Transformer-based deep learning model for predicting fNIRS short-channel signals

Functional near-infrared spectroscopy (fNIRS) enables portable and noninvasive monitoring of cerebral hemodynamics, but hemodynamic changes originating from extracerebral tissues may influence the signals. To avoid this, short-channel regression (SCR) is widely used, yet physical short-separation detectors are not always available or optimally positioned due to hardware limitations or the experimental setup. In such cases, a virtual, data-driven alternative to physical short-channel detectors may be a viable solution.

We aimed to (i) develop a transformer-based deep learning model to predict short-separation optical density (OD) signals from long-separation channels and (ii) evaluate whether these virtual signals enable effective SCR.

We trained the model on a resting-state fNIRS dataset (69 subjects) with paired short- and long-separation recordings. Dual-wavelength OD signals in segmented time windows were used as input for a transformer encoder trained to reconstruct the extracerebral hemodynamic component measured by short channels. Model performance was evaluated using 3 independent datasets: a holdout subset of the same resting-state dataset (23 subjects), a second dataset acquired using a different system (40 subjects), and a task-based finger-tapping dataset (4 subjects). A wavelet coherence-based channel rejection step was optionally applied during preprocessing. Predictions were evaluated using signal similarity metrics (mean squared error [MSE], normalized MSE [NMSE], and Pearson correlation [ $r$ ]) and denoising efficacy (residual variance after regression).

Predicted short-channel signals showed high correspondence with ground-truth measurements in OD (median $r=0.70$ and $\mathrm{NMSE}=0.047$ ) and concentration data (up to $r=0.67$ ). When used for SCR, virtual regressors effectively denoise long-channel data. Performance was robust across all datasets, with greater accuracy when low-coherence channels were excluded. In motor task blocks, predicted regressors preserved task-evoked activations and reduced residual variance.

Transformer-based models accurately reconstruct extracerebral hemodynamic signals from long-separation fNIRS data, providing a virtual alternative to physical short channels and supporting standardized, hardware-independent preprocessing.