Why fNIRS Still Lacks a General-Purpose Foundation Model, and What One Would Require

Functional near-infrared spectroscopy (fNIRS) is almost absent from the foundation-model literature. On OpenNeuro, the largest open BIDS-structured neuroimaging repository, fNIRS accounts for 1.0% of participants and 1.6% of datasets, against 28.6 times more participants for EEG. We ask why, and what a first fNIRS foundation model would require. The paper is a review and a roadmap: its contribution is an inventory, a synthesis of evidence from adjacent modalities, and a set of requirements, not new model results.We argue that the barrier is not primarily data structure. The fNIRS data standards already require sensor positions to be recorded, so the laborious step of reconstructing where each sensor sat, which dominates the equivalent EEG models, is largely unnecessary here. The barrier is corpus scale, compounded by the absence of any clinical archive comparable to the Temple University Hospital EEG corpus, and by the recency of NIRS-BIDS, published in 2025 against 2019 for EEG. Because no single fNIRS corpus is large enough, aggregation across heterogeneous montages is the most plausible route to pretraining scale, which makes montage-agnostic architecture a precondition rather than an optimization.Reviewing four EEG foundation models that solve montage heterogeneity, we find that the two available ablations of both components rank spatial aggregation above the choice of positional encoding. Several models without a unifying spatial abstraction fall to near-chance under stringent cross-montage evaluation, mainly when encoders are frozen or tested on layouts they were not trained on. In hemodynamic and multimodal work, the one foundation model borrowed from another domain gave no uplift on fNIRS, while joint pretraining with EEG did. In the one available ablation, though, that benefit tracked the mutual information between modalities and, at the larger shot counts, fell below fNIRS-only pretraining at the lowest level tested. From functional magnetic resonance imaging (fMRI) we take a direction fNIRS has not yet followed: recent work there has moved from signal reconstruction toward latent prediction, because the signal is noisy and off-the-shelf evaluation matters. That rationale may be particularly relevant to fNIRS, where some of the systemic physiological confounds, particularly the Mayer wave, overlap the band in which the hemodynamic signal is analyzed. We chart the nine fNIRS-relevant works we located. We identify three especially consequential properties of fNIRS with no electrophysiological counterpart (source–detector geometry, chromophore choice and hemodynamic timescale) and set out eight requirements for a first fNIRS foundation model, on which work is under way in our laboratory.