance Across Neurological Disorders

Stroke and Large Hemispheric Infarction

In stroke populations, particularly large hemispheric infarction (LHI), both PSD and BSI demonstrate significant prognostic value. A 2022 prospective study of 38 LHI patients found that BSIfast calculated for the C3-C4 electrode pair (BSIfastC3-C4) independently predicted 3-month mortality with an area under the curve (AUC) of 0.805 [43]. Multivariable analysis confirmed BSIfastC3-C4 as an independent predictor (OR=1.059, 95% CI 1.003-1.119, p=0.039), with predictive power further enhanced when combined with Glasgow Coma Scale and infarct volume (AUC=0.840, p=0.002) [43].

Recent research has also elucidated distinct qEEG signatures based on infarct topography. A 2025 study comparing large hemispheric infarction (LHI) and brainstem infarction (BSI) revealed that BSI patients exhibited significantly elevated delta-band absolute and relative power alongside attenuated alpha/beta power compared to LHI patients [42]. Notably, in the non-ipsilesional occipital region, enhanced delta/beta activity demonstrated positive correlations with favorable clinical outcomes, while increased theta/alpha activity showed inverse prognostic associations [42].

Disorders of Consciousness (DoC)

For patients with disorders of consciousness, qEEG biomarkers provide valuable prognostic information beyond standard clinical assessment. A 2025 retrospective study of 97 DoC patients found that visual EEG assessment showed moderate predictive accuracy for survival (AUC=0.77), while qEEG-based models demonstrated comparable performance with slightly higher (though not statistically significant) AUC values [45]. Most impressively, combining qEEG features with clinical prognostic factors significantly improved predictive accuracy, particularly for neurological recovery (AUC improved from 0.729 to 0.936; p<0.001) [45].

Specific qEEG patterns have emerged as favorable prognostic indicators in DoC populations. Patients with reactive EEG signals to external stimuli, preserved higher-frequency bands (alpha and beta), and stronger global functional connectivity are more likely to experience positive outcomes 3-6 months post-injury [44]. Machine learning approaches applying discriminant analyses to EEG-based functional connectivity and dominant frequency have achieved accuracies of 83.3% for predicting clinical outcome in nontraumatic patients and 80% in traumatic patients [44].

Alzheimer's Disease and Mild Cognitive Impairment

In neurodegenerative conditions, qEEG biomarkers detect early functional alterations often preceding structural changes. Alzheimer's disease patients consistently demonstrate a "slowing" of EEG activity, characterized by increased power in slow-frequency bands (delta and theta) and decreased power in high-frequency bands (alpha and beta) [40]. Specific relative power ratios, particularly the (δ+θ)/(α+β) ratio (DTABR), have shown significant utility in differentiating AD from normal aging and detecting early risk [40].

For mild cognitive impairment (MCI) patients, the most prominent EEG changes include decreased beta power and increased delta and theta power, most pronounced in temporal regions [40]. These spectral alterations likely reflect early disruptions in thalamocortical oscillatory coordination and synaptic efficiency, potentially driven by Aβ/tau-mediated neurodegeneration [40].

Table 2: Prognostic Performance of qEEG Biomarkers Across Neurological Conditions

Condition	Biomarker	Predictive Value	Performance Metrics	Reference
Large Hemispheric Infarction	BSIfastC3-C4	3-month mortality	AUC=0.805; OR=1.059 (1.003-1.119)	[43]
Large Hemispheric Infarction	BSIfastC3-C4 + GCS + infarct volume	3-month mortality	AUC=0.840	[43]
Disorders of Consciousness	qEEG + clinical factors	Neurological recovery	AUC=0.936	[45]
Nontraumatic DoC	Functional connectivity	Clinical outcome	Accuracy=83.3%, Sensitivity=92.3%, Specificity=60%	[44]
Traumatic DoC	Functional connectivity + dominant frequency	Clinical outcome	Accuracy=80%, Sensitivity=85.7%, Specificity=71.4%	[44]

Comparative Diagnostic Accuracy: qEEG Versus Other Neuroimaging Modalities

qEEG Versus Structural and Functional MRI

When evaluating diagnostic and prognostic capabilities, qEEG offers distinct advantages and limitations compared to established neuroimaging techniques like fMRI. While fMRI provides unparalleled spatial resolution (millimeter-level) and access to subcortical structures, its temporal resolution is constrained by the hemodynamic response, typically lagging 4-6 seconds behind neural activity with sampling rates generally between 0.33-2 Hz [5]. qEEG, in contrast, captures neural dynamics at millisecond temporal resolution but with more limited spatial resolution (centimeter-level) and restricted to cortical surfaces [41] [40].

From a practical standpoint, qEEG systems offer significant advantages in cost-effectiveness, portability, and tolerance to movement artifacts, enabling brain monitoring in naturalistic settings and at the bedside [40]. These characteristics make qEEG particularly suitable for longitudinal monitoring, critically ill patients, and populations challenged to remain motionless (e.g., children, cognitively impaired individuals) [5] [40].

qEEG Versus fNIRS

Functional near-infrared spectroscopy (fNIRS) shares qEEG's advantages of portability and motion tolerance but measures hemodynamic responses similar to fMRI rather than direct electrical activity [41]. Recent technological advances, particularly time-domain fNIRS (TD-fNIRS), have improved sensitivity to brain activations, with one 2025 study demonstrating impressive classification of mild cognitive impairment (AUC=0.92) when combining neural metrics with behavioral and self-report features [46].

The complementary nature of EEG and fNIRS has prompted growing interest in multimodal integration approaches [41]. Simultaneous EEG-fNIRS recording capitalizes on EEG's superior temporal resolution and fNIRS's better spatial localization, providing a more comprehensive picture of brain function by capturing both electrical activity and hemodynamic responses [41]. This integration is physiologically grounded in the neurovascular coupling phenomenon, where neural activity is accompanied by hemodynamic fluctuations delivering oxygen and nutrients to activated regions [41].

Diagram 2: Comparative Characteristics of Neuroimaging Modalities. This comparison highlights the complementary strengths and limitations of different neuroimaging approaches, illustrating qEEG's distinct advantage in temporal resolution and cost-effectiveness.

Table 3: Essential Research Resources for qEEG Biomarker Studies

Resource Category	Specific Examples	Function & Application	Technical Notes
EEG Acquisition Systems	Nicolet Monitor (Natus), SOLAR1848, clinical-grade digital EEG systems	High-quality signal acquisition with sufficient sampling rates (≥250 Hz)	Ensure compatibility with analysis software; verify amplifier specifications [43] [42]
Analysis Software Platforms	MATLAB with EEGLAB toolbox, commercial qEEG software	Signal processing, spectral analysis, biomarker calculation	EEGLAB provides open-source environment for custom analysis pipelines [43] [42]
Standardized Electrode Arrays	10-20 system Ag/AgCl electrodes, high-density caps	Consistent spatial sampling across subjects	Maintain impedance <5-10 kΩ; consider specific montages for different conditions [43] [44]
Artifact Processing Tools	Independent Component Analysis (ICA), automated rejection algorithms	Identification and removal of non-neural signals	Combine automated and manual review for optimal artifact management [43] [42]
Clinical Assessment Tools	Glasgow Coma Scale, NIH Stroke Scale, Coma Recovery Scale-Revised	Clinical correlation and validation of biomarker findings	Standardized assessments essential for prognostic model development [43] [45]
Normative Databases	Age-matched healthy control data, disease-specific reference values	Benchmarking and z-score calculation	Consider population-specific norms; account for age-related changes [40]

The accumulating evidence firmly establishes qEEG biomarkers, particularly Power Spectral Density and Brain Symmetry Index, as valuable tools for neurological prognosis across diverse clinical conditions. Their non-invasive nature, cost-effectiveness, and real-time monitoring capabilities position them as practical alternatives or complements to traditional neuroimaging in both research and clinical settings. The demonstrated prognostic accuracy of these biomarkers—especially when integrated with clinical factors—underscores their potential to enhance predictive modeling and support therapeutic decision-making.

Future advancements in qEEG biomarkers will likely focus on several key areas: standardized acquisition and processing protocols to improve reproducibility, multimodal integration with complementary techniques like fNIRS and fMRI, and the application of advanced analytics including machine learning to extract increasingly sophisticated prognostic patterns from EEG data [41] [5] [40]. As these developments mature, qEEG biomarkers hold promise for transforming neurological care through objective, accessible, and dynamic assessment of brain function and recovery potential.

Functional near-infrared spectroscopy (fNIRS) is emerging as a powerful neuroimaging tool for studying prefrontal cortex (PFC) function in naturalistic settings where traditional modalities face limitations. This comparison guide objectively analyzes fNIRS performance against functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) for monitoring PFC dynamics in Parkinson's disease (PD) and psychiatric disorders. We synthesize experimental data demonstrating fNIRS's unique capability to capture cortical activation patterns during complex motor-cognitive tasks in PD and its diagnostic accuracy in differentiating psychiatric conditions. The portability, motion tolerance, and cost-effectiveness of fNIRS enable research paradigms closer to real-world conditions, providing complementary insights to fMRI and EEG while addressing different methodological constraints.

The quest for objective biomarkers in neurological and psychiatric disorders has driven extensive utilization of neuroimaging technologies in both research and clinical settings. Functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS) each offer distinct capabilities and face specific limitations for probing brain function.

fMRI is considered the gold standard for localized brain activation mapping through blood-oxygen-level-dependent (BOLD) signals, providing high spatial resolution but requiring stringent motion restrictions [47]. EEG measures electrical brain activity with millisecond temporal resolution but limited spatial accuracy [48]. fNIRS occupies a unique middle ground, measuring hemodynamic responses through near-infrared light with better spatial resolution than EEG and greater tolerance for movement than fMRI [47] [48].

For disorders such as Parkinson's disease that involve dynamic motor symptoms, and psychiatric conditions that manifest during complex cognitive tasks, fNIRS offers particular advantages for naturalistic assessment. This guide systematically compares these modalities with experimental data supporting their relative strengths and limitations in clinical research applications.

Technical Comparison of Neuroimaging Modalities

Measurement Principles and Performance Characteristics

Table 1: Technical comparison of fMRI, fNIRS, and EEG

Feature	fMRI	fNIRS	EEG
What it Measures	BOLD signal (magnetic properties of hemoglobin)	Hemodynamic response (oxygenated/deoxygenated hemoglobin)	Electrical potentials from neuronal firing
Temporal Resolution	Low (seconds)	Moderate (seconds)	High (milliseconds)
Spatial Resolution	High (millimeter-level)	Moderate (centimeter-level, cortical only)	Low (centimeter-level)
Depth of Measurement	Whole brain	Outer cortex (1-2.5 cm)	Cortical surface
Sensitivity to Motion	High	Low to moderate	High
Portability	Low (fixed scanner)	High (wearable systems)	High (wearable systems)
Naturalistic Setting Suitability	Low	High	Moderate
Cost	High	Moderate	Low to moderate
Best Use Cases	Localizing precise activation foci; deep brain structures	Naturalistic studies; clinical populations; children	Fast cognitive tasks; sleep studies; brain-computer interfaces

fNIRS detects neural activity indirectly through neurovascular coupling, where active brain regions experience increased blood flow and oxygen consumption [47]. It utilizes near-infrared light (650-1000 nm) which penetrates biological tissues and is absorbed primarily by oxygenated (HbO) and deoxygenated hemoglobin (HbR) at different rates [49]. The fNIRS signal therefore reflects changes in hemoglobin concentration, with active cortical regions typically showing increased HbO and decreased HbR [49].

Unlike fMRI which measures the same hemodynamic response but requires a constrained environment, fNIRS is minimally restrictive and significantly more tolerant of subject movement [47]. While EEG provides direct measurement of neural electrical activity with superior temporal resolution, it struggles with spatial localization due to the skull's dampening effect on electrical signals [48].

Comparative Advantages and Limitations

fNIRS advantages include its portability for real-world applications, relative affordability compared to fMRI, tolerance for movement making it suitable for clinical populations and children, silent operation avoiding auditory interference, and ability to measure both HbO and HbR simultaneously [47] [50]. Recent methodological advances exploiting the dual nature of fNIRS signals through Riemannian geometry have further improved brain-state classification accuracy [50].

Key limitations of fNIRS include restricted penetration depth (surface cortex only), limited spatial resolution compared to fMRI due to light scattering, and lack of inherent anatomical information requiring co-registration with structural imaging [47]. Additionally, fNIRS signals can be contaminated by systemic physiological artifacts such as blood pressure changes, heart rate, and scalp blood flow [51].

Complementary use of these modalities is increasingly common, with simultaneous fNIRS-fMRI measurements validating fNIRS signals and fNIRS-EEG combinations providing both hemodynamic and electrophysiological information [51] [48].

fNIRS Experimental Protocols and Methodologies

Standardized Experimental Paradigms

fNIRS Experimental Workflow

Verbal Fluency Task (VFT) has been extensively used in psychiatric research. Participants generate words belonging to a specific category (e.g., words beginning with a certain letter) within a time limit, typically 60 seconds. This paradigm engages executive functions and language processing, primarily activating frontal and temporal regions [52]. The standardized protocol includes pre-task baseline (30 seconds), task period (60 seconds), and post-task rest period [52].

Stroop Task evaluates executive function and cognitive control through color-word interference. Participants must name the color of ink in which conflicting color words are printed (e.g., the word "RED" printed in blue ink). The incongruence between word meaning and ink color creates cognitive conflict requiring inhibitory control [53]. This paradigm effectively engages the dorsolateral prefrontal cortex (DLPFC) and is particularly sensitive to fronto-executive deficits in Parkinson's disease [53].

Dual-Task Paradigms assess cognitive-motor integration by requiring simultaneous performance of motor and cognitive tasks. For Parkinson's research, this typically involves walking or marching while performing cognitive tasks like arithmetic calculations [54]. These paradigms reveal the compensatory prefrontal mechanisms employed when automatic motor control is impaired.

Naturalistic Tasks include ecologically valid activities such as video game play (e.g., Dance Dance Revolution) that engage multiple cognitive domains in realistic contexts [51]. Such paradigms leverage fNIRS's tolerance for movement and portability, enabling measurement of brain activation during full-body movements not feasible in fMRI environments.

fNIRS Signal Acquisition and Processing

fNIRS systems consist of light sources (emitting at typically 760 nm and 850 nm) and detectors arranged on the scalp surface with specific source-detector distances (typically 3 cm) to probe cortical regions [49]. Raw light intensity measurements are converted to oxygenated and deoxygenated hemoglobin concentration changes using the modified Beer-Lambert law [49].

Standard preprocessing includes filtering to remove physiological noise (cardiac pulsation ~1 Hz, respiration ~0.3 Hz, Mayer waves ~0.1 Hz), motion artifact correction, and baseline correction. For block designs, the hemodynamic response is typically averaged across trials, and statistical comparisons are made between active and rest conditions [51] [52].

Advanced analytical approaches include functional connectivity analysis examining correlations between different cortical regions [53], and machine learning classification of brain states using features derived from the hemodynamic response [55] [50].

fNIRS Applications in Parkinson's Disease Research

Prefrontal Cortex Activation Patterns in PD

Table 2: fNIRS studies in Parkinson's disease

Study Design	Participants	Key Findings	Clinical Implications
Dual-task marching with arithmetic [54]	58 PD patients, 42 healthy controls	PD patients showed widespread PFC activation during single tasks with no significant increase during dual tasks; healthy controls showed selective PFC activation during single tasks with global engagement during dual tasks	PD patients exhibit "ceiling effect" in PFC resources, indicating reduced neural adaptability
Stroop task across cognitive stages [53]	45 PD patients (6 PD-NC, 22 PD-MCI, 17 PDD), 14 healthy controls	PD-MCI showed hypoactivation in DLPFC, M1, and PMC; PDD showed increased activation in mPFC, OFC, and DLPFC; increased DLPFC activation correlated with poorer executive function	fNIRS can characterize cognitive impairment stages in PD; differentiated PD subgroups with 83.3% accuracy when combined with SVM
Functional connectivity during Stroop task [53]	45 PD patients, 14 healthy controls	PD-NC and PD-MCI had enhanced interhemispheric connectivity compared to HCs; PDD showed reduced connectivity among PMC, VLPFC, and OFC	PD-MCI may employ compensatory cortical networking; PDD shows network breakdown

Parkinson's disease research has particularly benefited from fNIRS capabilities due to the combined motor and cognitive aspects of the disorder. Patients with PD demonstrate abnormal prefrontal recruitment during cognitive-motor tasks, reflecting compensatory mechanisms for impaired basal ganglia function [54].

