Riemannian geometry boosts functional near-infrared spectroscopy-based brain-state classification accuracy

Background: Functional near-infrared spectroscopy (fNIRS) has recently gained momentum as a reliable and accurate tool for assessing brain states based on the vascular response to neural activity. This increase in popularity is due to its robustness to movement, non-invasive nature, portability, and user-friendly application. However, compared with other hemodynamic functional brain-imaging methods such as functional magnetic resonance imaging (fMRI), fNIRS is constrained by its limited spatial resolution and coverage with a particularly limited penetration depth. In addition, due to comparatively fewer methodological advancements, the performance of fNIRS-based brain-state classification still lags behind more prevalent methods such as fMRI. Methods: We introduce a classification approach grounded in Riemannian geometry for the classification of kernel matrices, leveraging the temporal and spatial relationships between channels and the inherent duality of fNIRS signals, specifically oxygenated and deoxygenated hemoglobin. For the Riemannian-geometry-based models, we compared different kernel matrix estimators and two classifiers: Riemannian Support Vector Classifier and Tangent Space Logistic Regression. These were benchmarked against four models employing traditional feature extraction methods. Our approach was tested on seven participants in two brain-state classification scenarios based on the same fNIRS dataset: an eight-choice classification, which includes seven established plus an individually selected imagery task, and a two-choice classification of all possible 28 two-task combinations. Results: This approach achieved a mean eight-choice classification accuracy of 65%, significantly surpassing the mean accuracy of 42% obtained with traditional methods. In addition, the best-performing model achieved an average accuracy of 96% for two-choice classification across all task combinations, compared with 78% with traditional models. Conclusion: To our knowledge, we are the first to demonstrate that the proposed Riemannian-geometry-based classification approach is both powerful and viable for fNIRS data, substantially increasing the accuracy in binary and multi-class classification of brain activation patterns.