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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Functional near-infrared spectroscopy (fNIRS) offers noninvasive monitoring of brain activity.
• Extracerebral hemodynamic signals can contaminate fNIRS data, necessitating correction methods like short-channel regression (SCR).
• Physical short-separation detectors for SCR may be limited by hardware or experimental setup.

Purpose of the Study:

• Develop a transformer-based deep learning model to predict short-separation optical density (OD) signals from long-separation fNIRS channels.
• Evaluate the efficacy of these predicted virtual signals for performing SCR and improving data quality.

Main Methods:

• A transformer encoder model was trained on resting-state fNIRS data with paired short- and long-separation recordings.
• The model predicted short-channel OD signals from long-channel data.
• Model performance was validated on three independent datasets, including resting-state and task-based paradigms.
• Signal similarity and denoising efficacy were assessed using metrics like MSE, NMSE, and Pearson correlation.

Main Results:

• Predicted short-channel signals demonstrated high correspondence with ground-truth measurements (median r = 0.70 for OD).
• Virtual regressors derived from the model effectively denoised long-channel fNIRS data.
• The model's performance was robust across diverse datasets and improved with the exclusion of low-coherence channels.
• In motor tasks, predicted regressors preserved task-evoked brain activity and reduced residual variance.

Conclusions:

• Transformer-based deep learning models can accurately reconstruct extracerebral hemodynamic signals from fNIRS data.
• This provides a viable virtual alternative to physical short channels for SCR.
• The approach supports standardized, hardware-independent preprocessing for fNIRS data.