A key finding across studies is the ceiling effect in PD patients' PFC activation. Unlike healthy controls who modulate PFC engagement according to task demands, PD patients show widespread prefrontal activation even during simple tasks, with limited capacity for further recruitment when task complexity increases [54]. This pattern suggests maximal utilization of cognitive resources for ordinary tasks, leaving limited reserve for additional challenges.

The progression of cognitive impairment in PD is associated with distinct cortical activation patterns. Patients with PD-mild cognitive impairment (PD-MCI) show reduced activation in dorsolateral prefrontal, primary motor, and premotor cortices during executive tasks like the Stroop test [53]. In contrast, those with Parkinson's disease dementia (PDD) demonstrate hyperactivation in medial prefrontal, orbitofrontal, and dorsolateral prefrontal regions, possibly reflecting inefficient neural processing or compensatory recruitment [53].

Functional Connectivity Alterations in PD

fNIRS enables investigation of how different cortical regions interact during cognitive processing. Research shows that functional connectivity patterns change across PD cognitive stages [53]. Both cognitively normal PD patients (PD-NC) and those with PD-MCI exhibit enhanced interhemispheric connectivity compared to healthy controls, suggesting adaptive networking to maintain function [53]. However, as dementia develops, this connectivity breaks down, with PDD patients showing reduced connectivity among premotor, ventrolateral prefrontal, and orbitofrontal regions [53].

These findings support the theory that early cognitive impairment in PD is marked by compensatory mechanisms that eventually fail as neurodegeneration progresses, leading to widespread network disruption in dementia stages.

fNIRS Applications in Psychiatric Disorders

Diagnostic Differentiation of Psychiatric Conditions

Table 3: fNIRS studies in psychiatric disorders

Study	Participants	Task	Key Findings	Classification Accuracy
MDD vs Healthy Controls [52]	105 MDD patients, 105 healthy controls	Verbal Fluency Task	MDD patients had significantly smaller HbO changes in frontal and temporal regions	75.2% sensitivity, 74.3% specificity (frontal); 76.5% sensitivity, 76.7% specificity (temporal)
MDD vs GAD vs Comorbid [55]	75 GAD, 75 MDD, 71 CMG, 75 healthy controls	Verbal Fluency Task	All patient groups showed reduced PFC activation vs HC; GAD > MDD in PFC activation; distinct vlPFC patterns in comorbid vs pure GAD	60.47% four-way classification accuracy; 77.19% three-way classification (without HC)
Schizophrenia [49]	Multiple studies	Various executive tasks	Consistently reduced frontal hemodynamic responses across verbal fluency, Stroop, and other executive tasks	Not reported

fNIRS has demonstrated particular utility in differentiating major depressive disorder (MDD) from healthy controls and other psychiatric conditions. During verbal fluency tasks, patients with MDD consistently show reduced frontal and temporal cortex activation compared to healthy individuals [52]. This hypofrontality is thought to reflect the executive dysfunction characteristic of depression.

The diagnostic accuracy of fNIRS for MDD is well-established, with one large study reporting 75.2% classification accuracy using frontal lobe activation and 76.5% using temporal lobe activation [52]. These findings have been replicated across different languages and cultures, supporting the robustness of fNIRS biomarkers for depression [52].

When comparing different psychiatric disorders, fNIRS can identify distinct prefrontal activation patterns. Patients with generalized anxiety disorder (GAD) show higher prefrontal activation during cognitive tasks compared to those with MDD, particularly in the left ventrolateral prefrontal cortex (vlPFC) [55]. These differential activation patterns enable moderate classification accuracy between disorders, with machine learning approaches achieving 60.47% accuracy for four-way classification (GAD, MDD, comorbid, healthy) and 77.19% for three-way classification excluding healthy controls [55].

Limitations in Psychiatric Applications

Despite promising results, fNIRS has limitations for psychiatric diagnosis. One study reported concordance rates as low as 44.0% for bipolar disorder and 38.2% for major depressive disorder when comparing fNIRS classifications with clinical diagnoses [56]. Challenges include symptom overlap between disorders, heterogeneity within diagnostic categories, medication effects on hemodynamic responses, and task selection variability [56].

These limitations suggest that fNIRS may be more effective as a state marker for conditions like depression rather than a definitive diagnostic tool for differentiating similar psychiatric disorders [56]. Combined use with other assessment methods likely enhances diagnostic accuracy.

Essential Research Reagents and Materials

Table 4: Essential research reagents and solutions for fNIRS research

Item	Function	Application Notes
fNIRS System (52-channel example) [52]	Measures cortical hemodynamic responses	Channel count should be determined by brain coverage needs; portability required for naturalistic studies
Task-Specific Software (e.g., Stepmania for DDR) [51]	Presents controlled experimental paradigms	Enables modification of timing, graphics, and music for specific research questions
Prefrontal Cortex Cap (22-channel example) [54]	Optode placement for prefrontal monitoring	Should cover key PFC subregions: lFPC, mFPC, rFPC, lDLPFC, BA8, rDLPFC
3D Digitalization System [47]	Records precise optode locations on scalp	Enights co-registration with anatomical brain atlas for spatial accuracy
Signal Processing Software (e.g., MATLAB toolboxes)	Analyzes raw fNIRS data	Should include motion correction, filtering, and hemodynamic response extraction algorithms
Verbal Fluency Task Materials [52]	Standardized cognitive challenge	Includes letter or category prompts with standardized administration instructions
Stroop Task Materials [53]	Executive function assessment	Computerized presentation recommended for precise timing of incongruent/ congruent stimuli

Successful fNIRS research requires careful selection of equipment and materials tailored to the specific research questions. The fNIRS system itself must provide sufficient channel density to cover the brain regions of interest, with studies typically employing 22-52 channels for adequate prefrontal coverage [54] [52].

Task presentation software must be synchronized with fNIRS data acquisition to enable event-related or block-design analyses. For naturalistic paradigms, specialized software may be needed, such as modified video games that can present controlled stimuli while allowing movement [51].

Proper optode placement is critical, requiring caps designed for specific brain regions of interest. Prefrontal studies typically use caps with extensive frontal coverage including dorsolateral, frontopolar, and ventrolateral prefrontal regions [54]. 3D digitalization systems record precise optode locations for accurate spatial registration with brain anatomy [47].

fNIRS occupies a unique and valuable position in the neuroimaging toolkit, particularly for studying prefrontal cortex function in naturalistic settings and clinical populations. While fMRI remains superior for precise spatial localization of deep brain structures, and EEG provides unparalleled temporal resolution, fNIRS offers an optimal balance of mobility, tolerability, and moderate spatial resolution for cortical monitoring.

In Parkinson's disease research, fNIRS has revealed characteristic prefrontal cortex activation patterns and functional connectivity changes across disease stages, providing insights into cognitive-motor integration deficits. For psychiatric disorders, fNIRS shows promise as an adjunct diagnostic tool, consistently identifying hypofrontality in depression and distinct activation patterns in anxiety disorders.

The future of fNIRS likely involves increased multimodal integration with EEG for comprehensive electrophysiological and hemodynamic assessment, enhanced analytical approaches using machine learning, and continued refinement of naturalistic paradigms that capture brain function in ecologically valid contexts. As technological advances address current limitations in spatial resolution and depth penetration, fNIRS is poised to play an increasingly important role in both clinical neuroscience research and diagnostic applications.

The quest for greater clinical diagnostic accuracy in understanding and treating neurological disorders is driving the adoption of multimodal neuroimaging. Single-modality techniques provide limited windows into brain function, leading to an increasing reliance on combining complementary technologies. This guide objectively compares two powerful multimodal approaches: the integration of functional Near-Infrared Spectroscopy (fNIRS) with Electroencephalography (EEG) to directly probe neurovascular coupling, and the combination of fNIRS with functional Magnetic Resonance Imaging (fMRI) to achieve spatiotemporal validation of hemodynamic responses. These frameworks allow researchers to correlate the brain's electrical activity with its metabolic and hemodynamic processes, providing a more complete picture of neural function in health and disease [41] [11]. For researchers and drug development professionals, understanding the capabilities, experimental requirements, and performance data of these integrated systems is crucial for selecting the appropriate tools for specific clinical research questions, from investigating neurodegenerative diseases like Alzheimer's to mapping functional networks in stroke recovery.

Comparative Performance Analysis of Multimodal Neuroimaging Platforms

The following table summarizes the core technical capabilities and performance metrics of the fNIRS-EEG and fMRI-fNIRS multimodal platforms, providing a data-driven comparison for research planning.

Table 1: Performance and Capability Comparison of Multimodal Neuroimaging Platforms

Feature	fNIRS-EEG Platform	fMRI-fNIRS Platform
Primary Research Application	Direct investigation of neurovascular coupling dynamics; Brain-Computer Interfaces (BCIs); studying Cognitive-Motor Interference [57].	Spatial validation and cross-modal calibration; mapping hemodynamic response localization [7] [21].
Temporal Resolution	Very High (EEG: millisecond; fNIRS: ~0.1-1 second) [41] [11].	Moderate (fMRI: ~1-2 seconds; fNIRS: ~0.1-1 second) [7].
Spatial Resolution	Moderate (fNIRS: ~1-3 cm; EEG: Low, requires source reconstruction) [41] [22].	High (fMRI: ~1-3 mm; fNIRS: ~1-3 cm) [21].
Key Quantitative Findings	Decreased neurovascular coupling strength in dual-task vs. single-task conditions (theta, alpha, beta rhythms) [57]. fNIRS structure–function coupling resembles slower-frequency EEG coupling at rest [22].	Spatial correspondence between fNIRS channels and fMRI activation clusters in motor cortex; classification accuracy for brain "fingerprinting" up to 98% with fNIRS [7].
Portability & Patient Tolerance	High. Robust to motion artifacts, suitable for natural environments, bedside, and pediatric populations [41] [11] [58].	Low to Very High (asynchronous). fMRI requires immobilization; fNIRS component is portable for asynchronous setups [21].
Best-Suited Clinical Contexts	Long-term monitoring, naturalistic task studies, patient groups unsuitable for fMRI (e.g., with implants, claustrophobia), neonates and infants [11] [58] [57].	Pre-surgical mapping, precise localization of functional deficits, validation of fNIRS probe placement and analysis pipelines [7] [21].

Experimental Protocols for Multimodal Integration

Concurrent fNIRS-EEG Protocol for Neurovascular Coupling

This protocol is designed to simultaneously capture electrophysiological and hemodynamic activities to study their coupling, for instance, during cognitive-motor tasks [57].

A. Equipment and Setup:

• fNIRS System: A continuous-wave system (e.g., NIRSport, NIRx) with sources emitting light at 760 and 850 nm, and detectors creating channels with a 3 cm source-detector separation [57]. Short-distance detectors (e.g., 8 mm) are recommended to remove systemic physiological confounds [21].
• EEG System: A standard EEG cap with active or passive electrodes (e.g., 30 electrodes arranged in the 10-5 system) [22].
• Integration Setup: A custom helmet or cap that houses both fNIRS optodes and EEG electrodes, ensuring stable placement and minimal interference. 3D-printed or thermoplastic custom-fit helmets provide the best integration [11].

B. Data Acquisition:

• Synchronization: Use a unified processor or a shared trigger signal to synchronize fNIRS and EEG data acquisition with microsecond precision [11].
• Task Paradigm: Employ a block-design or event-related design. Example: A study on Cognitive-Motor Interference (CMI) may include:
  • A Single Motor Task (SMT): e.g., grip force tracking.
  • A Single Cognitive Task (SCT): e.g., auditory number detection.
  • A Dual Task (DT): simultaneous performance of SMT and SCT [57].
• Data Preprocessing:
  • fNIRS: Convert light intensity to optical density. Prune low-SNR channels. Correct for motion artifacts using hybrid (spline interpolation + wavelet) or PCA-based algorithms. Band-pass filter (e.g., 0.02-0.5 Hz) and convert to oxygenated (HbO) and deoxygenated (HbR) hemoglobin concentrations via the Modified Beer-Lambert Law [41] [57].
  • EEG: Apply band-pass filtering (e.g., 0.5-40 Hz). Re-reference to average reference. Remove artifacts (e.g., ocular, cardiac) using Independent Component Analysis (ICA) [57].

C. Data Fusion and Analysis:

• Component Extraction: Apply Task-Related Component Analysis (TRCA) to both fNIRS and EEG data to extract components that are maximally reproducible across trials [57].
• Neurovascular Coupling Analysis: Calculate the correlation between the temporal traces of the task-related fNIRS components (e.g., HbO) and EEG components (in theta, alpha, and beta frequency bands) to quantify NVC strength [57].
• Statistical Comparison: Use repeated-measures ANOVA to statistically compare NVC strength between different task conditions (e.g., SMT vs. DT) [57].

Simultaneous and Asynchronous fMRI-fNIRS Protocol for Validation

This protocol is used to validate fNIRS signals against the gold-standard spatial resolution of fMRI, either simultaneously or asynchronously [7] [21].

A. Equipment and Setup for Simultaneous Recording:

• fMRI Scanner: A 3T MRI scanner equipped with a head coil.
• fNIRS System: An MRI-compatible fNIRS system (e.g., NIRScout) with optical fibers that are non-magnetic and non-conductive. The probe should cover the region of interest (e.g., motor cortex) [7].
• Probe Co-registration: Digitize the 3D positions of fNIRS optodes and key anatomical landmarks (e.g., Nz, Cz, Iz) using a tracking system. Co-register these positions with the subject's anatomical MRI scan using software like AtlasViewer or Brainstorm to map fNIRS channels to specific brain regions [7] [22].

B. Data Acquisition:

• Paradigm: A block-design motor task (e.g., finger tapping or motor imagery) is commonly used for its robust hemodynamic response [21]. For resting-state studies, participants are instructed to remain still with eyes closed for several minutes [7].
• Simultaneous Preprocessing:
  • fMRI: Standard preprocessing including slice-timing correction, motion realignment, spatial normalization, and smoothing. A band-pass filter (0.009-0.08 Hz) is typically applied for resting-state functional connectivity [7].
  • fNIRS: Similar preprocessing as in the fNIRS-EEG protocol, with additional steps to remove MRI-specific artifacts (e.g., scanner gradient artifacts) using template-based subtraction [7].

C. Data Analysis for Spatial Correspondence:

• For Task-Based Studies: A General Linear Model (GLM) is applied to both fMRI and fNIRS data. The subject-specific fNIRS time series (HbO, HbR) from a region of interest can be used as a regressor in the fMRI analysis to identify voxels with significant activation corresponding to the fNIRS signal [21].
• For Resting-State Studies: Calculate resting-state functional connectivity (rsFC) matrices for both modalities using Pearson correlation between the time series of different brain regions. "Brain fingerprinting" can then be performed by using a simple linear classifier to identify individuals based on their unique rsFC patterns [7].

Figure 1: Experimental workflows for fNIRS-EEG and fMRI-fNIRS multimodal integration, illustrating the key stages from setup to analytical output.

The Scientist's Toolkit: Key Research Reagents and Materials

Successful execution of multimodal neuroimaging studies requires specific hardware, software, and analytical tools. The following table details these essential components.

