🧠 neuroscience

Independent Component Analysis Outperforms Seed-Based Approach in Detecting fNIRS-based Resting-State Functional Connectivity

This study demonstrates that Independent Component Analysis (ICA) outperforms seed-based approaches in detecting fNIRS-based resting-state functional connectivity by providing higher accuracy and greater consistency between oxygenated and deoxygenated hemoglobin signals across the motor network.

Original authors: Kotsogiannis, F., Raible, S., Pereira, J., Heinecke, A., Klinkhammer, S., Sorger, B., Lührs, M.

Published 2026-02-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kotsogiannis, F., Raible, S., Pereira, J., Heinecke, A., Klinkhammer, S., Sorger, B., Lührs, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain is a massive, bustling city. Even when you aren't doing anything specific (like solving a puzzle or reading a book), the different neighborhoods in this city are still talking to each other. This constant, low-level chatter is called Resting-State Functional Connectivity (RSFC). Understanding how these neighborhoods connect helps doctors and scientists figure out how the brain works and why it might get sick.

For a long time, scientists used a giant, expensive machine called an MRI to listen to this chatter. But MRI machines are loud, scary for some people, and hard to use with kids or patients who can't stay still.

Enter fNIRS (functional near-infrared spectroscopy). Think of fNIRS as a smart, wearable headband that uses light to listen to the brain. It's quiet, portable, and great for everyone. However, the light it uses picks up two types of signals:

HbO (Oxygenated blood): Like the "fresh" delivery trucks bringing oxygen.
HbR (Deoxygenated blood): Like the "used" trucks taking waste away.

The big question this paper asks is: "When we use this headband to listen to the brain's chatter, what is the best way to decode the message?"

The Two Detectives: SBA vs. ICA

The researchers tested two main "detectives" (methods) to figure out which brain areas are talking to each other.

1. The Seed-Based Detective (SBA)

The Analogy: Imagine you want to know who is friends with a specific person in a crowded room. You pick one person (the "seed") and ask, "Who is talking to you?" You then map out everyone who is having a conversation with that one person.

The Problem: You have to guess who the "seed" is beforehand. If you pick the wrong person, your map is wrong. Also, this method assumes everyone is just talking to that one person, ignoring the fact that the whole room might be having a complex group conversation.

2. The Independent Component Detective (ICA)

The Analogy: Imagine you are in a crowded room with a terrible sound system where everyone is talking at once. It's a mess of noise. ICA is like a super-smart audio engineer who can separate the audio tracks. Instead of asking one person who they are talking to, the engineer listens to the whole room and says, "Ah, I can hear the 'Motor Team' talking over here, and the 'Frontal Team' talking over there."

The Benefit: It doesn't need you to pick a starting point. It finds the natural groups (networks) on its own.

The Experiment: The Finger-Tapping Test

The researchers put 38 people in a room and had them:

Relax (Resting state).
Tap their fingers (to light up the "Motor" part of the brain, like turning on a specific neighborhood in our city analogy).

They used the fNIRS headband to record the brain's activity while the participants relaxed. Then, they ran the data through both "detectives" (SBA and ICA) to see which one could best identify the Motor Network (the part of the brain that controls movement).

The Results: Who Won?

The study compared the detectives using two rules:

The Map Rule: Did the detective find the right neighborhoods based on a standard city map?
The Activity Rule: Did the detective find the neighborhoods that actually lit up when people tapped their fingers?

The Winner: ICA (The Audio Engineer)

Accuracy: ICA was much better at finding the correct connections. It was like the audio engineer who perfectly separated the "Motor Team" from the background noise.
Consistency: ICA gave very similar results whether it was looking at the "fresh delivery trucks" (HbO) or the "used trucks" (HbR). This is huge because it means the method is robust and reliable.
The HbR Surprise: For a long time, scientists thought the "used trucks" (HbR) were too noisy to be useful. This study proved that HbR actually contains just as much useful information as HbO, as long as you use the right detective (ICA) to listen to it.

