Light on Broken Networks: Resting-State fNIRS as a Tool for Connectivity Mapping

This study demonstrates that portable resting-state fNIRS can effectively map large-scale brain connectivity and network organization comparable to fMRI, validating its translational utility while highlighting that partial correlations improve edge-level agreement but may compromise broader network characterization.

The Gist

Imagine your brain is a massive, bustling city. To understand how this city works, scientists usually look at the "traffic patterns" between different neighborhoods. When people are just sitting quietly (not doing a specific task), these neighborhoods still talk to each other in rhythmic patterns. These conversations are called Resting-State Networks.

For a long time, the only way to map these conversations was with fMRI (functional Magnetic Resonance Imaging). Think of fMRI as a giant, expensive, noisy satellite camera that can see the whole city from space. It's incredibly detailed, but it has problems:

• It's huge and costs a fortune.
• You have to lie perfectly still inside a loud tube (like a giant metal coffin).
• If you move even a little, the picture gets blurry.
• You can't easily take it to a patient's home or check on them every day.

Enter fNIRS (functional Near-Infrared Spectroscopy). This is like a pair of smart glasses or a headband with little lights. It shines light through the scalp to see the blood flow in the top layer of the brain. It's portable, quiet, cheap, and you can move around while wearing it.

The Big Question:
Can these "smart glasses" (fNIRS) see the same city traffic patterns as the "satellite camera" (fMRI)? Or is the fNIRS view too blurry and distorted to be useful for doctors?

The Experiment: A Tale of Two Cities

The researchers in this paper decided to find out. They didn't just look at one street; they tried to map the whole city.

1. The Setup: They took two groups of healthy people. One group wore the fNIRS headband, and the other group went into the fMRI machine. (They matched the groups by age and gender, like comparing two similar classes of students).
2. The Task: Everyone just sat there, staring at a cross on a screen, letting their minds wander.
3. The Comparison: They took the "traffic maps" from the headband and the satellite camera and tried to overlay them to see if they matched.

What They Found: The Good, The Bad, and The "It Depends"

1. The Big Picture vs. The Details (The Map Analogy)

• The Big Picture (Group Level): When they looked at the average traffic of the whole city, the two maps looked surprisingly similar. The fNIRS headband successfully identified the major "neighborhoods" (like the Default Mode Network, which is active when you daydream, and the Visual Network).
  • Analogy: If you look at a map of a country, the headband and the satellite both correctly show where the major states and highways are. They agree on the big picture.
• The Details (Individual Level): When they looked at specific, tiny streets or individual connections between two specific buildings, the maps disagreed a lot.
  • Analogy: If you zoom in to see if a specific side-street is open or closed, the headband might say "open" while the satellite says "closed." For a single person, the headband isn't perfect yet.

2. The "Noise" Filter (Bivariate vs. Partial Correlation)

The researchers tried two different ways of calculating the connections:

• Method A (Bivariate): This is like listening to a conversation in a noisy room. You hear everything, including the background chatter (noise). This method showed some differences between the two devices.
• Method B (Partial Correlation): This is like putting on noise-canceling headphones. You filter out the background chatter to hear only the direct conversation between two people.
  • The Twist: When they used the "noise-canceling" method, the two maps looked even more similar at the street level! However, this method made it harder to see the big neighborhoods clearly. It was great for fine-tuning specific connections but less good for seeing the whole city layout.

3. The "Neighborhoods" (Modules)

The brain is organized into teams (networks). The researchers checked if the headband could identify these teams correctly.

• Result: Yes! The headband successfully found the same teams as the satellite camera, such as the "Visual Team" (seeing), the "Motor Team" (moving), and the "Thinking Team" (executive function).
• Caveat: The headband sometimes merged two distinct teams into one big group, whereas the satellite camera kept them separate. It's like the headband seeing "The West Side" as one big area, while the satellite sees "West Side North" and "West Side South" as distinct.

Why This Matters for Real Life

This study is a huge step forward for clinical neurology (helping sick people).