Table 2: Essential Research Tools for Multimodal Neuroimaging Studies

Tool Category	Specific Examples & Functions	Key Considerations
Integrated fNIRS-EEG Caps/Helmets	Custom 3D-printed helmets; Cryogenic thermoplastic sheets; Modified elastic EEG caps with fNIRS fixture ports [11].	Custom helmets offer best fit and stability but at higher cost. Elastic caps are more affordable but may lead to variable optode pressure and placement [11].
fNIRS Hardware	Continuous-Wave (CW) systems (e.g., NIRSport, NIRx); Sources (LEDs/Lasers at 760nm & 850nm); Silicon Photodiode (SiPD) detectors [21] [57].	CW-fNIRS is most common due to cost and simplicity. Source-detector distance of ~3 cm is standard, with short-distance detectors (~8 mm) recommended for denoising [41] [21].
EEG Hardware	Active/Passive electrode systems (e.g., BrainAMP); Amplifiers compatible with fNIRS environment [11].	Ensure no electromagnetic interference with fNIRS. The number and placement of electrodes should be coordinated with fNIRS optode layout [11].
Data Acquisition & Synchronization	Unified processors (e.g., integrated systems from specific vendors); Separate systems synchronized via TTL triggers [11].	Unified processors offer superior synchronization precision, which is critical for analyzing fast neural events and their hemodynamic correlates [11].
Core Analytical Software	HOMER2/3, NIRS Brain AnalyzIR, MNE, Brainstorm, SPM, EEGLAB, in-house MATLAB/Python scripts [7] [22] [21].	Software choice depends on preprocessing needs, fusion analysis type (e.g., TRCA, DCM, GLM), and user expertise. Open-source toolboxes promote reproducibility.

Discussion and Clinical Relevance

The integration of fNIRS with EEG and fMRI is more than a technical exercise; it addresses fundamental physiological questions and holds significant promise for clinical application. The neurovascular coupling mechanism, which forms the basis for interpreting hemodynamic signals from both fNIRS and fMRI, is itself implicated in the pathogenesis of several neurological disorders. For example, impaired NVC is a recognized sign of Alzheimer's disease and stroke [41] [59]. The fNIRS-EEG platform allows for the direct, non-invasive investigation of this coupling in patient populations, potentially serving as a biomarker for early diagnosis and treatment monitoring [57].

Furthermore, the regional heterogeneity of the brain's structure-function relationship, which follows a unimodal (sensory) to transmodal (association) gradient, can be differentially captured by these modalities. Research shows that fNIRS structure–function coupling resembles the slower-frequency coupling seen in EEG, with both modalities revealing stronger coupling in the sensory cortex and greater decoupling in the association cortex [22]. Discrepancies between EEG and fNIRS, particularly in the frontoparietal network, highlight their sensitivity to different physiological processes and underscore the value of a multimodal approach for a comprehensive assessment of brain network integrity [22].

For clinical researchers and drug developers, these integrated platforms offer powerful tools. The fNIRS-EEG system is ideal for longitudinal studies, bedside monitoring, and investigating patient groups like those with Parkinson's disease or infants, where fMRI is impractical [11] [58] [57]. Conversely, the fMRI-fNIRS platform provides a pathway to translate well-established fMRI paradigms to more accessible and portable fNIRS setups, once the spatial correspondence is validated, thereby expanding the scope of clinical neuroimaging from the scanner to the clinic and natural environments [21].

The objective comparison presented in this guide demonstrates that the choice between an fNIRS-EEG and an fMRI-fNIRS platform is not a matter of superiority, but of strategic alignment with research goals. The fNIRS-EEG integration is the premier choice for directly investigating the temporal dynamics of neurovascular coupling in naturalistic or clinical settings, offering high temporal resolution and portability. The fMRI-fNIRS integration serves as a critical tool for the spatiotemporal validation of fNIRS signals, leveraging the high spatial resolution of fMRI to ground-truth hemodynamic measurements. Together, these multimodal approaches are pushing the boundaries of clinical neuroscience, providing richer, more validated data that promises to enhance diagnostic accuracy, guide therapeutic interventions, and accelerate drug development for neurological disorders. Future directions will involve standardizing analysis pipelines, improving hardware integration, and further leveraging these tools to decipher the complex pathophysiology of brain diseases.

Navigating Technical Challenges and Optimizing Protocol Efficacy

Addressing Motion Artifacts and Physiological Noise in fNIRS and EEG

Functional near-infrared spectroscopy (fNIRS) and electroencephalography (EEG) have emerged as powerful, non-invasive neuroimaging tools that offer a compelling balance of portability, cost-effectiveness, and temporal resolution for both research and clinical applications. Unlike fMRI, these modalities allow for more naturalistic study designs and long-term monitoring, making them particularly valuable for studying neurological disorders and cognitive processes in real-world contexts [60] [61]. However, their significant vulnerability to motion artifacts and physiological noise presents a critical challenge that can compromise data quality and clinical diagnostic accuracy [60] [62].

The fundamental difference in what each technology measures—electrical activity for EEG versus hemodynamic responses for fNIRS—results in distinct artifact profiles and mitigation requirements. EEG signals, characterized by their microvolt-scale electrical potentials, are exceptionally susceptible to contamination from muscle activity, eye movements, and electrode displacement [63] [62]. Meanwhile, fNIRS, which measures changes in oxy- and deoxy-hemoglobin concentrations, suffers from motion-induced perturbations in optode-scalp coupling and systemic physiological oscillations that can mask neural signals [60] [64]. For researchers and clinicians aiming to employ these technologies in drug development or neurological disorder diagnosis, understanding and addressing these artifacts is not merely methodological but essential for generating reliable, interpretable data.

Technology Comparison: Fundamental Principles and Noise Profiles

Comparative Technical Specifications

Table 1: Technical comparison between fNIRS and EEG

Feature	EEG	fNIRS
Measured Signal	Electrical potentials from synchronized neuronal firing	Hemodynamic changes (HbO/HbR) via neurovascular coupling
Temporal Resolution	Millisecond range (high)	~0.1-1 second (moderate)
Spatial Resolution	~1-10 cm (limited)	~1-3 cm (moderate)
Primary Artifact Sources	Motion: Electrode displacement, cable movement [62]. Physiological: EMG, EOG, ECG [63] [62].	Motion: Optode-scalp decoupling, head movements [65] [64]. Physiological: Cardiac, respiratory, blood pressure cycles [64].
Key Artifact Characteristics	High-amplitude, high-frequency spikes; baseline shifts [62]	Baseline shifts, spike-like artifacts [64]; physiological noise is temporally correlated and spatially global [64]
Dominant Frequency Bands	Delta (0.5-4 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz), Gamma (>30 Hz) [24]	Very low frequencies (<0.1 Hz) for hemodynamic response; cardiac (~1 Hz) and respiratory (~0.3 Hz) oscillations [64]

Motion Artifact Characteristics and Origins

The tables above capture the fundamental differences in how motion manifests in each modality. For EEG, motion artifacts primarily arise from changes in the electrode-skin interface. Even minor head movements can cause electrode polarization changes, cable sway, or alterations in skin potential, generating high-amplitude, abrupt signal changes [62]. In mobile EEG (mo-EEG) scenarios, gait-related artifacts introduce particularly challenging periodic amplitude bursts synchronized with heel strikes [62].

For fNIRS, motion artifacts stem mainly from mechanical forces disrupting the light path between optodes and the scalp. This can include direct optode movement, changes in pressure on the scalp, or alterations in scalp blood flow beneath the optodes [65] [64]. These disruptions manifest as both rapid spike-like artifacts and slower baseline shifts that can mimic or obscure the true hemodynamic response [64]. Dry EEG systems, while offering faster setup times, are particularly prone to motion artifacts compared to gel-based systems due to the lack of a conductive gel that provides mechanical stabilization [63].

Artifact Correction Methodologies: From Traditional to Deep Learning Approaches

EEG Artifact Removal Techniques

EEG artifact correction has evolved from simple filters to sophisticated hybrid and deep-learning approaches that leverage both temporal and spatial information.

• Temporal Filtering & Blind Source Separation: Traditional approaches include temporal filters (high-pass, low-pass, notch) for removing frequency-specific noise [62] [66]. Blind Source Separation methods, particularly Independent Component Analysis (ICA), have been widely adopted to separate neural signals from artifact-contaminated components based on statistical independence [63] [66]. These components can then be manually or automatically rejected before signal reconstruction.

• Spatial Filtering Approaches: Spatial filtering techniques like SPHARA (Spatial Harmonic Analysis) leverage the spatial distribution of EEG electrodes to suppress noise that exhibits distinct spatial patterns from neural signals [63]. This approach is particularly effective for multi-channel systems and can be combined with temporal methods.

• Deep Learning Architectures: Recent advances have introduced sophisticated deep learning models specifically designed for EEG denoising. The Multi-Scale Temporal Propagation Network (MSTP-Net) uses multi-scale temporal blocks to capture both local and global temporal features, explicitly enlarging the receptive field to improve noise reduction performance [66]. Motion-Net employs a subject-specific U-Net architecture trained with visibility graph features to effectively remove motion artifacts from individual EEG recordings [62].

• Hybrid Methods: Combined approaches often yield superior results. Recent research demonstrates that integrating ICA-based methods (Fingerprint + ARCI) with spatial filtering (SPHARA) significantly outperforms either method alone for dry EEG, reducing standard deviation of signals from 9.76 μV to 6.15 μV [63].

fNIRS Artifact Correction Strategies

fNIRS artifact correction must address both motion-induced disturbances and physiological confounds, with methods typically categorized as univariate (single-channel) or multivariate (multi-channel).

• Spline Interpolation & Wavelet Methods: Spline interpolation models motion artifact segments using cubic splines and subtracts them from the signal, particularly effective for baseline shifts [64]. Wavelet filtering identifies and removes outlier coefficients in the wavelet domain that correspond to artifacts, excelling at removing spike-like disturbances [64]. A hybrid spline + wavelet approach has demonstrated a remarkable 94.1% channel improvement rate by leveraging the complementary strengths of both methods [64].

• CNN-Based Frameworks: The 1DCNNwP (1D Convolutional Neural Networks with Penalty) architecture represents a significant advance in real-time fNIRS processing, achieving an 11.08 dB improvement in signal-to-noise ratio and processing samples in just 0.53 ms on average [65]. This model combines convolutional layers for temporal feature extraction with a penalty network that acts as a regularization mechanism to enhance robustness.

• Hardware-Assisted Solutions: Some approaches incorporate additional hardware such as accelerometers to detect motion events or short-separation channels to measure superficial contaminants [65] [64]. While effective, these solutions increase system complexity and cost, making purely algorithmic approaches attractive for many applications.

Performance Comparison: Quantitative Analysis of Correction Methods

Correction Efficacy Metrics

Table 2: Performance comparison of artifact removal methods

Method	Modality	Key Performance Metrics	Comparative Advantages
Fingerprint + ARCI + improved SPHARA [63]	Dry EEG	SD: 9.76→6.15 μV; SNR: 2.31→5.56 dB	Superior for combined physiological artifact reduction and spatial denoising
MSTP-Net [66]	EEG	CC: 0.8816→0.9221; SNR: 10.46→12.76 dB	Excellent at capturing non-stationary EEG characteristics via multi-scale receptive fields
Motion-Net [62]	Mobile EEG	Artifact reduction: 86% ±4.13; SNR improvement: 20 ±4.47 dB	Subject-specific approach effective with smaller datasets; incorporates visibility graph features
1DCNNwP [65]	fNIRS	SNR improvement: >11.08 dB; Processing time: 0.53 ms/sample	Real-time capability; minimal prior data requirement; individual subject adaptation
Spline + Wavelet Hybrid [64]	fNIRS	Channel improvement rate: ~94.1%	Comprehensive correction for both spike-like artifacts and baseline shifts

Experimental Protocols for Method Validation

Experimental Design: Eleven healthy volunteers performed a motor execution paradigm involving left/right hand, feet, and tongue movements while 64-channel dry EEG was recorded. Movements were cued by visual stimuli with randomized inter-trial intervals.

Processing Pipeline:

• Reference Preprocessing: Bandpass filtering (1-30 Hz) and notch filtering (50 Hz)
• ICA-Based Cleaning: Application of Fingerprint and ARCI algorithms to identify and remove physiological artifacts
• Spatial Filtering: Implementation of improved SPHARA with additional zeroing of artifactual jumps
• Quality Assessment: Calculation of Standard Deviation (SD), Signal-to-Noise Ratio (SNR), and Root Mean Square Deviation (RMSD)
• Statistical Analysis: Generalized Linear Mixed Effects (GLME) modeling to quantify significant changes in signal quality parameters

Validation Approach: The 1DCNNwP method was validated using both simulated data (from the balloon model) and semi-simulated experimental data with real motion artifacts.

Training Strategy:

• Data Preparation: Simulated fNIRS signals with known artifacts or semi-simulated data combining experimental recordings
• Network Architecture: Seven-layer 1D CNN with pooling/up-sampling layers combined with a parallel penalty network for enhanced robustness
• Training Scheme: Moving window approach with input data augmentation to maximize limited training data
• Performance Benchmarking: Comparison against spline-interpolation, wavelet-based, TDDR, and spline-SG methods using SNR and contrast-to-noise ratio (CNR) metrics
• Real-Time Capability Assessment: Measurement of processing time per sample to verify real-time applicability

Impact on Diagnostic Accuracy in Neurological Disorders

The efficacy of artifact correction methods directly translates to improved diagnostic accuracy in neurological and psychiatric disorders. EEG-based classification studies have demonstrated remarkable performance in distinguishing between conditions such as Alzheimer's disease, schizophrenia, mild cognitive impairment (MCI), and depression [24]. For instance, Linear Discriminant Analysis classifiers have achieved 100% accuracy distinguishing healthy controls from Alzheimer's patients and 84.67% accuracy in three-way classification between depression, MCI, and schizophrenia when using properly cleaned EEG data [24].

However, fNIRS faces greater challenges in psychiatric differential diagnosis. Clinical experience shows concordance rates between fNIRS diagnoses and psychiatric diagnoses of only 44.0% for bipolar disorder and 38.2% for major depressive disorder [56]. This highlights the critical need for improved artifact correction methods, particularly for disorders with overlapping symptoms like depression, which can manifest similarly across different psychiatric conditions and mask subtle neural signatures unique to each disorder [56].

The frontal lobe channels frequently emerge as critical biomarkers across multiple neurological disorders in EEG studies [24], underscoring the importance of maintaining signal integrity in these regions through effective artifact removal.

Research Reagent Solutions: Essential Tools for Artifact Management

Table 3: Essential research tools for fNIRS and EEG artifact management

Tool/Category	Specific Examples	Function/Purpose
Dry EEG Systems	waveguardtouch dry EEG cap (64-channel) [63]	Rapid application without conductive gel; ideal for ecological studies
Conductive Media	NeuroPrep gel, Ten20 paste [60]	Optimize electrode-scalp interface; reduce impedance in wet EEG systems
Reference Electrodes	Gel-based ground/reference electrodes on mastoids [63]	Provide stable reference point; impedance kept below 50 kΩ
Motion Tracking	Accelerometers [65] [64]	Detect head movements for motion artifact reference in fNIRS
Peripheral Physiology	EOG, ECG, EMG electrodes [62]	Record reference signals for physiological artifact removal
Software Toolboxes	Homer2/3 [64]	Comprehensive fNIRS processing including artifact correction algorithms
Deep Learning Frameworks	MSTP-Net, Motion-Net, 1DCNNwP [65] [62] [66]	Advanced artifact removal using neural networks

Signaling Pathways and Experimental Workflows

Neuroimaging Artifact Correction Workflow

The systematic comparison of motion artifact and physiological noise correction methods for fNIRS and EEG reveals a rapidly evolving landscape where traditional signal processing techniques are being augmented—and in some cases surpassed—by sophisticated deep learning approaches. The optimal strategy depends critically on the specific application constraints: real-time BCI applications benefit from the computational efficiency of methods like 1DCNNwP, while clinical diagnostic applications may prioritize the superior artifact rejection of hybrid methods like Fingerprint + ARCI + SPHARA.

