fNIRS noise regression with GLM and temporally embedded canonical correlation analysis

• fNIRS preprocessing can use a multimodal extension of the General Linear Model with temporally embedded Canonical Correlation Analysis (CCA) for noise regression.¹
• The approach improves physiological time-series quality for non-invasive BCI and neuroimaging; BU Neurophotonics Center reported the method.¹
• Methods are directly applicable to neural data analysis and fNIRS pipelines.^{1 1}

Gardner updates

• fNIRS preprocessing via GLM with temporally embedded canonical correlation analysis for noise regression improves physiological time-series quality for non-invasive BCI and neuroimaging. ¹

Weekly enrichment (2026-07-20)

• The underlying method was published in NeuroImage volume 208 (2020) by von Lühmann and colleagues as a temporally embedded canonical correlation analysis (tCCA) extension of the fNIRS General Linear Model.^{2 3}
• tCCA flexibly integrates any number of auxiliary signals — short-separation fNIRS channels, blood pressure, respiration, photoplethysmogram, and accelerometer — using temporal embedding (time-shifted copies) to capture non-instantaneous and non-constant coupling that standard short-separation regression cannot.³
• The identified optimal parameters were a step size of 0.08 s, a maximum time lag of 3 s, and a correlation threshold of 0.3.³
• Validation used 5 minutes of resting-state data per participant (14 subjects), synthetic hemodynamic response functions (time-to-peak 6 s, 16.5 s duration), and two-fold cross-validation.³
• tCCA gave statistically significant physiological noise reduction across all frequency bands, including near-complete rejection of the ~1 Hz cardiac peak (paired t-test, p ≪ 0.001).³
• At the lowest contrast-to-noise ratio (20% HRF, single-trial recovery), tCCA improved correlation by 45%, reduced RMSE by 55%, and increased the F-score 3.25-fold versus the standard short-separation GLM.³
• Even when limited to short-separation channels alone, tCCA beat the conventional short-separation GLM (correlation +13%/+5%, RMSE −29%/−14%, F-score +31%/+34% for HbO/HbR).³
• On real visual-stimulation data, tCCA yielded significantly more activated channels than the standard method (paired t-test, p = 0.02).³
• The method is distributed as the hmrR_tCCA function in the Homer3 toolbox, inserted directly before the GLM, with short-separation channels and head accelerometer noted as the most valuable auxiliary inputs for BCI/neuroimaging pipelines.⁴

Footnotes

1. https://news.google.com/rss/articles/CBMiuwJBVV95cUxPX1V0NmQtUjUyY2Jtd1VkQ2FSRWtraFlzVjBQQ3oyVUxjRFlyS0NJM0tQLUFaMmxNellwaTUzbFpHSXVTVHhCWTIxaGQ1RWdZMGR6aER0UHRZOUhRTXlXdUxTTE5GYXk2UHN5emtoXzdPcEZJMmtBX05UdG43MHlrRGRrejhjcldldjlhcURaVlo3akJsODIwSmNiTmZPdGZRWkJHOTlFZlhqbnh0SG9tX1pwbGU4NWZWdVFFNS1vc25EUkNQdGYyR2FGdVdNUnhEYnkxYVZXdUEyRnBjb0R5V19kLXYzTjlpUWJDVlpsd3pkSlVPenFKdEJWRm56R3pRSGNqNzZOWWUxWnVZMU56NnhiSnd4OUFISmRkLS1iaGlacjR1dzF1V2JSSkl0bWo0eGJETnlsM0JIc0k?oc=5 ↩ ↩² ↩³ ↩⁴ ↩⁵

2. https://doi.org/10.1016/j.neuroimage.2019.116472 ↩

3. https://pmc.ncbi.nlm.nih.gov/articles/PMC7703677/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

4. https://github.com/BUNPC/Homer3/wiki/hmrR_tCCA ↩