• The Problem: Currently, if a patient has a brain tumor or epilepsy, doctors want to know how their brain networks are changing over time. But you can't put a patient in an fMRI machine every week; it's too expensive, too scary, and they might be too sick to lie still.
• The Solution: The fNIRS headband is like a portable, friendly doctor's tool.
  • You can wear it at home.
  • You can wear it while a patient is in a wheelchair.
  • You can check them every day to see if a treatment is working.

The Bottom Line

The paper concludes that fNIRS is a reliable tool for seeing the "big picture" of brain networks.

• If you want to know: "Is the patient's brain network generally healthy or broken?" -> Use the fNIRS headband. It works well!
• If you want to know: "Is this exact tiny connection between two specific neurons working?" -> Wait for more research. The headband is still a bit fuzzy compared to the satellite camera for that level of detail.

In short: The "smart glasses" might not have the same super-high resolution as the "satellite camera," but they are good enough to see the major landmarks, they are much easier to use, and they can go where the satellite can't. This makes them a game-changer for monitoring patients in the real world.

Technical Summary

1. Problem Statement

Resting-state functional connectivity (RSFC) and resting-state networks (RSNs) are critical for understanding large-scale brain organization and its disruption in neurological diseases. While functional magnetic resonance imaging (fMRI) is the gold standard for mapping these networks, its clinical utility is limited by high costs, sensitivity to motion, confinement to hospital settings, and logistical barriers to repeated longitudinal measurements.

Functional near-infrared spectroscopy (fNIRS) offers a portable, cost-effective, and motion-tolerant alternative. However, its reliability for mapping large-scale RSNs compared to fMRI remains insufficiently established. Previous studies have been limited by small sample sizes, restricted cortical coverage, lack of critical preprocessing (e.g., short-channel regression), and a focus on single analytical levels (e.g., only edge-wise or only graph metrics). The central problem addressed is whether fNIRS can reliably capture the same large-scale RSFC patterns and network topologies as fMRI across multiple analytical scales (edge, node, graph, and module levels) to support clinical translation.

2. Methodology

The study employed a comprehensive, multi-scale comparative framework using two independent cohorts of healthy participants ($N=31$ per cohort).

• Data Acquisition:
  • fNIRS: Collected using NIRSport2 devices with near whole-head coverage (32 sources, 28 detectors, 134 channels). Participants underwent a 5-minute resting-state scan. Data were preprocessed using Satori and Python, including motion artifact correction, short-channel regression (to remove superficial scalp noise), and bandpass filtering (0.01–0.1 Hz).
  • fMRI: Data were obtained from the Human Connectome Project (HCP) and matched to the fNIRS cohort by age and gender. Participants underwent a 15-minute resting-state scan. Data were preprocessed using the standard HCP pipeline (including ICA-FIX denoising).
• Mapping & Alignment:
  • fNIRS channels were mapped to MNI coordinates using the fOLD toolbox.
  • Spherical Regions of Interest (ROIs) were created at these coordinates, and corresponding BOLD signals were extracted from the fMRI data to ensure anatomical alignment between modalities (82 matched ROIs).
• Connectivity Estimation:
  • Functional connectivity matrices were generated for fMRI, fNIRS-HbO (oxygenated hemoglobin), and fNIRS-HbR (deoxygenated hemoglobin).
  • Two correlation types were computed: Bivariate (standard Pearson) and Partial (controlling for global signal/other nodes).
• Analytical Levels:
  1. Edgewise: Independent t-tests and Two One-Sided T-Tests (TOST) to assess statistical equivalence of connection strengths.
  2. Matrix-level: Pearson correlations between group-averaged connectivity matrices, assessed via permutation testing ($N=100,000$).
  3. Nodal/Graph-level: Comparison of graph metrics (Net Strength, Normalized Strength, Assortativity, Local Efficiency, Path Length, Betweenness Centrality) at subject and group levels.
  4. Modular: Community detection using the Louvain algorithm to identify RSNs and calculate Jaccard similarity between fMRI and fNIRS modules.

