Functional Magnetic Resonance Imaging (fMRI) is an imaging technique widely used to study human brain activity. fMRI signals in areas across the brain transiently synchronise and desynchronise their activity in a highly structured manner, even when an individual is at rest. These functional connectivity dynamics may be related to behaviour and neuropsychiatric disease. To model these dynamics, temporal brain connectivity representations are essential, as they reflect evolving interactions between brain regions and provide insight into transient neural states and network reconfigurations. However, conventional graph neural networks (GNNs) often struggle to capture long-range temporal dependencies in dynamic fMRI data. To address this challenge, we propose BrainATCL, an unsupervised, nonparametric framework for adaptive temporal brain connectivity learning, enabling functional link prediction and age estimation. Our method dynamically adjusts the lookback window for each snapshot based on the rate of newly added edges. Graph sequences are subsequently encoded using a GINE-Mamba2 backbone to learn spatial-temporal representations of dynamic functional connectivity in resting-state fMRI data of 1,000 participants from the Human Connectome Project. To further improve spatial modeling, we incorporate brain structure and function-informed edge attributes, i.e., the left/right hemispheric identity and subnetwork membership of brain regions, enabling the model to capture biologically meaningful topological patterns. We evaluate our BrainATCL on two tasks: functional link prediction and age estimation. The experimental results demonstrate superior performance and strong generalization, including in cross-session prediction scenarios.
Figures & tables
Figure 1: Pipeline for dFC construction and adaptive graph sequence generation. (a) The original 4D fMRI BOLD signals X∈RN×P×Q×T (with N , P , Q , T denoting ROIs, subjects, imaging sessions, and time points). (b) We construct dFC Gp,q as a graph sequence of length K for subject p and session q , all subjects’ dFC set is denoted as G . (c) For training the models, we designed adaptive lookback strategy as a function of the temporal novelty index (Eq. 1 ) as a heuristic to determine temporal context Dkp,q dynamically for each graph snapshot at time step k , see detailed computation of adaptive lookback ℓkp,q in Section 3.3.1 .
Figure 2: Illustration of our proposed BrainATCL framework for adaptive temporal brain connectivity embedding and two downstream tasks. (a) The BrainATCL is trained by minimizing the contrastive triplet loss. It combines a GINEConv-based GNN to encode structural information with edge attributes, followed by a Mamba-2 model for temporal aggregation across the graph sequence. The output of Mamba-2 is fed to linear layers to represent each node as a multivariate Gaussian distribution. (b) After pre-training the BrainATCL, the input graph sequence is processed and its node embeddings are used for link prediction classifier and age prediction regressor. These downstream task models are trained separately using binary cross-entropy (BCE) and mean-squared error (MSE) losses.
Within-session
Model Variant
MAP ↑
MRR ↑
AUC ↑
PR-AUC ↑
F1 ↑
w/o Edge Features
0.7313 ± 0.0024
0.6061 ± 0.0015
0.9728 ± 0.0004
0.7143 ± 0.0016
0.6830 ± 0.0011
w/ Edge Features
0.7681 ± 0.0012
0.6265 ± 0.0009
0.9746 ± 0.0001
0.7468 ± 0.0013
0.7050 ± 0.0008
Table 1: Ablation study on the effect of edge features (including network and hemisphere labels of the two endpoint nodes) for temporal link prediction under within-session and cross-session settings.
Within-session
Model Variant
MAP ↑
MRR ↑
AUC ↑
PR-AUC ↑
F1 ↑
VGAE Reconstruction
0.7054 ± 0.0019
0.5982 ± 0.0021
0.9701 ± 0.0004
0.6752 ± 0.0012
0.6488 ± 0.0011
Contrastive (BrainATCL)
0.7681 ± 0.0012
0.6265 ± 0.0009
0.9746 ± 0.0001
0.7468 ± 0.0013
0.7050 ± 0.0008
Table 2: Ablation study on training objectives by replacing the proposed contrastive loss with a VGAE-style reconstruction objective for temporal link prediction under the within-session setting.
Within-session
Model
MAP ↑
MRR ↑
AUC ↑
PR-AUC ↑
F1 ↑
MAE ↓
MSE ↓
GCN
0.2838 (0.0018)
0.2226 (0.0009)
0.9101 (0.0000)
0.2679 (0.0001)
0.3557 (0.0001)
3.09 (0.16)
13.77 (1.89)
GraphSAGE
0.3245 (0.0234)
0.2576 (0.0534)
0.9279 (0.0000)
0.3017 (0.0001)
0.4054 (0.0002)
3.29 (0.58)
16.58 (5.42)
GAT
0.3601 (0.0027)
0.2045 (0.0038)
0.9112 (0.0000)
0.3408 (0.0001)
0.4233 (0.0002)
3.16 (0.49)
15.78 (2.45)
GIN
0.3564 (0.0124)
0.2308 (0.0210)
0.9178 (0.0000)
0.3326 (0.0001)
0.4150 (0.0000)
3.15 (0.07)
14.70 (1.23)
GCN-GRU
0.2991 (0.0055)
0.2351 (0.0037)
0.9172 (0.0000)
0.2785 (0.0000)
0.3622 (0.0000)
3.08 (0.54)
14.68 (0.77)
Table 3: Benchmarking BrainATCL against multiple baselines for temporal link and age prediction under within-session and cross-session settings. MAP, MRR, AUCROC, PR-AUC, and F1 are used for link prediction, while MAE and MSE are used for age prediction. Reported values are means with standard deviations shown in parentheses. All comparisons are conducted under the same experimental configuration (stride=10, window=100, embedding=128, threshold=0.5).
