FroHyNet: frozen temporal priors and trainable hypergraph fusion for brain time-series classification

Functional brain time-series classification is constrained by small clinical cohorts. Handcrafted features are sample-efficient but discard temporal detail; end-to-end networks must estimate many temporal parameters from limited labels; and frozen time-series foundation models reduce that estimation burden but typically encode each channel independently, leaving coordinated activity across brain regions unmodeled.

We introduce the Frozen Temporal-Hypergraph Network (FroHyNet), an asymmetric parallel framework that preserves channel-wise temporal representations from a frozen MantisV2 encoder while a trainable spatio-temporal hypergraph branch learns topology-aware spatial corrections. Bottleneck-mix Top-K (BM-TopK) then residual-injects these corrections only along selected channel and latent-feature axes, treating the spatial stream as a selective update rather than a dense overwrite of the temporal representation. We conduct five-fold cross-validation across two functional near-infrared spectroscopy (fNIRS) and three functional magnetic resonance imaging (fMRI) datasets.

FroHyNet achieves the best average accuracy and area under the receiver-operating-characteristic curve (AUC) among methods evaluated on all five datasets, while keeping the trainable-parameter count well below those of from-scratch temporal networks. Gains are concentrated on fNIRS, with more modest improvements on the smaller fMRI cohorts. Branch ablations indicate dataset-dependent complementarity between the two streams, and fusion ablations on HC-D favor BM-TopK over the evaluated dense and recurrent alternatives.