fNIRS single-trial decoding improves systematically with higher optode density, model-based noise regression, and...

Objective.Advances in high-density (HD) diffuse optical tomography (HD-DOT) promise to overcome long-standing performance limitations in classification of sparse functional near-infrared spectroscopy (fNIRS) signals, but their combined impact on single-trial brain decoding and generalization remains largely unquantified. Here, we systematically evaluate how probe density, physiology removal via short-separation (SS) regression within a general linear model (GLM), and image-space feature representations aligned with brain parcellation schemes shape single-trial decoding accuracy.Approach.To enable a structured investigation and validation via realistic ground truth data, we introduce a flexible, easy-to-use framework that allows users to augment their own channel space resting-state fNIRS data with configurable synthetic hemodynamic response functions (HRFs) on target areas of the brain, using a state-of-the art diffuse optical forward model. Using three HD fNIRS datasets-a whole-head resting-state recording augmented with synthetic HRFs and two motor ball-squeezing datasets-we derive sparse-to-HD optode subsets, integrate GLM-based SS regression into cross-validation, and compare channel-space and parcel-space features derived from HD-DOT image reconstructions.Main results.HD configurations consistently and significantly improve classification accuracy and robustness to focal activations. SS correction yields systematic gains of approximately 4% in within-subject decoding and more than 10 percentage points in cross-dataset transfer. Parcel-space features outperform channel-space features at matched dimensionality, enabling robust leave-one-subject-out decoding (mean accuracy 79%) and cross-dataset generalization across different probe layouts (72% with SS correction).Significance.All methodology is implemented and available in the open-sourceCedalionframework. Together, these results demonstrate that HD-DOT, GLM-based SS regression, and parcel-space representations jointly enable significantly more accurate, robust, and probe-independent fNIRS classification pipelines.