fNIRS: Functional Near-Infrared Spectroscopy

• fNIRS is a non-invasive neuroimaging technique that infers localized brain activity by monitoring changes in oxygenated and deoxygenated hemoglobin.
• It employs continuous-wave, frequency-domain, and time-domain modalities alongside the modified Beer–Lambert law for precise quantification of hemodynamic responses.
• Widely applied in neuroscience, clinical diagnostics, and brain-computer interfaces, fNIRS balances affordability, portability, and spatial resolution.

Functional near-infrared spectroscopy (fNIRS) is a non-invasive optical neuroimaging technique that measures cortical hemodynamics by tracking changes in the absorption of near-infrared light, typically in the 650–900 nm range, as it propagates through biological tissue. fNIRS infers localized brain activity by quantifying dynamic concentration changes in chromophores, primarily oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (HbR). Its affordability, portability, and relatively high spatial resolution over superficial cortex have enabled widespread applications in neuroscience, neurology, BCI, human–machine interaction, and neuroergonomics.

1. Physical Principles and Core Measurement Theory

fNIRS hinges on differential absorption characteristics of HbO and HbR within the near-infrared spectrum. Incident light is injected into the scalp via sources (laser diodes, LEDs) and detected a few centimeters away after traversing the tissue. The quantitative foundation is the modified Beer–Lambert law:

$$A(\lambda) = -\log_{10}\left( \frac{I(\lambda)}{I_0(\lambda)} \right)$$

$$\Delta A(\lambda) = \epsilon_{\mathrm{HbO}}(\lambda)\cdot\Delta[\mathrm{HbO}]\cdot L \cdot \mathrm{DPF} + \epsilon_{\mathrm{HbR}}(\lambda)\cdot\Delta[\mathrm{HbR}]\cdot L \cdot \mathrm{DPF}$$

where $I_0(\lambda)$ and $I(\lambda)$ are incident and detected intensities, $\epsilon$ are molar extinction coefficients, $L$ is source–detector distance, and DPF is the differential pathlength factor accounting for scattering. With two wavelengths, concentration changes $\Delta[\mathrm{HbO}]$ and $\Delta[\mathrm{HbR}]$ are recovered by solving a 2×2 linear system (Sharmin et al., 2024, Ghalavand et al., 15 May 2025, Adamic et al., 2024).

System designs include continuous-wave (CW), frequency-domain, and time-domain modalities. CW-fNIRS measures steady-state attenuation, FD-fNIRS derives amplitude/phase changes of modulated light, and TD-fNIRS analyzes photon time-of-flight (DTOF/TPSF) for enhanced depth discrimination (Adamic et al., 2024, Frias et al., 5 Dec 2025).

2. Instrumentation: Hardware, Cap Design, and Data Acquisition

Typical fNIRS devices comprise multi-wavelength light sources and photodetectors coupled via optodes to the scalp. Channel counts range from 4 to >60, with source–detector separations of 2.5–3.5 cm probing up to ~2 cm into cortex. Sampling rates vary from 5–14 Hz CW to >1 kHz TD. Headgear is fixed-grid or customizable, sometimes 3D-printed for high-density or individualized placement (Sharmin et al., 2024, Kim et al., 26 May 2025).

Block and event-related paradigms are supported. Device examples include Artinis OxyMon (740, 860 nm), NIRSport-2, NIRSIT, ETG-4000, and open-source platforms such as OpenNIRScap (dual-wavelength, 24 channels, <\$500 cost). Advanced systems integrate analog multiplexing, microcontrollers (e.g. STM32L476), and on-board filtering to enable real-time, mobile data streams (Kim et al., 26 May 2025).

Acquisition protocols commonly employ millisecond-precision event-marker synchronization—traditionally via PsychoPy and TTL/serial triggers—which can be simplified using Python-based marker generation and GUI control without specialized hardware (Sharmin et al., 2024).

3. Signal Processing: Preprocessing, Feature Engineering, and Noise Suppression

Preprocessing pipeline steps include:

• Conversion of raw intensity to ΔOD and hemoglobin concentrations via MBLL.
• Band-pass filtering (e.g., 0.01–0.2 Hz Butterworth) to remove cardiac, respiratory, and very slow drifts.
• Artifact removal: ICA for denoising (cardiac artifact), wavelet-based filtering, spline interpolation, short-channel regression for superficial signal subtraction (Mirbagheri et al., 2020, Sharmin et al., 2024).
• Epoching and baseline correction aligned to stimulus onsets.

Feature engineering spans statistical moments (mean, variance, skewness), frequency-domain power (via FFT), slopes and first-derivative maxima, principal-component projections, and spatial–temporal patterns (e.g., via Gramian Angular Field transforms assembling all channels into 2D images) (Ghalavand et al., 15 May 2025, Li et al., 26 Feb 2025). Channel selection can be optimized using Pearson-correlation pruning to retain only minimally redundant, maximally discriminative pairs, supporting peak accuracies using just two channels (Li et al., 26 Feb 2025).