For researchers and drug development professionals, these advances in signal integrity directly translate to enhanced diagnostic accuracy and more reliable assessment of therapeutic interventions. Future progress will likely emerge from multimodal integration, where complementary strengths of fNIRS and EEG are leveraged simultaneously, and from continued refinement of subject-specific deep learning models that can adapt to individual neurophysiological signatures. As these technologies become increasingly central to neurological disorder research and clinical practice, robust artifact management will remain foundational to generating meaningful, interpretable data that can advance both basic neuroscience and therapeutic development.

Functional Magnetic Resonance Imaging (fMRI) has long been a cornerstone of human brain research, providing unparalleled spatial resolution for localizing brain activity deep within cortical and subcortical structures. However, its prohibitive cost, immobility, and sensitivity to motion artifacts significantly limit its accessibility and applicability in naturalistic settings or with clinical populations. This guide objectively compares the performance of emerging alternatives—functional Near-Infrared Spectroscopy (fNIRS) and Electroencephalography (EEG)—and details how their integration, alongside innovative computational methods, is overcoming these limitations while simultaneously addressing fNIRS's primary constraint: limited penetration depth for subcortical imaging.

fNIRS vs. EEG vs. fMRI: A Technical Comparison

The following table summarizes the core technical specifications and performance metrics of fMRI, fNIRS, and EEG, highlighting their complementary strengths and weaknesses.

Table 1: Comprehensive comparison of non-invasive neuroimaging modalities.

Feature	fMRI	fNIRS	EEG
What It Measures	Blood-Oxygen-Level-Dependent (BOLD) signal [5]	Hemodynamic response (oxy-Hb and deoxy-Hb) [67] [5]	Electrical potentials from neuronal firing [67]
Spatial Resolution	High (millimeter-level) [5]	Moderate (1-3 cm) [5] [68]	Low (centimeter-level) [67] [69]
Temporal Resolution	Low (seconds) [5]	Moderate (seconds) [67]	Very High (milliseconds) [67] [69]
Depth of Measurement	Whole brain (cortical & subcortical) [5]	Superficial cortex (1-2.5 cm depth) [67] [68]	Cortical surface [67]
Portability & Motion Tolerance	Very low; requires immobility [5]	High; resistant to motion artifacts [70] [67]	High (especially wireless systems) [67]
Approximate Cost	Very High [71] [5]	Moderate to High [67]	Generally Lower [67]
Best Use Cases	Precise spatial localization, deep brain structures [5]	Naturalistic studies, clinical populations, bedside monitoring [70] [67] [5]	Fast cognitive processes, brain-computer interfaces, sleep studies [67]

Experimental Validation: Protocols and Key Findings

Validating fNIRS Sensitivity to Cognitive Load

Experimental Protocol: A classic n-back working memory task was administered to participants while their prefrontal cortex activity was monitored using fNIRS [68]. In this paradigm, participants are presented with a sequence of stimuli and must indicate when the current stimulus matches the one presented 'n' trials back.

• Load Conditions: The task included multiple levels of difficulty (e.g., 1-back, 2-back, 3-back) to systematically increase working memory load.
• fNIRS Setup: Optical probes were placed over the frontal cortex, targeting regions like the dorsolateral prefrontal cortex (dlPFC), a key area for working memory [68].
• Measurement: The primary measured outcomes were changes in oxygenated hemoglobin (HbO) concentration and functional connectivity between brain regions.

Key Findings: The study demonstrated that fNIRS is highly sensitive to cognitive state and load. Specifically, activation in the dlPFC scaled linearly with increasing working memory load [68]. Furthermore, functional connectivity between frontal and parietal brain regions also increased with higher cognitive loads. This confirms fNIRS's capability to detect graded changes in brain activity, a crucial requirement for sophisticated cognitive neuroscience research.

Inferring Deep-Brain Activity from Cortical fNIRS

Experimental Protocol: To overcome fNIRS's inability to directly measure subcortical activity, a novel computational approach was developed and validated using simultaneous fNIRS-fMRI [71].

• Simultaneous Recording: Brain activity was recorded concurrently with fNIRS and fMRI while participants performed cognitive tasks (e.g., go/no-go, face recognition).
• Computational Modeling: A support vector regression (SVR) learning algorithm was trained to predict the fMRI-measured activity in deep-brain regions (e.g., fusiform cortex) using only the fNIRS signals from the cortical surface [71].
• Validation: The predicted deep-brain activity from the fNIRS-based model was compared against the actual fMRI recordings.

Key Findings: This method successfully inferred deep-brain activity from cortical fNIRS measurements. When using the cortical fNIRS signals, the model achieved a prediction accuracy with a correlation coefficient of up to 0.7 for the top 15% of predictions [71]. This demonstrates the feasibility of extending fNIRS applications to estimate activity in regions below its direct penetration depth.

Visualizing the Multimodal Integration Workflow

The integration of EEG and fNIRS creates a synergistic system that captures both electrical and hemodynamic brain activity. The following diagram illustrates the workflow of a simultaneous fNIRS-EEG experiment.

Diagram Title: Simultaneous fNIRS-EEG Experimental Workflow

The Scientist's Toolkit: Key Research Reagent Solutions

For researchers aiming to implement these protocols, the following table details essential hardware and analytical "reagents" and their functions.

Table 2: Essential materials and analytical solutions for fNIRS and EEG research.

Item Category	Specific Example / Method	Function in Research
Low-Cost fNIRS Hardware	HEGduino V2 [72]	A low-cost, single-channel fNIRS device that democratizes access to functional neuroimaging by reducing financial barriers.
Integrated Helmets	3D-Printed Custom Helmets [69]	Ensures precise, stable, and replicable placement of both EEG electrodes and fNIRS optodes on the scalp, improving data quality.
Quantitative EEG (qEEG) Parameters	Power Ratio Index (PRI), Brain Symmetry Index (BSI) [70]	Computerized EEG metrics that serve as prognostic biomarkers for motor and cognitive recovery, e.g., post-stroke.
Computational Inference	Support Vector Regression (SVR) [71]	A machine learning algorithm used to infer hemodynamic activity in deep-brain structures from cortical fNIRS signals.
Data Fusion Analysis	Joint Independent Component Analysis (jICA) [67]	A statistical method used to identify common underlying components from simultaneously acquired EEG and fNIRS data.

The landscape of clinical neuroimaging is evolving beyond a reliance on fMRI. While fNIRS and EEG individually address key limitations of cost, accessibility, and motion tolerance, their true transformative potential is unlocked through multimodal integration. By combining EEG's millisecond temporal resolution with fNIRS's robust, localized hemodynamic measurements, researchers can obtain a richer, more comprehensive picture of brain function in real-world settings. Furthermore, advanced computational methods are now actively overcoming the subcortical penetration barrier of fNIRS, opening new frontiers for non-invasive, affordable, and clinically viable brain imaging in neurological disorder research and drug development.

Data Fusion Complexities and Signal Processing Pipelines for Multimodal Data

In the pursuit of precise clinical diagnostics for neurological disorders, the limitations of unimodal neuroimaging have become increasingly apparent. No single technique can fully capture the brain's complex spatiotemporal dynamics, leading to an imperative for multimodal integration. Functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS) each offer distinct advantages: fMRI provides high spatial resolution for localizing deep brain structures, EEG captures millisecond-scale electrical neural activity, and fNIRS offers portable hemodynamic monitoring with good temporal resolution [11] [5]. The complementary nature of these modalities creates unprecedented opportunities for enhancing diagnostic accuracy in conditions such as epilepsy, stroke, and neurodegenerative diseases [11] [2].

However, the integration of these heterogeneous data streams presents significant computational and methodological challenges. Data fusion complexities arise from fundamental differences in temporal resolution (milliseconds for EEG versus seconds for hemodynamic responses), spatial coverage (superficial cortical for fNIRS/EEG versus whole-brain for fMRI), and the physiological origins of the signals measured by each technique [11] [5]. Furthermore, the absence of standardized processing pipelines and the sensitivity of results to analytical choices compound these challenges, potentially affecting the reproducibility and clinical translation of findings [73]. This review systematically compares multimodal fusion approaches, their experimental validation, and their impact on diagnostic accuracy in neurological disorder research.

Fundamental Techniques and Fusion Strategies

Core Fusion Architectures

Multimodal data fusion strategies are broadly categorized into three architectural paradigms, each with distinct advantages and limitations for clinical application. The choice of architecture significantly impacts how well complementary information is leveraged and how the system handles noisy or missing data, which are common challenges in clinical environments.

• Early Fusion (Feature-level Fusion): This approach combines raw or minimally processed data from multiple modalities before feature extraction. For instance, EEG electrical potentials might be concatenated with fNIRS hemoglobin concentrations into a unified input vector [74]. While this method preserves the richest information and allows the model to discover complex cross-modal relationships, it requires precise temporal synchronization and is highly sensitive to differences in sampling rates between modalities. In clinical practice, this approach is challenging due to the difficulty in perfectly aligning data streams with fundamentally different temporal characteristics.

• Intermediate Fusion (Model-level Fusion): Intermediate strategies process each modality separately in initial stages before combining extracted features at intermediate layers of a processing model. A common implementation uses dedicated neural network branches for EEG spectral features and fNIRS hemodynamic trends, merging their outputs in a shared representation layer [74] [75]. This architecture effectively balances modality-specific processing with cross-modal interaction, making it particularly suitable for deep learning approaches that can learn optimal fusion points automatically from data.

• Late Fusion (Decision-level Fusion): In this approach, each modality is processed independently through complete analysis pipelines, generating separate predictions or decisions that are combined at the final stage through methods such as weighted voting or averaging [74] [75]. Late fusion offers practical advantages in clinical settings, including robustness to missing modalities and flexibility in handling asynchronous data streams. However, it may fail to capture nuanced cross-modal interactions that occur at the feature level, potentially limiting the synergistic benefits of multimodal integration.

Comparative Analysis of Fusion Approaches

Table 1: Comparison of Multimodal Data Fusion Strategies for Neuroimaging

Fusion Strategy	Technical Implementation	Advantages	Limitations	Clinical Applicability
Early Fusion	Concatenation of raw/preprocessed signals before feature extraction [74]	Maximizes information preservation; Enables discovery of complex cross-modal relationships	Requires perfect data alignment; Highly sensitive to noise/artifacts	Low - due to synchronization challenges and sensitivity to clinical artifacts
Intermediate Fusion	Joint representation learning in shared latent space; Deep learning architectures with cross-modal attention [74] [75]	Balances modality-specific and cross-modal processing; Adaptable to signal characteristics	Complex model training; Requires substantial computational resources	Medium-High - increasingly used with deep learning models for BCI and monitoring
Late Fusion	Independent processing with decision integration via voting/averaging [74] [75]	Robust to missing data or failed modalities; Flexible for asynchronous data	May miss subtle cross-modal interactions; Limited synergistic potential	High - practical for clinical environments with variable data quality

Experimental Protocols and Validation Frameworks

Representative Experimental Design: Motor Imagery Classification

A rigorous experimental protocol demonstrating effective EEG-fNIRS fusion was developed for motor imagery (MI) classification, a critical application in brain-computer interfaces (BCIs) for neurorehabilitation. The methodology achieved an average accuracy of 83.26%, representing a 3.78% improvement over state-of-the-art unimodal methods on the TU-Berlin-A dataset [75].

The experimental workflow incorporated specialized processing for each modality. For EEG signals, researchers employed dual-scale temporal convolution and depthwise separable convolution to extract spatiotemporal features, complemented by a hybrid attention mechanism to enhance sensitivity to salient neural patterns. For fNIRS signals, spatial convolution across all channels identified activation differences between brain regions, while parallel temporal convolution combined with a Gated Recurrent Unit (GRU) captured the temporal dynamics of the hemodynamic response [75].

At the fusion stage, the protocol implemented a sophisticated uncertainty modeling approach. Decision outputs from both modalities were quantified using Dirichlet distribution parameter estimation to model prediction uncertainty. A two-layer reasoning process based on Dempster-Shafer Theory (DST) then fused evidence from basic belief assignments, effectively combining the complementary strengths of both modalities while accounting for the confidence in each signal's contribution [75].

Experimental Visualization: Multimodal Fusion Workflow

The following diagram illustrates the complete experimental workflow for multimodal EEG-fNIRS data fusion, from data acquisition through final decision integration:

Diagram 1: Experimental workflow for EEG-fNIRS fusion in motor imagery classification

Comparative Performance in Clinical Applications

Diagnostic Accuracy Across Neurological Disorders

Multimodal fusion approaches have demonstrated significant improvements in diagnostic accuracy and patient monitoring capabilities across various neurological conditions. The synergistic combination of electrophysiological (EEG) and hemodynamic (fNIRS/fMRI) information provides complementary insights into brain dysfunction, enabling more precise characterization of disease mechanisms.

Table 2: Clinical Application Performance of Multimodal Neuroimaging

Neurological Disorder	Multimodal Combination	Reported Diagnostic Improvement	Key Clinical Applications
Epilepsy	EEG-fNIRS [11] [2]	Improved seizure focus localization; Better differentiation of epileptiform activity from artifacts	Pre-surgical mapping; Seizure detection and monitoring; Treatment efficacy assessment
Stroke & Neurorehabilitation	fNIRS-EEG for BCI [76] [75]	83.26% classification accuracy for motor imagery [75]	Motor recovery tracking; Neurofeedback therapy; Rehabilitation outcome prediction
Traumatic Brain Injury (TBI)	fNIRS-EEG [11] [2]	Enhanced monitoring of cerebral oxygenation and neural function in ICU settings	Bedside monitoring; Cerebral autoregulation assessment; Prognostic evaluation
Attention-Deficit/Hyperactivity Disorder (ADHD)	fNIRS-EEG [11]	Better characterization of prefrontal cortex dysfunction during cognitive tasks	Differential diagnosis; Treatment response monitoring; Neurodevelopmental tracking
Neurodegenerative Disorders (Alzheimer's, Parkinson's)	fMRI-fNIRS [5] [77]	Correlation of deep brain structures (fMRI) with cortical activity (fNIRS)	Early detection; Disease progression monitoring; Cognitive reserve assessment

Modality Complementarity and Technical Specifications

Each neuroimaging modality provides unique and complementary information about brain function, with technical specifications that determine their optimal combination for specific clinical applications.

Table 3: Technical Comparison of Neuroimaging Modalities for Multimodal Fusion

Modality	Spatial Resolution	Temporal Resolution	Measured Parameters	Primary Clinical Strengths	Key Limitations
fMRI	1-3 mm (whole brain) [5]	1-2 Hz (limited by hemodynamic response) [5]	Blood Oxygen Level Dependent (BOLD) signal [5]	Localization of deep brain structures; Excellent spatial resolution	Expensive; Non-portable; Sensitive to motion artifacts
EEG	1-10 cm (limited to cortex) [11]	1-10 ms (direct neural activity) [11]	Electrical potentials from synchronized neural firing [11]	Captures direct neural activity with millisecond precision; Low cost	Limited spatial resolution; Sensitive to physiological artifacts
fNIRS	1-3 cm (cortical surfaces only) [5] [2]	100 ms - 1 Hz [5]	Concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) [2]	Portable; Good temporal resolution; Resilient to motion artifacts [2]	Limited to cortical regions; Sensitive to extracerebral physiology

Signal Processing Pipelines and Reproducibility Challenges

Standardized Processing Workflows

The complexity of multimodal data fusion necessitates rigorous, standardized processing pipelines to ensure reproducibility and reliability of findings. The FRESH (fNIRS Reproducibility Study Hub) initiative, which involved 38 independent research teams analyzing identical datasets, revealed that nearly 80% of teams agreed on group-level results when hypotheses were strongly supported by literature [73]. However, individual-level agreement was significantly lower and highly dependent on data quality.

