Advanced Signal Noise Processing in fNIRS: Methodologies and Clinical Applications for Biomedical Research

Abstract

This article provides a comprehensive examination of signal noise processing techniques in functional near-infrared spectroscopy (fNIRS) and their critical importance in clinical applications. Covering foundational principles to advanced methodologies, we explore the major sources of physiological and motion artifacts that compromise signal quality and detail established and emerging preprocessing pipelines for effective noise suppression. The content addresses practical challenges in optimizing spatial specificity and signal quality for real-time applications, while presenting validation frameworks and comparative analyses demonstrating how robust processing enables reliable clinical use in neurological disorders, neurorehabilitation, and drug development. Targeted at researchers and pharmaceutical professionals, this review synthesizes current best practices and future directions for transforming fNIRS into a validated clinical tool.

Understanding fNIRS Signal Noise: Sources, Challenges, and Clinical Implications

Fundamental Principles of fNIRS and Neurovascular Coupling

Frequently Asked Questions (FAQs): Principles and Applications

Q1: What physiological phenomena does fNIRS measure, and how is this related to neurovascular coupling?

fNIRS measures changes in the concentration of oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) in the blood of the brain's cortex [1]. This provides functional contrast similar to the functional MRI BOLD (blood oxygen level dependent) signal, reflecting changes in regional blood flow to areas of the brain involved in processing functional tasks [2]. This relationship is the foundation of neurovascular coupling, a key physiological mechanism where neuronal activation actively enhances regional cerebral blood flow and volume to meet increased metabolic demands [3] [4]. fNIRS detects the hemodynamic response function (HRF), which is an indirect marker of neural activity characterized by a peak in HbO concentration approximately 4–5 seconds after neuronal activation, followed by a drop in HbR [4].

Q2: What are the primary strengths and limitations of fNIRS in a clinical research setting?

• Strengths:

  • Portability and Tolerability: fNIRS is portable, non-invasive, and can be used in versatile experimental settings, including bedside monitoring and during physical activities like walking [3] [4] [2]. It is an ideal tool for studying epilepsy, autonomic function, and physiological phenomena [3].
  • Safety and Cost: It uses low-power light-emitting diodes (LEDs), is cost-effective compared to fMRI, and is safe for patients with non-MR compatible implants like pacemakers [3] [1].
  • Hemoglobin Specificity: Unlike fMRI, which measures a BOLD signal based on the ratio of HbO to HbR, fNIRS measures the two hemoglobin species separately [3].
  • Temporal Resolution: It offers a better temporal resolution than fMRI for tracking hemodynamic changes.

• Limitations:

  • Spatial Resolution and Depth Sensitivity: fNIRS is restricted to measuring the sub-cranial cortex, leaving out activations in deep brain structures [4]. Its spatial resolution is limited to a couple of centimeters [1].
  • Signal Contamination: The signal is highly sensitive to motion artifacts and physiological confounding factors (PCFs) originating from the scalp, such as heart rate and respiration [4].
  • Absolute Quantification: Most common continuous-wave (CW) fNIRS systems typically only provide relative changes in hemoglobin concentration, not absolute values [5] [1].

Q3: Which brain areas are accessible for study with fNIRS?

fNIRS can best monitor cortical areas not obscured by hair. This includes the prefrontal cortex (e.g., Brodmann Areas 10, 11, 46, 47), the motor cortex, and the visual cortex [1]. Areas like the ventral-medial prefrontal cortex (which folds inward) and the temporal cortex (due to the ears) are less accessible [1].

Troubleshooting Guides: Common Experimental Issues

Q1: How can I minimize motion artifacts during fNIRS data acquisition?

Motion artifacts (MAs) are a major source of noise, particularly in studies involving movement.

• Pre-Acquisition Strategies:

  • Secure Probe Placement: Ensure the fNIRS cap or probe is snugly and securely fitted to the subject's head.
  • Use Ancillary Sensors: Integrate an accelerometer to record head movement. This data can later be used in adaptive filtering to clean the signal [1].
  • Subject Instructions: Advise participants to move as little as possible during measurements, especially in resting-state studies [6].

• Post-Processing Solutions:

  • Apply specialized motion artifact reduction (MAR) algorithms during data preprocessing. A comparison of eight MAR algorithms suggested that single-channel algorithms followed by bandpass filtering and principal component analysis (PCA) are effective [4].

Q2: My fNIRS signal appears noisy. How can I distinguish and remove physiological confounds?

Physiological confounds like heart rate (~1 Hz), respiration (~0.3 Hz), and Mayer waves (~0.1 Hz) can obscure the hemodynamic response of interest (~0.1 Hz or lower) [4].

• Instrumental Solutions:

  • Short-Separation Channels: Use detectors placed very close (<1 cm) to light sources. These channels are predominantly sensitive to systemic physiological noise in the scalp and can be used to regress this noise out of the standard long-separation channels [3] [7].
  • Auxiliary Recordings: Simultaneously record physiology like pulse and respiration with dedicated sensors for use in noise regression models [7].

• Signal Processing Solutions:

  • Bandpass Filtering: Apply a bandpass filter (e.g., 0.01 - 0.2 Hz) to retain the frequencies of the hemodynamic response while attenuating higher frequency noise like heart rate [6] [4].
  • Principal Component Analysis (PCA): PCA can identify and remove global physiological noise patterns that are uniform across the scalp [4] [7].
  • Advanced Automated Denoising: Newer methods combine PCA with a general linear model (GLM) using information from both long- and short-separation channels to effectively remove physiological noise in whole-head montages [7].

Q3: What are the best practices for designing a robust fNIRS experiment?

The choice of experimental design is critical for eliciting a detectable and interpretable hemodynamic response.

• Block Design: Presents periods of task performance alternated with periods of rest. This design enhances the signal-to-noise ratio of the HRF and is common for functional localization [6].
• Event-Related Design: Presents discrete, randomized trials. This is suitable for studying the response to individual stimuli and avoiding habituation [6].
• Resting-State Design: Measures intrinsic, low-frequency brain activity without tasks. Key recommendations include [6]:
  • Perform measurements in a quiet, light-dimmed, and comfortable setting.
  • Ensure an appropriate measurement duration (often several minutes) to establish stable functional connectivity metrics.
  • Choose filtering methods carefully to avoid removing the low-frequency neural signals of interest along with physiological noise.

Quantitative Data and Methodologies

Noise Category	Specific Source	Typical Frequency	Impact on Signal	Recommended Correction Methods
Motion Artifacts	Head movement, sensor displacement	Variable, abrupt	Large amplitude spikes/shifts	Accelerometer + adaptive filtering [1], MAR algorithms [4]
Physiological Confounds	Cardiac pulsation	~1 Hz	High-frequency oscillation	Bandpass filtering [4], Short-separation channel regression [7]
	Respiration	~0.2-0.3 Hz	Low-frequency oscillation	Bandpass filtering, PCA [4]
	Mayer waves	~0.1 Hz	Very close to HRF frequency	Short-separation channel regression [7], PCA/GLM denoising [7]
Systemic Hemodynamic	Systemic blood pressure changes	<0.1 Hz	Can mimic brain activation	Short-separation channel regression [7], Ancillary physiological monitoring [7]

Table 2: Key fNIRS Signal Processing Pipelines for Clinical Applications

Processing Step	Purpose	Common Techniques & Notes
Motion Artifact Reduction (MAR)	Remove signal components caused by head movement.	Spline interpolation [4], wavelet-based methods [4]. Single-channel algorithms are often recommended [4].
Bandpass Filtering	Attenuate frequency-based physiological noise.	Cut-off values are critical; a typical passband is 0.01 - 0.2 Hz to include HRF and exclude cardiac/respiratory noise [4].
Physiological Noise Correction	Remove systemic physiological fluctuations from scalp and brain.	Short-separation channel regression [7], Principal Component Analysis (PCA) [4] [7], General Linear Model (GLM) with auxiliary signals [7].
Hemodynamic Response Extraction	Quantify the task-related brain activation.	Block averaging across trials or GLM fitting using a canonical HRF model [4].

Experimental Protocol: A Motor Task Paradigm

The following methodology, adapted from a study assessing fNIRS signal processing pipelines, provides a robust protocol for eliciting a measurable hemodynamic response [4]:

• Participants: Recruit a cohort (e.g., 23 young adult volunteers) with no known neurological conditions.
• Task Design: Employ a block design consisting of alternating periods:
  • Task Block (e.g., 20 seconds): Participants perform a motor task such as repetitive hand grasping.
  • Rest Block (e.g., 30 seconds): Participants remain still and relax.
  • Repeat this cycle for multiple runs (e.g., 10-15 cycles) to obtain sufficient data for averaging.
• fNIRS Acquisition:
  • Use a continuous-wave (CW) fNIRS system.
  • Apply a cap with optodes placed over the primary motor cortices (contralateral to the moving hand).
  • Include short-separation channels (e.g., <1 cm) to assist with physiological noise correction.
  • Record at a standard sampling rate (e.g., 10 Hz).
• Data Processing:
  • Process the raw intensity signals using a pipeline such as: MAR → Bandpass Filtering (0.01-0.2 Hz) → PCA.
  • Convert the processed optical density data to relative concentration changes of HbO and HbR using the Modified Beer-Lambert Law.
  • Perform block averaging or GLM analysis to extract the HRF for each channel and condition.
• Outcome: The result is a map of significant activation (typically an increase in HbO and a decrease in HbR) in the contralateral motor area [4].

Signaling Pathways and Workflows

Neurovascular Coupling to fNIRS Signal

fNIRS Signal Processing Workflow

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Materials and Tools for fNIRS Research

Item	Category	Function & Application Notes
Continuous-Wave (CW) fNIRS System	Hardware	Most common and cost-effective system type. Measures light attenuation to calculate relative changes in HbO and HbR concentration [5] [4].
Short-Separation Optodes	Hardware	Detectors placed <1 cm from a source. Critical for measuring and subsequently regressing out superficial, systemic physiological noise from the scalp to reveal the underlying brain signal [3] [7].
Accelerometer	Auxiliary Sensor	Measures head movement. The signal is used in adaptive filtering algorithms to identify and remove motion artifacts from the fNIRS data [1].
HOMER3 / NIRS Toolbox	Software	Open-source MATLAB toolboxes for fNIRS data analysis. They provide a suite of scripts for processing, including filtering, optical conversion, and statistical analysis [5].
Modified Beer-Lambert Law (mBLL)	Algorithm	The core equation used to convert changes in detected light intensity into relative changes in chromophore (HbO, HbR) concentrations [5] [2].

Functional Near-Infrared Spectroscopy (fNIRS) is a powerful, non-invasive neuroimaging technique that measures cerebral hemodynamic activity by using near-infrared light to detect changes in oxygenated and deoxygenated hemoglobin concentrations [8] [3]. Despite its advantages in portability, cost-effectiveness, and ecological validity, fNIRS signals are susceptible to various noise sources that can compromise data quality and interpretation. Effective signal processing is paramount for extracting meaningful neural information, especially in clinical applications and drug development research where accurate biomarkers are critical. This guide categorizes the major noise sources into three primary domains: physiological, motion, and instrumental artifacts. For each category, we provide troubleshooting methodologies, quantitative comparisons, and frequently asked questions to support researchers in optimizing their experimental outcomes.

Physiological Noise: Identification and Mitigation

Physiological noise originates from systemic bodily functions and is a dominant confounding factor in fNIRS signals. Understanding its components is the first step toward effective correction.

FAQ on Physiological Noise

Q1: What are the main types of physiological noise in fNIRS? The main physiological noise sources can be categorized by their temporal characteristics and origins [9]:

• Cardiac pulsations (~1 Hz): Caused by the heartbeat.
• Respiratory waves (R-band, ~0.3 Hz): Induced by the breathing cycle.
• Mayer waves (M-band, ~0.1 Hz): Low-frequency oscillations linked to blood pressure regulation.
• Very Low-Frequency Oscillations (VLFOs/A-band, 0.02-0.08 Hz): Often associated with local autoregulatory processes in the skin and brain, and can be task-evoked.

Q2: How does physiological noise affect the fNIRS signal? Physiological noise can obscure the task-related hemodynamic response, reducing the sensitivity and specificity of fNIRS to cerebral signals. In the worst case, physiological fluctuations synchronized with the experimental task can lead to false positives in activation maps [9]. The impact is typically stronger on oxygenated hemoglobin (HbO) than on deoxygenated hemoglobin (HbR) [9].

Q3: What are the most effective methods for correcting physiological noise? A combination of techniques is often required:

• Band-pass filtering: Effective for removing noise outside the expected frequency range of the hemodynamic response (e.g., high-frequency cardiac pulsations) [8].
• Advanced regression models: Using General Linear Models (GLM) with physiological regressors derived from concurrent recordings of heart rate, blood pressure, and skin blood flow [9].
• Wavelet-based methods: Effective for decomposing signals and identifying coherent physiological noise across different time scales [9].
• Short-Separation Channels: Using a dedicated optode pair with a short source-detector distance (e.g., <1 cm) to measure systemic physiology in the scalp. This signal can be used as a regressor to remove superficial noise from the standard channels [3].

Experimental Protocol: Physiological De-noising with GLM

This protocol outlines a method for characterizing and removing physiological noise using auxiliary measurements [9].

Objective: To identify and regress out the main physiological noise components from fNIRS signals recorded on the forehead.

Materials and Reagents:

• Time-domain or continuous-wave fNIRS system.
• Peripheral physiology recording units for Mean Arterial Blood Pressure (MAP) and Heart Rate (HR).
• Laser Doppler Flowmetry (LDF) probe for recording Skin Blood Flow (SBF) on the forehead.

Procedure:

• Setup: Position fNIRS optodes over the prefrontal cortex. Concurrently, attach the MAP, HR, and SBF sensors according to manufacturer guidelines.
• Data Acquisition: Record a resting-state baseline for at least 10 minutes, followed by task-related data acquisition.
• Signal Processing:
  • Process the fNIRS data to convert light intensity changes into HbO and HbR concentration changes.
  • Extract fluctuations from the MAP, HR, and SBF recordings.
• Wavelet Coherence Analysis (WCA):
  • Perform WCA between the fNIRS signals (HbO/HbR) and each peripheral physiological recording (MAP, HR, SBF).
  • This analysis quantifies the impact of each physiological process on the fNIRS signal across different time scales (R-band, M-band, A-band).
• GLM with Physiological Regressors:
  • Design a GLM where the fNIRS signal is the dependent variable.
  • Create regressors of non-interest from the WCA results, representing the noise components coherent with respiration, Mayer waves, and skin blood flow.
  • Include your task regressor(s) in the model.
  • Model fit will estimate the contribution of each noise component, effectively removing them and yielding a cleaner estimate of the cerebral hemodynamic response.

Research Reagent Solutions

Table: Essential Materials for Physiological Noise Investigation

Item	Function/Description
fNIRS System	Core device for measuring cortical hemodynamics via near-infrared light.
Electrocardiogram (ECG)	Provides a precise recording of heart rate and rhythm for cardiac noise regression.
Respiratory Belt	Measures chest or abdominal expansion to track the respiratory cycle.
Finapres or Similar	Non-invasive device for continuous measurement of Mean Arterial Blood Pressure (MAP).
Laser Doppler Flowmetry (LDF)	Monitors local Skin Blood Flow (SBF) in the scalp near the fNIRS measurement site.
Short-Separation Optodes	Specialized optodes placed <1 cm from a source to measure systemic physiology in the scalp.

Motion Artifacts: Correction and Evaluation

Motion artifacts (MAs) are a significant source of noise, particularly in studies involving movement, vulnerable populations, or naturalistic settings.

FAQ on Motion Artifacts

Q1: What causes motion artifacts, and what do they look like? MAs are caused by a temporary decoupling of the optodes from the scalp [10]. Common causes include head movements (nodding, shaking), facial movements (brow and jaw movement during speech), and body movements [11]. In the data, they manifest as high-amplitude spikes, rapid baseline shifts, or slower low-frequency variations that can mimic a hemodynamic response [10].

Q2: Is it better to reject trials with motion artifacts or to correct them? Evidence strongly supports correction over rejection. Trial rejection is only feasible if the number of artifacts is low and the total number of trials is high. For studies with limited trials (e.g., with infants or patients), correction is essential to retain statistical power [10].

Q3: Which motion correction algorithm is the most effective? The "best" algorithm can depend on the artifact type, but several comparative studies have identified top performers:

• Wavelet Filtering: Consistently ranks among the most effective methods, particularly for task-related, low-frequency artifacts [10] [12].
• Temporal Derivative Distribution Repair (TDDR): A robust method that operates on the signal's temporal derivatives and has shown superior performance in functional connectivity analysis [12].
• Other Common Methods: These include Spline Interpolation, Correlation-Based Signal Improvement (CBSI), Kalman Filtering, and Principal Component Analysis (PCA). Their performance varies [10] [12].

Quantitative Comparison of Motion Correction Algorithms

Table: Performance Comparison of Common Motion Artifact Correction Methods

Algorithm	Key Principle	Best For	Performance Notes
Wavelet Filtering [10] [12]	Decomposes signal and thresholds wavelet coefficients dominated by MAs.	Various artifact types, especially those correlated with the task.	Highly effective; reduced artifact area in 93% of cases in one study [10]. Superior for functional connectivity [12].
TDDR [12]	Uses a robust estimator to weight temporal derivatives, reducing influence of large MA-driven fluctuations.	Online processing and functional connectivity analysis.	Top performance in recovering original FC patterns; suitable for real-time use [12].
Spline Interpolation (MARA) [10] [12]	Identifies artifact segments and interpolates over them using spline functions.	Offline analysis with clear, discrete artifacts.	Performance depends on accurate artifact detection and level correction [12].
CBSI [10] [12]	Exploits the negative correlation between HbO and HbR during neural activation.	Simple, model-based correction without auxiliary data.	Assumes strict negative HbO-HbR correlation, which may not always hold [12].
PCA [12]	Removes first few principal components assumed to be dominated by MAs.	Removing large, sporadic artifacts that explain high variance.	Risks removing physiological signal of interest if it also has high variance [12].
Kalman Filtering [12]	Models the fNIRS signal as an autoregressive process and uses a state-space model to filter MAs.	-	Performance can be variable compared to top methods like TDDR and wavelet [12].
Accelerometer-Based [11]	Uses accelerometer data as a noise reference for adaptive filtering.	Real-time rejection when auxiliary hardware is available.	Requires additional hardware; effectiveness depends on coupling between accelerometer and optode motion [11].

Experimental Protocol: Evaluating Motion Correction with Real Data

This protocol describes a method to quantitatively evaluate the efficacy of different MA correction algorithms on real task data [10].

Objective: To compare the performance of multiple MA correction techniques (e.g., Wavelet, TDDR, CBSI) using objective metrics derived from the physiology of the hemodynamic response.

Materials and Reagents:

• fNIRS system with probes placed on the scalp.
• (Optional) Accelerometer for auxiliary motion tracking.
• Software toolboxes implementing various MA correction algorithms (e.g., Homer2, NIRS Brain AnalyzIR).

Procedure:

• Data Acquisition: Collect fNIRS data during a cognitive or motor task known to produce hemodynamic responses. Ensure the data contains real, task-related motion artifacts (e.g., from jaw movement during a vocal response task).
• Preprocessing: Apply initial standard processing (conversion to optical density, then to HbO/HbR).
• Motion Correction: Apply several different MA correction algorithms (Wavelet, TDDR, Spline, etc.) to the same preprocessed dataset.
• Hemodynamic Response Extraction: For each condition and corrected dataset, extract the average hemodynamic response.
• Performance Evaluation: Compare the corrected responses using objective metrics. Since the "true" HRF is unknown, use metrics related to physiological plausibility:
  • HbO/HbR Anti-Correlation: Calculate the correlation between the group-averaged HbO and HbR responses. A stronger negative correlation is expected after successful denoising.
  • Response Morphology: Assess the shape of the recovered HRF (e.g., a canonical HbO increase followed by a post-stimulus undershoot).
  • Contrast-to-Noise Ratio (CNR): Calculate the CNR of the task-related response in the corrected signal.

Instrumental and Environmental Artifacts

While physiological and motion artifacts are most common, instrumental and environmental factors also introduce noise.

FAQ on Instrumental and Environmental Noise

Q1: What are the common sources of instrumental noise? Instrumental noise can stem from the fNIRS hardware itself, including [3]:

• Detector Noise: Thermal and shot noise in the photodetector.
• Source Noise: Fluctuations in the intensity of the light emitters.
• Electronic Noise: From the analog-to-digital converters and other electronic components.

Q2: How does the experimental environment affect fNIRS signals? Ambient light is a major environmental contaminant. If it reaches the detectors, it can swamp the weak light signals that have passed through the head, dramatically reducing the signal-to-noise ratio. Ensuring proper optode contact and using black, opaque caps and coverings are essential to block out ambient light.

Q3: What are the spatial limitations of fNIRS? fNIRS primarily measures cortical activity directly beneath the optodes. Its sensitivity decreases with depth, and it cannot reliably access subcortical structures. The spatial resolution is fundamentally limited by the density of the optode array and the source-detector separation [3] [13]. Co-registration with anatomical MRI can help improve spatial accuracy.

Visualizing the fNIRS Noise Troubleshooting Workflow

The following diagram outlines a logical workflow for diagnosing and addressing the major noise sources discussed in this guide.

The Impact of Signal Quality on Clinical Data Interpretation

Troubleshooting Guide: Signal Quality Assessment

How can I automatically assess the quality of my fNIRS channels before processing?

Automated signal quality metrics are essential for efficiently identifying poor-quality channels, especially with large datasets or high-density systems. The table below summarizes common algorithms [14].

Table 1: Automated fNIRS Signal Quality Assessment Algorithms

Algorithm Name	Underlying Principle	Pros	Cons	Typical Thresholds
Coefficient of Variation (CV)	Measures relative variability in raw light intensity (Standard Deviation / Mean) * 100% [14].	Simple and straightforward to implement [14].	Cannot distinguish physiological fluctuations from motion artifacts; may accept flat signals [14].	Requires individual tuning; high CV indicates instability [14].
Scalp Coupling Index (SCI)	Assesses optode-scalp contact by correlating cardiac oscillations between the two wavelength signals [14].	Simple idea; strong correlation indicates good signal quality [14].	Susceptible to motion artifacts that affect both wavelengths equally [14].	SCI ≥ 0.75 - 0.80 for good channel inclusion [14].
PHOEBE	Combines SCI with spectral analysis of the cardiac peak in the cross-correlation between wavelengths [14].	Improved sensitivity by isolating the heartbeat component, even with motion [14].	More complex than SCI [14].	Good quality shows a clear, dominant cardiac peak [14].
Signal Quality Index (SQI)	Sophisticated algorithm focusing on the cardiac component across three rating stages [14].	Provides a nuanced 1-5 scale; higher performance than SCI and PHOEBE [14].	Complex implementation [14].	1 (Poor) to 5 (Excellent) [14].

Experimental Protocol for Signal Quality Check:

• Extract Raw Data: Use a segment of resting-state or task-based raw light intensity data.
• Calculate Metric: Choose an algorithm (e.g., SCI) and calculate the metric for each channel.
• Apply Threshold: Classify channels based on established thresholds (e.g., exclude channels with SCI < 0.75).
• Visual Inspection: Manually verify a subset of flagged channels to ensure the algorithm is performing as expected [14].

Why is my HbO signal showing activation, but my HHb signal is not? Could this be a false positive?

This is a common issue and your concern is valid. A robust hemodynamic response is typically characterized by a concurrent increase in oxygenated hemoglobin (HbO) and a decrease in deoxygenated hemoglobin (HHb) [15]. When this expected pattern is absent, it can indicate a false positive driven by systemic physiological noise rather than neuronal activity.

HbO is a high-contrast signal but is more susceptible to systemic interference like changes in blood pressure, heart rate, or skin blood flow. HHb is generally more robust to these confounds and offers better spatial specificity, but has a lower magnitude and thus less statistical power [15]. Relying solely on HbO increases the risk of false positives [15].

Troubleshooting Protocol:

• Inspect Both Signals: Always report and interpret both HbO and HHb signals together [16] [15].
• Employ Denoising Techniques: Use signal processing to remove systemic confounds.
  • Short-Separation Channels: Incorporate short-separation detectors (~8 mm) to regress out superficial, systemic components [7].
  • Advanced Processing: Apply methods like Principal Component Analysis (PCA) and General Linear Models (GLM) with physiological regressors (e.g., heart rate, respiration) to isolate cerebral signals [7].
• Use Combined Signals: Consider deriving signals that integrate information from both HbO and HHb, such as the Hemodynamic Phase Correlation (HPC) or Correlation-Based Signal Improvement (CBSI), which have shown improved robustness against false positives [15].

Diagram: A workflow for investigating potential false positives in fNIRS data by jointly interpreting HbO and HHb signals.

What is the most effective way to handle motion artifacts in clinical data?

Motion artifacts are a major source of signal quality degradation. The optimal handling strategy depends on your research context (offline analysis vs. real-time application) and the nature of the artifact.

Table 2: Common Methods for Handling Motion Artifacts and Physiological Noise

Noise Type	Handling Method	Key Implementation Details	Considerations for Clinical Use
Motion Artifacts	Algorithmic Correction (e.g., wavelet, PCA, movement-based) [17].	Identify artifact segments via large signal deviations or accelerometer data; reconstruct signal [17].	Offline: Wide range of algorithms available. Real-time: Requires robust, fast methods (e.g., movement-based).
Systemic Physiological Noise (Heart, Respiration)	Bandpass Filtering [17].	Use a bandpass filter (e.g., 0.01–0.2 Hz) to retain HRF and remove higher/lower frequency noise [17].	Simple but crude; may not fully separate brain from systemic physiology.
	Short-Separation Regression [7].	Use signals from short-separation channels (~8mm) as regressors to remove systemic components in long channels.	Highly effective; requires specific hardware setup with short-distance detectors.
	General Linear Model (GLM) with Physiological Regressors [7].	Record heart rate & respiration; use as noise regressors in GLM to statistically remove their influence.	Very effective but requires additional equipment to record physiology.

Experimental Protocol for Motion Artifact Handling:

• Detection: Use automated algorithms (e.g., based on signal slope or amplitude) or visual inspection to identify motion-contaminated segments.
• Correction/Rejection:
  • For offline analysis, apply a validated correction algorithm (e.g., wavelet filtering).
  • For severe, uncorrectable artifacts, mark the segment for rejection. In real-time applications, implement a quality check to pause feedback if excessive motion is detected [18].
• Documentation: Clearly report the specific motion artifact detection and correction methods used in your publication [16].

Frequently Asked Questions (FAQs)

We are getting inconsistent results across study sites. How can we improve reproducibility?

Reproducibility in fNIRS is significantly influenced by data quality, analysis pipeline variability, and researcher experience [19]. A recent large-scale initiative (FRESH) found that while agreement at the group level can be high, individual-level results vary more, especially with lower data quality [19].

Key Recommendations:

• Standardize Pipelines: Pre-register your analysis plan and use standardized, shared processing pipelines across sites to minimize analytical variability [19].
• Ensure Data Quality: Implement rigorous, standardized protocols for optode placement and signal quality checks at each site [20].
• Report Thoroughly: Adhere to best practices for fNIRS publications, detailing all acquisition parameters, preprocessing steps, and statistical models [16].
• Share Data: Use standardized data formats like SNIRF and BIDS to facilitate data sharing and reanalysis [20].

For clinical applications like BCIs, how can I ensure my real-time signal is of sufficient quality?