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Temporal Edge Appearance (TEA) plots for subject 10, session 1 under different threshold values. For each time point, the blue bars indicate the number of repeated edges (i.e., edges that have appeared in any previous time point), while the orange bars represent newly appeared edges. The red line shows the novelty index at each time point, quantifying the proportion of new edges relative to total edges. The average novelty across the time series is reported in the title of each subplot.
Figure 4: The Top panel shows ablation study on the effect of hyperparameters (threshold, embedding size and stride) on within-session (blue) and cross-session (red) temporal link prediction performance in terms of MAP (solid lines) and MRR (dashed lines). Solid/dashed lines denote the mean performance, while the shaded areas correspond to the standard deviation across five runs. The bottom panel shows MAP and MRR scores across different lookback strategies and stride sizes for the link prediction task. The red dashed line shows the average novelty score corresponding to each stride setting. The x-axis denotes the stride size ( s ), with the number of temporal graphs per subject indicated in parentheses.
Figure 5: Distribution of cross-subject cosine distances between time-averaged ROI embeddings of the two subjects with the largest age difference (15 years), grouped by major functional networks. All subnetworks are defined based on the Schaefer parcellation Schaefer et al. (2018)
Cross-site out-of-distribution (OOD) generalization in resting-state functional magnetic resonance imaging (rs-fMRI) often relies on learning task-discriminative representations from full-scan functional connectivity (FC) graphs and promoting invariance across source sites. However, FC graphs are estimated from finite, temporally correlated blood-oxygen-level-dependent (BOLD) sequences. Cross-site agreement therefore does not necessarily imply that predictive evidence remains supported under FC re-estimation within the same scan. In this paper, we propose Brain Network Re-estimation-Informed OOD Learning (BRIO), a framework that uses within-scan FC re-estimation to guide cross-site alignment. BRIO maps fullscan graphs and their re-estimates into consistently indexed connectome factors, enabling comparisons of their predictive contributions. It assesses re-estimation support from changes in these contributions relative to within-class subject variability and class separation. For each source-site pair and class, this task-calibrated support from both sites is combined with predictive relevance to form pairwise qualifications, which determine relative factor weights and overall alignment strength. Leave-one-site-out experiments on four real-world datasets (ABIDE, REST-metaMDD, SRPBS, and ABCD) show that BRIO consistently outperforms competitive baselines, with relative improvements of up to 3.8% in accuracy. These gains also persist under an alternative brain parcellation on ABIDE.
Yingxu Wang, Kunyu Zhang, Yanwu Yang3 +4
Mohamed bin Zayed University of Artificial Intelligence · Zhengzhou University · University Hospital Tübingen +3
Functional magnetic resonance imaging (fMRI) is a widely used technique for studying the brain. Recent methods that utilize graph neural networks (GNNs) for analysis of brain functional connectivity have shown great potential for the classification of brain disorders, such as Alzheimer's disease (AD). However, these methods often assume a preset number of functional modules across all subjects, which overlooks inter-subject variability. In addition, the discovered modules are rarely used to directly guide the learned connectivity patterns. Here, to address these issues, we propose a Meta Probabilistic Pooling GNN (MPP-GNN). We frame the model's task as a coupled, bilevel optimization that performs adaptive graph partitioning hierarchically to discover subject-specific modules and then uses the discovered brain modules as an explicit prior to guide edge refinement and representation learning. We validate MPP-GNN on two public datasets for AD classification, achieving the highest AUC in comparison to established baselines for both datasets. Furthermore, our analysis demonstrates that MPP-GNN shows significant alignment with the canonical functional-network organization defined by the Yeo brain atlas and reveals a network-level dedifferentiation pattern for AD.
Modeling resting-state functional magnetic resonance imaging (rs-fMRI) data is crucial for understanding brain-wide neural activity. However, traditional methods struggle to capture complex temporal dynamics over long horizons, to account for the brain's anatomical spatial structure, and to model high-dimensional ambient signals that lie on a low-dimensional intrinsic subspace. We propose FAST-Brain, a unified flow-aligned spatio-temporal surrogate brain model that addresses all three challenges. At its core is a flow-aligned generative framework that directly predicts the clean blood-oxygen-level-dependent (BOLD) signal, paired with a graph convolutional network that captures spatial structural constraints and a Transformer that models long-range temporal dependencies. Theoretically, we show that under a low-dimensional subspace assumption, the approximation error of our model scales with the intrinsic dimension rather than the ambient dimension, which justifies our direct modeling of the BOLD signal. Extensive experiments on synthetic and Human Connectome Project datasets demonstrate that FAST-Brain achieves state-of-the-art performance in recovering functional connectivity, effective connectivity, and the implicit low-dimensional signal subspace.
Shucheng Liu, Changchun Shi, Kai Zhang +1
University of North Carolina at Chapel Hill · London School of Economics and Political Science