A representative standardized pipeline for fNIRS data includes three critical stages: movement artifact reduction (MAR) using algorithms such as wavelet-based or correlation-based methods; bandpass filtering (BPF) to isolate the hemodynamic response (typically 0.01-0.5 Hz); and principal component analysis (PCA) to remove global physiological noise [78]. For EEG data, standard preprocessing includes bandpass filtering (0.5-45 Hz), artifact removal (ocular, cardiac, muscle), and re-referencing [11]. The integration of these pipelines requires careful temporal alignment and often incorporates short-separation regression for fNIRS to remove superficial contaminants, a technique that remains underutilized despite its potential [76] [78].

Visualization of Standardized Processing Pipeline

The following diagram illustrates a robust signal processing workflow that addresses common artifacts in multimodal neuroimaging data:

Diagram 2: Standardized signal processing pipeline for multimodal neuroimaging data

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of multimodal fusion research requires specialized tools and methodologies. The following table details essential components of the multimodal research toolkit:

Table 4: Essential Research Toolkit for Multimodal Neuroimaging Studies

Tool Category	Specific Tools/Techniques	Primary Function	Implementation Considerations
Hardware Integration	Custom 3D-printed helmet systems [11]; Cryogenic thermoplastic sheets [11]	Precise co-registration of EEG electrodes and fNIRS optodes	Ensure consistent optode-scalp coupling; Minimize cross-modality interference
Data Acquisition Systems	Synchronized EEG-fNIRS systems (e.g., NIRScout with BrainAMP) [11]; Unified processors [11]	Simultaneous multimodal data acquisition with precise timing	Choose between separate synchronized systems vs. unified hardware
Signal Processing Libraries	MATLAB Toolboxes (Homer2, NIRS-KIT); Python libraries (MNE, NiLearn)	Implement artifact removal and feature extraction pipelines	Ensure compatibility across modalities; Standardize sampling rates
Artifact Removal Algorithms	Movement Artifact Reduction (MAR) algorithms [78]; Short-separation channel regression [76]	Remove motion artifacts and physiological confounders	Select algorithms based on artifact type and data quality
Fusion Algorithms	Deep learning architectures (CNN-RNN hybrids) [75]; Attention mechanisms [74]	Integrate multimodal features at different processing stages	Balance model complexity with interpretability for clinical translation
Statistical Validation Tools	Dempster-Shafer Theory [75]; Cross-validation frameworks	Model uncertainty and validate multimodal classifiers	Account for multiple comparisons; Control for false discovery rates

The integration of fMRI, EEG, and fNIRS through advanced data fusion strategies represents a paradigm shift in clinical neuroimaging, offering unprecedented insights into brain function and neurological disorders. The complementary nature of these modalities—combining fMRI's spatial precision, EEG's temporal resolution, and fNIRS's portability—enables more comprehensive monitoring and diagnosis than any single modality can achieve independently. However, widespread clinical adoption remains constrained by significant challenges, including hardware integration complexities, the absence of standardized processing pipelines, and computational demands of fusion algorithms.

Future progress in the field will depend on collaborative efforts to develop standardized data formats, validated processing pipelines, and open-source benchmarking datasets to enhance reproducibility [73]. Advancements in machine learning approaches, particularly deep learning architectures with inherent multimodal fusion capabilities, show promise for automatically learning optimal integration strategies from data rather than relying on manually engineered pipelines [75]. Additionally, hardware innovations creating more compact, integrated multimodal systems will lower implementation barriers in clinical settings. As these technical and methodological challenges are addressed, multimodal fusion approaches are poised to significantly enhance diagnostic accuracy, treatment monitoring, and therapeutic outcomes across the spectrum of neurological disorders.

Leveraging Machine Learning for Feature Selection and Diagnostic Classification

Neurological disorders (ND) represent a significant global health challenge, profoundly affecting cognitive functions, emotional regulation, and quality of life for millions worldwide [24]. Disorders such as Alzheimer's disease, mild cognitive impairment (MCI), schizophrenia, and depression often present with overlapping symptoms, complicating accurate and timely diagnosis using traditional clinical assessment methods [24]. Early and precise detection is crucial for effective intervention, yet conventional diagnostic approaches like clinical examinations, neuropsychological assessments, and advanced neuroimaging (MRI, fMRI, PET) often involve subjective interpretation, high costs, limited accessibility, and sometimes invasive procedures [24] [5].

The integration of machine learning (ML) with neuroimaging technologies has emerged as a transformative approach for enhancing diagnostic accuracy in neurological care. By analyzing complex patterns within brain data that may elude human observers, ML algorithms can identify subtle biomarkers of neurological dysfunction, potentially enabling earlier intervention and personalized treatment strategies [24] [79]. This review provides a comprehensive comparison of how ML leverages features from three prominent neuroimaging modalities—fMRI, EEG, and fNIRS—for diagnostic classification of neurological disorders, examining their respective capabilities, optimal applications, and performance metrics within clinical neuroscience research and drug development contexts.

Neuroimaging Modalities: Technical Specifications and Clinical Applications

Table 1: Technical Specifications and Clinical Applications of Neuroimaging Modalities

Specification	fMRI	EEG	fNIRS
Spatial Resolution	High (millimeter-level) [5]	Low (5-9 cm) [80]	Moderate (2-3 cm) [80]
Temporal Resolution	Low (0.33-2 Hz) [5]	High (>1000 Hz) [80]	Moderate (≤10 Hz) [80]
Penetration Depth	Whole head (deep structures) [5] [80]	Brain cortex [80]	Superficial cortex [5] [80]
Portability	None [80]	High (portable systems available) [80]	Yes [5] [80]
Cost	High [80]	Low [80]	Low [80]
Primary Measured Signal	Blood Oxygen Level Dependent (BOLD) [5]	Electrical activity from neuronal firing [81]	Hemodynamic changes (HbO/HbR) [82] [5]
Key Clinical Strengths	Localizing deep brain structures; whole-brain network analysis [5] [83]	Capturing rapid neural dynamics; epilepsy monitoring [24]	Bedside monitoring; naturalistic settings; patient-friendly [5] [84]
Motion Artifact Robustness	Limited [5] [80]	Limited [82] [80]	Very good [82] [80]

Complementary Strengths and Synergistic Potential

Each modality offers distinct advantages for neurological assessment. fMRI provides unparalleled spatial resolution for mapping brain activity across both cortical and subcortical structures, making it invaluable for disorders involving deep brain regions [5] [83]. EEG captures neuronal electrical activity with millisecond precision, ideal for studying brain dynamics and transient events like seizures, though it suffers from limited spatial resolution and susceptibility to electrical noise [24] [82]. fNIRS measures hemodynamic responses through near-infrared light, offering a balance between spatial and temporal resolution with superior motion tolerance, enabling studies in naturalistic environments and with challenging populations [5] [84].

The complementary nature of these signals has motivated research into multimodal integration, particularly combining EEG's temporal precision with fNIRS's spatial stability, or fNIRS's portability with fMRI's anatomical precision [82] [5]. Such integration aims to overcome individual limitations while providing a more comprehensive picture of brain function for diagnostic classification [82] [80].

Machine Learning Approaches for Feature Selection and Classification

Feature Extraction and Selection Methodologies

Effective feature selection is crucial for handling the high dimensionality of neuroimaging data and improving model generalizability by reducing feature redundancy [24] [83].

EEG Feature Engineering: EEG analysis typically extracts features across multiple domains: frequency-based (power spectral density, band power ratios), time-based (event-related potentials), entropy measures, and complexity indices [24]. The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm has demonstrated particular effectiveness for EEG feature selection, efficiently identifying the most discriminative channels and features while reducing dimensionality [24]. Studies indicate frontal lobe channels are frequently selected across neurological disorders, suggesting their critical role as biomarkers [24].

fNIRS Feature Extraction: fNIRS features primarily focus on hemodynamic response characteristics, including mean oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) concentrations, signal slope, standard deviation, and temporal patterns [84] [80]. Advanced approaches incorporate brain network topological features (degree centrality, betweenness centrality, local efficiency) derived from functional connectivity patterns, which have shown strong discriminative power in mental health classification [84].

fMRI Connectome Analysis: Functional connectivity (FC) features derived from resting-state or task-based fMRI represent brain region interactions as connectomes [83]. These high-dimensional features (often thousands of connections per subject) require robust selection methods. Comparative studies show LASSO achieves superior feature stability (Kuncheva index: 0.74) and classification performance compared to Relief and ANOVA selection methods [83].

Multimodal Feature Fusion: Hybrid approaches combining EEG and fNIRS leverage their complementary nature through feature-level fusion (concatenating feature vectors) or decision-level fusion (combining classifier outputs) [82] [80]. Feature-level fusion typically yields higher accuracy but requires careful handling of dimensional disparities, while decision-level fusion offers better noise resistance [82].

Classification Algorithms and Performance

Table 2: Classifier Performance Across Neurological Disorders and Modalities

Disorder(s) / Task	Modality	Feature Approach	Classifier	Accuracy	Citation
Alzheimer's vs. Control	EEG	Multiple feature domains + LASSO selection	Linear Discriminant Analysis (LDA)	100%	[24]
Depression, MCI, Schizophrenia	EEG	Multiple feature domains + LASSO selection	Linear Discriminant Analysis (LDA)	84.67%	[24]
Four-class ND Classification	EEG	Multiple feature domains + LASSO selection	Linear Discriminant Analysis (LDA)	57.89%	[24]
Mental Workload (n-back)	EEG-fNIRS Hybrid	Functional Brain Connectivity (FBC) + HbO/HbR	Machine Learning	77-83%	[80]
Motor Imagery Task	EEG-fNIRS Hybrid	Deep Learning + Evidence Theory	Custom Fusion Model	83.26%	[75]
Mental Arithmetic Task	EEG-fNIRS Hybrid	Multi-domain features + Multi-level learning	Ensemble Classification	98.42%	[82]
Mental Health States (3-class)	fNIRS	Temporal + Network Topological features	MLP Stacking Ensemble	94%	[84]
Schizophrenia Classification	fMRI	Functional Connectivity Features	Logistic Regression + LASSO	91.85%	[83]

Classifier Selection Considerations: Linear Discriminant Analysis (LDA) frequently achieves strong performance with neuroimaging data, particularly for binary classification tasks, due to its effectiveness with continuous features and computational efficiency [24] [80]. For more complex, multi-class problems (distinguishing between multiple neurological disorders), ensemble methods and advanced neural networks typically yield superior results [82] [84]. Deep learning approaches offer automatic feature learning but require large datasets and substantial computational resources, potentially limiting clinical applicability [24] [85].

Experimental Protocols and Methodologies

Protocol 1: EEG-Based Classification of Multiple Neurological Disorders

Dataset: 40 Alzheimer's patients, 43 healthy controls, 42 schizophrenia patients, 28 MCI patients, and 28 depression patients [24]. EEG recordings utilized 19 electrodes following the international 10-20 system at 500 Hz sampling rate [24].

Preprocessing: Signals were normalized (0-1 range), filtered with a 50 Hz notch filter to eliminate mains interference, and a Butterworth bandpass filter (1-30 Hz) was applied [24].

Feature Extraction and Selection: Multiple features were extracted from time, frequency, entropy, and complexity domains. The LASSO algorithm was employed for feature selection to identify the most discriminative EEG channels and features [24].

Classification: Two-class (disease-disease, healthy-control-disease), three-class, and four-class classifications were conducted using LDA and other ML classifiers with selected features [24].

Key Findings: LDA achieved perfect separation (100% accuracy) between healthy controls and Alzheimer's patients. Multi-class classification performance decreased with increasing class complexity (84.67% for three-class, 57.89% for four-class), reflecting the challenge of differentiating disorders with overlapping symptoms [24].

Protocol 2: Hybrid EEG-fNIRS for Mental Workload Classification

Dataset: 26 healthy participants completing n-back tasks (0-, 2-, and 3-back) with simultaneous 30-channel EEG and 36-channel fNIRS recordings [80].

Preprocessing: EEG data was resampled to 200 Hz and processed using improved weight-adjusted second-order blind identification. fNIRS signals were processed at 10 Hz sampling rate, converting optical density to HbO and HbR concentrations using the modified Beer-Lambert equation, then filtered (0-0.04 Hz Butterworth filter) and baseline-corrected [80].

Feature Extraction: EEG features included univariate spectral power and bivariate functional brain connectivity in delta, theta, and alpha bands. fNIRS features included HbO and HbR biomarkers. A 5-second sliding window with 1-second steps was used for analysis [80].

Classification and Findings: Combining EEG-based functional connectivity with fNIRS hemodynamic features significantly improved classification performance (77% accuracy for 0-back vs. 2-back; 83% for 0-back vs. 3-back) compared to either modality alone. Topographic analysis revealed different discriminative regions for EEG (posterior area, particularly POz in alpha band) and fNIRS (right frontal region, AF8) [80].

Protocol 3: fMRI Connectome-Based Classification with Feature Stability Analysis

Dataset: 27 schizophrenia patients and 27 healthy controls (Dataset 1); additional dataset with 172 subjects (Dataset 2) [83].

Feature Extraction: Functional connectomes were constructed from fMRI data, representing functional connectivity between brain regions as a matrix. Each subject's connectome was vectorized into a high-dimensional feature vector (3,403 features in Dataset 1; 13,366 in Dataset 2) [83].

Feature Selection: Three feature selection methods were compared: LASSO (embedded method), ANOVA (filter-based), and Relief (wrapper-based). Feature stability was quantified across cross-validation folds using Kuncheva and Jaccard indices [83].

Classification and Findings: LASSO achieved superior classification accuracy (91.85%) and feature stability (Kuncheva index: 0.74) compared to other methods. Stable feature selection identified reproducible functional connectivity biomarkers, particularly involving prefrontal and temporal regions, contributing to reliable schizophrenia classification [83].

Visualizing Experimental Workflows

Multimodal Neuroimaging Classification Pipeline

Multimodal Fusion Strategies for EEG and fNIRS

Table 3: Essential Research Materials and Analytical Tools for Neuroimaging ML Research

Category	Item	Specification / Function	Representative Use
Data Acquisition	EEG System with Electrodes	19+ electrodes following 10-20 system; 500+ Hz sampling rate [24]	Recording electrical brain activity [24]
	fNIRS System with Optodes	Near-infrared light sources/detectors; measures HbO/HbR changes [82] [84]	Monitoring hemodynamic responses [84] [80]
	fMRI Scanner	High-field magnet (1.5T+); BOLD signal detection [5] [83]	Whole-brain functional connectivity mapping [83]
Experimental Paradigms	N-back Task	Working memory assessment with varying difficulty levels [80]	Mental workload classification [80]
	Motor Imagery (MI)	Imagination of movement without execution [82] [75]	Motor disorder rehabilitation research [82]
	Mental Arithmetic (MA)	Cognitive task involving arithmetic calculations [82] [84]	Assessing cognitive function in depression [82]
	Verbal Fluency Task (VFT)	Language production and executive function test [84]	Mental health state classification [84]
Data Processing	LASSO Feature Selection	Regularization method for high-dimensional data [24] [83]	Identifying discriminative neuroimaging features [24]
	Functional Brain Connectivity	Measuring statistical interdependence between brain regions [80] [83]	Network-based disorder classification [80]
	Modified Beer-Lambert Law	Converting optical density to HbO/HbR concentrations [80]	fNIRS signal quantification [80]
Classification Algorithms	Linear Discriminant Analysis	Finds linear combinations for class separation [24] [80]	High-accuracy binary classification [24]
	Multi-Layer Perceptron Ensemble	Neural network with multiple hidden layers [84]	Complex multi-class problems [84]
	Evidence Theory / DST	Modeling uncertainty in decision fusion [75]	Multimodal classifier combination [75]

Machine learning applied to neuroimaging data represents a paradigm shift in neurological disorder diagnosis, offering objective, data-driven approaches that complement traditional clinical assessment. The optimal modality selection depends on specific diagnostic requirements: fMRI for disorders requiring deep brain structure analysis, EEG for capturing rapid neural dynamics, and fNIRS for naturalistic settings and longitudinal monitoring. Hybrid systems, particularly EEG-fNIRS, demonstrate synergistic potential by combining complementary signal characteristics.