In real-time applications like Brain-Computer Interfaces (BCIs) or neurofeedback, you cannot correct data after acquisition, making proactive quality control paramount [18]. Running a system on poor-quality signals can render it ineffective and undermine user trust [18].

Key Recommendations:

• Real-Time Quality Metrics: Implement real-time calculations of metrics like SCI or CV to monitor optode coupling during the setup and throughout the session.
• Robust Real-Time Preprocessing: Incorporate simple, robust denoising techniques in your real-time pipeline, such as high-pass filtering or movement correction algorithms designed for low latency.
• Signal Quality Feedback: Provide a visual indicator to the experimenter (or user) if signal quality drops below a certain threshold, allowing for intervention (e.g., adjusting the cap).

What are the critical steps for designing a reliable fNIRS clinical experiment?

A well-designed experiment is the first defense against poor data interpretation.

Key Recommendations:

• Choose an Appropriate Design: For block designs, use blocks of sufficient length (e.g., 20-30 s) to allow the hemodynamic response to evolve. For event-related designs, use jittered inter-stimulus intervals to allow the response to return to baseline [21].
• Select a Proper Control Condition: The control condition should be matched to your task condition in all aspects except the cognitive process of interest (cognitive subtraction) [21].
• Plan for Physiological Confounds: If your task influences heart rate, respiration, or blood pressure (e.g., a stressful task), plan to measure these and account for them in your analysis [16] [7].

Diagram: A workflow for a robust fNIRS clinical experiment, from design to interpretation.

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Tools for fNIRS Signal Quality Assurance

Tool Category	Specific Examples	Function in Signal Quality
Signal Quality Algorithms	Scalp Coupling Index (SCI), Signal Quality Index (SQI), PHOEBE [14].	Automatically assess and quantify channel-level signal quality to guide channel inclusion/exclusion.
Denoising Software Tools	HomER2, NIRS-KIT, NIRS Brain AnalyzIR, FIELDTRIP [20].	Provide implementations of standard and advanced preprocessing algorithms (filtering, PCA, GLM, wavelet).
Hardware Add-ons	Short-Separation Detectors [7], Accelerometers [17], Auxiliary Physiological Recorders (Pulse Oximeter, Respiration Belt) [7].	Directly measure noise sources (superficial hemodynamics, motion, systemic physiology) for improved signal regression.
Standardized Data Formats	SNIRF (Shared NIR Format), BIDS-fNIRS (Brain Imaging Data Structure) [20].	Ensure data interoperability, facilitate sharing, and support the use of standardized processing pipelines.

Current Gaps in Standardization and Methodological Heterogeneity

FAQs: Addressing Common fNIRS Experimental Challenges

FAQ 1: What are the most critical factors affecting the reproducibility of fNIRS results? Reproducibility is significantly influenced by three interconnected factors: data quality, the choice of analysis pipeline, and researcher experience [19]. A large-scale reproducibility study found that nearly 80% of research teams agreed on group-level results when hypotheses were strongly supported by existing literature. However, agreement at the individual level was lower and improved markedly with higher data quality. Key sources of variability include how poor-quality data are handled, how hemodynamic responses are modeled, and the specific statistical methods employed [19].

FAQ 2: How does methodological heterogeneity impede comparisons between fNIRS studies? Substantial heterogeneity in test setups, task variants, and control conditions makes cross-study comparisons difficult [22]. For example, a 20-year review of fNIRS studies using the Stroop test found a lack of standardized control conditions to properly evaluate how the Stroop effect is resolved. Furthermore, most studies reported only oxygenated hemoglobin (HbO), while deoxygenated hemoglobin (HbR) analyses were predominantly confined to event-related designs, limiting the depth of interpretation [22]. This heterogeneity extends to occupational workload studies, where variations in optode placement and systemic artifact correction methods hinder the validity and comparability of findings [23].

FAQ 3: What are the primary technical limitations of fNIRS in clinical applications? The main technical limitations include limited spatial resolution (approximately 1-3 cm) and an inability to detect signals from deep brain regions beyond approximately 2 cm [24] [13] [25]. Furthermore, the fNIRS signal is confounded by systemic effects such as cardiovascular fluctuations and scalp blood flow. It is crucial to consider these confounding signals both during experimental design and data processing [20]. The community is actively developing standard protocols for in vivo validation and systemic artifact correction to mitigate these issues [20].

FAQ 4: Why is data quality so variable in real-world fNIRS studies? Data quality in real-world settings is affected by motion artifacts, environmental interference, and the broad-spectrum error inherent in the hardware [24] [26]. This error arises because compact light sources have a wide spectral width, and compact detectors are often incapable of distinguishing between wavelengths with perfect precision. Without careful selection of light sources and detectors, significant signal distortion can occur during the conversion of detected light intensities into neural activity signals [26].

Troubleshooting Guides

Troubleshooting Data Quality and Preprocessing

Table 1: Common fNIRS Data Quality Issues and Solutions

Problem	Root Cause	Impact on Signal	Corrective Action
High-Frequency Noise	Cardiac pulsation (~1 Hz), instrument noise [25].	Signal appears jittery or rapid, unnatural oscillations.	Apply a low-pass filter with a cut-off frequency below 0.5 - 1 Hz to isolate the slower hemodynamic response [25].
Low-Frequency Drift	Physiological cycles (e.g., Mayer waves ~0.1 Hz), instrument drift [25].	Signal exhibits a slow, wandering baseline.	Apply a high-pass filter with a cut-off of ~0.01 Hz or use polynomial detrending to remove slow baseline drifts.
Motion Artifacts	Subject movement, causing optode displacement [24].	Sudden, large spikes or shifts in the HbO/HbR signal.	Use motion-tolerant hardware with flexible probes and algorithm optimization (e.g., wavelet decomposition, correlation-based signal improvement) [24].
Systemic Confounds	Changes in systemic physiology (heart rate, blood pressure, scalp blood flow) [20].	HbO and HbR may become positively correlated, indicating non-neural origin.	Implement standard procedures for monitoring (e.g., with additional short-separation channels) and correcting for systemic effects during data processing [20].
Broad-Spectrum Error	Wide spectral width of light sources and non-specific detectors [26].	Inaccurate conversion of light intensity to hemoglobin concentration changes.	Select wavelength pairs that are robust to this error during system design, as identified through simulation [26].

Troubleshooting Experimental Design and Analysis

Table 2: Methodological Gaps and Standardization Strategies

Methodological Gap	Evidence of Heterogeneity	Proposed Standardization Strategy
Analysis Pipeline Variability	38 research teams used different pipelines on the same data, leading to variability in results [19].	Adopt community-driven frameworks like the FRESH Reproducibility Study Hub to benchmark analysis techniques. Report preprocessing and analysis steps in detail.
Data/Code Sharing	Lack of shared data and code limits transparency and direct comparison of methods [20] [19].	Use standardized file formats (SNIRF) and data organization (BIDS-fNIRS) to ensure interoperability and promote open science [20].
Handling of Poor-Quality Data	A key source of variability was how different teams identified and handled noisy channels or epochs [19].	Establish and report clear, pre-registered criteria for data inclusion/exclusion and artifact rejection.
Probe Placement & Registration	Only 26 out of 41 occupational studies used standardized optode placements [23].	Use standard procedures and tools for probe placement and anatomical registration to ensure consistency across labs [20].
Control Conditions & Tasks	Stroop test studies often lack sufficient conditions to evaluate how the cognitive effect is solved [22].	Design control tasks that are carefully matched to the experimental task to isolate the cognitive process of interest.

Experimental Protocols for Key Methodological Investigations

Protocol: Assessing the Impact of Analysis Pipelines on Reproducibility

This protocol is based on the design of the FRESH (fNIRS REproducibility Study Hub) initiative [19].

• Aim: To quantify how different analysis pipelines affect the results and conclusions derived from the same fNIRS dataset.
• Method:
  • Dataset Preparation: Provide multiple research teams with the same raw fNIRS dataset(s), which include both high and low data quality segments.
  • Hypothesis Testing: Teams are given a specific, pre-registered hypothesis to test (e.g., "Does the task condition show significantly higher HbO in the PFC than the control condition?").
  • Independent Analysis: Each team processes the data and tests the hypothesis using their own preferred analysis pipeline.
  • Comparison: The results (e.g., t-statistics, p-values, effect sizes) and final conclusions (significant vs. not significant) from all teams are collected and compared.
• Key Variables Measured: Inter-team agreement on group-level results, inter-team agreement on individual-level results, and self-reported confidence correlated with years of fNIRS experience.
• Interpretation: The protocol identifies which stages of the pipeline (e.g., motion correction, statistical modeling) contribute most to variability in results, highlighting targets for future standardization.

Protocol: Designing fNIRS Studies for Real-World Settings

This protocol synthesizes best practices for designing ecologically valid fNIRS studies [25].

• Aim: To capture meaningful brain activity in naturalistic, dynamic environments while maintaining scientific rigor.
• Method:
  • Design Selection: Choose between block designs (ideal for maximizing signal-to-noise ratio in controlled tasks) and event-related designs (necessary for irregularly timed, natural events).
  • Control Condition: Select a well-matched control condition that isolates the cognitive process of interest. In social neuroscience, this could be a non-interactive baseline.
  • Paradigm Flexibility: Leverage the portability and motion tolerance of fNIRS to design tasks that involve walking, social interaction, or instrument playing, which are impossible in an fMRI scanner.
  • Data Analysis: For dynamic tasks, use a General Linear Model (GLM) approach with a design matrix that connotes task events with the Hemodynamic Response Function (HRF), as this is more flexible than simple block averaging for irregular timings.
• Key Applications: Hyperscanning of multiple individuals during social interactions, monitoring brain function in occupational settings, and studying brain dynamics during artistic performance [25] [27].

Signaling Pathways and Experimental Workflows

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for fNIRS Research

Item / Tool	Function / Purpose	Key Considerations
fNIRS Hardware (CW, FD, TD)	Non-invasive measurement of HbO and HbR concentration changes in the cortex.	CW systems are most common; FD/TD provide better depth resolution. Select devices with wavelengths robust to broad-spectrum error [26].
Standardized Headcaps & Probes	Holds optodes in consistent positions on the scalp across subjects and sessions.	Use caps with pre-defined, measurable positions. Flexible probes help reduce motion artifacts [24] [23].
Anatomical Registration Tools	Co-registers fNIRS optode locations with standard brain atlas or individual anatomy.	Crucial for accurate spatial reporting and comparison with fMRI literature. Supported by best-practice guidelines [20].
Short-Separation Detectors	Placed ~0.8 cm from a source to measure systemic confounds from the scalp.	Essential for separating cortical activity from superficial, non-cerebral hemodynamic changes [20].
Tissue Phantoms	Stable, brain-like materials for system validation and performance testing.	Used to characterize instrument performance (e.g., MEDPHOT, BIP, nEUROPT protocols) [20].
Analysis Software & Toolboxes	For preprocessing, statistical analysis, and visualization of fNIRS data.	Many toolboxes exist (e.g., Homer2/3, Brainstorm, NIRS-KIT). Variability in pipelines is a key reproducibility challenge [19] [27].
SNIRF & BIDS-fNIRS	Standardized file format (SNIRF) and data organization structure (BIDS).	Ensures data interoperability, facilitates sharing, and enhances reproducibility [20].
Wireless fNIRS Systems	Enable brain monitoring in fully naturalistic, real-world environments.	Critical for studies of occupational workload, social interaction, and artistic performance [23] [25] [27].

The Role of fNIRS in Clinical Neuroscience and Drug Development

FAQs: Core Principles and Technical Specifications

What is fNIRS and what does it measure? Functional Near-Infrared Spectroscopy (fNIRS) is a non-invasive optical neuroimaging technique that measures changes in hemoglobin concentrations in the brain. It detects cortical activity by quantifying relative changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (Hbb) based on their distinct light absorption properties in the near-infrared spectrum (700-900 nm) [28] [29]. This provides an indirect measure of brain activity through the mechanism of neurovascular coupling, where neural activation triggers localized changes in blood flow and oxygenation [28].

What are the key technical specifications of fNIRS systems? Table: Key Technical Specifications of fNIRS

Parameter	Typical Range/Capability	Details
Penetration Depth	Up to 2-3 cm [30] [1]	Allows monitoring of the cerebral cortex.
Spatial Resolution	5-10 mm [30]	A low-resolution technique compared to fMRI.
Temporal Resolution	Up to 100 Hz (typical imaging rates: 3-25 Hz) [30]	Superior to fMRI, suitable for capturing hemodynamic responses.
Source-Detector Separation	1.5-5 cm [28]	Varies based on head size; critical for signal quality.

What are the primary strengths and limitations of fNIRS?

• Strengths: Portability, cost-effectiveness, tolerance for movement, safety (non-invasive, non-ionizing), and ability to provide long-term, real-time monitoring at the bedside [28] [29] [1]. It also allows for simultaneous measurement with other modalities like EEG [28].
• Limitations: Limited penetration depth (superficial cortex only), susceptibility to physiological noise, and inferior spatial resolution compared to fMRI [28] [29] [1]. It does not provide structural brain information [1].

FAQs: Troubleshooting Common Experimental Issues

How can I mitigate physiological noise in my fNIRS signal? Physiological noise from cardiac pulsation, respiration, and blood pressure changes (Mayer waves) is a major contaminant. A combination of pre-processing techniques is recommended [17] [7]:

• Frequency Filtering: Use bandpass filters (e.g., 0.01-0.2 Hz) to isolate the hemodynamic response from higher-frequency cardiac noise and lower-frequency drift [17].
• Short-Separation Channels: Incorporate optodes placed <1 cm apart. These channels predominantly capture systemic physiological noise from the scalp, which can be regressed out from the standard long-separation channels using techniques like General Linear Model (GLM) [7].
• Advanced Denoising: Methods like Principal Component Analysis (PCA) can identify and remove global superficial components. Using auxiliary measurements (e.g., pulse oximeters) can further improve noise correction [7].

What are the best practices for correcting motion artifacts? Motion artifacts are a common issue, though fNIRS is more robust than fMRI [1]. Correction strategies include:

• Pre-Processing Algorithms: Techniques such as wavelet filtering and smoothing filters (e.g., moving average, Savitzky–Golay) are frequently used to reduce motion noise [17].
• Auxiliary Sensors: Attaching an accelerometer to the subject's head provides a direct record of movement, which can be used in adaptive filtering to clean the signal [1].
• Automated Pipelines: For whole-head setups, automated denoising pipelines that combine multiple methods (PCA, GLM, short-separation channels) have shown superior performance in improving contrast-to-noise ratio [7].

How can I ensure accurate spatial localization of my fNIRS signal? Accurate co-registration of optodes with individual anatomy is crucial for reproducibility [28].

• Precise Optode Placement: Use standardized placement systems and measure sensor locations relative to cranial landmarks (e.g., nasion, inion).
• Anatomical Co-registration: For precise localization, correlate optode positions with the subject's structural MRI. Without an individual MRI, standardized brain templates can provide an approximation [1].
• High-Density Arrays: Using high-density whole-head optode arrays improves spatial resolution and sensitivity for mapping network-level brain activities [28] [31].

Experimental Protocols and Methodologies

Protocol: Prefrontal Cortex Assessment in Addiction Research This paradigm is used to investigate cue-reactivity and craving in substance use disorders [32] [33].

• Participant Preparation: Place the fNIRS probe on the forehead, ensuring the bottom is just above the eyebrows and the sides are not over hairy areas. The middle of the sensor should align with the nose [1].
• Experimental Paradigm: Use a block design alternating between rest and stimulus conditions.
  • Resting State (Baseline): 2-5 minutes of quiet rest to establish a baseline.
  • Addiction Induction (Task): Exposure to drug-related cues (e.g., images, videos, or paraphernalia) for a set period (e.g., 30-60 seconds) to elicit craving.
• Data Acquisition: Collect HbO and Hbb concentrations from prefrontal regions, particularly targeting the Orbitofrontal Cortex (OFC) and Dorsolateral Prefrontal Cortex (DLPFC), which are implicated in reward and cognitive control [32].
• Analysis: Compare HbO activation during the task block against the rest block. Machine learning classifiers (e.g., SVM, CNN) can differentiate hemodynamic patterns between users of different substances [32].

Protocol: Monitoring Treatment Response in Psychiatry fNIRS is increasingly used to track neurofunctional changes in response to pharmacological or therapeutic interventions [31].

• Study Design: Longitudinal or randomized controlled trial (RCT) designs are employed. Assessments are conducted before treatment initiation (baseline) and at predefined intervals during treatment.
• Task Paradigm: Cognitive tasks that engage the prefrontal cortex are commonly used. The verbal fluency task is a standard paradigm where patients generate words belonging to a category, effectively activating the DLPFC [31].
• Signal Acquisition: High-density fNIRS systems (>32 channels) are recommended for better spatial coverage. Consistent probe placement across sessions is critical.
• Outcome Measures: The primary outcomes are changes in HbO activation strength and pattern in target regions (e.g., increased DLPFC activity after antidepressant treatment). Functional connectivity between brain regions can also be a biomarker.

The Scientist's Toolkit: Essential Research Reagents

Table: Essential Materials and Analytical Tools for fNIRS Research

Item/Solution	Function/Purpose	Technical Notes
High-Density fNIRS System	Measures cortical hemodynamics with multiple source-detector pairs.	Enables whole-head mapping. Systems with >32 channels provide improved spatial resolution [31].
Short-Distance Channels	Measures and corrects for systemic physiological noise from the scalp.	Optodes placed <1 cm apart; critical for robust denoising [7].
Auxiliary Sensors (Accelerometer, Pulse Ox)	Records motion and physiological data (heart rate, SpO2) for noise correction.	Used in adaptive filtering to remove motion and cardiorespiratory artifacts [1].
Anatomical Co-registration System	Precisely maps optode locations onto individual brain anatomy.	Uses 3D digitizers or MRI; significantly improves spatial accuracy of the signal [28].
GLM & PCA Software Tools	Statistical processing and noise removal.	General Linear Model (GLM) for HRF estimation; Principal Component Analysis (PCA) for global noise removal [17] [7].

fNIRS Processing Pipelines: From Raw Data to Clinical Insights

Frequently Asked Questions (FAQs)

Q1: What are the fundamental principles behind converting raw light intensity into a measure of brain activity?

The conversion is a multi-step process based on physical principles of light propagation in biological tissues. Raw light intensity (in arbitrary units) is first converted to Optical Density (OD), a unitless measure of light attenuation. The OD is then converted into relative changes in hemoglobin concentration (in micromolar, µM) using the Modified Beer-Lambert Law (MBLL). Finally, these concentration changes are analyzed to infer brain activity based on the principle of neurovascular coupling, where active brain regions receive increased oxygenated blood flow [13] [15] [34].

Q2: Why are both oxygenated (HbO) and deoxygenated (HbR) hemoglobin concentrations important, and which one should I trust?

Both signals are crucial for a physiologically accurate interpretation. Functional brain activation is typically characterized by a coupled response: a statistically significant increase in HbO and a decrease in HbR [15].

• HbO: Has a larger amplitude and higher contrast, making it easier to detect statistically significant changes. However, it is more susceptible to systemic physiological interference (e.g., blood pressure changes, scalp blood flow) and can be prone to false positives [15].
• HbR: Has a smaller amplitude and lower statistical power but is generally more robust against systemic interference and offers better spatial specificity. Relying solely on HbO is not considered best practice [15] [35].

The most robust approach is to require that both signals show the expected pattern of change to confirm functional activation [15].

Q3: My calculated Optical Density (OD) values are extremely small (e.g., ±0.03). Is this normal, and what could be the cause?

Exceptionally small OD values can occur and are often traced to the input data. A common cause is that the raw intensity values are too close to the mean intensity used in the OD calculation. Since OD is calculated as the logarithm of the ratio (intensity / mean intensity), if this ratio is very close to 1, the resulting OD will be near zero. This can lead to subsequent hemoglobin concentration values that are also unrealistically small [36].

• Troubleshooting Steps:
  • Check Raw Intensity: Ensure your raw intensity data has a sufficient signal-to-noise ratio and is within the device's dynamic range.
  • Adjust Hardware Settings: If possible, increasing the LED power level can boost the detected intensity signal, potentially resolving the issue [36].
  • Verify Calibration: Confirm that your system is properly calibrated.

Q4: I encounter an "IndexError" when using the optical_density function in MNE-Python. How can I resolve it?

This error often arises when the input data structure does not meet the function's expectations, particularly when using custom-built hardware or non-standard data formats. The function expects the data channels to be correctly ordered and named [37].

• Troubleshooting Steps:
  • Verify Channel Information: Ensure the info object you provide to RawArray contains correct channel names and types. The function expects specific naming conventions to identify light sources and detectors.
  • Check Data Format: When using custom hardware, you must meticulously format your data to match the requirements of the analysis software. Consult the library's documentation for custom data input.
  • Consider Standard Formats: Using a standardized data format like SNIRF is highly recommended, as it ensures compatibility and reduces such errors [37] [38].

Troubleshooting Guides

Guide to Resolving Data Conversion Errors

This guide addresses common errors during the initial stages of fNIRS data conversion.

Problem: Errors when converting raw intensity to optical density (e.g., IndexError: index 0 is out of bounds for axis 0 with size 0).

Application Context: This typically occurs in programming environments like Python (with MNE-NIRS) or MATLAB (with Homer2/3) when the data structure is incompatible [37].

Investigation & Diagnosis:

• Cause 1: Incorrect Channel Ordering or Naming. The software cannot pair light sources and detectors because channel metadata is missing or incorrect [37].
• Cause 2: Custom Data Format. Data from custom-built hardware systems is not structured in a way the software recognizes [37].

Resolution Steps:

• Inspect Raw Data: Open your raw data file and confirm that all source-detector pairs are present and logged correctly.
• Validate Info Structure: In MNE-Python, check the info object of your RawArray to ensure all channels are listed and correctly defined.
• Use Standard File Formats: If your acquisition system supports it, save data in the SNIRF format, which is widely supported by modern toolboxes like Homer3, MNE-NIRS, and NIRS-KIT, preventing many common issues [37] [38].
• Consult Documentation: Refer to examples for custom data input in your chosen toolbox (e.g., NIRS-KIT provides a manual input function for reorganizing data into a specific .csv format) [38].

Guide to Addressing Low Signal-to-Noise Ratio (SNR) in Hemodynamic Signals

Problem: The final hemoglobin concentration signals are noisy, making it difficult to distinguish true brain activation from background interference.

Application Context: This is a pervasive challenge in fNIRS research, affecting the accuracy of both clinical and cognitive studies [15].

Investigation & Diagnosis:

• Cause 1: Physiological Noise. Signals from heart rate, respiration, and blood pressure changes can obscure the task-related brain signal [15] [34].
• Cause 2: Motion Artifacts. Subject movement can cause large, sudden shifts in the signal.
• Cause 3: Poor Optode Contact. Inadequate coupling between the optodes and the scalp results in a weak and noisy signal.
• Cause 4: Insufficient Source-Detector Distance. If the distance is too short (<25-35 mm for adults), the signal may originate primarily from the scalp and skull rather than the brain cortex [34].

Resolution Steps:

• Optimize Hardware Setup: Ensure optodes are securely attached, hair is parted, and use a source-detector distance of 25-35 mm for adult brain imaging [34].
• Apply Signal Processing: Use band-pass or high-pass filtering to remove slow drifts and cardiac pulsation. Employ algorithms (e.g., PCA, CBSI, wavelet filtering) available in toolboxes like Homer or NIRS-KIT to correct for motion artifacts and physiological noise [15] [38].
• Utilize Short-Channel Regression: If available, use short-separation detectors (<15 mm) to measure systemic noise and regress it out from the long-distance channels that record brain signals [34].
• Choose Robust Analysis Signals: For medium-low SNR data, consider using the Correlation-Based Signal Improvement (CBSI) method, which leverages the temporal correlation between HbO and HbR. For high-quality data, the Hemodynamic Phase Correlation (HPC) signal can provide excellent localization and robustness against false positives [15].

Table 1: Common fNIRS Data Conversion Errors and Solutions

Error Type	Symptoms	Likely Cause	Recommended Solution
Optical Density Conversion Error	IndexError; failure to compute OD [37] [36].	Incorrect data formatting or channel information.	Format data to meet software requirements; use SNIRF format; verify channel info structure.
Abnormally Small OD/Hb	OD values near zero (±0.03); Hb concentrations in the range of 1e-13 µM [36].	Raw intensity values are too stable and close to the mean.	Increase LED power; check device calibration; verify signal integrity.
Low Signal-to-Noise Ratio (SNR)	Noisy HbO/HbR signals; inability to detect task response [15].	Physiological noise, motion artifacts, poor optode contact.	Improve optode-scalp contact; use short-channel regression; apply CBSI processing [15].
Inconsistent HbO/HbR Responses	HbO increases without a concurrent decrease in HbR [15].	Systemic physiological interference or false positive.	Analyze both HbO and HbR; use signals like HPC that combine both; employ physiological monitoring.

Experimental Protocols for Signal Validation

Protocol: Validating the Hemodynamic Response with a Motor Task

This protocol provides a standard method to verify that your fNIRS system and analysis pipeline are correctly capturing a well-established brain activation pattern.

1. Objective To acquire and process a robust fNIRS signal from the primary motor cortex during a finger-tapping task, resulting in a canonical hemodynamic response (increase in HbO, decrease in HbR).

2. Materials and Reagents

• fNIRS system with optodes placed over the primary motor cortex (C3/C4 location according to the 10-20 system).
• Data acquisition software (e.g., NIRStar, Aurora).
• Data analysis toolbox (e.g., Homer3, NIRS-KIT, MNE-NIRS).
• Metronome or visual cueing software.

3. Step-by-Step Procedure Step 1: Subject Preparation. Position the fNIRS cap on the subject's head. Ensure optodes over the motor cortex have a source-detector distance of 25-35 mm and good contact with the scalp [34]. Step 2: Paradigm Design. Implement a block design:

• Rest Block: 30 seconds of inactivity.
• Task Block: 20 seconds of repetitive finger tapping (e.g., against the thumb) at a rate of 2Hz, guided by a metronome.
• Repeat: This rest-task cycle should be repeated at least 5-10 times to improve SNR. Step 3: Data Acquisition. Start recording and run the paradigm. Step 4: Data Processing. Process the data using a standard pipeline:
  • Convert raw intensity to Optical Density.
  • Convert Optical Density to HbO and HbR concentration changes using the MBLL.
  • Apply a band-pass filter (e.g., 0.01 - 0.2 Hz) to remove drift and high-frequency noise.
  • Segment data into epochs aligned to task onset.
  • Average epochs to generate the final hemodynamic response.

4. Expected Outcome The averaged signal from channels over the contralateral motor cortex should show a clear increase in HbO and a concurrent decrease in HbR during and shortly after the task blocks, peaking around 5-6 seconds post-stimulus [15] [34].

Workflow Diagram: fNIRS Signal Conversion and Analysis

The following diagram illustrates the complete pathway from data acquisition to interpretation, including key troubleshooting checkpoints.

Diagram 1: fNIRS Analysis Workflow

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 2: Essential Software Tools for fNIRS Data Analysis

Tool Name	Primary Function	Key Feature for Troubleshooting	Reference
Homer2 / Homer3	A comprehensive suite for fNIRS processing.	Allows users to build and customize their own processing stream, integrating custom algorithms. Supports SNIRF format.	[38] [39] [34]
NIRS-KIT	A MATLAB toolbox for both resting-state and task-based fNIRS.	Excellent compatibility; supports data from NIRS-SPM, Homer2, and SNIRF format. Provides batch processing.	[38]
MNE-NIRS	An open-source Python package for fNIRS analysis.	Integrates with the broader MNE ecosystem for EEG/MEG. Strong visualization and statistical capabilities.	[37]
NIRS Toolbox	A MATLAB-based toolbox with advanced statistical models.	Supports functional connectivity, multimodal analysis, and offers automatic import of 3-D probe information for NIRx data.	[39]
Turbo-Satori	A real-time fNIRS analysis software.	Enables real-time data visualization and processing, crucial for neurofeedback and Brain-Computer Interface (BCI) applications.	[39]

Advanced Analysis & Signal Optimization Techniques

For researchers requiring higher accuracy, advanced methods combine HbO and HbR to create more robust metrics.

