Novel fNIRS technique for brain-state detection

• A novel technique improves brain-state detection using fNIRS, advancing non-invasive neuroimaging relevant to BCI and neurofeedback.¹
• Improved state decoding supports next-generation non-invasive applications (Max-Planck). Tier-1.^{1 1}

Weekly enrichment (2026-07-20)

• The technique was published in Neurophotonics (2025, DOI 10.1117/1.NPh.12.4.045002) by a team led by Tim Näher at the Max Planck Institute for Biological Cybernetics in Tübingen, with corresponding author Bettina Sorger (Maastricht University).²³⁴
• The method applies Riemannian geometry to fNIRS kernel matrices, stacking oxygenated (HbO) and deoxygenated (HbR) hemoglobin signals into block-diagonal matrices to exploit their complementary, naturally anticorrelated spatial co-activation patterns rather than treating them as redundant.²
• Two Riemannian classifiers (a Riemannian Support Vector Classifier and Tangent Space Logistic Regression) were benchmarked against four traditional feature-based models (LDA, logistic regression, random forest, and SVC).²
• The dataset covered seven participants performing eight mental-imagery tasks with 12 trials each (96 trials per participant), evaluated with repeated (n = 5) five-fold stratified cross-validation and permutation testing.²
• For eight-choice classification the Riemannian approach reached a mean accuracy of 65% versus 42% for traditional methods (theoretical chance 12.5%); the best block-diagonal + Riemannian SVC model reached 96% mean accuracy on two-choice classification across all 28 task pairs versus 78% for traditional models (chance 50%).²⁵
• “Mental talking” was the most separable task (70% correctly classified), followed by mental calculation, an individually selected “own paradigm,” and spatial-navigation imagery (each 68%), tennis imagery (62%), and mental singing/drawing (58%).²
• In a proof-of-concept for disorders of consciousness, the paradigm correctly identified responsiveness in every case and unresponsiveness in nine of ten cases.⁴⁵
• Riemannian models are computationally efficient, require less training data than CNN/LSTM deep-learning approaches, and are implemented with the open-source pyriemann and scikit-learn stack, making them practical for fNIRS-based brain-computer interfaces.²
• The authors caution that the lack of a dedicated train-test split and of short-separation channel regression means cross-validation accuracy may overestimate real-world performance, and that some HbO/HbR divergence could reflect systemic (non-neural) artifacts.²

Footnotes

1. https://news.google.com/rss/articles/CBMif0FVX3lxTFB6VXJvS2VwMmxGd19SaDBCcW5JRzVuUEZQYkxDZFdZSWh0QnpWNUlGdGp5VmJ6eTI1Q3BqeXQwQlQ5bFFJRGdGWU1VaS14a0ZVdHZjU1VzZG1kLXJDajFaZUpwZThiRXZUcThoNDhxY0RPT2dURG5SUWxKNkNIRmM?oc=5 ↩ ↩² ↩³

2. https://pmc.ncbi.nlm.nih.gov/articles/PMC12523035/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://doi.org/10.1117/1.nph.12.4.045002 ↩

4. https://www.mpg.de/25678245/novel-technique-improves-brain-state-detection ↩ ↩²

5. https://medicalxpress.com/news/2025-11-mathematics-based-approach-brain-state.html ↩ ↩²