The Runner-Up: Correlation SBA

The "Seed-Based" detective wasn't terrible, especially a version that just looked for simple correlations. It was faster and easier to use, but it wasn't as accurate or consistent as ICA.
The other versions of SBA (which tried to use complex math models) struggled a bit more, especially when looking at the "used trucks" (HbR).

Why Does This Matter?

Better Tools for Doctors: If we want to use fNIRS headbands in hospitals to help diagnose mental health issues or brain injuries, we need to know exactly how to analyze the data. This paper says: "Use ICA." It gives the most reliable map of the brain's connections.
No More Guessing: You don't need to guess which part of the brain to start with. ICA finds the patterns automatically.
Double the Data: We can now trust both types of blood signals (HbO and HbR) to tell us about brain health, effectively doubling the information we get from these wearable devices.

The Bottom Line

Think of the brain as a complex symphony.

SBA is like trying to understand the song by listening to just one instrument and guessing who else is playing with it.
ICA is like having a conductor who can instantly separate the strings, the brass, and the percussion to hear the whole symphony clearly.

This study shows that for listening to the brain's resting state with light-based headbands, the Conductor (ICA) is the best choice. It finds the music more accurately, works with all the instruments (both blood signals), and helps us standardize how we study the brain in the future.

1. Problem Statement

Resting-state functional connectivity (RSFC) is a critical metric for understanding brain function and diagnosing neurological disorders. While functional near-infrared spectroscopy (fNIRS) offers significant advantages over fMRI for clinical and pediatric populations (e.g., mobility, lower cost, tolerance to motion), the optimal analytical strategy for estimating RSFC from fNIRS data remains unclear.

Key gaps in current literature include:

Methodological Uncertainty: It is unknown whether Seed-Based Analysis (SBA) or Independent Component Analysis (ICA) is superior for detecting reliable connectivity patterns, particularly across both oxygenated (HbO) and deoxygenated (HbR) hemoglobin signals.
HbR Neglect: Many studies rely exclusively on HbO due to its higher signal-to-noise ratio (SNR), potentially ignoring valuable information contained in HbR.
Lack of Standardization: Previous comparisons often suffered from small sample sizes, restricted brain coverage, manual/subjective component selection, and a lack of rigorous statistical comparison between methods (e.g., SBA-GLM vs. SBA-Correlation).

2. Methodology

Participants and Data Acquisition:

Sample: 38 healthy participants (after quality control from an initial 57) were recruited.
Hardware: Near whole-head fNIRS coverage (134 channels) using NIRSport2 devices, based on the 10-10 system.
Protocol:
- Resting State: 5 minutes of eyes-open fixation.
- Task: Bilateral finger-tapping (5 min 50s) used as a functional localizer to identify motor seeds and generate ground-truth activation maps.
- Physiology: Respiratory signals were recorded simultaneously for regression analysis.

Preprocessing:

Standardized pipeline using Satori and Python.
Steps included: Optical density conversion, Modified Beer-Lambert Law (mBLL) conversion, bandpass filtering (0.01–0.1 Hz for SBA; 0.01–0.2 Hz for ICA), spike removal, TDDR (Temporal Derivative Distribution Repair), and short-channel regression to remove superficial noise.
Exclusion: Participants were excluded if >30% of channels had a Scalp Coupling Index (SCI) < 0.70.

Analytical Approaches Evaluated:
The study compared five distinct methods for estimating RSFC:

SBA-GLM: General Linear Model using the seed time-course as a predictor.
SBA-GLM-Resp: GLM including respiratory signal regression.
SBA-Correlation (SBA-Corr): Pearson correlation between seed and all other channels.
ICA-Skew: Independent Component Analysis maximizing skewness (replicating prior work).
ICA-LogCosh: ICA using the $\log(\cosh(x))$ contrast function (robust approximation of negentropy).