Table 3: Advanced Derived Signals for Improved Analysis

Signal Name	Calculation Formula	Application Context	Advantage
Total Hemoglobin (HbT)	HbT = HbO + HbR	An indicator of total blood volume changes.	Less commonly used alone but can provide complementary information.	[15]
Hemoglobin Difference (HbD)	HbD = HbO - HbR	A measure of blood oxygenation.	Higher amplitude than HbO alone during activation.	[15]
Correlation-Based Signal Improvement (CBSI)	A model based on the temporal anti-correlation of HbO and HbR.	Medium to low SNR data.	Effective at suppressing motion artifacts and physiological noise.	[15]
Hemodynamic Phase Correlation (HPC)	A model combining HbO and HbR within a General Linear Model (GLM) framework.	High SNR data.	Provides high accuracy in localizing activation and is robust against false positives.	[15]

Functional near-infrared spectroscopy (fNIRS) is a non-invasive neuroimaging technique that monitors cerebral hemodynamics by measuring changes in oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) concentrations. Despite its advantages in portability and ecological validity, fNIRS signals are highly susceptible to motion artifacts (MAs), which are significant noise components introduced by participant movement. These artifacts arise from the disruption of optode-scalp coupling due to head, body, or facial movements (e.g., nodding, talking, jaw movement), leading to signal spikes, baseline shifts, and low-frequency variations that can obscure the true hemodynamic response [10] [11]. Effective motion artifact reduction is therefore a critical preprocessing step for ensuring the validity and reliability of fNIRS data in both research and clinical applications, particularly in vulnerable populations such as infants, children, and patients with neurological disorders where movement is common [10] [40].

Multiple algorithms have been developed to correct for motion artifacts in fNIRS data. They can be broadly categorized into hardware-based and algorithmic (software-based) methods [11]. Hardware-based methods utilize additional sensors, such as accelerometers or cameras, to record motion and inform artifact removal [41] [42]. Algorithmic methods operate on the captured fNIRS signal itself and are more widely used due to their general applicability. The table below summarizes the most prevalent algorithmic techniques, their principles, advantages, and limitations.

Table 1: Comparison of Primary Motion Artifact Correction Algorithms for fNIRS

Algorithm	Underlying Principle	Key Advantages	Key Limitations
Spline Interpolation (MARA) [10] [43]	Models artifact segments using cubic splines and subtracts them from the signal.	Effective for correcting baseline shifts [43]. Simple and practical [43].	Performance depends on accurate artifact detection; user-defined parameters required [43].
Wavelet Filtering [12] [10]	Uses discrete wavelet transform to identify and remove outlier coefficients associated with artifacts.	Powerful for removing spike-type artifacts; does not require auxiliary hardware [12] [40].	Parameter selection (threshold) is critical; can be ineffective against baseline shifts alone [43].
Temporal Derivative Distribution Repair (TDDR) [12]	Assumes motion-free fluctuations are normally distributed; uses a robust estimator to down-weight large temporal derivatives from MAs.	Effective online method; superior performance in functional connectivity analysis [12].	Performance depends on adherence to statistical assumptions [12].
Principal Component Analysis (PCA) [12] [10]	Removes the first few principal components that account for the highest variance, assumed to be dominated by MAs.	Data-driven; good for artifacts appearing across multiple channels.	Risk of removing physiological signal; limited by the number of channels [12].
Correlation-Based Signal Improvement (CBSI) [12] [10]	Exploits the negative correlation between HbO and HbR concentrations during neural activity to cancel out MAs.	Simple calculation; requires no parameters.	Relies on the negative correlation assumption, which may not always hold [12].
Kalman Filtering [12] [10]	Uses an autoregressive model and a state-space approach to predict the motion-free signal state.	Can be effective for specific artifact types.	Requires precise modeling; build-up of errors over time is possible [44].
Deep Learning (DAE) [44]	A denoising autoencoder (DAE) learns to map noisy fNIRS signals to clean ones using a synthetic training dataset.	Assumption-free; no manual parameter tuning; high computational efficiency once trained [44].	Requires a large, diverse dataset for training; complex model design [44].

Performance and Selection Guidance

Comparative studies have evaluated these algorithms on various metrics, including their ability to recover a known hemodynamic response and improve functional connectivity (FC) analysis.

Table 2: Performance Summary of Key MAR Algorithms from Comparative Studies

Algorithm	Performance in Functional Connectivity & Topology Analysis [12]	Performance on Task-Based Data with Low-Frequency Artifacts [10]	Performance on Pediatric Data [40]
Wavelet Filtering	One of the most effective methods; superior denoising and best ROC curve.	The most effective approach, correcting 93% of artifacts.	Ranked among the best methods for child data.
TDDR	One of the most effective methods; enhanced ability to recover original FC patterns.	Not assessed in the reviewed study.	Not assessed in the reviewed study.
Spline Interpolation	Performance not significantly different from most others.	Effective, but less so than wavelet filtering.	Effective, but less so than moving average and wavelet.
Moving Average (MA)	Not assessed in the reviewed study.	Not assessed in the reviewed study.	Ranked among the best methods for child data.
CBSI	Performance not significantly different from most others.	Effective, but less so than wavelet filtering.	Showed lower efficacy.
PCA	Performance not significantly different from most others.	Effective, but less so than wavelet filtering.	Showed lower efficacy.
Kalman Filtering	Performance not significantly different from most others.	Effective, but less so than wavelet filtering.	Not assessed in the reviewed study.

Based on this evidence, Wavelet filtering and TDDR are highly recommended for general use, particularly for FC analysis [12]. For challenging, low-frequency artifacts that resemble the hemodynamic response, such as those induced by speaking, wavelet filtering has been shown to be particularly powerful [10]. A hybrid approach combining spline interpolation and wavelet filtering has also been proposed to leverage the strengths of both methods—spline for baseline shifts and wavelet for spikes—achieving a channel improvement rate as high as 94.1% [43].

Experimental Protocols for Algorithm Evaluation

To ensure the robustness and reliability of MAR algorithms, they should be evaluated using well-established experimental protocols. The following methodologies are commonly used in the literature.

Protocol 1: Simulated Data with Added Ground-Truth Hemodynamic Response

Objective: To quantitatively evaluate the performance of different MAR algorithms by comparing the processed signal to a known, simulated hemodynamic response [10].

Procedure:

• Data Collection: Collect resting-state fNIRS data, which contains real motion artifacts but no task-induced neural activity.
• Add Simulated HRF: Add a simulated hemodynamic response function (HRF), typically a gamma function, to the resting-state data at predetermined time points [44]. This creates a semi-simulated dataset where the "true" signal is known.
• Apply MAR Algorithms: Process the contaminated data with various MAR algorithms.
• Performance Metrics: Calculate quantitative metrics, such as the Mean-Squared Error (MSE) and Pearson's Correlation Coefficient (R²), between the recovered HRF and the original simulated HRF [10]. This provides an objective measure of each algorithm's efficacy in recovering the true signal.

Protocol 2: Controlled Head Movement Tasks

Objective: To characterize the specific impact of different head movements on fNIRS signal quality and to validate MAR algorithms using ground-truth movement data [45].

Procedure:

• Participant Instructions: Participants perform controlled head movements along three rotational axes: pitch (nodding yes), yaw (shaking no), and roll (bending neck sideways). Movements can be further categorized by speed (fast/slow) and amplitude (half/full/repeated rotation) [45].
• Synchronized Recording:
  • fNIRS: Record whole-head fNIRS signals.
  • Motion Tracking: Record the movements using video and analyze them frame-by-frame with a deep neural network (e.g., SynergyNet) to compute precise head orientation angles [45]. Alternatively, accelerometers or gyroscopes can be used [11].
• Data Analysis: Correlate the extracted movement metrics (amplitude, speed) with identified artifacts (spikes, baseline shifts) in the fNIRS signals. This helps identify which movements most compromise signal quality in different brain regions [45].

Implementation Workflows and Signaling Pathways

The following diagrams illustrate the logical workflow for MAR algorithm implementation and the neurovascular pathway that fNIRS signals aim to capture.

Figure 1: Motion Artifact Reduction Workflow

Figure 2: Neurovascular Coupling & Motion Artifact Impact

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Essential Materials for fNIRS Research and Motion Artifact Management

Item	Function / Purpose	Example Use-Case
fNIRS System	A continuous-wave (CW), frequency-domain (FD), or time-domain (TD) device that emits near-infrared light and detects attenuated light to compute HbO and HbR concentrations.	Core neuroimaging hardware for all fNIRS studies.
Standard Optodes & Headgear	Sources and detectors are typically held in a flexible cap or rigid headband. Standard Velcro-based or foam-mounted arrays are common but can be prone to motion artifacts.	Standard setup for most cognitive and motor task studies in controlled environments [41].
Collodion-Fixed Fibers	Miniaturized optical fiber tips fixed to the scalp with collodion adhesive. This method significantly improves optode-scalp coupling and reduces motion artifacts.	Essential for long-term clinical monitoring (e.g., in epilepsy patients) or during tasks involving excessive movement, reducing motion artifact magnitude by ~90% [41].
Auxiliary Motion Sensors	Accelerometers, gyroscopes, or magnetometers integrated into the fNIRS headgear or worn by the participant.	Provides ground-truth movement data to inform hardware-based motion correction algorithms like AMARA or ABMARA [11] [42].
Software Toolboxes (Homer2/3)	Open-source fNIRS data processing packages for MATLAB. Include built-in functions for motion artifact detection (e.g., hmrMotionArtifact) and correction (e.g., wavelet, spline).	The standard starting point for data preprocessing, including MAR application and HRF estimation [40] [43].
Computer Vision Systems	Video cameras and software (e.g., SynergyNet deep neural network) for frame-by-frame analysis of head orientation.	Provides detailed, markerless ground-truth movement data for characterizing artifact morphology and validating MAR methods [45].

Frequently Asked Questions (FAQs) and Troubleshooting

Q1: I am analyzing fNIRS data from children, which is notoriously noisy. Which motion correction algorithm should I use? A: For pediatric data, studies have shown that Moving Average (MA) and Wavelet filtering methods yield the best outcomes [40]. Children's data often contains a heterogeneous mix of artifact types, and these methods have proven effective in handling this variability. It is always recommended to compare the performance of a few top algorithms on a subset of your data.

Q2: When I correct for motion artifacts, am I at risk of removing the real neural signal? A: Yes, this is a valid concern. Most MAR algorithms operate on the principle that motion artifacts have characteristics distinct from the true hemodynamic response (e.g., larger amplitude, higher frequency). However, aggressive filtering or inappropriate parameter settings can distort the signal of interest. This is why validation using simulated data (where the true signal is known) is a critical step. Algorithms like TDDR and Wavelet filtering have demonstrated a strong ability to recover the original signal in functional connectivity analyses, making them robust choices [12].

Q3: My motion artifacts are very slow and look similar to the hemodynamic response. What is the best correction method? A: Low-frequency, task-correlated artifacts are particularly challenging. Research indicates that Wavelet filtering is the most effective technique for this specific type of artifact, successfully correcting over 90% of cases in studies involving vocalization tasks [10]. A hybrid method that first uses spline interpolation to correct baseline shifts and then applies wavelet filtering can also be highly effective [43].

Q4: Is it better to use a hardware-based solution or an algorithmic one? A: Both have their place. Hardware-based solutions (e.g., collodion-fixed fibers) are proactive and can reduce the magnitude of motion artifacts at the source by up to 90%, which is ideal for long-term monitoring or highly mobile subjects [41]. Algorithmic solutions are reactive but more flexible and widely applicable, as they do not require special equipment. For most standard experiments, a robust algorithmic approach (e.g., Wavelet, TDDR) is sufficient. For studies where movement is intrinsic to the task (e.g., walking, seizures), a combination of secure hardware and advanced algorithms is recommended.

Q5: I'm using the spline interpolation method. How can I improve its artifact detection? A: Accurate detection is crucial for spline interpolation. Instead of relying solely on the moving standard deviation method, consider using the novel kurtosis-based Wavelet Detection (kbWD) algorithm. This method uses the distribution of wavelet coefficients to identify artifacts and requires only a single threshold parameter (kurtosis), making it more adaptable to varying signal-to-noise ratios and reducing user bias [43].

Q6: Are there fully automated methods that don't require manual parameter tuning? A: Yes, this is an emerging area. The AMARA algorithm is a promising automatic method that uses accelerometer data [42]. Furthermore, Deep Learning approaches, such as Denoising Autoencoders (DAEs), are assumption-free and require no manual parameter tuning once the model is trained, showing great potential for accurate and efficient artifact removal [44].

FAQs on Filtering Physiological Noise in fNIRS

Q1: Why is filtering necessary for removing physiological noise in fNIRS? fNIRS signals are contaminated by structured, or "colored," physiological noise from systems like cardiac pulsation (around 1 Hz), respiration (around 0.3 Hz), and blood pressure oscillations known as Mayer waves (around 0.1 Hz) [9] [46]. Unlike white noise, this noise has strong temporal autocorrelation, meaning each data point is not independent of the next [46]. Filtering is a primary pre-processing step to attenuate these specific frequency components, which otherwise obscure the task-evoked hemodynamic response and can lead to false positives or negatives in the data [47] [17].

Q2: What are the standard frequency cut-offs for high-pass, low-pass, and band-pass filters in fNIRS? The choice of cut-off frequencies depends on the expected frequency of the hemodynamic response and the frequencies of the noise you aim to remove. The table below summarizes typical cut-offs and their purposes [17]:

Filter Type	Typical Cut-off Frequencies	Primary Purpose
High-Pass Filter	0.01 - 0.02 Hz [17]	Removes very slow drifts (e.g., from instrumental noise) on the scale of several minutes.
Low-Pass Filter	0.1 - 0.2 Hz [17]	Attenuates high-frequency noise like cardiac pulsation (~1 Hz) and respiration (~0.3 Hz).
Band-Pass Filter	0.01 - 0.2 Hz [17]	A combination that removes both very slow drifts and high-frequency physiological noise.

Q3: A band-pass filter removed most of the cardiac noise, but I still see strong, slow oscillations in my data. What else could be causing this? Band-pass filtering is effective for removing cardiac and respiratory noise, but it may not fully eliminate other physiological noises like Mayer waves (~0.1 Hz) or task-evoked systemic artifacts in the very low-frequency (VLFO) band (0.02-0.08 Hz) [9]. These noises often fall within the passband of a standard 0.01-0.2 Hz filter. To address this, consider these advanced methods:

• Short-Channel Regression: This is considered a best-practice method. It uses an additional, short source-detector separation channel (~8 mm) that is sensitive only to scalp hemodynamics. The signal from this short channel is used as a nuisance regressor to remove the superficial component from the long-channel signal [7] [48].
• Principal Component Analysis (PCA) / Independent Component Analysis (ICA): These blind source separation techniques can identify and remove dominant, spatially global noise components from your data [47] [7].
• General Linear Model (GLM) with Physiological Regressors: Record auxiliary signals like blood pressure, heart rate, and skin blood flow. These can be incorporated as nuisance regressors in a GLM to account for their influence on the fNIRS signal [9] [7].

Q4: After applying a high-pass filter, my hemodynamic response looks distorted. What might have happened? This can occur if the filter's cut-off frequency is set too high. The hemodynamic response is a slow signal, and its fundamental frequency components are very low. An excessively aggressive high-pass filter (e.g., with a cut-off above 0.05 Hz) can attenuate or distort these slow components, altering the shape and amplitude of the recovered response [17]. Always verify the effect of your filter parameters on a known or simulated hemodynamic response before applying it to your experimental data.

Experimental Protocol: Identifying and Filtering Physiological Noise

This protocol provides a methodology for characterizing physiological noise in a resting-state recording, which can inform the design of your filtering pipeline for task-based studies [9].

1. Objective To identify the dominant frequency components of physiological noise in your fNIRS setup and subject population to optimize filter parameters.

2. Materials and Reagents

• fNIRS System: A continuous-wave or time-domain fNIRS system with sources and detectors [47].
• Auxiliary Physiological Monitors (Recommended): Equipment to simultaneously record:
  • Electrocardiogram (ECG) or Pulse Oximeter: For heart rate.
  • Respiratory Belt: For respiration.
  • Blood Pressure Monitor: For continuous, non-invasive arterial pressure.
  • Skin Blood Flow Monitor: Laser Doppler flowmetry on the forehead [9].

3. Procedure

• Step 1: Data Acquisition. Set up your fNIRS probe on the subject's forehead (a region highly susceptible to physiological noise). Record a 10-minute resting-state fNIRS signal (e.g., oxygenated hemoglobin concentration) concurrently with the auxiliary physiological signals. Ensure a high sampling rate (>> 1 Hz) to adequately capture cardiac pulsations [9].
• Step 2: Spectral Analysis. Compute the power spectral density (PSD) of the recorded fNIRS signal. This can be done using methods like the Welch periodogram.
• Step 3: Identify Noise Peaks. In the PSD plot, identify prominent peaks and note their frequencies. Correlate these peaks with the frequencies of the auxiliary signals:
  • A peak at ~1 Hz is likely cardiac noise.
  • A peak at ~0.3 Hz is likely respiratory noise.
  • A peak at ~0.1 Hz is often associated with Mayer waves.
• Step 4: Apply and Validate Filter.
  • Based on your spectral analysis, design a band-pass filter (e.g., 0.01 - 0.2 Hz) or a set of notch filters to target the identified noise peaks.
  • Apply this filter to your resting-state data.
  • Generate a new PSD plot of the filtered data and confirm the attenuation of the target noise frequencies.

The following workflow summarizes the experimental protocol for identifying and addressing physiological noise:

The Scientist's Toolkit: Key Research Reagents & Solutions

The table below lists essential "reagents" or tools for an experiment focused on physiological noise correction.

Item / Solution	Function in Experiment
fNIRS System with Short-Separation Channels	The primary instrument. Short-separation channels (<1.5 cm) are crucial for measuring and regressing out scalp hemodynamic noise [7] [48].
Auxiliary Physiological Monitors	Devices to record ECG, respiration, and blood pressure. These signals are used to identify noise sources and can be included as regressors in a GLM for denoising [9] [7].
Signal Processing Software (e.g., MATLAB, Python, Homer2, nirsLAB)	Software environments used to implement filtering, GLM, PCA/ICA, and wavelet coherence analysis [47] [17].
Wavelet Coherence Analysis (WCA)	An analytical method to quantify the coupling between the fNIRS signal and auxiliary physiological measurements across different time scales, helping to identify the dominant noise processes [9].
General Linear Model (GLM)	A statistical framework used to estimate the task-evoked hemodynamic response while modeling out nuisance factors like physiological noise and motion artifacts [9] [46].

Functional near-infrared spectroscopy (fNIRS) has emerged as a valuable tool for non-invasive monitoring of cerebral cortical activity, with particular relevance for clinical research and drug development studies. Unlike functional MRI, fNIRS offers portability, compatibility with medical implants, and greater tolerance for subject movement, making it suitable for diverse patient populations and experimental settings [3] [4]. However, the fNIRS signal is notoriously susceptible to various noise sources, including motion artifacts, physiological interference from superficial tissues, and systemic physiological fluctuations [49] [4]. Effective denoising is therefore prerequisite for obtaining reliable hemodynamic data in both basic research and clinical applications.

This technical support guide focuses on two advanced denoising techniques: Principal Component Analysis (PCA) and Short-Separation Channel Regression. These methods address the critical challenge of separating cortical hemodynamic signals from confounding noise sources, thereby improving the accuracy and interpretability of fNIRS data in clinical research contexts.

Troubleshooting Guides

Guide 1: Addressing Poor Hemodynamic Response Recovery After Short-Separation Channel Regression

Problem: After applying short-separation channel regression, the recovered hemodynamic response function (HRF) appears non-physiological, shows unexpected negative dips in oxyhemoglobin (HbO), or lacks the characteristic response morphology.

Explanation: Short-separation regression aims to remove systemic physiological noise originating from superficial tissues (scalp, skull) by subtracting signals from short-distance channels (<1.5 cm) from conventional long-separation channels (~3 cm) [49] [50]. However, improper implementation can lead to inadequate noise removal or accidental removal of neural signals.

Solutions:

• Verify Short-Channel Proximity: Ensure short-separation channels are located within 1.5-2 cm of the long-separation channels they are regressing. Performance degrades significantly with greater distances due to spatial heterogeneity in scalp hemodynamics [50].
• Optimize Regressor Selection: Use multiple short-separation channels as regressors rather than a single global regressor. Spatial heterogeneity in scalp hemodynamics means different regions exhibit different noise characteristics [49].
• Check for Mayer Wave Influence: Incorporate specific regression of Mayer waves (approximately 0.1 Hz oscillations), which are a prominent source of physiological noise that affects short and long channels differently [49].
• Validate with Resting-State Data: Test your regression algorithm on resting-state data with synthetic hemodynamic responses added to quantify performance before applying to experimental data [50].

Table 1: Impact of Short-Separation Channel Distance on Signal Quality

Distance Between Short and Long Channel	Improvement in Contrast-to-Noise Ratio (CNR)	Practical Utility
≤ 1.5 cm	50% for HbO, 100% for HbR	High
1.5 - 2 cm	Moderate but variable	Moderate
> 2 cm	Mild to negligible	Low

Guide 2: Managing Motion Artifacts in fNIRS Signals

Problem: fNIRS signals contain sharp spikes, baseline shifts, or irregular patterns caused by subject movement, particularly in studies involving speech, chewing, or patient populations with difficulty remaining still.

Explanation: Motion artifacts disrupt the coupling between optodes and scalp, causing temporary signal disruptions that can mimic or obscure true hemodynamic responses [51] [4]. Jaw movements are particularly problematic for temporal and prefrontal measurements.

Solutions:

• Implement Mechanical Stabilization: Use an individually customized bite bar apparatus during experiments involving auditory or language tasks to suppress jaw movement artifacts [51].
• Apply specialized Motion Correction Algorithms: Implement algorithms such as:
  • Wavelet-Based Correction: Effective for spike-like artifacts but requires MATLAB Wavelet Toolbox [52]
  • PCA-Based Denoising: Removes motion artifacts through identification and rejection of noise-related components [51] [4]
  • Spline Correction: Particularly effective for correcting motion artifacts with sharp spikes [4]
• Combine Multiple Approaches: Use short-separation channels in conjunction with motion sensors (accelerometers) to better distinguish motion artifacts from physiological noise [4].

Table 2: Motion Artifact Reduction Algorithms Comparison

Algorithm Type	Best For	Implementation Considerations
Wavelet Transformation	Spike-like artifacts	Requires MATLAB Wavelet Toolbox [52]
PCA-Based Denoising	Various artifact types	Requires statistics toolbox [4] [52]
Spline Interpolation	Sharp, transient artifacts	Available in Homer2 and other packages [4]
Volatility Correction	Step-like artifacts	Computationally efficient [52]
CBSI Algorithm	Coupled changes in HbO and HbR	Doesn't require short channels [52]

Guide 3: Optimizing PCA Denoising Parameters

Problem: PCA denoising either removes insufficient noise or accidentally eliminates neural signals of interest, resulting in compromised data quality.

Explanation: PCA separates fNIRS signals into components based on variance, assuming that noise contributes to specific high-variance components. Incorrect component selection can lead to under- or over-correction [4].

Solutions:

• Determine Optimal Component Number: Use a combination of variance explained and visual inspection of component time courses to select components for removal. Physiological noise typically appears in high-variance components with characteristic frequency signatures [4].
• Combine with Bandpass Filtering: Apply bandpass filtering (typically 0.01-0.5 Hz) before or after PCA to enhance separation of hemodynamic signals from higher-frequency noise [4].
• Validate with Task Design: Compare component time courses with task timing to avoid removing task-related components. True hemodynamic responses should correlate with experimental paradigm timing [4].
• Implement Pipeline Sequencing: Place PCA after initial motion correction and filtering steps but before hemodynamic response function estimation for optimal results [4].

Frequently Asked Questions (FAQs)

Q1: Why does my fNIRS data show negative HbO responses during tasks, and how can I address this?

Negative HbO responses during expected activation can result from several factors: (1) Inadequate removal of systemic physiological noise, particularly Mayer waves; (2) Motion artifacts from jaw movements or head displacement; (3) Improper short-separation regression that removes neural signals along with noise; or (4) Genuine neurovascular uncoupling in certain patient populations. To address this, first verify your data quality and preprocessing pipeline using a motor task with known response properties before proceeding to novel paradigms [53]. Implement combined short-separation regression and PCA denoising, and ensure proper optode coupling throughout the experiment [49] [4].

Q2: How many short-separation channels do I need for a whole-head fNIRS montage?

There is no definitive rule, but current evidence suggests that multiple short-separation channels distributed across the head are necessary due to spatial heterogeneity in scalp hemodynamics [49] [50]. Aim for at least one short-separation channel per region of interest, positioned within 1.5 cm of each long-separation channel in that region. For whole-head coverage, a practical approach is to place short-separation channels around each source optode, as implemented in modern high-density arrays [49].

Q3: Can I use PCA denoising without short-separation channels?

Yes, PCA denoising can be implemented without short-separation channels and provides significant noise reduction, particularly for motion artifacts and global physiological fluctuations [4]. However, for optimal removal of superficial physiological noise, short-separation regression is superior. When both methods are available, they can be combined in a processing pipeline for enhanced denoising performance [4] [52].

Q4: What are the most critical steps to validate a denoising pipeline for clinical applications?

For clinical applications, validation should include: (1) Quantifying the contrast-to-noise ratio improvement using synthetic responses added to resting-state data [50]; (2) Demonstrating robust recovery of expected responses in canonical activation paradigms (e.g., motor tasks) [54]; (3) Assessing test-retest reliability within subjects; and (4) Verifying that denoising does not introduce biases in group comparisons relevant to your clinical research question [4].

Experimental Protocols & Methodologies

Protocol 1: Validating Denoising Performance with Synthetic Signals

This protocol evaluates the efficacy of denoising methods by adding known hemodynamic responses to resting-state data, enabling quantitative performance assessment [50].

• Acquire resting-state fNIRS data (15 minutes) from your subject population using your standard montage, including short-separation channels.
• Preprocess the data with standard steps (conversion to optical density, bandpass filtering) but without denoising algorithms applied.
• Create a synthetic hemodynamic response using a canonical HRF model (e.g., double-gamma function) with timing that matches a block or event-related design.
• Add the synthetic response to all long-separation channels at a specified amplitude, simulating brain activation.
• Apply your denoising algorithms (short-separation regression, PCA, or combined) to the data with added synthetic responses.
• Recover the HRF using general linear modeling or block averaging.
• Quantify performance by comparing the recovered response amplitude and contrast-to-noise ratio to the known input response.

Table 3: Key Performance Metrics for Denoising Validation

Metric	Calculation	Target Outcome
Contrast-to-Noise Ratio (CNR)	Response amplitude / background standard deviation	Maximum increase after denoising [50]
Signal-to-Noise Ratio (SNR)	Signal power / noise power	>50% improvement
Correlation with Canonical HRF	Pearson's r between recovered and model HRF	r > 0.7
Residual Motion Artifacts	Visual inspection and outlier detection	Minimal residual spikes

Protocol 2: Combined PCA and Short-Separation Regression Pipeline

This methodology outlines a comprehensive denoising approach combining both techniques, optimized for clinical research applications [4] [52].

