FroHyNet: frozen temporal priors and trainable hypergraph fusion for brain time-series classification.
- 1College of Biomedical Engineering, Sichuan University, Chengdu, China.
- 2Mental Health Center, West China Hospital, Sichuan University and Sichuan Clinical Medical Research Center for Mental Disorders, Chengdu, China.
Abstract
Background
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.Methods
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.Results
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.
Related
- Spatial-temporal graph convolutional attentional network for motor imagery EEG classification
- Comparing Temporal Averaging and Transformer Encoders for Video-Based Cattle Behavior Classification Under Severe Class Imbalance.
- A trainable language model with potential to modulate translation rates in non-model organisms by generating upstream untranslated region sequence libraries
- Modular core network constructed from Escherichia coli transcriptome datasets using a hypergraph-based pan-network approach
- Relation-Aware Graph-Hypergraph Networks for Antibody Sequence-Structure Co-Design
BibTeX
@article{gao2026frohynet,
title = {{FroHyNet: frozen temporal priors and trainable hypergraph fusion for brain time-series classification.}},
author = {Mingyao Gao and Ling He and Wen Li and Yajing Meng and Qing Li and Changjian Qiu and Jing Zhang},
journal = {Frontiers in medicine},
year = {2026},
month = oct,
doi = {10.3389/fmed.2026.1966334},
url = {https://europepmc.org/article/MED/42908832},
note = {CC BY},
}
RIS
TY - JOUR
TI - FroHyNet: frozen temporal priors and trainable hypergraph fusion for brain time-series classification.
AU - Mingyao Gao
AU - Ling He
AU - Wen Li
AU - Yajing Meng
AU - Qing Li
AU - Changjian Qiu
AU - Jing Zhang
PY - 2026
DA - 2026/10/07
JO - Frontiers in medicine
DO - 10.3389/fmed.2026.1966334
UR - https://europepmc.org/article/MED/42908832
AB - Background 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.Methods 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.Results 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.
SN - 2296-858X
ER -