Seed and Component Selection:

Seeds (for SBA): Identified via a group-level GLM on the finger-tapping task; the channel with the highest absolute t-value (S25-D23) was selected.
Components (for ICA): A data-driven, automated Leave-One-Out (LOO) approach was used. For each subject, ICs were correlated with a group motor map generated from all other subjects to avoid circularity. The IC with the highest spatial correlation was selected. To ensure stability, ICA was run 50 times per subject with different random seeds, and the median spatial map was computed.

Performance Evaluation:

ROC Analysis: Methods were evaluated against two "gold standard" reference maps:
1. Anatomical: fOLD atlas-based motor region map.
2. Functional: Binarized finger-tapping task activation map.
Metrics: Area Under the Curve (AUC) calculated via DeLong's test for statistical significance.
Consistency: Spatial correlation ( $r$ ) between HbO and HbR RSFC maps was calculated to assess cross-chromophore consistency.

3. Key Results

Performance (AUC Scores):

ICA Superiority: ICA methods (both Skew and LogCosh) consistently outperformed all SBA variations across both chromophores and reference standards.
- HbO (Functional Map): ICA-Skew (AUC = 0.826) > ICA-LogCosh (0.820) > SBA-Corr (0.738) > SBA-GLM (0.718).
- HbR (Functional Map): ICA-Skew (AUC = 0.922) > ICA-LogCosh (0.913) > SBA-Corr (0.658) > SBA-GLM (0.635).
Statistical Significance: DeLong's test confirmed that ICA significantly outperformed SBA-GLM and SBA-GLM-Resp. SBA-Corr significantly outperformed GLM-based methods in HbR but generally underperformed compared to ICA.
Respiratory Regression: Adding respiratory regression to GLM (SBA-GLM-Resp) did not significantly improve performance and, in some strict thresholds, slightly reduced it.

Cross-Chromophore Consistency:

ICA demonstrated significantly higher spatial similarity between HbO and HbR patterns ( $r = 0.90–0.92$ ) compared to SBA methods ( $r = 0.84–0.86$ ).
This indicates that ICA extracts more consistent functional networks across the two hemodynamic signals.

Qualitative Findings:

ICA successfully identified distinct motor and frontal networks.
SBA patterns were more diffused, particularly SBA-Corr.
HbR signals were found to contain robust RSFC information comparable to HbO when analyzed with appropriate methods (ICA).

4. Key Contributions

Systematic Comparison: First study to comprehensively compare SBA (GLM, GLM-Resp, Correlation) and ICA (Skew, LogCosh) using near whole-head fNIRS data with rigorous statistical testing (DeLong's test).
Validation of HbR: Demonstrated that HbR signals contain reliable RSFC information comparable to HbO, challenging the historical reliance on HbO alone.
Automated ICA Selection: Introduced a reproducible, data-driven Leave-One-Out approach for selecting motor-related ICs, eliminating the subjectivity of visual inspection common in previous studies.
Methodological Standardization: Provided evidence that ICA is the preferred method for accuracy and consistency, while identifying correlational SBA as a computationally efficient alternative when localizer tasks are unavailable.
Reference Standard Diversity: Showed that the choice of reference standard (Anatomical vs. Functional) impacts performance evaluation, with functional maps revealing more distinct performance differences between methods.

5. Significance and Conclusion

This study establishes Independent Component Analysis (ICA) as the superior framework for estimating fNIRS-based resting-state functional connectivity. ICA offers higher accuracy, better statistical performance, and greater consistency between HbO and HbR signals compared to Seed-Based Analysis.

Clinical Implication: The findings support the integration of fNIRS-RSFC into clinical practice for psychiatric and neurological assessment, particularly for populations where fMRI is impractical.
Recommendation: Researchers should prefer ICA for reliable, cross-chromophore consistent estimates. If computational resources are limited or a localizer task cannot be performed, Correlational SBA is a viable, efficient alternative, though it is less robust than ICA.
Future Directions: The authors advocate for further standardization, larger sample sizes, and the exploration of graph theory and machine learning approaches to further refine fNIRS-RSFC analysis.

In summary, the paper provides a critical step toward the methodological standardization of fNIRS, validating its utility for both basic neuroscience and clinical translation.