• Data Conversion and Initial Preparation:

  • Convert raw intensity to optical density [54]
  • Mark obviously noisy channels based on signal quality indices
  • Trim data to remove pre-experiment baseline if necessary

• Motion Artifact Correction:

  • Apply wavelet-based motion correction or spline interpolation
  • Mark motion-contaminated segments for potential exclusion in subsequent analysis

• Conversion to Hemoglobin Concentration:

  • Apply modified Beer-Lambert law to convert optical density to oxy- and deoxyhemoglobin concentration changes [54]

• Short-Separation Regression:

  • For each long-separation channel, identify the nearest short-separation channel (<1.5 cm distance)
  • Apply Kalman filtering or general linear model to regress out the short-separation signal from each long-separation channel [49] [50]

• PCA Denoising:

  • Perform PCA on the hemoglobin concentration data
  • Identify noise components based on frequency content, variance distribution, and visual inspection
  • Remove identified noise components and reconstruct the signal

• Additional Filtering and Epoch Extraction:

  • Apply final bandpass filter (0.01-0.5 Hz) to remove residual high-frequency noise
  • Extract epochs relative to experimental events
  • Perform statistical analysis on denoised epochs

Signaling Pathways and Workflows

fNIRS Denoising Logical Pathway

Short-Separation Regression Mechanism

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 4: Essential Materials for fNIRS Denoising Research

Item	Function/Application	Example/Specification
fNIRS System with Short-Separation Capability	Enables collection of superficial signals for regression	NIRSport, NIRScout with short-separation optodes [49] [52]
Customized Bite Bar	Controls jaw-related motion artifacts in auditory/language studies	Acrylic piece with dental putty molding [51]
MATLAB with Toolboxes	Implementation of advanced denoising algorithms	Statistics, Signal Processing, Wavelet Toolbox [52]
Homer2/Homer3 Software Package	Standard fNIRS processing pipeline including basic denoising functions	Open-source fNIRS analysis platform [52]
Accelerometers/Motion Sensors	Supplementary motion detection for artifact identification	3-axis accelerometers synchronized with fNIRS [4]
Anatomical Co-registration Tools	Mapping optode locations to brain anatomy for improved interpretation	MRI-compatible digitization systems [3]

Hemodynamic Response Function (HRF) Estimation and General Linear Model (GLM)

Core Concepts and Common Challenges in fNIRS Analysis

What are the HRF and GLM, and why are they crucial for fNIRS research?

The Hemodynamic Response Function (HRF) models the temporal coupling between neural activity and subsequent hemodynamic changes measured by fNIRS. It describes the typical delay and shape of the blood oxygenation response following neuronal firing [55]. This response is not fixed; its shape can vary significantly across different brain regions, between individual subjects, and even across trials within the same subject [56] [55].

The General Linear Model (GLM) is a statistical framework used to estimate the strength of brain activation in fNIRS data. It works by modeling the measured fNIRS signal as a linear combination of explanatory variables (regressors), plus an error term. The most critical regressor is typically constructed by convolving the experimental task timeline (a boxcar or delta function) with a model of the HRF [57] [58]. The GLM's primary advantage is its ability to simultaneously separate task-evoked brain activity from various confounding noise sources, thereby increasing the contrast-to-noise ratio (CNR) of the true neural signal [57].

Frequently Asked Questions and Troubleshooting

• FAQ: My estimated HRF amplitudes or optical density (OD) values are extremely small. What could be wrong? This is often a data quality issue at the acquisition stage. Extremely small OD values, for instance in the range of [-0.03, 0.03], can occur if the raw light intensity values are too close to the mean intensity used in the conversion calculation. This results in a logarithm that is near zero, subsequently producing negligible HRF concentrations [36].

  • Solution: First, verify the quality of your raw signals. Ensure proper optode-scalp contact. If possible, increase the source light intensity (e.g., LED power) during data collection to improve signal strength [36].

• FAQ: Why should I use a GLM instead of simple block averaging? While block averaging is straightforward, the GLM offers several key advantages for fNIRS analysis [57]:

  • Superior Noise Handling: The GLM can incorporate "nuisance regressors" to model and remove systemic physiological noise (e.g., cardiac, respiratory) and motion artifacts simultaneously with the HRF estimation.
  • Increased Statistical Power: It uses the entire temporal structure of the fNIRS time series, not just averages of pre-defined blocks.
  • Flexibility: It can easily model complex experimental designs with mixed trial types and varying durations.

• FAQ: My results are inconsistent or hard to replicate. What are the key factors affecting reproducibility? A large-scale reproducibility study (FRESH) found that agreement on fNIRS results is influenced by several factors [19]:

  • Data Quality: Higher quality data leads to better agreement between different analysis pipelines.
  • Researcher Experience: Teams with more years of fNIRS experience and higher self-reported confidence showed greater consensus.
  • Analytical Choices: The handling of poor-quality data, the specific model used for the HRF, and the choices for statistical analysis are major sources of variability across studies.

• FAQ: How can I improve the GLM's performance in removing physiological noise? The current best practice is to include short-separation (SS) channels as nuisance regressors in the GLM, as they capture systemic physiological artifacts present in the scalp [57]. For even better performance, consider advanced methods like temporally embedded Canonical Correlation Analysis (tCCA), which flexibly combines multiple auxiliary signals (like SS channels, heart rate, or respiration) to create optimal nuisance regressors. This approach has been shown to significantly improve HRF recovery, especially in low-CNR scenarios or when few trials are available [59].

The following table summarizes the performance improvements of advanced noise regression methods compared to standard GLM, as demonstrated in simulation and experimental studies.

Table 1: Quantitative Performance Improvement of GLM with tCCA vs. Standard GLM with Short-Separation Regression [59]

Metric	Improvement for HbO	Description
Correlation with True HRF	Increase of up to +45%	Better recovery of the true HRF shape.
Root Mean Squared Error (RMSE)	Reduction of up to -55%	Lower error in the estimated response.
F-Score	Increase of up to 3.25-fold	Improved balance between sensitivity and specificity in detecting activation.

Detailed Experimental Protocols for Robust HRF Estimation

Protocol 1: HRF Estimation using an Iterative Optimization Algorithm

This protocol is suitable for estimating subject- and region-specific HRF shapes without assuming a rigid canonical form [56].

• Signal Preprocessing: Begin by converting raw light intensities to optical densities (OD) and then to concentrations of HbO and HbR using the Modified Beer-Lambert Law.
• Define the HRF Model: Model the HRF using a combination of two Gamma functions to characterize the positive response and potential undershoot. This model has several free parameters (e.g., delay of response and undershoot, their dispersions, scaling, and baseline) [56].
• Formulate the Complete Model: Construct the measured signal as a linear combination of the evoked-HR (convolution of the task paradigm with the model from Step 2), a baseline component, and models of physiological noises (cardiac, respiratory, Mayer waves), whose frequencies and amplitudes are also treated as free parameters [56].
• Set Up the Optimization Problem: Define an objective function, typically the sum of squared residuals between the model and actual data, with constraints on all free parameters (up to 12 in this model) [56].
• Solve Iteratively: Use an iterative optimization algorithm (e.g., a simplex method like Nelder-Mead) to find the parameter values that minimize the objective function. Initialize the algorithm with physiologically plausible values from the literature [56].
• Validate the Model: Verify the algorithm's accuracy using simulated data with known ground-truth parameters before applying it to experimental data [56].

Protocol 2: GLM Analysis with Enhanced Physiological Noise Regression

This protocol outlines a standard GLM approach, enhanced with advanced noise regression techniques for improved single-trial analysis, which is critical for BCI and clinical applications [57] [59].

• Construct the Design Matrix (X):

  • Task Regressor: Create a boxcar function representing the "on" and "off" periods of your task. Convolve this with a chosen HRF model (e.g., canonical HRF) to generate the primary task-related regressor.
  • Nuisance Regressors: Incorporate regressors to account for non-neural signals. Best practices include:
    • Short-Separation Regressors: Include signals from short-distance channels to capture systemic physiology from the scalp [57].
    • tCCA Regressors: Use the GLM with tCCA method to generate optimal nuisance regressors from a combination of SS channels and other available auxiliary signals (e.g., heart rate, respiration, motion parameters) [59].
  • Other Confounds: Add polynomial drift terms to model slow signal drifts.

• Model Fitting: Solve the linear model Y = X * β + ε for the regression coefficients (β), which represent the strength of the brain activity attributed to the task. Y is the preprocessed fNIRS data (HbO/HbR), and ε is the error term [57].

• Statistical Inference: Perform hypothesis tests (e.g., t-tests) on the task-related β coefficients to determine if they are significantly different from zero, indicating statistically significant brain activation. Correct for multiple comparisons if testing across many channels [16].

The workflow for this GLM protocol, including the advanced tCCA step, is visualized below.

This table lists key software tools and methodological "reagents" essential for conducting HRF and GLM analysis in fNIRS research.

Table 2: Essential Tools and Resources for fNIRS HRF/GLM Analysis

Item Name	Function/Brief Explanation	Relevance to HRF/GLM
Canonical HRF Model	A standard mathematical model (often based on two Gamma functions) representing the typical hemodynamic response [56].	Serves as the default basis for constructing the task regressor in the GLM design matrix.
Short-Separation (SS) Channels	fNIRS channels with a small source-detector distance (e.g., ~0.8 cm) that predominantly measure systemic physiology in the scalp [57].	Crucial as nuisance regressors in the GLM to separate non-cerebral physiological noise from brain activity.
NIRS-SPM Toolbox	A public statistical toolbox for SPM-based analysis of NIRS data, incorporating GLM and wavelet analysis [56].	Provides a validated implementation of the GLM for fNIRS, including options for different basis functions.
HRfunc Tool	A Python-based tool for estimating HRF and performing deconvolution, designed to model HRF variability [55].	Enables estimation of subject- and context-specific HRFs, moving beyond the rigid canonical model.
tCCA Framework	A methodological framework (temporally embedded Canonical Correlation Analysis) for creating optimized nuisance regressors [59].	An advanced alternative to standard SS regression; significantly improves noise removal and HRF recovery in the GLM.
Homer2 / Homer3	Widely used fNIRS analysis packages providing a GUI and scripting environment for complete processing streams [39].	Common platforms for implementing custom preprocessing and GLM analysis pipelines.

Advanced Methodological Considerations

For clinical applications, it is vital to recognize that the HRF can be altered by pathology, age, or medication [55] [16]. Assuming a canonical HRF in these populations may lead to inaccurate results. Methodologies that allow the HRF shape to be estimated from the data itself (as in Protocol 1) or tools like HRfunc that leverage a database of HRFs from similar populations are highly recommended [55].

Furthermore, the entire analytical pipeline, including all preprocessing steps, parameter choices, and statistical thresholds, must be pre-registered or clearly documented to enhance reproducibility, especially given the known variability introduced by different analysis choices [19]. The following diagram illustrates the core-concept relationship between neural activity, the confounding factors in fNIRS, and how the GLM separates them.

Clinical Application Case Studies

Parkinson's Disease (PD): Early Diagnosis with Machine Learning

Experimental Protocol & Methodology A cross-sectional study investigated fNIRS for early PD diagnosis using machine learning. Researchers recruited 120 PD patients (60 in Hoehn and Yahr stage 1, 60 in stage 2) and 60 healthy controls [60]. Data acquisition utilized an ETG-4000 near-infrared brain function imaging instrument with 8 emitting and 7 detecting optodes placed 3 cm apart, forming 22 channels covering prefrontal regions including the Frontopolar Cortex (FPC), Dorsolateral Prefrontal Cortex (DLPFC), and Brodmann Area 8 [60]. The experimental design employed a block paradigm with a 10 Hz sampling frequency. Participants performed tasks while fNIRS monitored cerebral blood oxygen changes. Data processing involved a general linear model with β-value extraction, followed by analysis with four machine learning models: Support Vector Machine (SVM), K-Nearest Neighbors (K-NN), Random Forest (RF), and Logistic Regression (LR) [60]. Model interpretability was enhanced using SHapley Additive exPlanations (SHAP) technology.

Key Experimental Findings The SVM model demonstrated superior performance in differentiating PD patients from controls with 85% accuracy, an f1 score of 0.85, and an area under the ROC curve of 0.95 [60]. SHAP analysis identified channels CH01, CH04, CH05, and CH08 (located in the Frontopolar Cortex) as the most contributory to classification, highlighting this region's importance in early PD detection [60].

Table 1: Machine Learning Model Performance for PD Diagnosis

Model	Accuracy	F1 Score	AUC
SVM	85%	0.85	0.95
K-NN	Not Specified	Not Specified	Not Specified
RF	Not Specified	Not Specified	Not Specified
LR	Not Specified	Not Specified	Not Specified

A separate study investigating cognitive impairment in PD used fNIRS during Stroop task performance across 45 PD patients categorized into normal cognition (PD-NC), mild cognitive impairment (PD-MCI), and dementia (PDD) groups [61]. Results showed significant hypoactivation in the DLPFC, primary motor cortex (M1), and premotor cortex (PMC) in PD-MCI patients, while PDD patients showed increased activation in medial PFC, orbitofrontal cortex (OFC), and DLPFC [61]. Increased DLPFC activation correlated with poorer executive function, and fNIRS with SVM classification distinguished PD-MCI from cognitively normal individuals with 83.3% accuracy [61].

Stroke Rehabilitation: Monitoring Motor Recovery

Experimental Protocol & Methodology fNIRS has been widely applied to monitor post-stroke rehabilitation, particularly for upper limb motor recovery. Studies typically involve stroke patients performing motor tasks while fNIRS measures cortical activation patterns [62]. Research designs often include both affected and unaffected limb movements to assess hemispheric asymmetry and compensatory mechanisms. The technique monitors hemodynamic changes in motor and sensory cortices, assessing functional connectivity and activation progression during recovery [62]. Data processing involves calculating lateralization indices and analyzing functional connectivity between brain regions to track neuroplastic changes.

Key Experimental Findings Studies reveal that stroke patients exhibit different cortical activation patterns compared to healthy controls. Healthy individuals typically show contralateral hemisphere dominance during unilateral tasks (lateralization index of 0.268), while stroke patients demonstrate bilateral motor cortex activation (lateralization index of -0.009) [62]. This suggests compensatory involvement of the ipsilateral hemisphere. Research indicates that recovery of motor function correlates with improved symmetry in primary motor areas between hemispheres, with cortical activation patterns transitioning from the healthy to the affected hemisphere as motor function improves [62]. Patients with severe motor dysfunction show more extensive bilateral functional connectivity involving prefrontal, motor, and occipital areas compared to those with moderate dysfunction [62].

Table 2: fNIRS Applications in Neurological Rehabilitation

Condition	Key Findings	Clinical Utility
Parkinson's Disease	Decreased FPC activation in early PD; altered PFC activation patterns across cognitive stages	Early diagnosis; cognitive impairment tracking
Stroke Motor Recovery	Shift from bilateral to contralateral activation with recovery; enhanced ipsilateral compensation	Rehabilitation monitoring; treatment efficacy assessment
Cochlear Implant Outcomes	Distinct temporal cortex activation patterns between good and poor performers	Rehabilitation strategy optimization; outcome prediction

Cochlear Implant Research: Cortical Adaptation Patterns

Experimental Protocol & Methodology A comprehensive study investigating cortical factors in cochlear implant (CI) outcomes included 46 CI users and 26 normal-hearing controls [63] [64]. Participants completed a multimodal speech comprehension task using the German Matrix Sentence Test (OLSA) under four conditions: speech-in-quiet, speech-in-noise, audiovisual speech, and visual speech (lipreading) [63]. fNIRS recordings covered prefrontal, temporal, and visual cortices using a FOIRE-3000 continuous-wave system with 16 light sources and 16 detectors, sampling at 14 Hz [63] [64]. The experimental design included 13-second stimuli with each sentence repeated three times, followed by comprehension questions. Additional data included detailed metadata, patient history, hearing tests, behavioral measures, and spatially registered probe positions [63].

Key Experimental Findings The study revealed distinct brain activation patterns between CI users with good speech understanding (good performers, GP) and those with poor outcomes (poor performers, PP) [64]. GP participants showed brain activation patterns in temporal regions during listening tasks comparable to normal-hearing individuals, indicating successful hearing rehabilitation [64]. In contrast, PP participants relied more heavily on visual cues and showed altered neural resource allocation during audio-only conditions [64]. Both GP and PP groups demonstrated adaptive mechanisms during visual speech processing, but PP participants showed greater visual reliance potentially limiting rehabilitation success. These findings highlight the role of cortical factors in CI outcomes and suggest potential biomarkers for predicting rehabilitation success.

Troubleshooting Guides and FAQs: Addressing fNIRS Signal Noise

Frequently Asked Questions

Q: What are the main sources of signal noise in fNIRS data? A: fNIRS signals are affected by several noise sources: (1) Physiological noise from cardiorespiratory cycles (0.8-1.5 Hz heartbeat, 0.1-0.5 Hz respiration), blood pressure oscillations (~0.1 Hz), and very low-frequency vasomotion (<0.1 Hz); (2) Motion artifacts caused by subject movement creating optode-scalp displacement; (3) Instrumental noise from thermal noise, shot noise, and ambient light interference; (4) Poor optode-scalp coupling resulting in insufficient signal-to-noise ratio [65] [66].

Q: How can I identify noisy channels in my fNIRS data? A: Noisy channels can be identified through: (1) Visual inspection of raw signal for abnormal spikes, baseline shifts, or flatlines; (2) Standard deviation thresholding - channels exceeding specific standard deviation thresholds indicate poor signal quality; (3) Signal quality metrics including scalp coupling index, peak spectral power, and coefficient of variation; (4) Automatic detection algorithms that apply quantitative criteria to flag aberrant channels for exclusion or correction [65] [66].

Q: What is the recommended sequence for fNIRS signal processing steps? A: Research recommends this processing order: (1) Quality control to identify and remove noisy channels that could affect downstream analyses; (2) Artifact removal to correct portions affected by movement artifacts; (3) Noise removal using bandpass filters (typically 0.01-0.5 Hz) to remove physiological noise not of interest; (4) Conversion to hemoglobin concentrations using the modified Beer-Lambert law [65]. This sequence prevents highly noisy channels and artifacts from influencing general noise removal algorithms.

Q: How does fNIRS compare to other neuroimaging modalities for clinical populations? A: fNIRS offers unique advantages: (1) Superior motion tolerance compared to fMRI, making it suitable for patients with movement disorders like PD; (2) Non-invasive, silent operation compatible with cochlear implants and other devices; (3) Direct hemoglobin measurement without radiation exposure like PET; (4) Better spatial resolution than EEG and finer temporal resolution than fMRI; (5) Portability allowing ecological studies in clinical settings [62] [29] [61]. These characteristics make fNIRS particularly valuable for clinical populations requiring naturalistic assessment environments.

Troubleshooting Common fNIRS Signal Quality Issues

Problem: Excessive motion artifacts in patient populations

• Solution: Implement movement artifact correction algorithms (e.g., wavelet-based, ICA, spline interpolation). For tremor-prone patients (e.g., PD), consider shorter trial durations and secure optode fixation with additional padding. For clinical populations with involuntary movements, use motion-tolerant acquisition parameters and include rest periods between tasks [65] [61].

Problem: Poor signal-to-noise ratio in specific channels

• Solution: Check optode-scalp coupling and reapplying with more conductive gel. Identify noisy channels using quantitative metrics (standard deviation, scalp coupling index) and exclude them from analysis if they don't meet quality thresholds. Ensure proper optode positioning without hair obstruction and verify detector saturation levels during setup [65] [66].

Problem: Inconsistent hemodynamic responses across subjects

• Solution: Standardize preprocessing parameters across all subjects including filter ranges (typically 0.01-0.5 Hz for functional signals) and artifact detection thresholds. Implement quality control metrics to exclude subjects with insufficient signal quality. Ensure consistent experimental conditions including ambient lighting, noise levels, and subject instructions [65].

Problem: Physiological confounding in clinical populations

• Solution: Apply short-separation channels (typically 8-15 mm) to regress out superficial physiological noises. Use principle component analysis or ICA to remove global physiological fluctuations. For elderly or clinical populations with potentially altered neurovascular coupling, consider individualized hemodynamic response function estimation rather than canonical models [63] [64].

Experimental Workflows and Signaling Pathways

fNIRS Clinical Research Workflow

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Essential fNIRS Research Materials and Equipment

Item	Function/Purpose	Specifications/Notes
ETG-4000 fNIRS System	Continuous-wave fNIRS data acquisition	Uses 695 nm & 830 nm wavelengths; 10 Hz sampling rate; multi-channel mode [60]
FOIRE-3000 System	Continuous-wave fNIRS with multiple wavelengths	16 sources & 16 detectors; 780 nm, 805 nm, 830 nm; 14 Hz sampling [63] [64]
Optode Caps/Holders	Secure optode positioning on scalp	Customizable for brain regions; typically 3 cm source-detector distance; compatible with EEG caps [60] [63]
Short Separation Detectors	Superficial signal regression	8-15 mm source-detector distance; captures extracerebral physiological signals for noise correction [63]
3D Scanner (Structure Sensor Pro)	Spatial registration of probe positions	Anatomical accuracy; corrects for individual head size/shape differences; enables group analysis [63]
Modified Beer-Lambert Law	Conversion of optical density to hemoglobin	Calculates HbO and HbR concentration changes; requires differential pathlength factor [62] [29]
SHAP Analysis	Machine learning interpretability	Identifies most contributory channels/features to model decisions; enhances clinical interpretability [60]

Optimizing fNIRS Signal Quality: Practical Solutions for Clinical Settings

Ensuring Proper Optode Placement and Scalp Coupling

Frequently Asked Questions (FAQs)

Q1: Why is proper optode placement and scalp coupling so critical for fNIRS data quality? Proper optode placement ensures that the fNIRS system is accurately targeting the specific cortical region of interest (ROI) for your experiment. Incorrect placement can result in measuring brain activity from an entirely different area, leading to erroneous conclusions [18] [67]. Similarly, good scalp coupling is necessary to maximize the Signal-to-Noise Ratio (SNR). Poor coupling, often caused by hair obstruction, leads to weak optical signals, increased noise, and channels that are functionally unusable [68].

Q2: What are the common signs of poor scalp coupling in my data? The most direct indicator is a weak or absent cardiac pulsation (heartbeat) in the raw fNIRS signal [68] [69]. A consistently low amplitude signal across multiple wavelengths for a specific channel also suggests poor coupling. During real-time setup, software tools like PHOEBE can provide quantitative metrics like the Scalp Coupling Index (SCI) to visually flag optodes requiring adjustment [68].

Q3: How can I improve the consistency of optode placement across multiple sessions or subjects? Using standardized positioning systems like the 10-10 or 10-5 international systems provides a reliable framework [67]. For higher precision, especially in clinical applications, using neuro-navigation techniques based on individual anatomical (MRI) data can significantly improve placement accuracy and consistency. This ensures the same cortical area is targeted in every session [70].

Q4: Can systemic physiological changes confound my fNIRS signal even with perfect placement and coupling? Yes. fNIRS signals are susceptible to physiological confounders such as changes in heart rate, blood pressure, respiration (particularly CO₂ concentration), and autonomic nervous system activity. These changes can occur in both cerebral and extracerebral tissues and may mimic (false positive) or mask (false negative) the neuronally evoked hemodynamic response [71].

Troubleshooting Guides

Guide 1: Diagnosing and Resolving Poor Scalp Coupling

Problem: Low signal amplitude and poor SNR in one or several channels.

Step-by-Step Diagnosis:

• Inspect Raw Signals: Visually examine the raw light intensity or optical density data for all channels. Look for the presence of a clear cardiac pulsation waveform at around ~1 Hz [68] [69].
• Use Coupling Quality Metrics: If available, utilize a real-time coupling assessment tool like PHOEBE. This software calculates an objective SNR measure (like the Scalp Coupling Index) for each channel and visually maps the coupling status of each individual optode on a head model, instantly identifying which one needs adjustment [68].
• Check Physical Setup: Manually inspect the optodes indicated by the software or associated with poor-signal channels. Look for hair trapped under the optode tip, insufficient pressure against the scalp, or sweat that might be interfering with contact.

Solutions:

• Hair Management: Use a blunt-ended tool to part the hair and create a clear path for the optode to make direct contact with the scalp [68].
• Optode Adjustment: Reposition or reseat the flagged optode, ensuring firm but comfortable contact.
• Use Adequate Optode Paste: Ensure an appropriate amount of optode gel or paste is used to improve light conduction between the optode and skin.

Guide 2: Optimizing Optode Placement for Target Brain Regions

Problem: Uncertainty in placing optodes to maximize sensitivity to a specific brain region.

Step-by-Step Optimization:

• Define Your ROI: Precisely define the cortical region of interest based on its anatomical or functional coordinates (e.g., from fMRI meta-analyses) [70] [67].
• Use Probabilistic Mapping Tools: Employ toolboxes like the fNIRS Optodes' Location Decider (fOLD). These tools use photon transport simulations on standard head atlases to recommend the optode positions from the 10-5 system that provide the highest sensitivity to your pre-defined ROIs [67].
• Individualize with Neuroimaging (For High-Precision Studies): For the highest accuracy, especially in sparse optode layouts for Brain-Computer Interfaces (BCIs), use subject-specific anatomical (MRI) and/or functional (fMRI) data to guide optode placement. This accounts for individual anatomical and functional variability [70].

Table 1: Comparison of Optode Placement Guidance Approaches

Approach	Description	Best For	Advantages	Limitations
Literature-Based (LIT)	Placement based on coordinates from prior published studies [70].	Standard cognitive tasks with well-established activation loci.	Simple, fast, requires no additional data.	Does not account for individual anatomical/functional variability.
Probabilistic (PROB)	Uses standard head atlases and probabilistic activation maps [70] [67].	Studies without access to individual MRI/fMRI.	Accounts for general anatomy; more informed than LIT.	Less precise than subject-specific methods.
Individual fMRI (iFMRI)	Uses subject-specific anatomical and functional MRI data [70].	Clinical applications, BCIs, and studies requiring high precision.	Highest spatial specificity; targets individual functional anatomy.	Requires expensive and time-consuming MRI/fMRI scanning.

Experimental Protocols

Protocol 1: Real-Time Assessment of Scalp Coupling Using SCI

Objective: To ensure all fNIRS channels have adequate scalp coupling before commencing data acquisition.

Materials:

• fNIRS system with real-time data output capability.
• Computer with software like PHOEBE (compatible with certain NIRx systems) or similar custom scripts [68].

Methodology:

• After placing the fNIRS headgear, start data acquisition.
• Run the coupling assessment software (e.g., PHOEBE). The software will, in real-time:
  • Band-pass filter the raw photodetected signals between 0.5 Hz and 2.5 Hz to isolate the cardiac pulsation [68].
  • Compute the Scalp Coupling Index (SCI) for each channel. The SCI quantifies the prominence of the cardiac waveform in the signal [68].
  • Display a head model with each optode color-coded based on its coupling status (e.g., green for good, red for poor).
• Identify and physically adjust the optodes flagged with poor coupling.
• Repeat steps 2-3 until all optodes indicate good coupling quality.

The following diagram illustrates this real-time workflow:

Protocol 2: Designing an Optimized Optode Layout Using the fOLD Toolbox

Objective: To determine the most effective optode positions on the scalp for measuring brain activity from specific regions of interest.

