BCIwiki (bciwiki.com) — Researchers affiliated with the University of California, Berkeley, the University of Southern California and Microsoft Research introduced MtReNIRS, an fNIRS model that maps optical probe layouts to a shared anatomical reference. Posted to arXiv on October 26, 2026, the manuscript reports submission to IEEE OJSP. Pretraining used 12 public datasets covering 268 subjects and about 567 recording hours.
Functional near-infrared spectroscopy (fNIRS) monitors hemodynamic changes associated with neural activity. The study asks whether a single signal encoder can learn features usable across subjects, tasks, devices, and optical probe layouts. The proposed model, MtReNIRS, uses a montage-aware masked autoencoder for self-supervised feature learning. It was pretrained on 12 public datasets spanning 268 subjects and 567 recording hours.
MtReNIRS maps probe layouts to a shared anatomical reference, jointly encoding slow hemodynamic dynamics, chromophore identity and cortical location. Three unseen datasets were used to evaluate cross-subject transfer and within-subject calibration with two to eight labels per class. Cross-subject performance was not uniformly significantly better than fully supervised baselines; few-shot gains varied by dataset and comparator.
Device, task and montage differences co-occurred across the external datasets, so device effects were not isolated. The results support transfer and calibration on the three evaluated datasets, rather than establishing device-invariant performance across settings.