3. Key Contributions

• First Multi-Scale Evaluation: This is the first study to systematically compare fNIRS and fMRI across edge, node, graph, and modular levels simultaneously using near whole-head coverage.
• Methodological Rigor: The study utilized critical preprocessing steps often omitted in prior work, specifically short-channel regression to remove systemic noise, and compared both bivariate and partial correlation approaches.
• Quantitative Equivalence Testing: The application of TOST (Two One-Sided T-Tests) provided a formal statistical assessment of whether fNIRS and fMRI edges are practically interchangeable, moving beyond simple significance testing.
• Clinical Translation Framework: The study delineates which network metrics are robust across modalities, providing a guide for when fNIRS is suitable for clinical monitoring versus when fMRI remains necessary.

4. Key Results

A. Edgewise and Global Correspondence

• Bivariate Correlations: Showed substantial differences at the edge level (50–61% of edges differed significantly between fMRI and fNIRS).
• Partial Correlations: Dramatically reduced edge-level differences (<3% significantly different). However, this came at the cost of network sparsity and potential removal of true functional interactions.
• Global Similarity: Group-level connectivity patterns showed moderate but significant cross-modal similarity ($r \approx 0.37–0.39$) for both correlation types, indicating that the overall topology of functional connections is preserved.

B. Graph Metrics (Nodal Level)

• Robust Metrics: Normalized nodal strength and Assortativity showed the highest cross-modal agreement. Under bivariate correlations, normalized strength showed 0% nodal differences, and patterns were significantly correlated ($\rho \approx 0.33–0.51$).
• Vulnerable Metrics: Net strength (sensitive to negative weights) and Path-based metrics (Local Efficiency, Path Length) showed substantial divergence, particularly under partial correlations (100% of nodes differed significantly).
• Signal Type: HbO generally performed slightly better or comparably to HbR, with bivariate correlations yielding more stable metric distributions than partial correlations for nodal properties.

C. Modular Organization (Network Level)

• Community Detection: Both modalities successfully identified seven stable communities corresponding to canonical RSNs (Default Mode, Central Executive, Salience, Attentional, Sensorimotor, and Visual networks).
• Overlap: There was significant spatial overlap between fMRI and fNIRS modules (Mean Jaccard $\approx 0.35$ for bivariate, $\approx 0.30–0.36$ for partial).
• Specific Networks: The Default Mode Network (DMN), Visual Networks, and Attentional networks showed the strongest convergence. Finer-grained distinctions (e.g., separating central vs. peripheral visual networks) were less reliably captured by fNIRS.
• Correlation Type Impact: Bivariate correlations yielded more consistent modular overlap across chromophores (HbO/HbR) compared to partial correlations, which produced less stable community structures.

5. Significance and Conclusion

The study concludes that fNIRS is a valid and clinically relevant tool for population-level characterization of large-scale brain networks, particularly for longitudinal monitoring where repeated fMRI is impractical.

• Translational Utility: fNIRS reliably captures global RSFC patterns and major functional subnetworks (DMN, Visual, Attentional). This supports its use for tracking disease progression, treatment response, and rehabilitation in conditions like brain tumors, epilepsy, and neurodegenerative diseases.
• Methodological Guidance:
  • For Group/Population Studies: fNIRS is highly effective. Bivariate correlations are recommended for capturing robust network topology and modular organization.
  • For Individual/Precision Medicine: Caution is advised. Subject-level edge and nodal metrics (especially path-based ones) show significant modality dependence. fNIRS should not yet be used for fine-grained, subject-specific connectivity diagnostics without fMRI validation.
  • Metric Selection: Clinicians and researchers should prioritize Normalized Nodal Strength and Assortativity for cross-modal comparisons, as these are the most robust metrics.
• Future Directions: The findings highlight the need for simultaneous fNIRS-fMRI recordings and individualized optode digitization to further refine anatomical precision and improve the reliability of subject-level estimates.

In summary, while fNIRS does not perfectly replicate fMRI at the finest resolution, it captures the "broken networks" of clinical populations at a level sufficient for meaningful clinical insight and longitudinal monitoring, bridging the gap between laboratory research and bedside application.