Synthetic data generation with Monte Carlo photon migration (MCX), parametric multi-layer head models, and cloud-based workflows have established benchmark datasets with labeled ground truth for supervised machine learning and tomographic inverse solvers (Waks, 2024).

4. Analytical and Computational Methodologies

Statistical analysis frameworks include block averaging, general linear models (GLM) for hemodynamic response modeling, mixed-effects models, functional connectivity (e.g., interhemispheric Pearson correlation), and various t-tests, ANOVA, and nonparametric methods (Zhu et al., 2013, Sharmin et al., 2024).

Machine learning pipelines range from classical classifiers (LDA, SVM, ensemble trees) using feature vectors, to deep learning architectures: CNNs (1D for temporal signals; 2D for GAF images), LSTM and Bi-directional LSTM for modeling temporal dependencies, transformers (fNIRS-T), and hybrid CNN-LSTM models (Wickramaratne et al., 2021, Khan et al., 2024, Ghalavand et al., 15 May 2025, Khan et al., 2024). Metric learning can split embeddings into class and detector subspaces, enabling reliable exclusion of out-of-distribution artifacts alongside high in-distribution accuracy (Cao, 2024). Bayesian hierarchical modeling—via Dirichlet process priors—accounts for inter-participant variability, providing personalized classification in pain detection applications (Lopez-Martinez et al., 2019).

5. Practical Applications: BCI, Clinical Neuroscience, and Human Factors

fNIRS has demonstrated utility across:

• Brain–computer interface (BCI) systems for direct communication (rest/task classification, imagined speech, mental arithmetic, motor imagery) (Sereshkeh et al., 2018, Wickramaratne et al., 2021, Ghalavand et al., 15 May 2025).
• Cognitive workload, stress, and affect assessment based on PFC hemodynamics and machine/deep learning (Khan et al., 2024, Mirbagheri et al., 2020, Khan et al., 2024).
• Game-based learning: mapping engagement, workload, and expertise in realistic educational contexts (Sharmin et al., 2024).
• Clinical diagnostics: e.g., detecting pain via personalized hierarchical models (Lopez-Martinez et al., 2019), autism/connectivity biomarkers (Zhu et al., 2013).
• Real-world deployments in driving, VR, adaptive HMI, and mobile neuroergonomics with multimodal fusion of fNIRS, eye tracking, and behavioral data (Khan et al., 2024).

High-density tomography and time-domain fNIRS devices can achieve 1 cm spatial resolution, supporting fNIRS-based visual imagery reconstruction using state-of-the-art diffusion models (Adamic et al., 2024).

6. Limitations, Contemporary Challenges, and Innovations

Limitations of fNIRS include:

• Restricted penetration depth; sensitivity confined to superficial cortex.
• Signal contamination from scalp, skull, and CSF—addressed via dual-slope FD-NIRS and three-layer tissue modeling to isolate cerebral hemodynamics (Frias et al., 5 Dec 2025).
• Susceptibility to motion artifacts, ambient light, hardware drift; partial mitigation via artifact-correction pipelines and metric learning for OOD exclusion (Cao, 2024).
• Inter-subject and session variability in DPF and hemodynamic response function (HRF).
• Computational demands for real-time, on-device deep learning.

Innovations include correlation-based channel selection for ultralight wearable BCIs, multimodal integration (e.g. EEG+fNIRS), containerized cloud simulation infrastructures for data generation, and open-source, low-cost instrumentation such as OpenNIRScap (Kim et al., 26 May 2025).

7. Trends, Recommendations, and Future Directions

• Rapid publication growth post-2019, with expanding hardware diversity, wireless arrays, and wearable form factors (Sharmin et al., 2024, Kim et al., 26 May 2025).
• Shifts toward longitudinal, real-world, and multimodal experimental designs.
• Analytical convergence around standardized preprocessing protocols, artifact rejection, and model interpretability.
• Recommendations include increased cohort diversity, adaptive closed-loop paradigms, explainable AI, deeper integration with multimodal sensors, and validation in ecologically valid environments.
• Anticipated advances in spatial resolution, real-time BCI control, clinical translation (pain, neurodevelopment, disorders), and low-cost, open hardware for global access.

Functional near-infrared spectroscopy thus occupies a central and expanding position in technical neuroimaging, balancing accessibility, temporal resolution, and spatial specificity. Recent work demonstrates convergence of physiologically informed hardware, advanced computational pipelines, and robust machine learning, collectively enhancing the fidelity, interpretability, and applicability of functional hemodynamic mapping in neuroscience and real-world neurotechnology.