Materials:

• Computer with the fOLD toolbox installed [67].
• Definition of target brain regions (e.g., Brodmann Areas or MNI coordinates).

Methodology:

• Input Regions of Interest: Specify the cortical regions you intend to study within the fOLD toolbox.
• Run Simulation: The toolbox executes Monte Carlo simulations of photon transport on realistic head models (e.g., the Colin27 atlas) to model how light propagates through tissues [67].
• Calculate Sensitivity Profile: For potential optode locations (based on the 10-5 system), fOLD calculates a normalized sensitivity profile for each source-detector channel, indicating the brain volume it is sensitive to [67].
• Receive Recommendations: fOLD outputs a list of recommended optode positions that provide the highest collective sensitivity to your specified ROIs.

Signaling Pathways & Workflows

The following diagram summarizes the logical relationship between proper experimental setup and the final data quality, highlighting key sources of error and their solutions.

The Scientist's Toolkit: Essential Materials and Reagents

Table 2: Key Reagents and Solutions for fNIRS Setup

Item	Function / Explanation	Application Note
Optode Gel or Paste	Improves optical coupling between the optode tip and the scalp by filling micro-gaps and matching refractive indices. Reduces signal loss due to air gaps.	Use a non-abrasive, hypoallergenic formula suitable for prolonged skin contact [68].
Blunt-Ended Probe / Parting Tool	Used to gently part the hair underneath the optode, creating a direct path to the scalp. This is critical for minimizing light attenuation caused by hair [68].	Essential for achieving good coupling in subjects with thick or dark hair.
Isopropyl Alcohol Wipes	To clean the scalp area and optode tips before application. Removes oils and sweat that can degrade optical contact.	Ensures a clean interface and prevents residue buildup on optodes.
Measuring Tape & Marker	For identifying fiducial points (nasion, inion, pre-auricular) and measuring distances according to the 10-10 or 10-5 international systems for standardized optode placement [67].	Fundamental for reproducible cap placement across subjects and sessions.
Neuronavigation System	(For high-precision studies) Uses subject-specific MRI data to co-register optode positions with underlying cortical anatomy in real-time, ensuring accurate targeting of ROIs [70].	The gold-standard for maximizing spatial specificity in clinical and BCI applications.

Strategies for Improving Spatial Specificity in Cortical Mapping

Functional Near-Infrared Spectroscopy (fNIRS) has emerged as a valuable tool for cortical mapping in clinical and research settings, offering a portable, cost-effective alternative to functional Magnetic Resonance Imaging (fMRI) [3]. However, ensuring precise spatial localization of neural activity—known as spatial specificity—remains a significant challenge. Spatial specificity refers to the ability to accurately identify and consistently target specific cortical regions of interest (ROIs) across multiple measurement sessions [18]. This is particularly crucial for applications like neurofeedback and brain-computer interfaces where reliable targeting of the same brain region across sessions is essential for effective training outcomes [18]. The fundamental limitations stem from fNIRS's inherent characteristics: it measures superficial cortical regions only, typically with limited head coverage, and is susceptible to anatomical variations between subjects and inconsistent optode placement [18] [3]. This technical support guide addresses these challenges by providing targeted troubleshooting strategies and methodological refinements to enhance the spatial precision of your fNIRS experiments.

Frequently Asked Questions (FAQs)

Q1: Why is my fNIRS activation map inconsistent across multiple sessions with the same subject? Inconsistent activation maps often result from slight variations in optode placement between sessions. Even small displacements can significantly alter the cortical region being measured due to the complex curvature of the head. Furthermore, limited anatomical information and low head coverage of many fNIRS systems exacerbate this issue [18]. To mitigate this, implement precise optode positioning techniques using international 10-20 or 10-10 systems and employ anatomical co-registration to verify placement against individual structural scans.

Q2: How can I distinguish genuine cortical activation from superficial scalp hemodynamics? The confounding effect of systemic physiological noise from the scalp is a common problem that reduces spatial specificity [3]. These extracerebral signals can masquerade as brain activity. The most effective solution is to integrate short-distance channels (SDCs) with a source-detector separation of typically <1 cm [3] [72]. SDCs act as a reference, selectively measuring the superficial signal, which can then be regressed out from the standard long-separation channels that contain both cerebral and extracerebral components.

Q3: Which fNIRS chromophore (HbO or HbR) provides better spatial localization? Current evidence suggests that both oxygenated (HbO) and deoxygenated hemoglobin (HbR) can effectively localize neural activity, with no statistically significant superiority observed for one chromophore in terms of spatial correspondence with fMRI's BOLD signal [72]. The optimal approach is to analyze both signals. HbO typically shows a more robust response to neural activation, while HbR can provide complementary information and, in some cases, may more closely reflect the fMRI BOLD signal [3] [72].

Q4: Can I use fNIRS to accurately target deep brain structures or mesial cortical surfaces? No, fNIRS is fundamentally limited to sampling the superficial cerebral cortex. The near-infrared light has a limited penetration depth, and the technique cannot access deep gray matter nuclei, mesial cortex surfaces, or regions within deep scissures [3] [4]. For studies requiring whole-brain coverage or investigation of subcortical structures, fMRI remains the gold standard.

Troubleshooting Guides

Problem: Poor Inter-Subject Consistency in ROI Localization

Diagnosis: When the same task produces activated regions that vary excessively across subjects, it is often due to the lack of individual anatomical guidance and standardized montage placement.

Solutions:

• Implement Anatomical Co-registration: Use a 3D magnetic space digitizer (e.g., Polhemus Fastrak) to record the 3D coordinates of each optode relative to anatomical landmarks (nasion, inion, pre-auricular points) [73]. Co-register these positions to individual (or template) MR images to generate subject-specific forward models of light propagation [3].
• Utilize High-Density Arrays: Move beyond sparse optode arrangements. High-density whole-head optode arrays improve spatial sampling and enable the use of image reconstruction techniques that can provide more accurate localization of the underlying hemodynamic changes [3].

Problem: Low Signal-to-Noise Ratio (SNR) Obscuring Focal Activation

Diagnosis: A poor SNR makes it difficult to distinguish focal brain activation from noise, reducing the effective spatial resolution and specificity.

Solutions:

• Advanced Signal Processing: Employ a robust preprocessing pipeline. This should include:
  • Motion Artifact Reduction (MAR): Use algorithms like wavelet-based filtering or robust regression to identify and correct for motion-induced signal distortions [4].
  • Bandpass Filtering (BPF): Apply a filter (e.g., 0.01–0.2 Hz) to remove cardiac, respiratory, and very low-frequency drifts [4].
  • Principal Component Analysis (PCA): Use PCA to isolate and remove global physiological noise components shared across channels [4].
• Multimodal Integration: Simultaneously record with EEG. The high temporal resolution of EEG can help constrain and inform the analysis of the fNIRS signals, and data fusion methods like structured sparse multiset Canonical Correlation Analysis (ssmCCA) can help pinpoint brain regions consistently activated across both modalities [73].

Problem: Inaccurate Targeting of the Action Observation Network (AON)

Diagnosis: Studies investigating networks like the AON during motor execution, observation, or imagery often report inconsistent findings due to the challenges in accurately covering key parietal and motor regions.

Solutions:

• Probe Placement: Ensure your optode montage bilaterally covers sensorimotor and inferior parietal cortices, key hubs of the AON. Use digitization to verify coverage post-hoc [73].
• Multimodal Validation: Where possible, validate your fNIRS setup and findings with fMRI. Asynchronous or simultaneous fMRI-fNIRS recordings can be used to confirm that your fNIRS montage is correctly positioned to capture the intended hemodynamic responses in motor-task paradigms [72].

Experimental Protocols for Validation

Protocol: Validating fNIRS Spatial Specificity Against fMRI

Purpose: To quantitatively assess the spatial correspondence between fNIRS-derived hemodynamic responses and the fMRI BOLD signal in motor regions [72].

Materials: 3T fMRI scanner, continuous-wave fNIRS system (e.g., NIRSport2) with short-distance detectors, response box.

Procedure:

• Participant Setup: Place the fNIRS cap on the participant, ensuring coverage of bilateral motor areas. Use a digitizer to record optode positions.
• fMRI Acquisition: Acquire a high-resolution structural scan (MPRAGE). For functional scans, use an EPI sequence focused on motor areas (e.g., 26 slices, 3x3mm in-plane resolution, TR=1500ms).
• fNIRS Acquisition: Set up the fNIRS system with sources at 760 nm and 850 nm, sampling at >5 Hz.
• Task Paradigm: Employ a block design. A sample protocol is detailed below.
• Data Analysis:
  • fMRI: Preprocess data (motion correction, spatial smoothing, normalization). Model data using a General Linear Model (GLM) to define ROIs in primary motor (M1) and premotor (PMC) cortices.
  • fNIRS: Preprocess data (pruning bad channels, converting to optical density, then to concentration changes). Apply MAR, BPF, and GLM fitting.
  • Spatial Correlation: Use subject-specific fNIRS signals from motor channels to predict activation in the fMRI data and evaluate the overlap.

Table: Sample Block Design for Motor Task Validation [72]

Block Type	Duration	Participant Instruction
Baseline	30 sec	Rest, focus on a fixation cross.
Motor Action (MA)	30 sec	Execute bilateral finger tapping sequence (e.g., 1-2-1-4-3-4).
Baseline	30 sec	Rest.
Motor Imagery (MI)	30 sec	Imagine performing the same sequence without moving.
Repeat	8 min 30 sec total	4 blocks of MA and MI, interspersed with Baseline.

Protocol: Assessing tDCS-Induced Plasticity with fNIRS

Purpose: To quantify changes in sensorimotor cortical activation and connectivity following transcranial direct current stimulation (tDCS) using fNIRS [74].

Materials: fNIRS system (e.g., CW-6 from Techen Inc.), tDCS stimulator, EMG system, torque sensor.

Procedure:

• Baseline Measurement: Record fNIRS and EMG during a wrist flexion task before tDCS application.
• tDCS Application: Apply bi-hemispheric tDCS (e.g., anode over left M1, cathode over right M1) for a standard duration (e.g., 20 min) at a safe intensity (e.g., 2 mA).
• Post-Stimulation Measurement: Repeat the wrist flexion task during and immediately after tDCS.
• Data Analysis:
  • Activation Analysis: Compare the amplitude and topography of HbO and HbR responses in the sensorimotor cortex (M1, S1, SMA) across the three conditions (pre, during, post).
  • Connectivity Analysis: Calculate resting-state functional connectivity between channels within and across hemispheres before and after stimulation to assess tDCS-induced plasticity [74].

Signaling Pathways and Workflows

Diagram: fNIRS Cortical Mapping Workflow and Specificity Challenges

Research Reagent Solutions

Table: Essential Materials for High-Specificity fNIRS Research

Item	Specification / Example	Primary Function
fNIRS System	Continuous Wave (CW) system (e.g., NIRSport2, Hitachi ETG-4100) [73] [72]	Measures changes in light attenuation to calculate Δ[HbO] and Δ[HbR].
Short-Distance Detectors	Separation < 1 cm from source [72]	Measures extracerebral signals for regression, improving specificity.
3D Digitizer	Magnetic space digitizer (e.g., Polhemus Fastrak) [73]	Records precise 3D optode locations for anatomical co-registration.
EEG-fNIRS Integrated Cap	EGI EEG cap with embedded fNIRS optodes [73]	Enables simultaneous multimodal acquisition for improved validation.
tDCS Stimulator	Constant current, bi-hemispheric montage capable [74]	Modulates cortical excitability to study plasticity and its mapping.
Motion Sensors	Accelerometers	Provides reference signal for motion artifact reduction algorithms.
Task Performance Monitors	Torque sensor, EMG system, response box [74] [72]	Quantifies behavioral output and muscle activity, correlating with brain signals.

Real-Time Processing Challenges for Neurofeedback and BCI Applications

Functional Near-Infrared Spectroscopy (fNIRS) has emerged as a promising neuroimaging technology for brain-computer interface (BCI) and neurofeedback (NFB) applications, particularly in clinical settings. fNIRS measures cortical hemodynamic responses by detecting changes in oxygenated (HbO) and deoxygenated hemoglobin (HbR) concentrations, providing a non-invasive, portable, and relatively motion-tolerant method for monitoring brain activity [13] [75]. Unlike EEG, which measures electrical activity, fNIRS tracks hemodynamic changes linked to neural metabolism, offering better spatial resolution for localized brain function assessment [76]. This makes it particularly valuable for rehabilitation, assistive technology, and monitoring neurological diseases and disorders of consciousness [77] [13].

However, real-time processing of fNIRS signals presents significant challenges that can compromise signal quality and interpretation. Motion artifacts, physiological noise, and computational delays create substantial obstacles for reliable BCI and neurofeedback systems where timely, accurate signal processing is critical [77] [78] [75]. This technical support center addresses these challenges through troubleshooting guides and FAQs designed to help researchers optimize their experimental protocols and overcome common implementation barriers.

Frequently Asked Questions (FAQs)

Q1: What are the most significant sources of noise in fNIRS signals for real-time applications?

fNIRS signals contain multiple noise sources that pose particular challenges for real-time processing:

• Motion artifacts (MAs): Caused by head movements that disrupt optode-scalp coupling, creating high-frequency spikes and baseline shifts [77] [79]. These are particularly problematic for real-time applications as they can mimic true hemodynamic responses.
• Physiological noise: Includes cardiac pulsation (1-1.5 Hz), respiration (0.2-0.5 Hz), and low-frequency Mayer waves (approximately 0.1 Hz) from systemic physiology [78] [80].
• Global components (GC): Task-dependent systemic hemodynamic activity from extracerebral tissues (scalp, skull) that shares the same frequency spectrum as neural signals, making it impossible to remove with standard filtering [78].

Q2: Why is the negative correlation between HbO and HbR important for signal quality assessment?

During genuine neural activation, HbO typically increases while HbR decreases, creating a characteristic negative correlation pattern [79]. Motion artifacts and other noise sources often disrupt this relationship, making it a valuable quality indicator. Research shows that motion noise causes HbO and HbR signals to become more positively correlated [79]. Monitoring this correlation in real-time provides a practical method for identifying contaminated signal segments before they affect BCI output or neurofeedback.

Q3: What computational challenges affect real-time fNIRS processing?

Real-time processing faces several computational hurdles:

• Baseline establishment: fNIRS measures relative hemoglobin changes, requiring appropriate baseline definition at experiment start [77].
• Processing latency: Complex denoising algorithms must complete within strict time constraints (ideally <100-500ms) to maintain effective feedback loops [77] [75].
• Image reconstruction: For high-density DOT systems, generating 3D hemodynamic images requires computationally intensive matrix inversions that challenge real-time operation [77].
• Channel count: Systems with hundreds of channels (e.g., ~750 in advanced systems) multiply computational demands [77].

Q4: How does fNIRS compare to EEG for neurofeedback and BCI applications?

fNIRS and EEG offer complementary strengths and limitations:

Table: Comparison of fNIRS and EEG for BCI/Neurofeedback Applications

Characteristic	fNIRS	EEG
Spatial Resolution	Better (~1-2 cm) [76]	Lower (difficult to localize signals) [76]
Temporal Resolution	Slower (hemodynamic response over seconds) [76] [75]	Excellent (milliseconds) [76]
Motion Tolerance	Relatively tolerant [76]	Highly sensitive to movement [76]
Signal Origin	Hemodynamic (blood oxygenation) [76]	Electrical (neuronal firing) [76]
Setup Complexity	Generally simpler [76]	Often requires precise electrode placement [76]

Q5: What are the key reproducibility challenges in fNIRS research?

Reproducibility remains challenging due to analytical flexibility. A recent large-scale initiative (FRESH) found that while nearly 80% of research teams agreed on group-level results for strong hypotheses, individual-level results showed greater variability [19]. Key factors affecting reproducibility include:

• How poor-quality data is handled and excluded
• Variations in response modeling approaches
• Differences in statistical analysis methods
• Researcher experience with fNIRS analysis [19]

Troubleshooting Guides

Motion Artifact Correction

Problem: Motion artifacts corrupting fNIRS signals, making them unreliable for real-time BCI control or neurofeedback.

Symptoms:

• Abrupt, high-amplitude spikes in hemodynamic signals
• Baseline shifts that persist after movement cessation
• Unphysiological HbO-HbR correlations (positive instead of negative) [79]

Solutions:

• Deep Learning Approaches: Implement denoising autoencoders (DAE) with sliding window strategies for real-time motion correction. This approach has demonstrated capability to process approximately 750 channels simultaneously with low latency [77].
• Correlation-Based Methods: Apply algorithms that maintain negative correlation between HbO and HbR, effectively reducing spike-like noise in both online and offline applications [79].
• Hardware Solutions: Use short-distance channels (0.5-1 cm source-detector separation) specifically designed to capture extracerebral signals for regression-based correction [78] [80].

Implementation Protocol for Real-Time DAE:

• Train DAE model on extensive HD-DOT datasets containing various motion artifacts
• Implement sliding window processing for real-time application
• Apply baseline calibration to establish reference points
• Validate performance with separate datasets augmented with artificial MAs [77]

Physiological Noise Reduction

Problem: Physiological fluctuations (cardiac, respiratory, blood pressure) obscuring task-related hemodynamic responses.

Symptoms:

• Periodic oscillations in HbO and HbR signals at characteristic frequencies (0.1-1.5 Hz)
• Reduced signal-to-noise ratio for task-evoked responses
• Inconsistent activation patterns across subjects and sessions [78] [80]

Solutions:

• Systemic Physiology-Augmented fNIRS (SPA-fNIRS): Integrate peripheral physiological measurements (respiration, EDA, PPG, ECG) to model and regress out non-neural variance [80].
• Advanced Filtering: Combine band-pass filtering with wavelet-based detrending (e.g., wavelet-MDL) to address both high-frequency and slow-drift noise [78].
• Multimodal Signal Processing: Use temporally embedded Canonical Correlation Analysis (CCA) within an extended General Linear Model (GLM) framework to improve physiological noise regression [80].

Table: Physiological Noise Sources and Mitigation Strategies

Noise Source	Frequency Range	Mitigation Strategies
Cardiac Pulsation	1-1.5 Hz [78]	Band-pass filtering, Kalman filtering
Respiration	0.2-0.5 Hz [78]	SPA-fNIRS with respiration monitoring
Mayer Waves	~0.1 Hz [78]	Wavelet detrending, adaptive filtering
Systemic Circulation	<0.1 Hz	Global component removal, short-channel regression

Implementation Protocol for SPA-fNIRS:

• Collect synchronized peripheral physiological signals (respiration, EDA, PPG, ECG)
• Use multimodal signal processing to separate systemic physiological components
• Apply regression techniques to remove physiological noise from cerebral signals
• Validate with short-distance channels where available [80]

Real-Time Processing Optimization

Problem: Excessive processing latency compromising BCI responsiveness and neurofeedback efficacy.

Symptoms:

• Delayed or sluggish system response to brain activity
• Feedback occurring too late to reinforce targeted cognitive states
• Inability to process high-density channel arrays in real-time [77] [75]

Solutions:

• Streamlined Processing Pipeline: Implement optimized algorithms for real-time operation with pre-calculated parameters where possible (e.g., inverse Jacobian matrix for image reconstruction) [77].
• Sliding Window Strategies: Process data in overlapping temporal windows to maintain real-time performance while applying advanced denoising techniques [77].
• Hardware Acceleration: Utilize GPU processing or specialized hardware for computationally intensive operations like image reconstruction in DOT systems.

Implementation Protocol:

• Establish baseline period at beginning of recording session
• Implement sliding window processing with 1-2 second updates
• Pre-calculate computationally intensive matrices (e.g., inverse Jacobian)
• Set processing priority to minimize latency for critical signal pathways [77]

Experimental Protocols for Key Methodologies

Validating Motion Correction Algorithms

Purpose: Evaluate the efficacy of motion artifact correction methods for real-time fNIRS applications.

Procedure:

• Data Collection: Record fNIRS data during alternating blocks of motor execution/imagery and intentional head movements [78] [79].
• Signal Quality Metrics: Calculate correlation coefficients between HbO and HbR signals - true neural activation should show negative correlation [79].
• Performance Comparison: Apply multiple correction methods (band-pass filtering, wavelet-MDL, DAE) to the same dataset.
• Quantitative Assessment: Compare mean squared error and correlation to MA-free data for each method [77].

Analysis:

• Use spatial specificity measures (inter-channel correlations) to identify improved localization after correction
• Evaluate temporal consistency between HbO and HbR changes
• Assess processing latency to ensure real-time feasibility [78]

Implementing Real-Time Neurofeedback

Purpose: Establish a robust protocol for fNIRS-based real-time neurofeedback.

Procedure:

• System Setup: Configure fNIRS device with real-time processing capability (e.g., Turbo-Satori software) [81].
• Baseline Calibration: Record 5-10 minutes of resting-state data to establish individual baseline hemodynamics [77].
• Feedback Design: Create simple visual feedback (e.g., moving bar, animation) linked to target hemodynamic changes [76].
• Task Protocol: Implement blocked or continuous paradigm with appropriate inter-trial intervals accounting for hemodynamic delay [75].
• Quality Monitoring: Continuously monitor signal quality metrics (HbO-HbR correlation, signal strength) throughout session [79].

Troubleshooting:

• If feedback appears unresponsive, verify processing pipeline latency
• If signal quality degrades during session, check optode-scalp coupling
• If physiological noise dominates, implement additional filtering or SPA-fNIRS approaches [80]

Signaling Pathways and Experimental Workflows

fNIRS Real-Time Processing Pipeline

Physiological Noise Contamination and Correction

The Scientist's Toolkit: Research Reagent Solutions

Table: Essential Tools and Methods for fNIRS BCI/Neurofeedback Research

Tool/Method	Function	Implementation Example
Denoising Autoencoder (DAE)	Deep learning-based motion artifact correction	Real-time MA removal across ~750 channels with sliding window [77]
SPA-fNIRS Platform	Systemic Physiology Augmentation for noise removal	NIRxWINGS2 module for synchronized physiology recording [80]
Short-Distance Channels	Extracerebral signal reference for global component removal	Source-detector pairs at 0.5-1 cm separation [78] [80]
Wavelet-MDL Detrending	Physiological noise removal without precise frequency cutoffs	NIRS_SPM implementation for optimal trend removal [78]
Turbo-Satori Software	Real-time fNIRS analysis platform	Compatible with NIRScout and NIRSport systems [81]
Correlation-Based Signal Improvement	Noise reduction leveraging HbO-HbR anticorrelation	Algorithm enforcing negative correlation during processing [79]
Inverse Jacobian Pre-calculation	Accelerated image reconstruction for DOT	Pre-computed matrix for real-time 3D hemodynamic imaging [77]
MNE-Python Framework	Open-source fNIRS analysis package	GLM and group-level analysis implementation [81]

Real-time fNIRS processing for BCI and neurofeedback applications faces significant but addressable challenges in motion artifact correction, physiological noise removal, and computational efficiency. By implementing the troubleshooting strategies, experimental protocols, and analytical frameworks outlined in this technical support center, researchers can enhance signal quality and system reliability for clinical applications. Future advancements in deep learning approaches, standardized processing pipelines, and integrated hardware-software solutions will further strengthen fNIRS as a robust tool for real-time brain monitoring and intervention.

Handling Subject Variability and Movement in Clinical Populations

Frequently Asked Questions (FAQs)

FAQ 1: What are the most effective methods to handle motion artifacts in clinical fNIRS data? Motion artifacts are a primary concern in clinical fNIRS. A combination of automated algorithms and visual inspection is recommended. Wavelet-based methods are particularly effective for correcting high-frequency spikes, while spline interpolation can better handle baseline shifts. Some advanced methods combine both approaches. It is critical to visually inspect the algorithm's performance on a subset of your data, as default parameters may not be suitable for all datasets and improper use can distort results [82].

FAQ 2: How can we differentiate brain activity from systemic physiological noise in patients? Systemic physiological noise (e.g., from heart rate, blood pressure, respiration) can mimic or mask neurovascular coupling. To enhance specificity for cerebral signals, several strategies are employed:

• Short-Channel Regression: Using a short source-detector separation channel (e.g., <1 cm) to regress out signals originating predominantly from the scalp [82] [4].
• Principal Component Analysis (PCA): A data-driven approach to remove global physiological components that are common across channels [82] [4].
• Additional Physiological Monitoring: Incorporating measurements of heart rate, blood pressure, or respiration as regressors in the General Linear Model (GLM) to account for their influence [16] [82].

FAQ 3: What are the key channel quality metrics for data from clinical populations? Ensuring good signal quality is a crucial first step. Key metrics include:

• Source-Detector Distance: Channels too close together (<1 cm) are sensitive mainly to the scalp and should be excluded from brain activity analysis [54].
• Scalp Coupling Index (SCI): Quantifies the presence of cardiac pulsation in the optical signal. Channels with an SCI below a threshold (e.g., 0.5) are often considered of poor quality and marked as bad [54].
• Visual Inspection: Inspecting the optical density signal for the presence of a cardiac rhythm in the time or frequency domain, though this can be subjective and time-consuming [82].

FAQ 4: Which filter types and parameters are optimal for clinical fNIRS data? Filtering is used to remove unwanted physiological noise. The hemodynamic response has frequency content predominantly below 0.7 Hz. A common practice is to use a band-pass filter with cutoffs between 0.01-0.02 Hz (high-pass) and 0.5-0.7 Hz (low-pass) to remove very slow drifts and higher-frequency cardiac noise, respectively [54] [17] [69]. Both Finite Impulse Response (FIR) and Infinite Impulse Response (IIR) filters like Butterworth are widely used, with some research suggesting high-order FIR filters are optimal within a GLM framework [17] [69].

FAQ 5: How should analysis pipelines be adapted for neurodegenerative patients (e.g., Parkinson's Disease)? When studying clinical populations like Parkinson's disease patients, consider:

• Clinical Variability: Account for expected alterations in behavioral, neuronal, and vascular responses when interpreting results [16].
• Increased Artifacts: Be prepared for potentially increased noise and motion artifacts, and ensure the artifact rejection procedure is well-documented [16] [60].
• Task Design: Adapt experimental paradigms to the motor and cognitive capabilities of the patient group [60].

Troubleshooting Guides

Problem 1: Excessive Motion Artifacts Corrupting Data

Issue: The fNIRS signal contains large, abrupt spikes or baseline shifts due to subject movement, which is common in children, elderly, or patients with motor symptoms.

Solution: Implement a robust motion artifact correction (MAR) pipeline.

• Identify Bad Segments: First, visually inspect the raw or optical density data to identify channels and time periods severely corrupted by motion.
• Select a Motion Correction Algorithm: Apply an automated algorithm. The following table compares common methods:

Method	Principle	Best For	Considerations
Wavelet Filtering [82]	Uses wavelet transforms to identify and correct artifacts in specific time-frequency components.	High-frequency spikes.	Powerful but parameters may need tuning for your data.
Spline Interpolation [82]	Identifies artifact segments and interpolates over them using spline functions.	Slow baseline shifts.	May perform better when combined with a smoothing filter.
PCA-Based Methods [4]	Removes principal components that explain the most variance, often associated with global motion.	General purpose, multi-channel data.	Can be effective as part of a larger preprocessing pipeline.
tCCA & MSCA [82]	Advanced methods that can separate motion artifacts from brain signals based on their statistical properties.	Complex, mixed artifacts.

• Validate: Always plot the data before and after correction on a per-channel basis to ensure the algorithm did not distort the underlying hemodynamic signal [82].