Critical success factors include appropriate feature selection strategies (LASSO showing particular promise for stability and performance), classifier selection aligned with diagnostic complexity, and sufficient dataset sizes for model training. As multimodal integration methodologies advance and computational power increases, ML-driven diagnostic classification is poised to enhance early detection, personalized treatment planning, and objective monitoring of therapeutic efficacy in neurological care, ultimately improving patient outcomes across the spectrum of brain disorders.

Validating Diagnostic Efficacy and Comparative Performance Across Modalities

Validating the spatial specificity of functional Near-Infrared Spectroscopy (fNIRS) against the gold standard of functional Magnetic Resonance Imaging (fMRI) is a critical step for establishing its utility in both basic neuroscience and clinical diagnostics. While fMRI provides high-resolution spatial maps of brain activity, its cost, immobility, and sensitivity to motion artifacts limit its application in naturalistic settings and with certain patient populations [77] [47]. fNIRS, which measures the same hemodynamic response as fMRI via near-infrared light, offers a portable, cost-effective, and more robust alternative [86] [47]. However, its adoption hinges on a clear understanding of its spatial accuracy and limitations. This guide synthesizes evidence from concurrent fNIRS-fMRI studies to objectively compare their performance, with a specific focus on spatial specificity across motor and cognitive tasks, providing researchers with the experimental data and methodologies needed to inform their neuroimaging tool selection.

Comparative Performance Data: fNIRS vs. fMRI

The following tables summarize key quantitative findings from recent validation studies, highlighting the spatial relationship and performance metrics between fNIRS and fMRI.

Table 1: Spatial Correspondence in Motor and Cognitive Tasks

Study Task Modality	Key Brain Regions	Spatial Correlation Finding	Notes on Specificity
Motor Execution & Imagery [21]	Primary Motor Cortex (M1), Premotor Cortex (PMC)	fMRI activation clusters were successfully identified using subject-specific fNIRS signals as predictors. No significant difference in spatial correspondence between HbO and HbR.	The spatial correspondence was statistically equivalent for motor execution and motor imagery tasks.
Multiple Cognitive Tasks [19] [87]	Frontal and Parietal Cortices	The fNIRS signal correlated most strongly with the BOLD response from an elliptical region between the emitter and detector at an average depth of ~14 mm.	Correlation strength was influenced by scalp-brain distance and signal-to-noise ratio (SNR).
Resting-State Connectivity [7]	Broad cortical coverage (Frontal, Temporal, Parietal, Occipital)	Brain fingerprinting accuracy with fNIRS reached 75%–98%, approaching the 99.9% accuracy of fMRI, depending on the number of runs and brain regions used.	Demonstrated that fNIRS can capture unique, individual-specific functional connectivity patterns.

Table 2: Key Performance Characteristics and Limitations

Parameter	fNIRS	fMRI
Spatial Resolution	Lower (1-3 cm) [77]; limited to cortical surfaces [47]	High (millimeter-level); whole-brain including subcortical structures [77]
Temporal Resolution	Superior (millisecond-level precision) [77]	Lower (0.33-2 Hz), constrained by hemodynamic response lag [77]
Portability & Cost	Portable, cost-effective, easy to use [86] [47]	Immobile, expensive, requires specialized facilities and operators [77] [47]
Primary Signal Measured	Concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) [47]	Blood Oxygen Level-Dependent (BOLD) signal, primarily reflecting changes in deoxygenated hemoglobin [19]
Key Limitation	Limited penetration depth; inability to image subcortical structures [77] [47]	Sensitivity to motion artifacts; loud, restrictive environment; not suitable for all populations (e.g., with metal implants) [77] [47]

Experimental Protocols in Key Validation Studies

The robust spatial correspondence outlined in the tables above is derived from carefully controlled experimental protocols. Below are the methodologies from two pivotal studies that exemplify the approach to multimodal validation.

This study exemplifies a protocol designed to rigorously assess spatial correspondence in the motor network.

• Participants: 9 healthy volunteers.
• Paradigm: An asynchronous block design was used, where participants performed separate sessions for fMRI and fNIRS. The paradigm included:
  • Motor Action (MA): Bilateral finger tapping sequences.
  • Motor Imagery (MI): Imagining the same sequence without movement.
  • Baseline: Rest periods.
• fMRI Acquisition: Data were acquired on a 3T Siemens scanner. A high-resolution T1-weighted image was obtained for coregistration. Functional images (26 slices) were acquired with an echo-planar imaging sequence focused on motor areas.
• fNIRS Acquisition: A portable NIRSport2 system with 16 sources and 15 detectors was used, creating 54 channels over bilateral motor areas. Short-distance detectors (8 mm) were incorporated to mitigate extracerebral confounds.
• Data Analysis: The core analysis involved using the preprocessed, subject-specific fNIRS signals (HbO, HbR, HbT) as predictors in a General Linear Model of the concurrently acquired fMRI data. This innovative method tested the ability of fNIRS to identify fMRI activation clusters in individually defined primary motor and premotor cortices.

This pioneering work provided a comprehensive comparison across multiple cognitive domains using simultaneous data acquisition.

• Participants: 13 healthy adults.
• Paradigm: Participants performed a battery of four cognitive tasks during simultaneous fNIRS-fMRI recording:
  • Left Finger Tapping (tap): A block-design motor task.
  • Go/No-Go (nog): A response inhibition task.
  • Judgment of Line Orientation (jlo): A visuospatial processing task.
  • Visuospatial N-Back (vis): A working memory task.
• Simultaneous Acquisition: NIRS probes were placed over frontal and parietal regions while fMRI data was collected. This allowed for a direct, temporal-domain correlation between the signals.
• Data Analysis: The study investigated both temporal correlations and spatial specificity. Analyses focused on how factors like signal-to-noise ratio (SNR) and scalp-brain distance affected the correlation between NIRS (oxy-Hb and deoxy-Hb) and fMRI (BOLD) signals. The spatial area of highest correlation was mapped.

Signaling Pathways and Neurovascular Coupling

Both fNIRS and fMRI measure hemodynamic changes consequent to neural activity, a process known as neurovascular coupling. The following diagram illustrates the shared physiological pathway and the distinct physical principles each technology uses to measure it.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Equipment and Software for fNIRS-fMRI Validation Studies

Item Name	Function / Application	Example from Literature
Simultaneous fNIRS-fMRI System	Enables direct temporal correlation and validation of fNIRS signals against the fMRI gold standard within the same experimental session.	Studies use specialized fNIRS equipment compatible with the high-electromagnetic field environment inside the MRI scanner [19] [7].
Continuous-Wave fNIRS Device	Measures relative changes in hemoglobin concentrations using near-infrared light. The most common type used in cognitive studies.	NIRScout [7] and NIRSport2 [21] systems are cited, typically with dual wavelengths (~760 & 850 nm).
3T MRI Scanner	Provides high-resolution structural and functional (BOLD) images for spatial localization and correlation analysis.	A 3T Siemens Magnetom TimTrio scanner was used for its high signal-to-noise ratio and spatial precision [21].
Digitization & Probe Placement System	Critical for accurate spatial registration of fNIRS optodes to anatomical locations. Allows for projection of fNIRS channels onto cortical surfaces.	Magnetic motion tracking sensors or 3D digitalizers are used with software like AtlasViewer for co-registration based on the 10-20 system [21] [7].
Short-Distance Detectors	Optodes placed close to sources (<1 cm) to measure and subsequently remove signals originating from the scalp and skull, improving the specificity of brain signals.	Used in motor task studies to separate cerebral hemodynamics from superficial confounds [21].
Homer3 / AtlasViewer Software	A standard fNIRS data processing suite (HOMER) for converting raw signals, correcting for motion artifacts, and calculating hemoglobin changes. AtlasViewer is used for spatial registration.	Employed for preprocessing pipelines including motion correction and registration to an anatomical atlas [21] [7].
General Linear Model (GLM) Tools	Standard statistical framework (in software like SPM, BrainVoyager) for analyzing fMRI and fNIRS task-related activity and determining spatial overlap.	Used to model fMRI data with fNIRS predictors and to identify significant activation clusters in both modalities [21].

Accurately diagnosing Disorders of Consciousness (DOC), such as Minimally Conscious State (MCS) and Unresponsive Wakefulness Syndrome (UWS), represents a major challenge in clinical neurology. Behavioral misdiagnosis rates remain unacceptably high, approximately 41%, due to patient motor impairments and fluctuating vigilance [26]. Functional neuroimaging offers an objective path forward. While functional Magnetic Resonance Imaging (fMRI) and electroencephalography (EEG) have proven valuable, functional Near-Infrared Spectroscopy (fNIRS) is emerging as a powerful, portable alternative. This guide objectively compares the diagnostic performance of fNIRS against fMRI and EEG in differentiating MCS from UWS, providing researchers with the experimental data and methodologies needed to evaluate these technologies.

Performance Comparison of Neuroimaging Modalities

The following table summarizes the key performance characteristics of fNIRS, fMRI, and EEG for DOC diagnosis, highlighting their respective advantages and limitations.

Table 1: Performance Comparison of Neuroimaging Modalities in DOC Diagnosis

Feature	fNIRS	fMRI	EEG
Spatial Resolution	Moderate (2-3 cm) [88]	High (1-2 mm) [26]	Low [89]
Temporal Resolution	Good (0.1-1 Hz) [88]	Low (0.5-1 Hz)	Excellent (ms range)
Portability/Bedside Use	Excellent [26] [88]	Poor [26]	Excellent
Tolerance to Motion	Good [26] [90]	Poor [26]	Moderate (susceptible to noise) [26]
Patient Compatibility	Excellent (tolerates metal implants) [26]	Poor (incompatible with many implants) [89]	Excellent
Equipment Cost	Low to Moderate [88]	Very High [91]	Low
Representative Classification Accuracy (MCS vs. UWS)	76.9% (Functional Connectivity) [26]	N/A in results	N/A in results

fNIRS in Practice: Experimental Data and Protocols

fNIRS demonstrates diagnostic utility through two primary paradigms: resting-state functional connectivity and active task-based assessments.

Resting-State Functional Connectivity

Resting-state fNIRS measures spontaneous brain activity without patient participation, making it widely applicable.

Table 2: Key Findings from Resting-State fNIRS Studies in DOC

Study Design	Key Findings	Classification Performance
52 DOC patients (26 MCS, 26 UWS) & 49 Healthy Controls (HC) [26]	VS/UWS patients showed significantly reduced functional connectivity compared to MCS in prefrontal, premotor, and sensorimotor regions, and within auditory, frontoparietal, and default mode networks (DMN).	- Channel 4-29 connectivity: 76.92% accuracy, AUC=0.818- Auditory network: 73.08% accuracy, AUC=0.803
60 pDOC patients [91]	Functional connectivity between frontal and occipital lobes was decreased in UWS/VS compared to MCS and EMCS (Escape MCS) groups.	Not specified, but reported as a significant diagnostic reference.

The experimental workflow for a typical resting-state fNIRS study is systematic and can be visualized as follows:

Active Task-Based Paradigms

Active paradigms instruct patients to perform mental tasks to detect volitional brain activity, which can reveal Cognitive Motor Dissociation (CMD)—where patients are consciously aware but unable to move.

Table 3: Key Findings from Active Task-Based fNIRS Studies

Task Paradigm	Study Population	Key Findings
Hand-Open-Close Motor Imagery [88]	70 prolonged DOC patients	Identified 7 CMD patients (4 from VS/UWS, 3 from MCS-). CMD patients had more favorable 6-month outcomes.
Tongue Motor Command [89]	36 acute DOC patients & 19 HC	Cortical activation was detected in 44% of patients (62% of MCS, 35% of UWS).

The following diagram outlines a standard active task-based fNIRS protocol:

The Scientist's Toolkit: Essential Reagents & Materials

Successful fNIRS experimentation requires a specific set of hardware, software, and assessment tools.

Table 4: Essential Research Reagent Solutions for fNIRS DOC Studies

Item	Specification / Example	Primary Function
fNIRS System	Continuous-wave systems (e.g., NirSmart-6000A, NirScan) [26] [91]	Emits NIR light and detects attenuated signal to measure hemodynamic changes.
Optode Cap	24 sources & 24 detectors (63 channels) based on 10-20 system [26]	Holds light sources and detectors in place for whole-brain coverage.
Data Acquisition Software	Manufacturer-specific (e.g., NirSpark) [91]	Records raw light intensity data at sampling rates ~11 Hz.
Preprocessing Toolbox	Homer2 [26], NIRS-SPM [90], NIRS-KIT [26]	Performs motion correction, filtering, and conversion to HbO/HbR.
Statistical Analysis Tool	MATLAB, Python, Atlas-Viewer [89]	Executes GLM, functional connectivity, and machine learning analyses.
Behavioral Scale	Coma Recovery Scale-Revised (CRS-R) [26]	Provides gold-standard behavioral diagnosis for correlation with fNIRS data.
Short-Separation Channels	~0.8 cm source-detector distance [92]	Measures systemic superficial noise for enhanced signal regression.

Technical Considerations and Noise Reduction

A significant challenge in fNIRS is distinguishing neuronal signals from physiological noise (e.g., pulse, blood pressure changes). Advanced processing is critical. One effective method uses Principal Component Analysis (PCA) spatial filtering to remove global systemic artifacts, which has been shown to reveal neural activity in Broca's area during overt speech that was otherwise obscured [90]. Furthermore, incorporating short-separation channels and auxiliary measurements (e.g., pulse, respiration) within a General Linear Model (GLM) framework provides an automated, robust denoising pipeline for whole-head recordings, significantly improving the detection of focal brain activations [92].

fNIRS establishes itself as a diagnostically valuable and practical tool for differentiating MCS from UWS. It bridges a critical gap between the high spatial resolution of fMRI and the high temporal resolution of EEG, while offering unparalleled portability for bedside assessment. Evidence shows fNIRS can achieve classification accuracies exceeding 76% based on resting-state functional connectivity [26] and can successfully identify covert consciousness in behaviorally unresponsive patients [88] [89]. For researchers and clinicians, fNIRS provides a viable, cost-effective, and reliable modality to enhance diagnostic accuracy, guide prognosis, and ultimately improve patient care in the complex field of disorders of consciousness.

The rising global prevalence of neurological and psychiatric disorders underscores an urgent need for objective, accurate, and accessible diagnostic tools. Alzheimer's disease (AD), schizophrenia, and depression represent a significant portion of this disease burden, often presenting with overlapping symptoms that complicate differential diagnosis [24]. Traditional diagnostic methods, which rely heavily on clinical examinations and subjective neuropsychological assessments, are frequently supplemented by neuroimaging. Techniques like functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) offer valuable insights but are limited by high cost, low portability, and operational complexity [93] [5]. In this landscape, electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) have emerged as promising, non-invasive, and cost-effective alternatives. Both modalities are increasingly paired with machine learning (ML) to develop computer-aided diagnostic systems. This review objectively compares the reported classification accuracies of EEG and fNIRS ML models for AD, schizophrenia, and depression, detailing experimental protocols and situating these findings within the broader pursuit of clinical diagnostic accuracy.

Performance Comparison: EEG vs. fNIRS in Disease Classification

Direct, within-study comparisons of EEG and fNIRS on the same patient cohort are rare. However, an analysis of recent high-quality studies reveals the performance benchmarks each modality can achieve independently. The following tables summarize quantitative results from key experiments, providing a clear basis for comparison.

