Why collapsing task phases loses signal
Most connectome models turn a task-fMRI scan into one pairwise graph or one functional connectivity (FC) matrix and learn a single representation of the entire session. The paper argues this aggregation attenuates condition-wise signatures: FC reconfigures across task phases, and a pairwise graph only passes messages along two-ROI edges, so it cannot directly aggregate information over a set of regions.
CoHyFuse is the paper's answer to that gap. It is a condition-aware, ROI-centered hypergraph framework: each hyperedge connects multiple regions of interest (ROIs) rather than a pair, and the membership structure is rebuilt for each task condition. The same ROI can therefore belong to different multi-ROI hyperedges in different phases of the same paradigm.
How a condition becomes a hyperedge set
For each condition, the method derives an FC-profile embedding per ROI and selects its top-K_q neighbors; those neighborhoods define the hyperedges, and the resulting memberships form a condition-specific incidence matrix. The neighborhood size K_q is itself condition-specific, so hyperedge scale can differ between encoding, distraction, and recall phases instead of being fixed globally.
Each condition's hypergraph is passed through a hypergraph neural network branch, and the condition embeddings are combined by multi-head attention. That fused condition representation is then integrated with a complementary whole-session FC branch before regression or classification. The design keeps two views: phase-resolved ROI-set structure and session-level connectivity.
Regression gains on two AABC tasks
In the AABC cohort (N=1,074), CoHyFuse reached 7.83±0.10 MAE and 0.439±0.026 R² on FACENAME Fluid Cognition Composite prediction, and 7.52±0.37 MAE and 0.592±0.022 R² on VISMOTOR age prediction. These were the best mean out-of-fold results among the evaluated baselines, which included SVM, MLP, GCN, GAT, BrainGNN, HGNN, HNHN, ED-HNN, ALTER, TA-GAT, and STNAGNN, all run on the same subject-wise folds, preprocessing, parcellation, and validation-based model selection.
The margins are modest but paired bootstrap tests supported them: the FACENAME MAE gain over TA-GAT was 0.34 (95% CI [0.11, 0.57], p=0.004), and the VISMOTOR gain over ALTER was 0.39 (95% CI [0.18, 0.60], p=0.018). The authors note that FACENAME spans a broad adult lifespan cohort, so they report its R² as unadjusted total out-of-fold performance rather than a covariate-adjusted effect size. A confound check found an age/sex/FD-only model reached R²=0.25 versus 0.439 for CoHyFuse, and the FACENAME association remained significant after fold-wise adjustment (partial r=0.34, p<0.001).
Which components carry the gain
Ablations on FACENAME separate the contributions. Removing all condition branches caused the largest drop (ΔMAE +0.48, ΔR² −0.065), and removing the whole-session FC branch also hurt (ΔMAE +0.37, ΔR² −0.064). Among single-condition removals, dropping the DIST (distraction) condition produced the largest MAE increase (+0.34), ahead of ENC (+0.27) and REC (+0.25).
Within-framework controls test whether the hypergraph operator itself matters. CoHyFuse-MLP, which uses vectorized condition-wise FC matrices with an MLP predictor, reached 8.11 MAE and 0.404 R²; swapping in the incidence-based HGNN improved this to 7.83 and 0.439. CoHyFuse-GCN and CoHyFuse-GAT kept the same condition-wise inputs, K_q search space, fusion structure, and whole-session branch but used matched pairwise propagation, and both scored worse than the hypergraph version. The authors read this as support for condition-indexed incidence construction and hyperedge-based ROI-set aggregation adding value beyond simple concatenation or pairwise propagation.
Condition-specific neighborhood sizes matter
The K_q search compares shared and condition-specific settings on FACENAME. Shared-K schemes such as (5,5,5), (10,10,10), and (50,50,50) produced MAE between 7.98 and 8.05 with R² around 0.405 to 0.408. Condition-specific settings were better: (50,45,5) gave 7.83 MAE and 0.439 R², with (45,50,5) and (50,50,10) close behind at 7.85 and 7.84 MAE.
This is a within-paper comparison, not a claim that these exact values transfer to other tasks or cohorts. The paper's own limitation section notes that condition-wise FC estimation needs long condition windows and that condition-wise K_q increases hyperparameter search cost.
An auxiliary classification check and the distraction driver
On the auxiliary CMI-HBN ADHD classification benchmark (N=223), CoHyFuse obtained 72.0±2.1% macro-AUC and 74.2±2.9% accuracy. This is a separate task and cohort from the AABC regression results, so the two sets of numbers should not be read as one combined performance claim.
Occlusion analysis identified the Distraction condition as the primary driver of model prediction, pointing toward Salience/Ventral Attention Network (SAN)–FrontoParietal Network (FPN) and within-SAN hyperedge-defined ROI-set motifs as candidate model-relevant patterns. The paper also reports that the Distraction condition yielded the strongest association between hyperedge-guided subnetwork strength and FCC (r=0.36), with dominant network-pair contributions in SAN–SAN and SAN–FPN connectivity. These are described as candidate patterns and associations, not established causal mechanisms.
Where the evidence stops
The predictive claims are bounded by the evaluated baselines, the two public datasets, and the specific preprocessing, parcellation, and validation protocol shared across models. The authors list further constraints: stable condition-wise FC requires long condition windows, and condition-wise K_q adds hyperparameter search cost. They propose learnable neighborhood selection, temporally adaptive condition modeling, and broader validation across tasks, cohorts, atlases, and connectivity estimators as future work.
The interpretability results are occlusion-based candidate patterns rather than validated neural mechanisms, and the ADHD benchmark is described as auxiliary. Readers should treat the FACENAME and VISMOTOR gains as evidence that state-resolved ROI-set structure adds complementary information to whole-session FC within this setup, not as a general guarantee across paradigms.
The priority this paper sets
The paper's most decision-relevant point is a comparison, not a single score: condition-wise incidence construction plus hyperedge aggregation outperformed both condition-wise FC concatenation (CoHyFuse-MLP) and matched pairwise propagation (CoHyFuse-GCN/GAT) under the same inputs and fusion structure. If you are choosing between modeling task phases as separate pairwise graphs and modeling them as condition-specific ROI-set hypergraphs, this paper's evidence favors the latter within its evaluated cohorts.
The second priority is the distraction condition. It produced the largest single-condition ablation effect and the strongest reported association with FCC, so any follow-up that reuses this framework should check whether the distraction-phase hyperedges are carrying the signal rather than assuming all phases contribute equally.
Sources used in this answer
CoHyFuse: Condition-wise Hypergraph Fusion with Global Connectome in Task-fMRI
CoHyFuse builds condition-specific ROI hypergraphs and fuses them with a whole-session FC branch, achieving the best mean out-of-fold performance among evaluated baselines on AABC FACENAME (7.83 MAE, 0.439 R²) and VISMOTOR (7.52 MAE, 0.592 R²), plus 72.0% macro-AUC on auxiliary ADHD classification.