Problem 2: Poor Signal-to-Noise Ratio (SNR) Despite Filtering

Issue: The hemodynamic response is weak or indistinguishable after standard preprocessing, potentially due to deep physiological confounds.

Solution: Employ advanced denoising techniques to improve cortical specificity.

• Band-Pass Filtering: Apply a band-pass filter (e.g., 0.01-0.7 Hz) as a foundational step [54].
• Use Short-Separation Channels: If your setup includes short-separation channels (typically 0.8 cm), use them as regressors in a GLM to subtract the superficial signal component from standard channels (3 cm) [82] [4].
• Global Average Regression or PCA: If short channels are unavailable, consider using PCA to remove the first few components that represent global systemic physiology shared across the array [82].
• Incorporate Physiological Regressors: Record heart rate and respiration if possible. These signals can be included as confound regressors in a GLM to account for physiological noise in the fNIRS data [16] [82].

Problem 3: Inconsistent or Atypical Hemodynamic Responses in Patients

Issue: The observed HbO/HbR response morphology in a clinical group does not match the canonical shape, making statistical modeling difficult.

Solution: Adapt the statistical model to account for response variability.

• Model Flexibility: In a GLM framework, consider using a more flexible basis set (e.g., Fourier basis or finite impulse response models) instead of a single, fixed canonical hemodynamic response function (HRF). This allows the model to capture variations in the response shape specific to the clinical population [16] [69].
• Block Averaging: For an initial, model-free assessment of the response, use block averaging. This involves segmenting the data around stimulus events and averaging across trials, which can reveal the true average response shape in your population without assuming a prior shape [54] [17].
• Check for Signal Quality: Rule out that the atypical response is not caused by residual motion artifacts or poor channel quality by revisiting the preprocessing steps [54] [82].

Detailed Methodology: fNIRS in Parkinson's Disease Research

The following protocol is adapted from a study investigating early-stage Parkinson's disease (PD) using fNIRS and machine learning [60].

• Participants: 120 PD patients (Hoehn & Yahr stages 1-2) and 60 age-matched healthy controls.
• Equipment: ETG-4000 fNIRS system, with probes placed over the prefrontal cortex. The probe holder had 8 emitters and 7 detectors (3 cm separation), forming 22 channels.
• Experimental Design (Block Design):
  • Task: A motor task (e.g., finger tapping) or cognitive task.
  • Cycle Structure: Each cycle consisted of a pre-task baseline (10 s), a task period, and a post-task rest period.
  • Repetitions: The cycle was repeated multiple times to build up the number of trials.
• Data Processing & Analysis:
  • Preprocessing: Conversion of raw intensity to optical density, then to HbO/HbR concentrations via the modified Beer-Lambert law.
  • GLM Analysis: A General Linear Model was applied to the HbO/HbR data to estimate the β-weights for the task condition versus baseline for each channel.
  • Machine Learning: The β-values from all channels were used as features to train a Support Vector Machine (SVM) classifier to discriminate between PD patients and controls.

The table below summarizes frequency parameters commonly used in fNIRS studies for removing physiological noise [54] [17] [69].

Filter Type	Typical Cut-off Frequencies	Purpose	Common Filter Types Used
High-Pass	0.01 - 0.05 Hz	Remove very slow drifts (e.g., vasomotion, instrument drift).	Butterworth, FIR
Low-Pass	0.1 - 0.7 Hz	Remove high-frequency noise (e.g., cardiac pulsation ~1 Hz, respiration ~0.3 Hz).	Butterworth, FIR
Band-Pass	0.01 - 0.7 Hz	Standard range to isolate the hemodynamic response.	Butterworth, FIR

The Scientist's Toolkit: Key Research Reagents & Materials

Item	Function in fNIRS Research
Continuous Wave (CW) fNIRS System	The most common type of system; measures light intensity attenuation to calculate changes in hemoglobin concentrations [4].
Optode Cap / Probe Holder	A headgear that securely holds light sources (emitters) and detectors (optodes) in a predefined array on the participant's scalp [60].
Short-Separation Channels	Special optode pairs with a small separation (e.g., 0.8 cm) used to measure and regress out signals from superficial tissues (scalp, skull) [82] [4].
Modified Beer-Lambert Law (mBLL)	The algorithm used to convert the raw, attenuated light intensity signals into changes in oxy- and deoxy-hemoglobin concentrations [54] [17].
General Linear Model (GLM)	A statistical framework used to model the fNIRS time series and estimate the strength (β-weights) of the task-evoked hemodynamic response [54] [82] [17].

Signaling Pathways & Workflows

fNIRS Clinical Preprocessing Workflow

Physiological Noise Separation Logic

For researchers and clinicians using functional near-infrared spectroscopy (fNIRS), ensuring high-quality signal acquisition is paramount for generating reliable data in both research and clinical trial settings. fNIRS is increasingly valued in clinical neuroscience and drug development for its portability, cost-effectiveness, and tolerance to movement, making it suitable for diverse populations from pediatric to geriatric patients [83] [3]. However, the technique is susceptible to signal contamination from various sources, including motion artifacts, physiological noise (e.g., cardiac and respiratory cycles), and poor optode-scalp coupling [83] [4]. Two critical metrics for quantifying signal integrity are the Signal-to-Noise Ratio (SNR) and the Scalp Coupling Index (SCI). Proper understanding and application of these metrics are essential for robust data collection, especially in longitudinal clinical studies or trials where consistent, high-quality data is necessary for assessing intervention effects.

Key Quality Metrics Explained

What is the Scalp Coupling Index (SCI)?

The Scalp Coupling Index (SCI) is an objective, quantitative measure of the signal-to-noise ratio for an individual fNIRS channel. It specifically quantifies the prominence of the cardiac waveform within the recorded signal [84] [68].

• Physiological Basis: The SCI leverages the fact that a clear photoplethysmographic (PPG) signal from the cardiac pulse is a strong indicator of good optode-scalp coupling [68]. This pulsatile component is primarily attributed to blood circulation in the scalp and is a prerequisite for successful fNIRS measurement [68].
• Calculation Method: The SCI is computed as the zero-lag cross-correlation between the normalized raw intensity signals at two different wavelengths (e.g., 760 nm and 850 nm), after they have been band-pass filtered within the cardiac frequency band (typically 0.5 - 2.5 Hz, corresponding to 30-150 beats per minute) [84] [68]. This correlation measures how well the cardiac signals from the two wavelengths align.

What is Signal-to-Noise Ratio (SNR) in fNIRS?

While SCI is a specific form of SNR measurement, SNR in fNIRS can be assessed more broadly. It refers to the ratio of the power of the brain signal of interest (e.g., the hemodynamic response) to the power of contaminating noise.

• Noise Sources: fNIRS signals are contaminated by multiple noise sources [83] [3] [4]:
  • Motion Artifacts (MAs): Sudden shifts in signal due to head or body movement.
  • Physiological Confounding Factors (PCFs): Includes cardiac (~1 Hz), respiratory (~0.3 Hz), and very low-frequency Mayer waves (~0.1 Hz) [85] [7].
  • Poor Scalp Coupling: Caused by hair, skin pigmentation, or improper optode placement [85].

The table below summarizes the core metrics and their interpretation:

Table 1: Key fNIRS Signal Quality Metrics

Metric	Definition	Typical Calculation	Quality Interpretation
Scalp Coupling Index (SCI)	Measures coupling quality via cardiac pulse correlation [84] [68]	Cross-correlation of two wavelength signals in cardiac band [84]	Good (Green): ≥ 0.8 or 0.9 [85] [86]Medium (Orange): 0.7 - 0.89 [86]Poor (Red): < 0.7 or 0.8 [85] [86]
Peak Power (PP)	Assesses presence of cardiac signal via power in cardiac band [86]	Power spectral density in cardiac frequency band	Higher power indicates stronger cardiac component and better coupling [86]. Thresholds are system-specific.
General SNR	Ratio of desired brain signal power to noise power	Varies (e.g., based on raw intensity)	A higher value indicates a cleaner signal. The minimum acceptable SNR is study-dependent.

Experimental Protocols for Quality Assessment

Protocol 1: Calculating SCI with MNE-NIRS

This protocol outlines how to compute the SCI using the open-source MNE-NIRS toolbox in MATLAB/Python, which is a common method for post-acquisition quality control [86].

Table 2: Reagents and Tools for SCI Assessment

Item Name	Function/Description
MNE-NIRS Toolbox	Open-source software for fNIRS data analysis. Provides functions for scalp_coupling_index and peak_power [86].
fNIRS System	A continuous-wave (CW) fNIRS instrument that provides raw light intensity data at multiple wavelengths.
Raw Intensity Data	The unprocessed light intensity signals (e.g., .wl1, .wl2 files from a NIRx system) are the primary input [86].
PHOEBE Software	A tool for real-time optode placement guidance, which can compute SCI during setup [68].

Step-by-Step Workflow:

• Data Import: Load the raw fNIRS data into the MNE-NIRS environment. This typically involves reading the data files and their associated probe layout information [86].
• Data Conversion (Optional): Convert the raw intensity data to optical density (OD). While SCI is often computed on raw intensity, some pipelines perform this conversion first [86].
• SCI Computation: Use the scalp_coupling_index function. The function will automatically [86]:
  • Band-pass filter the raw intensity signals for both wavelengths into the cardiac frequency band (e.g., 0.5-2.5 Hz).
  • Calculate the zero-lag cross-correlation between these two filtered signals for each channel.
• Result Interpretation: The function returns an SCI value for each channel. Channels can then be classified and marked as "bad" if they fall below a predetermined threshold (e.g., 0.7 or 0.8) [86].
• Visualization: Plot the distribution of SCI values across all channels or visualize the quality over time using a timechannel_quality_metric plot [86].

Figure 1: Workflow for calculating the Scalp Coupling Index (SCI).

Protocol 2: Real-Time Quality Assessment with PHOEBE

For optimizing data collection setup, the PHOEBE tool provides real-time feedback. This is crucial for clinical studies to minimize setup time and ensure data quality from the start [68].

Step-by-Step Workflow:

• System Setup: Don the fNIRS headgear on the participant and connect it to the acquisition system.
• Launch PHOEBE: Start the PHOEBE software and load the corresponding probe geometry file.
• Initiate Data Stream: Begin data acquisition from the fNIRS instrument. PHOEBE will receive the data stream in real-time.
• Monitor Optode Status: PHOEBE calculates a channel-level SNR measure (based on SCI and other power features) and uses graph theory to resolve this to the optode level. It then displays the coupling status of each individual optode on a head model [68].
• Adjust Optodes: The visual display immediately shows which optodes require adjustment. The experimenter can then reposition or adjust these specific optodes until the display indicates good coupling for all.
• Begin Experiment: Once all optodes show good coupling, the formal data acquisition can commence.

Troubleshooting Guide & FAQ

This section addresses common problems researchers encounter regarding fNIRS signal quality.

Frequently Asked Questions

Q1: What is an acceptable SCI threshold for my study? There is no universal threshold, but a common benchmark is 0.8 [85]. Channels with an SCI ≥ 0.8 or 0.9 are generally considered of good quality (often marked green in software), while those below 0.7 or 0.8 are considered poor (marked red) and should be excluded from analysis [85] [86]. The exact threshold may depend on your specific research question and fNIRS hardware.

Q2: Why are my SCI values low even though the signal looks fine? The raw fNIRS signal is dominated by low-frequency components (e.g., the hemodynamic response itself). A signal can appear smooth but lack the high-frequency cardiac pulsation, which is what the SCI detects. A low SCI indicates that the cardiac component is weak or absent, likely due to poor optode-scalp contact, even if the low-frequency trend seems stable [68].

Q3: How does hair type affect signal quality and what can I do? Hair, particularly thick, curly, or dark hair, is a significant challenge [85].

• Mechanical Blocking: Hair prevents the optode from making direct contact with the scalp.
• Optical Absorption: Melanin in dark hair is a strong absorber of near-infrared light, reducing the amount of light that reaches the detector [85].
• Solutions: Use a blunt tool to part the hair and create a clear path to the scalp. Ensure optodes have long enough stems to reach through the hair. Consistently using SCI during setup can help verify that these measures are effective [85].

Q4: Can I use fNIRS on participants with dark skin pigmentation? Yes, but with important considerations. Melanin in the epidermis is a dominant absorber of NIR light. Higher melanin concentrations can lead to greater light absorption, systematically attenuating the signal and potentially causing inaccuracies or underestimation of hemoglobin concentration changes [85]. While relative measures (ΔHbO, ΔHbR) are still valid, extra care must be taken to ensure excellent optode coupling and to account for potentially lower signal levels.

Q5: What should I do with channels that have consistently poor SCI? Channels consistently showing poor SCI after repeated optode adjustment should be marked as "bad" and excluded from subsequent analysis [86] [68]. Some advanced techniques, like generative deep learning models, are being explored to reconstruct missing or damaged channels, but exclusion remains the standard practice [87].

Troubleshooting Common Signal Quality Issues

Table 3: Troubleshooting fNIRS Signal Quality Problems

Problem	Potential Causes	Solutions & Corrective Actions
Low SCI/Poor SNR	1. Hair between optode and scalp2. Loose headgear3. Excessive hair product4. Skin pigmentation	1. Part hair carefully with a blunt tool [85]2. Tighten headgear, check stability3. Clean optodes after use4. Ensure optimal coupling; use SCI for verification
Motion Artifacts	1. Participant movement (coughing, talking, fidgeting)2. Loose headgear shifting	1. Instruct participant to remain still2. Secure headgear snugly3. Use motion artifact reduction (MAR) algorithms in post-processing (e.g., in NIRS-KIT, Homer) [38] [4]
Physiological Noise	1. Cardiac signal (~1 Hz)2. Respiration (~0.3 Hz)3. Mayer waves (~0.1 Hz)	1. Use band-pass filtering (e.g., 0.01 - 0.2 Hz) to isolate hemodynamic response [87]2. Employ short-separation channel regression to remove systemic superficial signals [7] [4]3. Use PCA/ICA to remove global physiological components [7] [4]
Channel Overexposure	1. Optode too close to skin2. Light skin complexion/scalp3. Source-Detector distance too short	1. Check optode positioning, use distance guards if available [85]2. Readjust instrument gain/calibration during setup [85]

Figure 2: A logical flowchart for troubleshooting poor fNIRS signal quality.

FAQs & Troubleshooting Guides

System Selection & Fundamental Principles

What is the core physical difference between CW-fNIRS and TD-fNIRS?

CW-fNIRS systems measure the attenuation of continuous light intensity as it passes through biological tissue. The calculated changes in hemoglobin concentration are relative, as the precise pathlength of light is unknown [88]. In contrast, TD-fNIRS uses pulsed laser sources and time-resolved detectors to measure the temporal distribution of photon time-of-flight (DTOF). This allows TD-fNIRS to differentiate between short and long photon pathlengths, providing a means for better depth resolution and the ability to quantify absolute tissue optical properties [89] [90].

How deep into the brain cortex can fNIRS measure?

For functional brain imaging, the typical probing depth of fNIRS is about 3 centimeters [30]. This depth is sufficient to reach the cerebral cortex in adults. The effective depth is also influenced by the source-detector separation, with a typical separation of 3 cm for adults [13].

Which system typically offers a higher temporal resolution?

Both CW and TD systems can achieve high sampling rates. fNIRS systems can sample brain signals at intervals of up to 0.1 seconds (10 Hz), which is a higher temporal resolution than fMRI and PET [88]. The actual imaging frame rate can vary depending on the number of source steps in a particular application [30].

Performance & Signal Quality

Why is TD-fNIRS often considered the "gold standard" in non-invasive optical brain imaging?

TD-fNIRS is highly regarded because its time-resolved measurements enable improved and more quantitative estimates of both oxy- and deoxy-hemoglobin concentrations [89]. The pathlength-resolved measurement capability theoretically gives TD-fNIRS a higher sensitivity for detecting dynamics in deep tissue compared to CW-fNIRS [90].

Under what conditions might CW-DCS currently be superior to TD-DCS for functional brain detection?

A 2024 simulation study highlighted that despite the theoretical advantages of TD-Diffuse Correlation Spectroscopy (TD-DCS), the current limitations of optoelectronic devices can impact its performance in real-world applications. Factors such as the instrument response function (IRF), the finite coherence length of the light source, and photon detection efficiency can degrade its detection ability. This same simulation found that at an 830 nm wavelength, CW-DCS was more sensitive than TD-DCS in detecting functional changes in cerebral blood flow, given the same incident light power and detection fiber [90].

Which system typically demonstrates better test-retest reliability?

Recent research with a specific TD-fNIRS system (Kernel Flow2) has demonstrated high test-retest reliability across multiple time points and even when using different headsets. This high reliability was observed in features derived from resting state (e.g., hemoglobin concentrations, functional connectivity) and during task-based activations (e.g., auditory and cognitive tasks) [89]. This supports its potential for clinical applications where stable, repeatable measurements are essential.

Clinical & Research Applications

For large-scale clinical trials or longitudinal treatment monitoring, which system is more practical?

CW-fNIRS systems have historically been more common due to their lower cost, relative simplicity, and robustness. However, modern, scalable TD-fNIRS systems are now being developed and validated for such applications. The portability, participant-friendliness, and high reliability of these new TD systems support their potential for use in precision medicine for diagnosis, treatment selection, and monitoring of neuropsychiatric disorders [89] [88].

How does susceptibility to physiological artifacts compare?

All fNIRS systems are susceptible to physiological artifacts from the scalp (e.g., blood pressure changes). However, the depth-resolving capability of TD-fNIRS can potentially make it more resilient to such superficial confounds, as it offers a means to separate signals originating from the cortex from those originating from the scalp [89] [90]. Both modalities require careful data processing and motion correction procedures to mitigate these artifacts [88].

Quantitative System Comparison

Table 1: Key Technical Specifications of CW-fNIRS and TD-fNIRS

Feature	Continuous Wave (CW-fNIRS)	Time-Domain (TD-fNIRS)
Basic Principle	Measures attenuation of continuous light.	Measures temporal spread of short light pulses (DTOF).
Measured Quantities	Relative changes in HbO and HbR concentration [88].	Absolute quantification of tissue properties & hemoglobin concentrations is possible [89].
Depth Resolution	Limited; relies on multi-distance measurements and modeling.	Superior; inherent depth sensitivity from photon time-of-flight [90].
Sensitivity to Deep Tissue	Less sensitive to deep tissue dynamics; signal is heavily weighted toward superficial layers [90].	Theoretically higher sensitivity to deep tissue (e.g., cortex) via selection of long-pathlength photons [90].
Typical Penetration Depth	~3 cm (brain imaging) [30].	~3 cm (brain imaging) [30].
Spatial Resolution	5-10 mm [30].	Potentially higher than CW due to depth gating, on the order of 5-10 mm.
Temporal Resolution	High (up to 10 Hz typical) [88] [30].	Can be high, but may be traded for better signal-to-noise ratio in time-gating.
Current Cost & Complexity	Generally lower cost and more established.	Historically bulky and expensive; newer commercial systems aim for scalability [89].
Resilience to Scalp Hemodynamics	Lower; requires short-separation channels to regress out artifacts.	Higher; potential to discriminate scalp and brain signals based on photon pathlength [89] [90].

Table 2: Suitability for Different Research and Clinical Applications

Application Context	CW-fNIRS Suitability	TD-fNIRS Suitability	Key Considerations
Resting-State Functional Connectivity	High (if using short-separation regression) [89].	Very High (demonstrated high test-retest reliability) [89].	TD-fNIRS's inherent noise resilience is beneficial for stable connectivity metrics.
Task-Based Activation Studies	High (widely used in cognitive, sensory, and motor tasks) [88].	Very High (highly reliable activation patterns in sensory and cognitive tasks) [89].	Both are valid; TD may provide more localized and quantitative activation maps.
Treatment Response Monitoring	High (used in pharmacological, psychotherapeutic, and neuromodulatory monitoring) [88].	High (supported by high reliability, ideal for longitudinal studies) [89] [88].	Reliability across days and devices is critical, a key strength of modern TD systems.
Bedside & Ecological Monitoring	Very High (portable, robust, motion-tolerant) [88] [13].	High (newer systems are portable and participant-friendly) [89].	CW has a longer track record, but TD is rapidly catching up in portability.
Disorders of Consciousness (DoC)	High (valuable for bedside assessment of residual function) [13].	Promising (potential for better discrimination of neural vs. physiological signals) [89].	The non-invasive, portable nature of both is a major advantage over fMRI in DoC.

Experimental Protocols for System Validation

Protocol 1: Assessing Test-Retest Reliability

This protocol is derived from a study investigating the reliability of a TD-fNIRS system [89].

• Objective: To quantify the test-retest reliability (TRR) of brain metrics derived from an fNIRS system across multiple days and different hardware devices.
• Participants: 49 healthy adults.
• Design: A repeated-measures design where each participant completes two study visits.
• Visit Structure:
  • Stage 1: Resting-state session followed by a passive auditory task.
  • The headset is removed, and participants rest and complete surveys.
  • Stage 2: A second resting-state session followed by a Go/No-Go inhibitory control task.
• Device Variable: Half of the participants used the same headset for both stages within a visit ("STAY" cohort), while the other half used a different headset. On the second visit, the STAY cohort used the alternative headset. This design allows for the separation of variability due to time, device removal, and the use of different hardware.
• Key Metrics: Hemoglobin concentrations (HbO, HbR), head tissue light attenuation, amplitude of low-frequency fluctuations (ALFF), and functional connectivity.

Protocol 2: Functional Detection of Cerebral Blood Flow (CBF) using DCS

This protocol is based on simulation and experimental work comparing CW-DCS and TD-DCS [90].

• Objective: To compare the sensitivity of CW-DCS and TD-DCS in detecting functional changes in cerebral blood flow.
• Method: A simulation approach using a realistic, multi-layered head model (scalp, skull, CSF, gray matter, white matter).
• Controlled Parameters:
  • Incident light wavelength: 830 nm.
  • Incident light power: 75 mW (within safe limits for human tissue).
  • Detection: Single-mode detection fiber.
  • Source-Detector Distance: Varied (e.g., 20 mm to 30 mm) to assess depth sensitivity.
• Simulated Functional Change: A defined increase in blood flow index (BFI) within the cerebral cortex.
• Output Measurement: The sensitivity of each technique is quantified by the magnitude of the change in the measured autocorrelation function (g2) decay rate in response to the simulated cortical BFI change.

Signaling Pathways & Experimental Workflows

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Components for fNIRS Experiments

Item / Solution	Function / Rationale	Specification Notes
High-Density fNIRS System	Core data acquisition unit. Enables comprehensive cortical coverage.	Systems with >32 channels are common. Choose CW for cost-effectiveness or TD for quantitative depth-resolution [89] [88].
Validated Cognitive Paradigms	Elicit robust and specific neural activation patterns.	Passive auditory tasks and Go/No-Go inhibitory control tasks have shown high test-retest reliability and are well-established [89].
Short-Separation Detectors	Critical for measuring and regressing out systemic physiological artifacts from the scalp.	Typically placed 0.8 - 1.5 cm from a source. Essential for CW-fNIRS; also beneficial for TD-fNIRS as a complementary reference [90].
Motion Correction Algorithms	Software tools to identify and correct for motion artifacts in the signal.	Only 44.7% of clinical fNIRS studies report using motion correction, making it a crucial step for data quality [88].
Prefrontal Cortex (PFC) Montage	A standard headset configuration targeting the PFC.	The PFC is a key region for higher-order cognition and is the most frequently targeted area in clinical fNIRS studies (e.g., psychiatry, disorders of consciousness) [88] [13].
Standardized Data Processing Pipeline	A consistent set of steps for data filtering, conversion, and analysis.	Reduces methodological variability and enhances the reproducibility of results across studies [88].

Validating fNIRS Processing Methods: Performance Metrics and Clinical Correlation

Troubleshooting Guides and FAQs

FAQ: Understanding and Improving Measurement Reliability

Q1: Our fNIRS data shows poor test-retest reliability. What are the primary factors affecting measurement stability between sessions?

The reliability of fNIRS measurements is significantly influenced by optode placement consistency and proper signal processing. Test-retest reliability is excellent when the cap remains in place (ICC ≥ 0.78 for PFC, PMC, SSC) but deteriorates substantially after cap removal (ICC as low as 0.00 for PMC) [91]. Oxyhemoglobin (HbO) consistently demonstrates higher reproducibility across sessions compared to deoxyhemoglobin (HbR) [35]. Increased shifts in optode position between sessions directly correlate with reduced spatial overlap in measured brain activity [35].

Q2: What practical steps can we take to improve the contrast-to-noise ratio in our fNIRS recordings?

Implement a comprehensive preprocessing pipeline addressing both motion and physiological artifacts. Start with movement artifact reduction (MAR) algorithms, followed by bandpass filtering (0.01-0.1 Hz recommended) and principal component analysis (PCA) to remove physiological confounding factors [4]. For optimal results, use single-channel MAR algorithms rather than multichannel variants, which may cause overcorrection [4]. Incorporating short-distance channels (<1 cm source-detector separation) helps regress out superficial scalp contributions [3].

Q3: How does the choice of fNIRS technology impact the quality and reliability of derived metrics?

Time-Domain (TD-fNIRS) systems generally provide more reliable and quantitative measurements compared to Continuous Wave (CW-fNIRS) systems. TD-fNIRS demonstrates high test-retest reliability for resting-state features including hemoglobin concentrations, amplitude of low-frequency fluctuations, and functional connectivity [89]. These systems enable depth-resolved measurements, offering improved discrimination between cerebral and extracerebral signals [89]. For clinical applications requiring robust, repeatable measurements, TD-fNIRS should be considered the gold standard [89].

Troubleshooting Guide: Common Experimental Challenges

Problem: Inconsistent cortical mapping across subjects despite using standard optode placements.

Solution: Implement sensitivity-based matching rather than geometrical matching for scalp-cortex correlation. Conventional point-to-point geometrical mapping ignores the effect of light scattering in head tissues and individual anatomical differences [92]. The sensitivity-based matching method incorporates broad spatial sensitivity profiles of probe pairs using subject-specific head models from structural MRI, providing more accurate targeting of brain regions of interest [92].

Problem: Excessive motion artifacts compromising data quality in clinical populations.

Solution: Combine multiple artifact reduction strategies. Begin with MAR algorithms (spline interpolation, wavelet-based, or correlation-based), then apply bandpass filtering (0.01-0.1 Hz), and use PCA to remove physiological noise [4]. For studies with cooperative subjects, consider using a general linear model (GLM) approach with motion parameters as regressors [4]. When possible, integrate short-separation channels to specifically identify and remove superficial artifacts [3].

Problem: Low signal-to-noise ratio in task-evoked responses.

Solution: Optimize experimental design and processing parameters. Use block designs with sufficient trials (7 trials of 20s task/20s rest provides 80% true positive rate) to enhance detection power [91]. For analysis, employ vector diagram analysis to better define the initial dip in the hemodynamic response [3]. When analyzing data, focus on HbO as it demonstrates superior reproducibility compared to HbR [35].