Table 1: Key Performance Metrics of EEG-based Machine Learning Classification

Disorders Classified	Feature Extraction Methods	Classifier(s)	Highest Reported Accuracy	Reference
AD vs. Control	Frequency axis features, power, coherence; SVM-RFE, PCA	ANN, SVM	86% (ANN)	[24]
Depression vs. Control	Power Spectral Density (PSD), Higuchi Fractal Dimension	Logistic Regression	92%	[24]
Schizophrenia vs. Control	Statistical + Wavelet features	Decision Tree	97.98%	[24]
Depression vs. MCI vs. Schizophrenia	Time, frequency, entropy, complexity features; Lasso feature selection	LDA	84.67%	[24]
Depression vs. MCI vs. Schizophrenia vs. AD	Time, frequency, entropy, complexity features; Lasso feature selection	LDA	57.89%	[24]

Table 2: Key Performance Metrics of fNIRS-based Machine Learning Classification

Disorders Classified	Task Paradigm & Brain Region	Classifier/Model	Highest Reported Accuracy / AUC	Reference
MCI vs. Healthy Control	N-Back, Verbal Fluency; Whole-head neural activity	Machine Learning Model	AUC: 0.92	[46]
Anxiety vs. Depression	Emotional Autobiographical Memory Task (EAMT); dlPFC	Cascaded Feedforward Neural Network	95.2%	[94]
Health vs. Anxiety vs. Depression	Emotional Autobiographical Memory Task (EAMT); dlPFC	Cascaded Feedforward Neural Network	90.4%	[94]
MCI vs. AD	Various task-based and resting-state paradigms	Machine Learning Models	Up to 90%	[93]

Experimental Protocols and Methodologies

The high accuracy of ML models is contingent on the underlying experimental design, including the choice of paradigm, brain regions of interest, and feature extraction methods.

EEG Experimental Protocols

A representative study achieving high multi-class classification performance [24] employed a rigorous methodology. The protocol used a publicly available dataset with recordings from 19 electrodes according to the international 10-20 system. Participants underwent EEG monitoring during periods of rest with eyes open and closed.

• Preprocessing: Raw signals were normalized, a 50 Hz notch filter was applied to remove mains interference, and a Butterworth bandpass filter (1–30 Hz) was used to isolate relevant frequency bands.
• Feature Extraction: A comprehensive set of features was extracted from the EEG signals across multiple domains: time, frequency, entropy, and complexity measures.
• Feature Selection: The Least Absolute Shrinkage and Selection Operator (Lasso) algorithm was used for effective feature selection, identifying the most informative EEG channels and mitigating overfitting. Frontal lobe channels were frequently selected, highlighting their critical role in classifying these neurological disorders.
• Classification: The selected features were used to train various ML algorithms, with Linear Discriminant Analysis (LDA) emerging as the top performer for multi-class tasks.

This study's focus on disease-to-disease classification, rather than just patient versus healthy control, provides a more clinically relevant assessment of diagnostic specificity [24].

fNIRS Experimental Protocols

fNIRS studies often leverage specific cognitive tasks to elicit hemodynamic responses in targeted brain regions. A notable study on MCI [46] used a Time-Domain fNIRS (TD-fNIRS) system, which offers better sensitivity to brain activations compared to continuous-wave systems.

• Task Paradigm: Participants performed the N-Back task (to probe working memory) and the Verbal Fluency Task (VFT) (to probe language production). These domains are frequently affected in MCI and AD.
• Data Acquisition: The TD-fNIRS headset measured changes in oxygenated hemoglobin (HbO) concentration across the whole head during task performance.
• Feature Integration: The ML model was trained using a combination of input features: self-reported impairment, behavioral task performance (e.g., accuracy, reaction time), and neural activation metrics derived from the fNIRS signals.
• Result: The model achieved the strongest classification performance (AUC=0.92) only when neural metrics were included, demonstrating the value of fNIRS-derived brain activity data beyond behavioral measures alone [46].

For depression and anxiety, a novel approach [94] used an Emotional Autobiographical Memory Task (EAMT) to stimulate brain activity specific to emotional processing.

• Targeted Brain Region: The study focused specifically on the dorsolateral prefrontal cortex (dlPFC), a region with high specificity for emotional regulation.
• Feature and Model: A Cascaded Feedforward Neural Network was trained on the average oxyhemoglobin (Avg_HbO) signal from the dlPFC. The use of a targeted brain region and a lightweight model aimed to balance high accuracy with low computational redundancy.

Technical Comparison and Clinical Applicability

EEG and fNIRS measure fundamentally different physiological signals: EEG records electrical activity from populations of neurons, while fNIRS measures hemodynamic responses (changes in HbO and deoxygenated hemoglobin (HbR)) correlated with neural activity. This core difference dictates their respective strengths, limitations, and suitability for clinical environments.

Table 3: Technical and Practical Comparison of EEG and fNIRS

Parameter	EEG	fNIRS
Measured Signal	Electrical potential from synaptic activity	Hemodynamic (blood oxygenation) response
Temporal Resolution	Very High (milliseconds)	High (0.1 - 10 Hz)
Spatial Resolution	Low (source localization challenges)	Moderate (1 - 3 cm)
Portability	Moderate	High
Tolerance to Motion Artifacts	Moderate	High
Key Clinical Strength	Excellent for capturing rapid neural oscillations; lower cost.	Suitable for naturalistic settings and populations prone to movement (e.g., children, elderly).
Key Clinical Limitation	Susceptible to muscle and eye movement artifacts.	Limited to monitoring superficial cortical regions; cannot probe subcortical structures.

The portability and motion tolerance of fNIRS make it particularly suitable for bedside monitoring, longitudinal studies in naturalistic settings, and use with populations that may find fMRI confinement challenging, such as children or individuals with AD [95] [5]. EEG, while also portable, is more susceptible to artifacts from muscle activity and head movement [95]. A critical challenge for fNIRS is its limited depth penetration, restricting its measurement to the cerebral cortex and preventing direct investigation of subcortical structures implicated in many neurological disorders [5].

The Scientist's Toolkit: Essential Research Reagents and Materials

The following table details key materials and computational tools essential for conducting EEG or fNIRS research in this domain.

Table 4: Essential Research Reagents and Solutions for Neuroimaging Studies

Item Name	Function/Brief Explanation	Example/Note
EEG Cap with Electrodes	Holds electrodes in standard positions (10-20 system) for consistent scalp recording.	Typically includes 19 or more electrodes.
fNIRS Probes (Source-Detector Pairs)	Emits near-infrared light and detects its attenuation after passing through brain tissue.	Configurations can be customized for specific cortical regions of interest.
Conductive Electrode Gel	Reduces impedance between EEG electrodes and the scalp, improving signal quality.	Essential for high-fidelity EEG data acquisition.
NIRx NIRScout / TechEN CW6	Examples of commercially available fNIRS systems for conducting experiments.	Other brands include Hamamatsu, Kernel Flow2 [46].
Butterworth Bandpass Filter	Standard digital filter for removing noise outside the frequency range of interest.	Used in both EEG (e.g., 1-30 Hz) and fNIRS (e.g., 0.01-0.1 Hz) preprocessing.
Lasso (L1 Regularization)	A feature selection algorithm that penalizes less informative features, enhancing model generalizability.	Critical for managing high-dimensional EEG data [24].
Linear Discriminant Analysis (LDA)	A classic classifier that finds a linear combination of features to best separate classes.	Often a strong performer in neuroimaging classification tasks [24].

Experimental and Analytical Workflow

The process of developing a machine learning classifier for neurological disorders, common to both EEG and fNIRS, follows a structured pipeline. The key stages from data acquisition to model validation are outlined below.

Both EEG and fNIRS, when combined with machine learning, demonstrate considerable promise for the objective classification of Alzheimer's disease, schizophrenia, and depression. The choice between them is not a matter of one being universally superior, but rather depends on the specific clinical or research context.

• EEG has proven capable of high accuracy, particularly in distinguishing schizophrenia from controls and in complex multi-class scenarios. Its main advantages are excellent temporal resolution and lower cost.
• fNIRS offers a compelling combination of portability, tolerance to motion, and good spatial resolution for cortical regions, achieving high classification accuracy, especially when using advanced systems like TD-fNIRS and targeted cognitive or emotional paradigms.

Future progress in the field hinges on several key factors: the development of standardized analysis pipelines to improve reproducibility [73], the fusion of multimodal data (e.g., combined EEG-fNIRS) to leverage their complementary strengths [69], and a focus on collecting larger, shared datasets to train more robust and generalizable models. As these technologies and analytical techniques continue to mature, they hold the potential to significantly enhance diagnostic precision and enable earlier intervention for patients suffering from debilitating neurological and psychiatric disorders.

The pursuit of objective biomarkers for neurological disorders represents a central challenge in modern neuroscience. For researchers and drug development professionals, selecting the appropriate neuroimaging tool requires a critical balance between diagnostic accuracy and practical considerations for clinical deployment. Functional magnetic resonance imaging (fMRI), electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS) have emerged as leading non-invasive techniques, each with distinct technical and operational profiles. This guide provides a systematic comparison of these modalities, focusing on their performance characteristics, validation status, and practical implementation in both research and clinical settings. Understanding these trade-offs is essential for designing efficient studies, allocating resources effectively, and accelerating the translation of neuroimaging biomarkers from laboratory discovery to clinical application.

Technical Performance and Clinical Utility

Core Technical Characteristics

The fundamental operating principles of fMRI, EEG, and fNIRS dictate their specific performance advantages and limitations in clinical research.

• fMRI measures brain activity indirectly through the blood-oxygen-level-dependent (BOLD) signal, which reflects changes in blood oxygenation and flow coupled to neural activity. It provides high spatial resolution (millimeter range) throughout the entire brain, including deep subcortical structures such as the hippocampus and amygdala [96] [5]. However, its temporal resolution is limited (typically 0.33-2 Hz) by the slow hemodynamic response, which lags 4-6 seconds behind neural events [5].

• EEG records electrical potentials generated by synchronized postsynaptic activity of cortical pyramidal neurons. Its key strength is exceptional temporal resolution (milliseconds), enabling the capture of rapid neural dynamics, such as those seen in epilepsy. Its main limitation is poor spatial resolution (centimeter-level) due to the blurring effect of the skull and scalp on electrical signals [41] [97].

• fNIRS utilizes near-infrared light to measure concentration changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) in the superficial cortex. It offers a favorable balance, with better spatial resolution than EEG and better temporal resolution than fMRI (seconds). Its primary constraint is a limited penetration depth (~1-2.5 cm), restricting measurement to the outer cortical layers [86] [2].

Diagnostic Accuracy and Validation

The diagnostic value of a technique is measured by its sensitivity (ability to correctly identify true positives) and specificity (ability to correctly identify true negatives).

• fMRI has shown promise as a dynamic biomarker, particularly for Alzheimer's disease (AD). It can detect alterations in brain network connectivity years before the onset of structural atrophy or significant cognitive decline [96]. In irritable bowel syndrome (IBS), a structural MRI biomarker achieved 70% classification accuracy (68% sensitivity, 71% specificity) in a holdout sample, demonstrating reliability but also highlighting that this accuracy is insufficient for standalone diagnostic use [98].

• EEG, especially when augmented with artificial intelligence (AI), demonstrates high diagnostic performance for specific conditions. In nonconvulsive status epilepticus (NCSE), AI-integrated EEG systems achieved a pooled sensitivity of 95% and a specificity of 83% in a meta-analysis [99]. This high sensitivity is critical for ruling out the condition in emergency settings.

• fNIRS, combined with machine learning, shows strong discriminatory power in early-stage neurodegenerative diseases. A study using an fNIRS-based Support Vector Machine (SVM) model to differentiate Parkinson's disease (PD) patients from healthy controls achieved an accuracy of 85% and an area under the ROC curve (AUC) of 0.95, indicating excellent diagnostic potential [86].

Table 1: Comparison of Core Technical and Diagnostic Performance Characteristics

Feature	fMRI	EEG	fNIRS
What It Measures	BOLD signal (blood oxygenation)	Electrical potentials from neurons	Hemoglobin concentration changes (HbO, HbR)
Spatial Resolution	High (millimeter-level)	Low (centimeter-level)	Moderate (1-3 cm)
Temporal Resolution	Low (0.33-2 Hz)	Very High (milliseconds)	Moderate (seconds)
Penetration Depth	Whole brain (cortical & subcortical)	Superficial cortex	Superficial cortex (1-2.5 cm)
Sample Diagnostic Accuracy	70% classification in IBS [98]	95% sensitivity, 83% specificity in NCSE [99]	85% accuracy, AUC=0.95 in PD [86]
Key Clinical Strength	Mapping deep brain networks; high-resolution structure	Tracking rapid neural dynamics (seizures)	Monitoring cortical activity in naturalistic settings

Practical Considerations for Clinical Deployment

Beyond diagnostic performance, the practical aspects of deployment significantly influence a technology's adoption in clinical and research pipelines.

Cost, Portability, and Operational Workflow

The operational burden and financial cost of a technology are critical for large-scale studies and routine clinical use.

• fMRI is the most resource-intensive option. It requires expensive infrastructure (the scanner itself, a shielded room, cryogens), highly trained personnel, and has significant operational costs. Its lack of portability and confinement to a controlled laboratory environment limit its use for bedside monitoring or studies involving naturalistic behaviors [5].

• EEG systems are generally more affordable and portable than fMRI. Wireless systems are available, offering good flexibility. However, setup complexity can be moderate, often requiring electrode gel and skin preparation to ensure good signal quality. A significant practical drawback is its high sensitivity to motion artifacts, restricting its use in active or mobile participants [97].

• fNIRS offers a compelling practical profile. Its hardware is more portable and cost-effective than fMRI, and it is increasingly available in wearable formats. A key advantage is its tolerance to motion artifacts compared to EEG, making it suitable for studies in real-world settings, with children, or during rehabilitation exercises [97] [5]. Setup complexity is similar to EEG, but without the need for conductive gels in many systems.

Table 2: Practical Considerations for Clinical and Research Deployment

Practical Factor	fMRI	EEG	fNIRS
Relative Cost	Very High	Low to Moderate	Moderate
Portability	None (fixed installation)	High (wireless available)	High (wearable available)
Tolerance to Motion	Low	Low	Moderate to High
Typical Environment	Controlled laboratory	Controlled laboratory	Laboratory, clinic, real-world
Setup Complexity	High (safety, training)	Moderate (gel, prep)	Moderate (optode placement)
Participant Burden	High (confinement, noise)	Moderate (sitting still)	Low (more natural movement)

The Scan Time vs. Sample Size Trade-Off

A critical cost-benefit consideration in study design, particularly for fMRI-based Brain-Wide Association Studies (BWAS), is the trade-off between scan duration per participant and total sample size. Empirical modeling reveals that prediction accuracy of phenotypes increases with the total scan duration, defined as the product of sample size (N) and scan time per participant (T) [100].

• Interchangeability: For shorter scans (≤20 minutes), sample size and scan time are initially interchangeable; a larger sample can compensate for a shorter scan time and vice versa to achieve a similar prediction accuracy [100].
• Diminishing Returns: Beyond a certain point (e.g., >30 minutes), increasing scan time yields diminishing returns compared to increasing the sample size. Starting from 200 participants with 14-minute scans, a 3.5x larger sample (N=700) increased accuracy more than a 4.1x longer scan (T=58 min) [100].
• Cost Efficiency: When accounting for overhead costs per participant (e.g., recruitment, travel), longer scans can be more cost-effective than larger samples for improving prediction performance. Analyses suggest that 30-minute scans are, on average, the most cost-effective, yielding ~22% savings compared to 10-minute scans [100].