Test-Retest Reliability Metrics Across fNIRS Studies

Table 1: Intraclass Correlation Coefficient (ICC) values for fNIRS reliability studies

Brain Region	Task Paradigm	ICC without Cap Removal	ICC with Cap Removal	Population	Citation
Prefrontal Cortex (PFC)	Postural Control	0.78	0.50	Older Adults	[91]
Premotor Cortex (PMC)	Postural Control	0.78	0.00	Older Adults	[91]
Somatosensory Cortex (SSC)	Postural Control	0.78	0.50	Older Adults	[91]
Hand Motor Region	Finger Tapping	0.66	0.38	Older Adults	[91]
Whole-Brain Resting State	TD-fNIRS Resting	0.71-0.89 (varies by metric)	N/A	Healthy Adults	[89]
Auditory Cortex	Passive Auditory	0.65-0.82	N/A	Healthy Adults	[89]
Right Prefrontal	Go/No-Go Task	0.70-0.85	N/A	Healthy Adults	[89]

fNIRS Processing Pipeline Performance Metrics

Table 2: Performance comparison of signal processing approaches for fNIRS

Processing Step	Method/Algorithm	Key Parameters	Performance Outcome	Recommendation
Motion Artifact Reduction (MAR)	Spline Interpolation	0.5s segment identification	4% mean energy reduction	Preferred for single-channel correction [4]
Motion Artifact Reduction (MAR)	Multichannel Variant	4-fold expansive windows	Overcorrection observed	Use with caution [4]
Bandpass Filtering	Butterworth Filter	0.01-0.1 Hz	Effective physiological noise removal	Standard for HRF [93] [4]
Physiological Noise Removal	Principal Component Analysis	N/A	Effective for systemic artifacts	Recommended after MAR and filtering [4]
Scalp Contribution Reduction	Short-Distance Channels	<1 cm separation	Improved brain specificity	Highly recommended when available [3]
Statistical Analysis	Generalized Linear Model	HRF convolution	Robust activation detection	Alternative to block averaging [4]

Experimental Protocols for Reliability Assessment

Protocol 1: Test-Retest Reliability for Motor Tasks

This protocol assesses the stability of fNIRS measurements across multiple sessions, specifically designed for evaluating processing pipeline performance.

Materials and Setup:

• fNIRS system with capability for whole-head coverage (minimum 102 channels recommended)
• Digitization equipment for precise optode localization
• Motor task apparatus (finger-tapping interface or virtual reality system)
• Standardized fNIRS cap with size options (54-60 cm range)

Procedure:

• Perform head circumference measurement and select appropriate cap size
• Identify anatomical landmarks (Cz, inion, nasion, pre-auricular points) for consistent cap placement
• Administer block design protocol: 7 trials of 20s rest followed by 20s movement
• For test-retest assessment: conduct two tests on day 1 (separated by 30min rest with cap in place)
• Remove cap completely and reattach on day 2 for third measurement at same time of day
• Digitize optode positions for each session to quantify placement shifts

Data Analysis:

• Calculate intraclass correlation coefficients (ICC) for HbO and HbR across sessions
• Perform spatial overlap analysis using digitized optode positions
• Compare reproducibility between HbO and HbR using repeated measures ANOVA
• Compute contrast-to-noise ratio for active vs. rest conditions [91] [35]

Protocol 2: Processing Pipeline Benchmarking

This protocol enables systematic comparison of different processing approaches for optimizing contrast-to-noise ratio.

Materials:

• fNIRS dataset with known activation patterns (motor or visual tasks recommended)
• Multiple processing toolboxes (Homer3, SPM-fNIRS, NIRS-SPM)
• Standardized performance metrics (contrast-to-noise ratio, activation effect size)

Procedure:

• Acquire fNIRS data during well-established activation paradigm (e.g., finger tapping, visual stimulation)
• Apply multiple motion artifact correction algorithms to the same dataset:
  • Spline interpolation
  • Wavelet-based correction
  • Principal component analysis
  • Correlation-based methods
• Process data through standardized pipeline with varied parameters:
  • Bandpass filtering with different cutoff frequencies (0.01-0.1 Hz vs. 0.02-0.2 Hz)
  • HRF estimation using both block averaging and GLM approaches
  • Spatial analysis at both channel and source levels
• Quantify performance metrics for each pipeline variant:
  • Contrast-to-noise ratio (mean activation amplitude/standard deviation of baseline)
  • Test-retest reliability (ICC) across sessions
  • Spatial reproducibility (overlap of activation maps)

Analysis:

• Compare pipeline performance using standardized metrics
• Identify optimal processing sequence for specific experimental conditions
• Generate receiver operating characteristic (ROC) curves for activation detection accuracy [93] [4] [94]

Visual Workflows

fNIRS Processing Pipeline

Reliability Factors in fNIRS

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key materials and tools for fNIRS reliability research

Item	Specification	Function/Purpose	Recommendation
fNIRS Systems	Time-Domain (TD-fNIRS)	Gold standard for depth-resolved measurements; provides more quantitative hemoglobin estimates	Kernel Flow2 system validated for high test-retest reliability [89]
fNIRS Systems	Continuous Wave (CW-fNIRS)	Cost-effective alternative; requires more extensive signal processing	Systems with short-separation channels for improved artifact removal [3]
Analysis Software	Homer3	Open-source MATLAB toolbox; intuitive GUI for processing pipeline creation	Recommended for beginners and experts; active community support [93]
Analysis Software	SPM-fNIRS	Statistical parametric mapping; integrates with fMRI analysis workflows	Suitable for group-level analysis and statistical inference [94]
Optode Digitizer	3D position digitizer	Precise localization of optodes on scalp; enables spatial overlap analysis	Critical for test-retest studies quantifying cap placement consistency [35]
HD-DOT Systems	High-density arrays	Improved spatial resolution and specificity compared to sparse fNIRS	Wearable HD-DOT shows superior spatial localization for infant studies [95]
Short-Distance Channels	<1 cm separation	Estimates and removes scalp blood flow contribution	Essential for improving brain specificity in CW-fNIRS studies [3]

FAQ: Technical Troubleshooting for fNIRS Integration

Q1: Our simultaneous EEG-fNIRS recordings show inconsistent results. What are the primary sources of this variability?

A1: Inconsistent results in simultaneous EEG-fNIRS studies primarily stem from three sources:

• Analysis Pipeline Variability: Different data processing methods significantly impact outcomes. A recent reproducibility study found that nearly 80% of research teams agreed on group-level results when using standardized approaches, but individual-level results varied substantially based on processing choices [19].
• Data Quality Issues: Signal quality dramatically affects reproducibility. Studies show that better data quality improves agreement across analysis methods, while poor-quality data introduces greater variability [19].
• Handling of Motion Artifacts: fNIRS is relatively motion-tolerant, but EEG is highly susceptible to movement artifacts. This discrepancy requires careful motion correction strategies when using both modalities simultaneously [96] [97].

Q2: How can we improve the reproducibility of our fNIRS findings in clinical studies?

A2: Improve reproducibility through these evidence-based strategies:

• Standardize Data Collection: Adopt consistent parameters for scan length (recommended minimum 12 minutes for resting-state), eye condition, and fixation symbols [98].
• Implement Quality Control: Use digitized optode positions for accurate source localization, as even small shifts in placement reduce spatial overlap across sessions [35].
• Prioritize HbO Measurements: Focus on oxygenated hemoglobin (HbO) signals, which demonstrate significantly higher reproducibility across sessions compared to deoxygenated hemoglobin (HbR) [35].
• Report Analysis Details: Document preprocessing steps, motion correction procedures, and statistical models thoroughly, as these choices substantially impact results [19].

Q3: What are the specific advantages of combining fNIRS with fMRI rather than using either modality alone?

A3: The fMRI-fNIRS combination provides unique complementary benefits:

• Spatiotemporal Resolution: fMRI offers high spatial resolution (millimeter-level) for deep brain structures, while fNIRS provides superior temporal resolution for cortical surfaces [99].
• Validation Capabilities: Simultaneous acquisition allows fNIRS signals to be validated against the fMRI gold standard, confirming the utility of fNIRS measures [99].
• Experimental Flexibility: fNIRS enables naturalistic study designs and follow-up assessments outside the scanner environment, addressing fMRI's mobility limitations [99].

Quantitative Comparison: fNIRS vs. Gold Standard Modalities

Table 1: Technical Specification Comparison of Neuroimaging Modalities

Feature	fNIRS	fMRI	EEG
Spatial Resolution	Moderate (1-3 cm)	High (millimeter-level)	Low (centimeter-level)
Temporal Resolution	Moderate (0.1-1s)	Slow (seconds)	Excellent (milliseconds)
Depth Penetration	Superficial cortex (1-2.5 cm)	Whole brain including subcortical	Cortical surface
Motion Tolerance	High	Low	Low
Portability	High	Low	High
Cost	Moderate	High	Low
Best Use Cases	Naturalistic studies, clinical monitoring, children	Deep brain mapping, precise localization	Rapid neural dynamics, ERP studies

Table 2: fNIRS Performance Characteristics for Clinical Applications

Parameter	Performance	Clinical Implications
HbO vs HbR Reproducibility	HbO significantly more reproducible than HbR [35]	Prefer HbO for longitudinal clinical monitoring
Session-to-Session Reliability	Variable; improves with source localization [35]	Requires careful optode placement protocols
Task Detection Accuracy	80% agreement at group level with proper analysis [19]	Suitable for group clinical studies, cautious individual diagnosis
Resting-State Scan Duration	Recommended ≥12 minutes for reliability [98]	Longer acquisitions improve data quality

Experimental Protocols for Multimodal Integration

Protocol 1: Simultaneous EEG-fNIRS Acquisition for Cognitive Tasks

This protocol is adapted from semantic decoding studies that successfully differentiated between imagined objects using combined EEG-fNIRS [100].

• Participant Preparation: Apply EEG cap using international 10-20 system. Integrate fNIRS optodes within EEG cap using compatible openings, ensuring no physical interference between sensors. Use digitized optode positioning for precise localization [96] [35].
• Hardware Synchronization: Connect EEG and fNIRS systems to shared trigger generator (TTL pulses or parallel port). Use integrated systems when available to minimize synchronization issues [96].
• Task Paradigm: Implement block or event-related designs. Example: Present visual stimuli (animal vs. tool images) for 3-5 seconds, followed by mental imagery tasks (visualizing, sound imagination, tactile imagination). Include adequate rest periods (30+ seconds) for hemodynamic recovery [100].
• Data Quality Monitoring: Real-time monitoring of both EEG signal quality (impedance checks) and fNIRS signal-to-noise ratio (detector coupling). Flag segments with excessive motion for post-processing attention [97].

Protocol 2: fMRI-fNIRS Validation Studies

This protocol leverages the complementary strengths of both hemodynamic modalities for method validation [99].

• Sequential Design: Conduct fMRI sessions first for precise spatial localization, followed by fNIRS sessions with identical task paradigms in naturalistic settings.
• Synchronous Data Acquisition: When simultaneous collection is possible, use MRI-compatible fNIRS probes positioned to target cortical regions of interest identified in prior fMRI scans.
• Data Alignment: Co-register fNIRS optode positions with structural MRI using digitized positioning or photogrammetry. Apply anatomical head models for accurate source localization [99].
• Cross-Modal Validation: Compare fNIRS hemodynamic responses (HbO, HbR) with fMRI BOLD signals in overlapping cortical regions to establish correlation metrics [99].

Signaling Pathways and Experimental Workflows

Figure 1: Neural Signaling and Multimodal Integration Workflow

Research Reagent Solutions: Essential Materials for fNIRS-EEG-fMRI Studies

Table 3: Essential Research Materials for Multimodal Neuroimaging

Material/Equipment	Function	Application Notes
High-Density fNIRS Systems (>32 channels)	Measures cortical hemodynamics with improved spatial resolution	Preferred for clinical applications; provides better coverage [88]
Integrated EEG-fNIRS Caps	Enables simultaneous acquisition without sensor interference	Use pre-defined compatible openings to maintain signal quality [96]
Digitized Optode Position Systems	Tracks precise sensor placement across sessions	Critical for reproducibility; reduces spatial overlap issues [35]
Motion Correction Algorithms	Removes movement artifacts from both EEG and fNIRS data	Essential for naturalistic studies; fNIRS more tolerant than EEG [96] [97]
Short-Separation Detectors	Measures superficial signals for global noise correction	Underutilized but effective for confounder removal [97]
Synchronization Hardware (TTL generators)	Coordinates timing across multiple acquisition systems	Necessary for temporal alignment of multimodal data [96]
Standardized Cognitive Paradigms	Provides comparable tasks across studies	Verbal fluency tasks common in clinical fNIRS; facilitates cross-study comparison [88]

Machine Learning Approaches for Signal Classification and Diagnostic Accuracy

Troubleshooting Guide: Resolving Common fNIRS-ML Challenges

Low Classification Accuracy

Problem: My machine learning (ML) or deep learning (DL) model for fNIRS signal classification is achieving low accuracy scores, potentially hindering diagnostic reliability.

Solutions:

• Implement Short-Channel Regression (SCR): Superficial hemodynamic changes from extracerebral tissues (scalp, skull) can contaminate your signal. Use data-driven methods like transformer-based models to predict and remove this extracerebral component if physical short-separation detectors are unavailable [48].
• Apply Motion Artifact Reduction (MAR): Movement artifacts can significantly degrade signal quality. Evaluate and apply appropriate MAR algorithms such as wavelet filtering for high-frequency spikes or spline interpolation for baseline shifts. Visually inspect the results to ensure the correction does not distort the physiological signal [4] [101].
• Enhance Feature Extraction: Move beyond manual feature engineering. For deep learning, employ hybrid models that capture both spatial and temporal features. Convolutional Neural Networks (CNNs) can extract spatial patterns, while Temporal Convolutional Networks (TCNs) or Gated Recurrent Units (GRUs) can model long-term temporal dependencies in the hemodynamic response [102] [103].

Model Overfitting

Problem: My model performs well on the training data but generalizes poorly to new, unseen fNIRS data from different subjects or sessions.

Solutions:

• Simplify Model Architecture: Use models with fewer parameters. A hybrid CNN-GRU model with only 3.23K parameters has been shown to achieve high accuracy (over 85% for motor imagery tasks) while reducing the risk of overfitting on smaller fNIRS datasets [102].
• Incorporate Spatial Attention Mechanisms: These mechanisms help the model focus on the most relevant brain regions for a given task, improving generalization by ignoring irrelevant or noisy channels [102].
• Apply Rigorous Validation: Use subject-wise cross-validation instead of random data splits. This ensures that data from the same subject is not present in both training and test sets, providing a more realistic measure of model performance for new individuals [16].

Inconsistent Signal Preprocessing

Problem: My results are difficult to reproduce or compare with literature due to inconsistencies in the fNIRS signal processing pipeline.

Solutions:

• Follow Standardized Preprocessing Checklists: Adhere to best practices for fNIRS publications. This includes clearly reporting signal quality metrics, channel rejection criteria, motion artifact handling, filtering parameters, and strategies for confounding signal removal [16].
• Adopt a Robust Preprocessing Pipeline: A typical effective pipeline includes:
  • Channel Rejection: Use automated methods like the coefficient of variation (CV) or PHOEBE to identify and exclude poor-quality channels [101].
  • Motion Correction: Apply algorithms like wavelet-based filtering [101].
  • Bandpass Filtering: Use a Butterworth filter (e.g., 0.01 - 0.2 Hz) to retain the hemodynamic response while removing high-frequency noise and very low-frequency drift [101].
  • Physiological Denoising: Employ techniques like Principal Component Analysis (PCA) or short-channel regression to remove systemic physiological noise [4] [101].
• Document All Parameters: Meticulously record every step and parameter (e.g., filter types, cut-off frequencies, MAR algorithm settings) in your methodology to ensure full reproducibility [16].

Frequently Asked Questions (FAQs)

Q1: What is the most effective deep learning architecture for fNIRS-based classification?

No single architecture is universally best, but hybrid models that capture both spatial and temporal features consistently show superior performance. Current research indicates:

• CNN-GRU Hybrids are effective for joint spatial-temporal feature extraction, achieving 73% accuracy in predicting cognitive performance from fNIRS data [103].
• CNN-TCN (Temporal Convolutional Network) models benefit from TCN's ability to efficiently handle long-term dependencies with parallel computation, outperforming many RNNs [102].
• Spatial Attention Mechanisms combined with CNNs help the model focus on the most relevant brain regions, improving both accuracy and interpretability [102].

Q2: How can I distinguish between cognitive effort and simple brain activation using fNIRS?

Cognitive effort provides a more nuanced measure than activation alone by relating neural activity to behavioral performance. Use these derived metrics:

• Relative Neural Efficiency (RNE): Indicates how efficiently a person completes a task. High RNE means good performance with less neural effort, while Low RNE indicates poor results despite high effort [103].
• Relative Neural Involvement (RNI): Reflects the level of engagement or motivation during a task [103]. To calculate these, you need both the fNIRS-derived hemodynamic response (neural activity) and a behavioral performance score from the task. A balanced RNE/RNI profile suggests an engaged and efficient learner [103].

Q3: My clinical population cannot tolerate long experiments. How can I improve SNR with limited data?

For clinical populations where data collection is challenging:

• Optimize Preprocessing: Prioritize robust motion artifact correction and physiological denoising to maximize the usable signal from shorter recordings [4].
• Use Transfer Learning: Pre-train your model on larger, public fNIRS datasets (e.g., from healthy subjects) and then fine-tune it on your smaller clinical dataset [102].
• Implement Data Augmentation: Carefully apply synthetic data generation techniques, such as Generative Adversarial Networks (GANs), to artificially expand your training set and improve model robustness [103].

Experimental Protocols for Key Cited Studies

Table 1: Protocol for Transformer-based Short-Channel Prediction

Protocol Component	Implementation Details
Objective	Develop a transformer-based model to predict short-separation signals from long-separation fNIRS channels for virtual short-channel regression [48].
Dataset	Training: Resting-state fNIRS from 69 subjects. Validation: Three independent datasets (resting-state, different system, task-based) [48].
Input Data	Dual-wavelength optical density signals in segmented time windows [48].
Model Architecture	Transformer encoder trained to reconstruct the extracerebral hemodynamic component [48].
Evaluation Metrics	Signal similarity (Mean Squared Error, Pearson correlation) and denoising efficacy (residual variance after regression) [48].

Table 2: Protocol for Hybrid CNN-GRU Model for Cognitive Effort

Protocol Component	Implementation Details
Objective	Estimate cognitive effort (via performance score prediction) from fNIRS signals in an educational game setting [103].
Task	Unity-based quiz game with 16 questions (30-second response limit each) [103].
Participants	16 subjects [103].
fNIRS Setup	Prefrontal cortex coverage, sampling rate: 10 Hz [103].
Model Architecture	Hybrid CNN (for spatial features) and GRU (for temporal features) [103].
Output & Metrics	Performance score prediction; Cognitive Effort analysis via RNE and RNI [103].

Signaling Pathways and Workflows

fNIRS-ML Processing Pipeline

CNN-GRU Hybrid Architecture

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Computational Tools for fNIRS-ML Research

Tool/Resource	Function & Application	Key Features
Transformer Encoders	Reconstructs extracerebral hemodynamic signals from long-separation fNIRS data, providing a virtual alternative to physical short channels [48].	Captures long-range temporal dependencies; enables hardware-independent preprocessing.
Hybrid CNN-GRU Models	End-to-end classification of fNIRS signals for BCI and cognitive state decoding [103].	Extracts joint spatial (CNN) and temporal (GRU) features; suitable for smaller datasets.
Temporal Convolutional Networks (TCN)	Processes fNIRS time-series data for classification and prediction tasks [102].	More accurate and lightweight than LSTM; allows for parallel computation.
Spatial Attention Mechanisms	Enhances model focus on the most relevant brain regions during feature extraction [102].	Improves classification accuracy and model interpretability by capturing remote contextual information.
Wavelet-Based Motion Correction	Identifies and removes motion artifacts from fNIRS signals during preprocessing [101].	Particularly effective for correcting high-frequency spikes in the signal.
Short-Channel Regression (SCR)	Removes hemodynamic changes originating from extracerebral tissues (scalp, skull) [48] [3].	Gold-standard method for improving cerebral signal specificity; can be physical or data-driven.

Technical Support Center: Troubleshooting Guides and FAQs

Frequently Asked Questions (FAQs)

Q1: What are the most common sources of noise in a combined fNIRS-EEG experiment?

The most prevalent noise sources stem from physiological and motion artifacts. fNIRS is highly sensitive to systemic physiological noise from cardiac pulsation (~1 Hz), respiration (0.2-0.5 Hz), and Mayer waves (~0.1 Hz) in blood pressure [78]. It is also contaminated by task-dependent systemic hemodynamic activity in the scalp, which shares the same frequency band as the neural signal of interest and cannot be removed by standard filtering [78]. EEG is primarily contaminated by ocular activity (EOG) and head/neck muscle activity (EMG) [97]. For both modalities, head motion can cause significant artifacts, particularly by disrupting the optode-scalp coupling in fNIRS [4].

Q2: Why do my fNIRS and EEG signals sometimes show seemingly unrelated activity even when recorded simultaneously?

This is expected because fNIRS and EEG capture fundamentally different physiological processes. EEG measures synchronous neuro-electrical activity on a millisecond timescale, while fNIRS measures the slow hemodynamic response (changes in blood flow) that occurs over seconds following neural activity [97] [28]. They are linked via neurovascular coupling, but this relationship is complex and non-instantaneous. A brief neural event captured by EEG will manifest as a delayed hemodynamic change in fNIRS [97]. Furthermore, the same physiological source (e.g., cardiac activity) can manifest as an electrical spike in EEG and a hemodynamic pulse wave in fNIRS, appearing different in each modality [97].

Q3: What is the best method for removing motion artifacts from fNIRS data?

There is no single "best" method, and the choice often depends on your data and system hardware. The following table summarizes common MAR algorithms and their applications based on a comparative study [4]:

Table: Comparison of fNIRS Motion Artifact Reduction (MAR) Approaches

Method Category	Example Algorithms	Key Principle	Reported Efficacy
Single-Channel MAR	Various (e.g., spline interpolation, wavelet)	Corrects artifacts based on the statistical properties of a single channel's signal.	Suggested as a primary method; different algorithms found similar activations in motor areas [4].
Multichannel MAR Variants	--	Uses information from multiple channels to correct artifacts.	Can lead to overcorrection if not carefully implemented [4].
Short-Separation Regression	--	Uses a dedicated short source-detector distance channel (~0.5-1 cm) to measure and regress out scalp hemodynamics.	Considered the most efficient method for removing extracerebral components, but requires specific hardware [78].
Statistical Correction	Singular Value Decomposition (SVD) & Gaussian Kernel	Uses spatial filtering and statistical models to separate global systemic components from local neural signals.	An effective alternative when short-separation channels are unavailable; improves spatial specificity [78].

Q4: Our multimodal results are inconsistent across research teams. How can we improve reproducibility?

A recent large-scale reproducibility study (FRESH initiative) found that agreement is highest when teams have more fNIRS experience and when hypotheses are strongly supported by existing literature [19]. Key strategies to enhance reproducibility include:

• Standardize and Pre-register Pipelines: Pre-define your analysis pipeline, including how poor-quality data will be handled.
• Ensure High Data Quality: Agreement at the individual level is significantly better with high-quality data [19].
• Report in Detail: Clearly document all steps, parameters, and software tools used for preprocessing, analysis, and statistics [19].

Troubleshooting Guide: Systematic Artifact Identification and Correction

Table: Troubleshooting Common Artifacts in fNIRS-EEG Recordings

Symptoms	Potential Cause	Corrective Actions
fNIRS: High-frequency, low-amplitude oscillations.EEG: Simultaneous spikes in all channels.	Cardiac Artifact	fNIRS: Apply a band-pass filter (e.g., 0.01 - 0.5 Hz) to remove high-frequency cardiac pulsation (~1 Hz) [78].EEG: Use independent component analysis (ICA) to identify and remove components correlated with the ECG.
fNIRS: Low-frequency drifts or large, abrupt signal spikes.EEG: Large, low-frequency drifts.	Head Motion	fNIRS: Implement a Motion Artifact Reduction (MAR) algorithm (see Table above). Using short-separation channels is highly effective [4] [78].EEG: Apply high-pass filtering and inspect data for rejectable segments.
fNIRS: A hemodynamic-like response in a single channel or region that is not physiologically plausible.EEG: No corresponding change.	Scalp Blood Flow (extracerebral artifact)	fNIRS: This is a critical confounder. Use short-separation channel regression. If not available, apply statistical correction methods like SVD-based spatial filtering [78].
EEG: Large, low-frequency deflections, primarily in frontal channels.fNIRS: No direct correlate.	Ocular Artifacts (EOG)	EEG: Use EOG electrodes to record eye movements and employ regression-based removal or ICA to eliminate these artifacts [97].
EEG: High-frequency, irregular noise, especially in temporal channels.fNIRS: No direct correlate.	Muscle Artifacts (EMG)	EEG: Apply a low-pass filter (e.g., cut-off at 40-50 Hz) and use visual inspection or automated methods to mark and remove contaminated epochs.

Detailed Experimental Protocol: A Multimodal Motor Paradigm

The following protocol, adapted from a study investigating motor execution, observation, and imagery, provides a robust framework for simultaneous fNIRS-EEG data collection [73].

Objective: To elucidate differences in neural activity during Motor Execution (ME), Motor Observation (MO), and Motor Imagery (MI) using simultaneous fNIRS-EEG.

Materials and Setup:

• fNIRS System: A 24-channel continuous-wave system (e.g., Hitachi ETG-4100) measuring oxygenated (HbO) and deoxygenated hemoglobin (HbR) at 695 nm and 830 nm. The bilateral probe is placed over sensorimotor and parietal cortices [73].
• EEG System: A high-density 128-electrode cap (e.g., Electrical Geodesics, Inc.). The fNIRS probe is embedded within the EEG cap [73].
• 3D Digitizer: A magnetic space digitizer (e.g., Polhemus Fastrak) to record the precise locations of fNIRS optodes and EEG electrodes relative to anatomical landmarks (nasion, inion, preauricular points) [73].
• Stimulus: A cup placed on a table between the participant and an experimenter. Audio commands ("Your turn," "My turn") are delivered via pre-recorded audio [73].

Procedure:

• Participant Preparation: Fit the combined fNIRS-EEG cap. Digitize all optode and electrode positions.
• Task Conditions: The participant sits facing an experimenter. The following conditions are presented in a block design:
  • Motor Execution (ME): On the audio command "Your turn," the participant uses their right hand to grasp, lift, and move the cup about two feet toward themselves [73].
  • Motor Observation (MO): On the audio command "My turn," the participant observes the experimenter perform the same cup-moving action [73].
  • Motor Imagery (MI): On the audio command "Your turn," the participant mentally rehearses moving the cup without any physical movement [73].
• Data Acquisition: Simultaneously record fNIRS (e.g., at 10 Hz) and EEG (e.g., at 500-1000 Hz) throughout all task blocks and rest periods.

Data Analysis Workflow: The analysis involves parallel processing streams for fNIRS and EEG that are later fused.

Research Reagent Solutions: Essential Materials for fNIRS-EEG Experiments

Table: Key Equipment and Analytical Tools for Multimodal Research

Item	Function / Description	Example Use Case
High-Density DOT/fNIRS System	Uses multiple source-detector separations to enable 3D image reconstruction of functional activation, achieving spatial resolution comparable to fMRI [97].	Mapping entire cortical surface for connectivity analysis or detailed functional localization [97].
Short-Distance Channels	fNIRS channels with very short source-detector separation (0.5-1 cm) that measure signals predominantly from the scalp. Used to regress out extracerebral physiological noise [78].	Critical for improving the accuracy of fNIRS signals by removing systemic scalp hemodynamics, a major confounder [78].
Structured Sparse Multiset CCA (ssmCCA)	A advanced data fusion algorithm that identifies multivariate associations between fNIRS and EEG data to find brain regions consistently active in both modalities [73].	Identifying a shared neural region (e.g., in the parietal lobe) activated during motor execution, observation, and imagery [73].
fNIRS Brain AnalyzIR Toolbox	A software toolbox (e.g., for MATLAB) specifically designed for fNIRS data analysis, including preprocessing, statistical modeling, and functional connectivity [104].	Implementing a standardized analysis pipeline for data cleaning, first/second-level statistical models, and image reconstruction [104].
3D Magnetic Space Digitizer	A device to precisely record the 3D locations of fNIRS optodes and EEG electrodes on a participant's head relative to anatomical landmarks [73].	Essential for co-registering measurement locations to individual anatomy, improving spatial accuracy and reproducibility [73].