Experimental Protocols and Validation Frameworks

Protocol for Multimodal fNIRS-EEG Integration

The complementary nature of fNIRS and EEG makes their combined use a powerful approach. The standard protocol for concurrent recording involves several key steps [41] [97]:

• Synchronized Setup: Use an integrated cap with pre-defined placements for both EEG electrodes and fNIRS optodes based on the international 10-20 system. Care must be taken to avoid physical interference.
• Hardware Synchronization: Employ an external hardware trigger (e.g., TTL pulses) or shared clock system to synchronize the data acquisition streams from both devices at the start of the experiment.
• Stimulus Presentation: Deliver experimental stimuli using software that sends precise event markers to both the EEG and fNIRS recording systems simultaneously.
• Parallel Data Processing: Process the raw data from each modality through separate, optimized pipelines before integration.
  • EEG Processing: Typically includes filtering, artifact removal (e.g., eye blinks, muscle activity), and epoching.
  • fNIRS Processing: Typically includes converting raw light intensity to optical density, filtering motion artifacts, and using the Modified Beer-Lambert Law to calculate hemoglobin concentration changes [41].
• Data Fusion: Integrate the processed features from both modalities for analysis using techniques such as joint Independent Component Analysis (jICA), canonical correlation analysis (CCA), or machine learning models that combine input features from both EEG and fNIRS [97].

Framework for Biomarker Validation

For a neuroimaging biomarker to be clinically useful, it must progress through a rigorous validation process. The following criteria provide a framework for evaluation [98]:

• Diagnosticity: The biomarker must demonstrate high diagnostic performance (sensitivity and specificity) in classifying patients versus controls. The effect size should be substantial.
• Interpretability: The model and the brain features it uses should be interpretable within existing neuroscientific knowledge and supported by converging evidence from other methods (e.g., animal models, lesion studies).
• Deployability: The classification model and testing procedure must be precisely defined and standardized so it can be prospectively applied to new data in different settings without flexibility that introduces bias.
• Generalizability: The biomarker must be validated prospectively across different laboratories, scanner types, populations, and variants of testing conditions to define its boundary conditions and ensure broad utility.

Research Reagent Solutions

The following table details key hardware and software solutions essential for conducting neuroimaging studies in neurological disorders.

Table 3: Essential Research Reagents and Tools for Neuroimaging

Item / Solution	Function / Application
ETG-4000 fNIRS System	A continuous-wave fNIRS instrument for measuring hemodynamic responses in the prefrontal cortex and other brain regions; used for diagnostic model development in Parkinson's disease [86].
Integrated EEG-fNIRS Caps	Head caps with pre-defined placements for both EEG electrodes and fNIRS optodes, enabling synchronized multimodal data acquisition while minimizing hardware interference [97].
SVM (Support Vector Machine)	A machine learning algorithm used for classification (e.g., patients vs. controls); demonstrated high accuracy (85%) when applied to fNIRS data for Parkinson's disease detection [86].
Kernel Ridge Regression (KRR)	A machine learning algorithm used for individual-level phenotypic prediction in large-scale fMRI studies (e.g., BWAS) to optimize the scan-time/sample-size trade-off [100].
SHAP (SHapley Additive exPlanations)	A technique to enhance model interpretability by identifying which input features (e.g., specific fNIRS channels) contribute most to a model's diagnostic decision [86].
Tauvid (Flortaucipir F18)	An FDA-approved PET tracer for imaging tau pathology (neurofibrillary tangles) in the brain; represents a gold-standard biomarker for Alzheimer's pathology and catalyzes AI-based PET analysis [101].

Signaling Pathways and Experimental Workflows

Neurovascular Coupling Pathway

The physiological link between neuronal activity and the signals measured by fMRI and fNIRS.

Neuroimaging Biomarker Validation Workflow

A critical pathway from discovery to clinically deployed biomarker.

The cost-benefit analysis of fMRI, EEG, and fNIRS reveals a landscape with no single superior technology, but rather a set of tools with complementary strengths. fMRI provides the gold standard for spatial localization and deep brain imaging but at a high cost and with low temporal resolution. EEG is unparalleled for tracking millisecond-scale neural dynamics and is cost-effective, but suffers from poor spatial resolution and motion sensitivity. fNIRS occupies a strategic middle ground, offering a portable, motion-tolerant solution for mapping cortical hemodynamics with applicability in real-world settings.

For researchers and drug development professionals, the optimal choice depends heavily on the specific research question, the clinical context, and the practical constraints of the study population and environment. Furthermore, the emerging paradigm is not one of competition, but of integration. Multimodal approaches, such as combined EEG-fNIRS or fMRI-fNIRS, leverage the strengths of each modality to provide a more comprehensive picture of brain function. As biomarker validation frameworks become more rigorous and computational methods like AI continue to advance, the strategic selection and integration of these neuroimaging tools will be pivotal in developing robust, clinically deployable diagnostics for neurological disorders.

Conclusion

The integration of fMRI, EEG, and fNIRS represents a paradigm shift in diagnosing neurological disorders, leveraging their complementary strengths to overcome individual limitations. fMRI provides unparalleled spatial resolution for deep brain structures, EEG offers millisecond temporal precision for dynamic brain activity, and fNIRS delivers a portable, robust solution for naturalistic monitoring and bedside applications. The future of clinical neuroimaging lies in standardized multimodal protocols, advanced hardware compatible with MRI environments, and AI-driven data fusion to create a comprehensive picture of brain function. These advancements promise to enhance early detection, personalize therapeutic strategies, and provide objective biomarkers for drug development, ultimately improving patient outcomes in neurology.

References

• 1. Functional Magnetic Resonance Imaging and Functional Near ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5563305/]
• 2. Functional Near-Infrared Spectroscopy and Its Clinical ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.00724/full]
• 3. Uses, misuses, new uses and fundamental limitations of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5003851/]
• 4. A Narrative Review on Clinical Applications of fNIRS - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7573058/]
• 5. Applications and advances of combined fMRI-fNIRs ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1542075/full]
• 6. Spatial correspondence of cortical activity measured with ... [https://www.sciencedirect.com/science/article/pii/S1053811924000648]
• 7. a simultaneous fNIRS-fMRI study [https://pmc.ncbi.nlm.nih.gov/articles/PMC9896013/]
• 8. The oxygenation response to functional stimulation [https://www.sciencedirect.com/science/article/abs/pii/S1053811907000304]
• 9. A systematic review on the role of EEG and fMRI ... [https://pubmed.ncbi.nlm.nih.gov/40300301/]
• 10. Influence of skin blood flow on near-infrared spectroscopy ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811911005003]
• 11. A Cross-Disciplinary Review of the fNIRS-EEG Dual-Modality ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11506513/]
• 12. Bimodal EEG-fNIRS in Neuroergonomics. Current ... [https://www.frontiersin.org/journals/neuroergonomics/articles/10.3389/fnrgo.2022.934234/full]
• 13. Electroencephalography Signal Processing [https://pmc.ncbi.nlm.nih.gov/articles/PMC10385593/]
• 14. Exploring easily accessible neurophysiological biomarkers for ... [https://alzres.biomedcentral.com/articles/10.1186/s13195-024-01607-4]
• 15. Biomarkers in neurological disorders: a fast-growing market [https://pmc.ncbi.nlm.nih.gov/articles/PMC8096602/]
• 16. Abnormal resting-state EEG neural oscillations and functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12463927/]
• 17. EEG Signal Processing for Dummies [https://www.neuroelectrics.com/blog/eeg-signal-processing-for-dummies]
• 18. EEG-based functional connectivity patterns during ... [https://www.nature.com/articles/s41598-025-19245-7]
• 19. A quantitative comparison of NIRS and fMRI across multiple ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3021967/]
• 20. The present and future use of functional near‐infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6367070/]
• 21. Multimodal assessment of the spatial correspondence ... [https://www.nature.com/articles/s41598-023-29123-9]
• 22. Comparing structure–function relationships in brain ... [https://www.nature.com/articles/s41598-024-79817-x]
• 23. fMRI complexity matches tau PET performance [https://pubmed.ncbi.nlm.nih.gov/40799842/]
• 24. Enhancing Diagnostic Accuracy of Neurological Disorders ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12427693/]
• 25. Predicting upper limb motor dysfunction after ischemic stroke [https://pmc.ncbi.nlm.nih.gov/articles/PMC12150403/]
• 26. Functional connectivity in whole-brain and network analysis ... [https://translational-medicine.biomedcentral.com/articles/10.1186/s12967-025-07181-z]
• 27. Linear predictive coding electroencephalography ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12265758/]
• 28. Recent advances in neuroimaging of Alzheimer's disease and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12417339/]
• 29. The GREENBEAN checklist for reporting studies evaluating ... [https://www.sciencedirect.com/science/article/pii/S1388245725006297]
• 30. Functional near-infrared spectroscopy for the assessment ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1524806/full]
• 31. Longitudinal Changes of Resting-State Functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3589816/]
• 32. Resting-state BOLD networks versus task-associated ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC2722694/]
• 33. A Comprehensive Review on the Role of Resting-State ... [https://www.mdpi.com/2035-8377/16/1/12]
• 34. Longitudinal functional connectivity during rest and task is ... [https://www.nature.com/articles/s41598-025-21596-0]
• 35. Resting-state connectivity and task-based cortical ... [https://www.sciencedirect.com/science/article/pii/S2666956025000042]
• 36. A Longitudinal Resting-State fMRI Study - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5907391/]
• 37. A group based network analysis for Alzheimer's disease ... [https://www.nature.com/articles/s41598-025-95190-9]
• 38. Brain functional connectivity analysis of fMRI-based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11880024/]
• 39. The Role of Quantitative EEG in the Diagnosis of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7175442/]
• 40. The role of quantitative EEG biomarkers in Alzheimer's ... [https://www.frontiersin.org/journals/aging-neuroscience/articles/10.3389/fnagi.2025.1522552/full]
• 41. Concurrent fNIRS and EEG for Brain Function Investigation [https://pmc.ncbi.nlm.nih.gov/articles/PMC9371171/]
• 42. Quantitative EEG analysis related to coma and prognosis in ... [https://eurjmedres.biomedcentral.com/articles/10.1186/s40001-025-03191-1]
• 43. Brain symmetry index predicts 3-month mortality in patients ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9704942/]
• 44. Accuracy of EEG Biomarkers in the Detection of Clinical ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9405912/]
• 45. Prognostic value of quantitative and visual ... - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC12558952/]
• 46. A functional neuroimaging biomarker of mild cognitive ... [https://www.nature.com/articles/s44400-025-00018-y]
• 47. Comparison of fNIRS with other neuroimaging modalities [https://www.artinis.com/blogpost-all/comparison-fnirs-vs-fmri]
• 48. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOoo4NrI8Tca9NdFqVyA1GrkVvwQeMGXKSYYvhj7Wi-SAPXxtPObs]
• 49. Research progress of functional near-infrared spectroscopy in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8330753/]
• 50. Novel Technique Improves Brain-State Detection [https://www.mpg.de/25678245/novel-technique-improves-brain-state-detection]
• 51. fMRI Validation of fNIRS Measurements During a ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4544944/]
• 52. Validating a functional near-infrared spectroscopy ... [https://www.nature.com/articles/s41598-020-66784-2]
• 53. Stage-specific functional networks associated with ... [https://www.frontiersin.org/journals/aging-neuroscience/articles/10.3389/fnagi.2025.1562203/full]
• 54. Observation on prefrontal cortex activation in patients with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12075535/]
• 55. Functional near-infrared spectroscopy (fNIRS) in patients ... [https://www.sciencedirect.com/science/article/pii/S1876201825000255]
• 56. Accuracy of functional near-infrared diffuse optical ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12370649/]
• 57. An EEG-fNIRS neurovascular coupling analysis method to ... [https://www.sciencedirect.com/science/article/abs/pii/S001048252300433X]
• 58. Dynamic causal modelling on infant fNIRS data [https://www.sciencedirect.com/science/article/pii/S1053811918303185]
• 59. Neurovascular Coupling: Scientometric Analysis of 30 Years ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12606022/]
• 60. Exploring Imagined Movement for Brain–Computer Interface ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12468679/]
• 61. fNIRS, EEG, ECG, and GSR reveal an effect of complex ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2025.1459653/full]
• 62. Motion Artifact Removal from EEG Signals Using the ... [https://www.sciencedirect.com/science/article/pii/S1746809425013886]
• 63. Combination of spatial and temporal de-noising and artifact ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1576954/full]
• 64. Enhancing fNIRS data analysis with a novel motion artifact ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809424005548]
• 65. Real-time motion artifact suppression using convolution ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1432138/full]
• 66. EEG noise reduction based on the Multi-Scale Temporal ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809425013151]
• 67. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOorCo_wcBIXTXUBZ-o3gaj3LQu4MXvkErfsyTjfPwbe9TlAAvgPw]
• 68. Sensitivity of fNIRS to cognitive state and load [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2014.00076/full]
• 69. A Cross-Disciplinary Review of the fNIRS-EEG Dual ... [https://www.mdpi.com/2076-3425/14/10/1022]
• 70. Current implications of EEG and fNIRS as functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11629311/]
• 71. Inferring deep-brain activity from cortical activity using ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4361422/]
• 72. Validating the reproducibility of a low-cost single-channel ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11076663/]
• 73. fNIRS reproducibility varies with data quality, analysis ... [https://www.nature.com/articles/s42003-025-08412-1]
• 74. Multimodal Data Fusion: Key Techniques, Challenges & ... [https://www.sapien.io/blog/mastering-multimodal-data-fusion]
• 75. EEG–fNIRS signal integration for motor imagery ... [https://www.sciencedirect.com/science/article/pii/S2772528625000299]
• 76. Multimodal fNIRS–EEG sensor fusion: Review of data-driven ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12592382/]
• 77. Applications and advances of combined fMRI-fNIRs ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11958174/]
• 78. Assessment of fNIRS Signal Processing Pipelines [https://www.mdpi.com/2076-3417/12/1/316]
• 79. Machine learning in early diagnosis of neurological diseases [https://www.sciencedirect.com/science/article/pii/S3050623925000355]
• 80. EEG/fNIRS Based Workload Classification Using ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9571712/]
• 81. Feature Extraction and Classification Methods for Hybrid ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6032997/]
• 82. Improved classification performance of EEG-fNIRS ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2022.973959/full]
• 83. Selection and Stability of Functional Connectivity Features ... [https://arxiv.org/html/2511.05531v1]
• 84. A cross-task fNIRS framework for multi-class mental health ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809425010705]
• 85. Neurophysiological data augmentation for EEG-fNIRS ... [https://www.sciencedirect.com/science/article/abs/pii/S0169260725000112]
• 86. Exploring the application and challenges of fNIRS ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10957543/]
• 87. A quantitative comparison of NIRS and fMRI across ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811910013777]
• 88. Detecting cognitive motor dissociation by functional near- ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1532804/full]
• 89. Near-Infrared Spectroscopy to Assess Covert Volitional ... [https://link.springer.com/article/10.1007/s12028-025-02301-5]
• 90. Signal processing of functional NIRS data acquired during ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5592780/]
• 91. A study on resting-state functional near-infrared spectroscopy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12171453/]
• 92. Correcting physiological noise in whole-head functional ... [https://www.sciencedirect.com/science/article/pii/S0165027021001977]
• 93. Functional near-infrared spectroscopy for identifying mild ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12425739/]
• 94. Accurate identification of anxiety and depression based on ... [https://www.sciencedirect.com/science/article/abs/pii/S174680942500014X]
• 95. Functional Near-Infrared Spectroscopy in Schizophrenia ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12413766/]
• 96. Functional magnetic resonance imaging as a dynamic ... [https://www.sciencedirect.com/science/article/abs/pii/S0301008211000864]
• 97. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOoreWxW7bqdwEBWhmqtHzyg-m7Pjv8ZsH4DCUl4JbJ2kY6N3esah]
• 98. Neuroimaging-based biomarker discovery and validation [https://pmc.ncbi.nlm.nih.gov/articles/PMC4504798/]
• 99. A Systematic Review and Meta-Analysis Evaluating the ... [https://pubmed.ncbi.nlm.nih.gov/40679314/]
• 100. Longer scans boost prediction and cut costs in brain-wide ... [https://www.nature.com/articles/s41586-025-09250-1]
• 101. Artificial Intelligence in Alzheimer's Disease Diagnosis and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12387124/]