Standardization Efforts and Reporting Guidelines for Clinical Translation

This technical support center provides troubleshooting guides and FAQs to help researchers address specific issues encountered during fNIRS experiments, with a focus on signal noise processing for clinical applications.

Frequently Asked Questions & Troubleshooting Guides

Q1: Our group-level fNIRS results are inconsistent across studies. What are the primary factors affecting reproducibility, and how can we improve it?

The reproducibility of fNIRS findings, especially at the group level, is influenced by several methodological factors. A large-scale initiative (FRESH) that had 38 research teams analyze the same datasets found that nearly 80% of teams agreed on group-level results when hypotheses were strongly supported by literature [19]. Key factors affecting reproducibility include:

• Data Quality: Reproducibility is higher when data quality is better. Teams showed greater agreement when analyzing datasets with higher signal quality [19].
• Researcher Experience: Teams with higher self-reported confidence, which correlated with more years of fNIRS experience, demonstrated greater agreement in their results [19].
• Analysis Pipeline Variability: The main sources of variability were identified as:
  • How poor-quality data and motion artifacts are handled.
  • How the hemodynamic response is modeled.
  • How statistical analyses are conducted [19].

To improve reproducibility, adhere to community-endorsed best practices and reporting guidelines, which help increase the traceability and reliability of work by ensuring methods and findings are presented comprehensively and transparently [105].

Q2: What is the recommended sequence for a basic fNIRS data processing pipeline to minimize noise introduction?

A standardized sequence for processing steps is critical to prevent noise in early stages from affecting downstream results. The recommended order is [65]:

• Quality Control: Identify and remove channels with insufficient signal-to-noise ratio first.
• Artifact Removal: Correct or remove signal portions affected by motion artifacts.
• Noise Removal: Apply filters to remove physiological and other noise.
• Conversion: Convert raw light intensity to optical density, then to hemoglobin concentrations.

Note that the Conversion step can sometimes be performed earlier, but you must ensure processing parameters are calibrated for the correct signal type (raw intensity, optical density, or hemoglobin concentration) [65].

Q3: What are "noisy channels," and what are the best methods for identifying them?

A noisy channel has an insufficient signal-to-noise ratio, preventing reliable examination of hemodynamic activity [66]. This can be caused by poor light-tissue coupling, hardware defects, or movement artifacts [66].

Identification methods can be manual (visual inspection), semi-automatic, or fully automatic. Automatic methods typically work by calculating signal quality metrics like the Scalp Coupling Index (SCI) or the standard deviation of the signal and then flagging channels that exceed a predefined threshold [65] [66]. Best practices recommend using quantitative methods over subjective evaluations to avoid bias [65].

Q4: For clinical applications, what specific artifact correction methods are most effective for whole-head fNIRS recordings?

For comprehensive whole-head montages, advanced denoising methods that combine multiple data sources are most effective. One established method uses:

• Short-Separation Channels: These help identify and remove a globally uniform superficial component using techniques like Principal Component Analysis (PCA) [7].
• Auxiliary Measurements: Simultaneously recorded data from motion, respiration, and pulse sensors are incorporated into a General Linear Model (GLM) to identify and remove physiological noise [7].

This automated approach has been shown to improve contrast-to-noise ratio and reliability compared to minimal preprocessing methods, enabling the detection of focal activations across the brain [7].

Experimental Protocols & Methodologies

Table 1: Common fNIRS Pre-Processing Techniques and Their Applications

Technique Category	Specific Method	Primary Function	Common Parameters / Considerations
Systematic Noise Removal	Bandpass Filter	Removes high-frequency (e.g., cardiac) and low-frequency (e.g., drift) noise [17].	Cutoff frequencies typically ~0.01-0.1 Hz (high-pass) and ~0.3-0.7 Hz (low-pass) [17].
	Wavelet Filter	Effective for removing physiological noise and specific types of motion artifacts [17].	Choice of wavelet family and decomposition level.
Artifact Correction	PCA/GLM with SSChannels	Removes global superficial signals and physiological noise; ideal for whole-head setups [7].	Requires short-separation channels and/or auxiliary recordings.
Signal Quality Assurance	Scalp Coupling Index (SCI)	Quantifies optode-scalp coupling quality based on cardiac signal presence [54].	Channels with SCI < 0.5 are often considered "bad" [54].
Hemodynamic Response Isolation	High-Pass Filter	Removes slow drifts from the signal [17].	Cutoff frequency typically ~0.01-0.05 Hz.
	Low-Pass Filter	Removes high-frequency heart rate noise from the haemodynamic signal [54].	Cutoff frequency typically ~0.5-0.7 Hz [54].

Table 2: Key Research Reagents & Computational Tools

Item Name	Type	Primary Function in fNIRS Research
Short-Separation Channels	Hardware/Data	Probes placed close to a source (<1 cm) to measure systemic physiological noise from the scalp, enabling its regression from standard channels [7].
Auxiliary Recorders (ECG, Respiration)	Hardware/Data	Devices that measure heart rate, blood pressure, and breathing to model and remove physiological noise via GLM [7].
Scalp Coupling Index (SCI)	Algorithm/Metric	A quantitative metric to automatically identify and reject noisy channels based on the presence of a cardiac signal [54].
General Linear Model (GLM)	Algorithm/Model	A statistical model used to estimate the task-related hemodynamic response while accounting for noise regressors (e.g., from motion or physiology) [17].
Modified Beer-Lambert Law	Algorithm	Converts pre-processed optical density data into relative concentrations of oxygenated and deoxygenated hemoglobin [17] [54].

Detailed Protocol: fNIRS Pre-processing for Clinical Motor Control Studies

This protocol summarizes a common workflow used in motor control research, which is highly relevant for clinical applications [17] [54].

• Data Import and Annotation: Load raw intensity data. Set meaningful annotation names and durations for experimental events (e.g., "Tapping/Left", 5 seconds) [54].
• Channel Selection: Pick channels with a source-detector distance typically greater than 1 cm to ensure sensitivity to cerebral cortex activity [54].
• Signal Conversion: Convert raw intensity to optical density, then apply the Modified Beer-Lambert Law to convert optical density to relative hemoglobin concentrations (HbO and HbR) [54].
• Quality Control - SCI Calculation: Compute the Scalp Coupling Index for each channel. Mark channels with an SCI below an acceptable threshold (e.g., 0.5) as "bad" and exclude them from subsequent analysis [54].
• Filtering: Apply a bandpass filter (e.g., 0.05 - 0.7 Hz) to the haemoglobin data to remove slow drifts and high-frequency cardiac noise [54].
• Epoching: Segment the continuous data into epochs time-locked to experimental events (e.g., from -5 s pre-stimulus to 15 s post-stimulus). Apply baseline correction [54].
• Artifact Rejection: Automatically or manually reject epochs that exceed a predefined amplitude threshold (e.g., HbO > 80e-6), indicating large motion artifacts or other noise [54].

Workflow Visualization

Figure 1: Standardized fNIRS data processing and noise reduction workflow. Steps in red are critical for noise mitigation, green steps are data conversions, and blue steps are for analysis [65] [54].

FAQs on fNIRS Diagnostic Accuracy and Signal Challenges

What is the fundamental diagnostic principle of fNIRS in psychiatric disorders?

fNIRS measures cortical hemodynamic responses by detecting changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR) concentrations, which are coupled with neuronal activation. In psychiatric disorders like schizophrenia, patients exhibit reduced and altered hemodynamic responses in specific brain regions (e.g., frontotemporal cortex) during cognitive tasks like the Verbal Fluency Test (VFT). These measurable changes serve as potential biomarkers for differentiation [106].

How does physiological noise impact fNIRS diagnostic reliability?

The fNIRS signal is contaminated by multiple noise sources:

• Physiological confounds: Heartbeat (1-1.5 Hz), respiration (0.2-0.5 Hz), and blood pressure fluctuations (Mayer waves, ~0.1 Hz) [78].
• Task-evoked systemic artifacts: Global hemodynamic changes in scalp and skull layers that correlate with the task frequency, making them indistinguishable via conventional filtering [78] [107]. These artifacts can cause false positives or false negatives if not properly corrected, significantly compromising diagnostic conclusions [107].

What is the demonstrated diagnostic accuracy of fNIRS for schizophrenia?

Recent studies indicate that fNIRS biomarkers can differentiate stable schizophrenia patients from healthy controls with good sensitivity:

• The integral value (IV) of the hemodynamic response in the temporal lobes achieved an area under the curve (AUC) of 0.781 [106].
• The β value of a specific prefrontal channel (Ch48) achieved an AUC of 0.762 [106]. These biomarkers also correlated with cognitive performance in processing speed, attention, and social cognition [106].

Can fNIRS reliably differentiate between depression subtypes?

Current evidence suggests significant challenges. One clinical study reported low concordance between fNIRS diagnoses and clinical diagnoses:

• 44.0% for bipolar disorder [108]
• 38.2% for major depressive disorder [108] This limited accuracy stems from symptom overlap (particularly depression), medication effects, and the heterogeneity within diagnostic categories [108]. fNIRS may currently serve better as a state marker rather than a trait marker for disorders presenting with depression [108].

Table 1: Demonstrated Diagnostic Performance of fNIRS Across Conditions

Condition Studied	Key Diagnostic Metric	Performance Value	Associated Cognitive Domains	Reference
Schizophrenia	Integral Value (IV) of temporal lobes during VFT	AUC: 0.781 (95% CI: 0.667-0.896)	Speed of processing, Attention/Vigilance, Social Cognition [106]
Schizophrenia	β-value of channel 48 during VFT	AUC: 0.762 (95% CI: 0.655-0.869)	Speed of processing [106]
Major Depressive Disorder	Concordance with clinical diagnosis	38.2%	Depression-associated symptoms [108]
Bipolar Disorder	Concordance with clinical diagnosis	44.0%	Depression-associated symptoms [108]

Table 2: Common Noise Sources and Recommended Correction Methods

Noise Category	Specific Sources	Frequency Range	Recommended Correction Methods
Non-evoked Physiological Noise	Heartbeat, Respiration, Mayer waves	0.01 - 1.5 Hz	Band-pass filtering, Wavelet MDL detrending [78]
Evoked Systemic Artifacts	Scalp blood flow, Task-evoked systemic physiology	Overlaps with task frequency	Short-distance channel regression, PCA/ICA, Multi-channel regression [78] [107] [7]
Motion Artifacts	Subject movement, Optode coupling changes	Non-periodic	Movement Artifact Reduction (MAR) algorithms, Correlation-Based Signal Improvement (CBSI) [106] [4]
Instrumental Noise	Electronic noise, External light sources	Variable	Proper shielding, Synchronization with other modalities [109] [110]

Experimental Protocols for Diagnostic fNIRS

Verbal Fluency Test Protocol for Schizophrenia

• Task Structure: 160-second total duration comprising:
  • 10-s pre-task baseline
  • 30-s repeated counting period
  • 60-s task period (three 20-s blocks for different characters)
  • 70-s post-task baseline [106]
• fNIRS Setup: 52-channel system (e.g., ETG-4100) covering prefrontal and bilateral temporal cortex with 3.0 cm inter-optode distance, positioned according to international 10-20 system [106]
• Data Acquisition: Sampling at 10 Hz using two wavelengths (695 nm and 830 nm) [106]
• Key Outcome Measures:
  • Integral Value (IV): Cumulative magnitude of hemodynamic response during 60-s task
  • Centroid Value (CV): Temporal midpoint of fNIRS signal change
  • β-value: Coefficient from general linear model (GLM) analysis [106]

Motor Execution-Imagery Protocol for Signal Processing Validation

• Task Design: Sequential finger-tapping task performed in:
  • Motor Execution (ME): Physical performance of finger tapping
  • Motor Imagery (MI): Imagined performance of the same task [78]
• Block Structure: Five 20-s trials with fixed tapping sequence, semi self-paced within the period [78]
• fNIRS Configuration: Multichannel setup covering motor cortex regions
• Processing Comparison: Testing of multiple preprocessing approaches on the same dataset [78]

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Signal Processing and Experimental Solutions

Tool Category	Specific Solution	Primary Function	Application Context
Hardware Solutions	Short-distance channels (0.5-1 cm separation)	Regress out superficial scalp contributions [78] [7]	Essential for separating cerebral vs. extracerebral hemodynamics
Algorithmic Corrections	Principal Component Analysis (PCA)	Remove globally uniform superficial components [4] [7]	Whole-head recordings, systems without short-separation channels
Algorithmic Corrections	Independent Component Analysis (ICA)	Separate statistically independent signal components [107]	Artifact identification and removal
Algorithmic Corrections	Multi-channel Regression	Remove common components across channels [107]	Systems with multiple long-distance channels
Advanced Filtering	Wavelet Minimum Description Length (MDL)	Detrending and removing non-evoked physiological noise [78]	Superior to conventional band-pass filtering
Advanced Filtering	Maximum Likelihood Generalized Extended Stochastic Gradient (ML-GESG)	Reduce multivariate disturbances (heartbeat, breathing, instrumental) [110]	Complex noise environments
Motion Correction	Correlation-Based Signal Improvement (CBSI)	Motion artifact reduction [106]	Studies with movement or poor cooperation
Statistical Modeling	General Linear Model (GLM)	Model hemodynamic response and estimate β-values [106] [7]	Quantitative analysis of task-related activation
Experimental Paradigms	Verbal Fluency Test (VFT)	Activate prefrontal and temporal cortex for cognitive assessment [106]	Psychiatric disorders (schizophrenia, depression)
Experimental Paradigms	Motor Execution/Imagery Tasks	Activate sensorimotor cortex with well-characterized response [78] [107]	Signal processing validation, neurorehabilitation

Signal Processing Workflows

fNIRS Diagnostic Signal Processing Pathway

fNIRS Diagnostic Biomarker Extraction Pathway

Conclusion

Effective signal noise processing is the critical gateway to reliable clinical application of fNIRS technology. Through systematic addressing of physiological and motion artifacts with robust preprocessing pipelines, fNIRS has demonstrated significant potential across diverse clinical domains including neurodegenerative disease monitoring, neurorehabilitation assessment, and prosthetic device development. Future advancements will depend on standardized processing protocols, enhanced multimodal integration with EEG and fMRI, development of portable systems for real-world monitoring, and improved machine learning algorithms for automated artifact rejection. As processing methodologies continue to mature, fNIRS is poised to transition from a research tool to an indispensable clinical asset for personalized medicine and therapeutic development, particularly benefiting populations unsuitable for traditional neuroimaging. The ongoing collaboration between signal processing experts and clinical researchers will be essential to fully realize the potential of fNIRS in routine clinical practice and pharmaceutical research.

References

• 1. What should I know about fNIRS imaging? [https://www.biopac.com/knowledge-base/fnir-faq/]
• 2. Functional Near-infrared Spectroscopy (fNIRS) of Brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3294084/]
• 3. Functional Near-Infrared Spectroscopy and Its Clinical ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7364176/]
• 4. Assessment of fNIRS Signal Processing Pipelines [https://www.mdpi.com/2076-3417/12/1/316]
• 5. Functional near-infrared spectroscopy [https://en.wikipedia.org/wiki/Functional_near-infrared_spectroscopy]
• 6. Which experimental designs to use in fNIRS — Resting ... [https://www.artinis.com/blogpost-all/which-experimental-designs-to-use-in-fnirs-resting-state-design]
• 7. Correcting physiological noise in whole-head functional ... [https://www.sciencedirect.com/science/article/pii/S0165027021001977]
• 8. fNIRS-based brain-computer interfaces: a review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC4309034/]
• 9. Identifying and quantifying main components of physiological ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3865602/]
• 10. Motion artifacts in functional near-infrared spectroscopy [https://pmc.ncbi.nlm.nih.gov/articles/PMC3762942/]
• 11. Motion artifacts removal and evaluation techniques for ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.878750/full]
• 12. Disentangling the impact of motion artifact correction ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11498316/]
• 13. Functional near-infrared spectroscopy for the assessment ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1524806/full]
• 14. Ensuring good fNIRS data quality - Automatic assessment ... [https://www.artinis.com/blogpost-all/automatic-assessment-of-signal-quality]
• 15. Investigation of functional near-infrared spectroscopy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9116886/]
• 16. Best practices for fNIRS publications - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC7793571/]
• 17. Data Processing in Functional Near-Infrared Spectroscopy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8151801/]
• 18. Optimizing spatial specificity and signal quality in fNIRS [https://www.frontiersin.org/journals/neuroergonomics/articles/10.3389/fnrgo.2024.1286586/full]
• 19. fNIRS reproducibility varies with data quality, analysis ... [https://www.nature.com/articles/s42003-025-08412-1]
• 20. Standardization and Open Science Committee [https://fnirs.org/about-us/committees/standardization-committee/]
• 21. How to design real-world functional near-infrared ... [https://www.spiedigitallibrary.org/journals/neurophotonics/volume-13/issue-S1/S10701/How-to-design-real-world-functional-near-infrared-spectroscopy-studies/10.1117/1.NPh.13.S1.S10701.full]
• 22. A Scoping Review of fNIRS Applications with the Stroop Task [https://www.sciencedirect.com/science/article/pii/S1053811925005932?dgcid=rss_sd_all]
• 23. The application of fNIRS in studies on occupational workload [https://pubmed.ncbi.nlm.nih.gov/40331113/]
• 24. Functional Near-Infrared Spectroscopy in Schizophrenia ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12413766/]
• 25. How to design real-world functional near-infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12618016/]
• 26. Broad-Spectrum Error Robust Wavelengths for fNIRS ... [https://link.springer.com/article/10.1007/s12541-024-01199-x]
• 27. Method for Using Functional Near-Infrared Spectroscopy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11946166/]
• 28. Functional Near-Infrared Spectroscopy and Its Clinical ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.00724/full]
• 29. A Narrative Review on Clinical Applications of fNIRS - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC7573058/]
• 30. FAQ on NIRS - fNIRS Questions Answered [https://nirx.net/faq]
• 31. Applications of Functional Near-Infrared Spectroscopy ... [https://www.mdpi.com/2077-0383/14/15/5197]
• 32. Prefrontal fNIRS-based clinical data analysis of brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8620253/]
• 33. Unfolding the real-time neural mechanisms in addiction [https://www.sciencedirect.com/science/article/pii/S2772392522000438]
• 34. Functional Near-Infrared Spectroscopy as a Personalized ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9982178/]
• 35. Reproducibility of fNIRS within Subject for Visual and Motor ... [https://www.sciencedirect.com/science/article/pii/S1053811925004951]
• 36. Issue with Small OD and HRF Concentrations in fNIRS Data [https://openfnirs.org/community/homer3-forum/issue-with-small-od-and-hrf-concentrations-in-fnirs-data/]
• 37. Optical Density Function Error - MNE Forum - Discourse [https://mne.discourse.group/t/optical-density-function-error/4935]
• 38. NIRS-KIT: a MATLAB toolbox for both resting-state and task ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7829673/]
• 39. fNIRS Analysis | NIRS Analysis Techniques, Homer2, NIRS ... [https://nirx.net/fnirs-analysis]
• 40. Comparison of motion correction techniques applied to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9900395/]
• 41. Reducing motion artifacts for long-term clinical NIRS ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3849205/]
• 42. An Automatic Motion-Based Artifact Reduction Algorithm ... [https://www.mdpi.com/1999-4893/16/5/230]
• 43. Enhancing fNIRS data analysis with a novel motion artifact ... [https://www.sciencedirect.com/science/article/abs/pii/S1746809424005548]
• 44. Deep learning-based motion artifact removal in functional ... [https://www.spiedigitallibrary.org/journals/neurophotonics/volume-9/issue-4/041406/Deep-learning-based-motion-artifact-removal-in-functional-near-infrared/10.1117/1.NPh.9.4.041406.full]
• 45. Characterizing motion artifacts in functional near-infrared ... [https://www.sciencedirect.com/science/article/pii/S1746809425007670]
• 46. Commentary on the statistical properties of noise and its ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4773699/]
• 47. Signal Processing in Functional Near-Infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5766679/]
• 48. Transformer-based deep learning model for predicting ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12618017/]
• 49. Short-channel regression in functional near-infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7523733/]
• 50. Short separation channel location impacts the performance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3254723/]
• 51. Controlling jaw-related motion artifacts in functional near-infrared... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10681683/]
• 52. GitHub - abinnquist/fNIRSpreProcessing: SCN Lab current... [https://github.com/abinnquist/fNIRSpreProcessing]
• 53. Preprocessing Troubleshooting - Forum [https://openfnirs.org/community/homer3-forum/preprocessing-troubleshooting/]
• 54. Preprocessing functional near-infrared spectroscopy ( fNIRS ) data [https://mne.tools/stable/auto_tutorials/preprocessing/70_fnirs_processing.html]
• 55. HRfunc: a tool for modeling hemodynamic response variability ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12634216/]
• 56. Optimal hemodynamic response model for functional near- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4468613/]
• 57. Using the General Linear Model to Improve Performance in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7040364/]
• 58. A novel GLM-based method for the Automatic IDentification of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5518772/]
• 59. Improved physiological noise regression in fNIRS [https://www.sciencedirect.com/science/article/pii/S1053811919310638]
• 60. Exploring the application and challenges of fNIRS ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10957543/]
• 61. Stage-specific functional networks associated with ... [https://www.frontiersin.org/journals/aging-neuroscience/articles/10.3389/fnagi.2025.1562203/full]
• 62. Applications of Functional Near-Infrared Spectroscopy (fNIRS ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11188690/]
• 63. An fNIRS dataset for Multimodal Speech Comprehension ... [https://www.nature.com/articles/s41597-025-05654-w]
• 64. an fNIRS study in cochlear implant users and normal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12425163/]
• 65. Current Status and Issues Regarding Pre ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7373176/]
• 66. Noisy channels [https://openfnirs.org/2024/01/01/noisy-channels/]
• 67. fNIRS Optodes' Location Decider (fOLD): a toolbox for ... [https://www.nature.com/articles/s41598-018-21716-z]
• 68. PHOEBE: a method for real time mapping of optodes-scalp ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5175555/]
• 69. Current Status and Issues Regarding Pre-processing of ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2018.00505/full]
• 70. Guiding functional near-infrared spectroscopy optode- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8211086/]
• 71. False positives and false negatives in functional near-infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4791590/]
• 72. Multimodal assessment of the spatial correspondence ... [https://www.nature.com/articles/s41598-023-29123-9]
• 73. Simultaneous multimodal fNIRS-EEG recordings reveal ... [https://www.nature.com/articles/s41598-023-31609-5]
• 74. Functional near-infrared spectroscopy maps cortical plasticity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3817936/]
• 75. What we can and cannot (yet) do with functional near ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4033094/]
• 76. fNIRS and EEG Neurofeedback: Combined Benefits for ... [https://www.mendi.io/blogs/brain-health/fnirs-and-eeg-neurofeedback-combined-benefits-for-enhanced-mental-health?srsltid=AfmBOooiL2xXjuqQ4Es-H22FRuNzcY-Kjajad3P7mJedODy_Ea68Qppu]
• 77. A Deep-Learning Empowered, Real-Time Processing ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12270413/]
• 78. Signal Processing in fNIRS: A Case for the Removal of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6769087/]
• 79. Functional near infrared spectroscopy (NIRS) signal ... [https://www.sciencedirect.com/science/article/abs/pii/S105381190901235X]
• 80. Beyond the Brain: Physiological Signals in fNIRS Research [https://nirx.net/blog-content/beyondthebrain-spafnirs-integrating-physiology-wings2]
• 81. ISMRM 2021 | fNIRS Systems | NIRS Devices [https://nirx.net/ismrm]
• 82. Automated Processing of fNIRS Data—A Visual Guide to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6428450/]
• 83. Optimizing spatial specificity and signal quality in fNIRS: an overview of potential challenges and possible options for improving the reliability of real-time applications - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11188482/]
• 84. Scalp coupling index [https://openfnirs.org/2024/01/01/scalp-coupling-index/]
• 85. Hair management tips for fNIRS - [https://www.cortivision.com/hair-management-tips-for-fnirs/]
• 86. Signal Quality Evaluation — MNE-NIRS 0.7.1 documentation [https://mne.tools/mne-nirs/stable/auto_examples/general/plot_22_quality.html]
• 87. Reconstructing damaged fNIRS signals with a generative deep learning model | Artificial Intelligence Review [https://link.springer.com/article/10.1007/s10462-024-11028-2]
• 88. Applications of Functional Near-Infrared Spectroscopy (fNIRS ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12347915/]
• 89. Reliability of brain metrics derived from a Time-Domain ... [https://www.nature.com/articles/s41598-024-68555-9]
• 90. Comparison between time domain and continuous wave ... [https://www.sciencedirect.com/science/article/abs/pii/S0030399224005589]
• 91. Test-retest reliability of functional near-infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10218660/]
• 92. Targeting brain regions of interest in functional near‐infrared ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8046049/]
• 93. fNIRS analysis toolbox series – Homer [https://www.artinis.com/blogpost-all/2021/fnirs-analysis-toolbox-series-homer]
• 94. SPM for fNIRS toolbox: help [https://www.nitrc.org/forum/forum.php?offset=50&max_rows=25&style=nested&forum_id=5530]
• 95. Functional imaging of the developing brain with wearable ... [https://www.sciencedirect.com/science/article/pii/S1053811920309757]
• 96. fNIRS vs EEG: Comparing Non-Invasive Brain Imaging ... [https://imotions.com/blog/learning/research-fundamentals/fnirs-vs-eeg/?srsltid=AfmBOopsWPqtfD1KYnysWoaEceXH_UG_gLnollh_6--TiVFDfrBd-wR-]
• 97. Multimodal fNIRS–EEG sensor fusion: Review of data-driven ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12592382/]
• 98. Shedding light on the brain: guidelines to address ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12014849/]
• 99. Applications and advances of combined fMRI-fNIRs ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2025.1542075/full]
• 100. Simultaneous EEG and fNIRS recordings for semantic ... [https://www.nature.com/articles/s41597-025-04967-0]
• 101. Automated Processing of fNIRS Data—A Visual Guide to ... [https://www.mdpi.com/1999-4893/11/5/67]
• 102. Classification Algorithm for fNIRS-based Brain Signals ... [https://www.sciencedirect.com/science/article/abs/pii/S0306452224000617]
• 103. Hybrid Deep Learning Model to Estimate Cognitive Effort ... [https://arxiv.org/html/2504.13883v3]
• 104. SfNIRS Newsletter March 2025 [https://fnirs.org/sfnirs-newsletter-march-2025/]
• 105. Best practices for fNIRS publications [https://fnirs.org/best-practices-for-fnirs-publications/]
• 106. Functional near-infrared spectroscopy as a diagnostic aid ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12497714/]
• 107. Signal Processing in Functional Near-Infrared ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2017.00641/full]
• 108. Accuracy of functional near-infrared diffuse optical ... [https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2025.1496675/full]
• 109. Frequently Asked (f)NIRS Questions [https://www.artinis.com/faq]
• 110. Multivariate disturbance filtering in auditory fNIRS signals ... [https://www.sciencedirect.com/science/article/pii/S0010482524009259]