Rethinking Functional Brain Connectome Analysis: Do Graph Deep Learning Models Help

Graph deep learning models, a class of AI-driven approaches employing a message aggregation mechanism, have gained popularity for analyzing the functional brain connectome in neuroimaging. However, their actual effectiveness remains unclear. In this study, we re-examine graph deep learning versus classical machine learning models based on four large-scale neuroimaging studies. Surprisingly, we find that the message aggregation mechanism, a hallmark of graph deep learning models, does not help with predictive performance as typically assumed, but rather consistently degrades it. To address this issue, we propose a hybrid model combining a linear model with a graph attention network through dual pathways, achieving robust predictions and enhanced interpretability by revealing both localized and global neural connectivity patterns. Our findings urge caution in adopting complex deep learning models for functional brain connectome analysis, emphasizing the need for rigorous experimental designs to establish tangible performance gains and perhaps more importantly, to pursue improvements in model interpretability.

💡 Research Summary

This paper conducts a systematic, large‑scale evaluation of graph deep learning (GDL) methods for functional brain connectome analysis, directly comparing them with classical machine‑learning (ML) approaches across four publicly available fMRI datasets: ABIDE (autism classification), PNC (gender classification), HCP (fluid intelligence regression), and ABCD (fluid intelligence regression). The authors first benchmark a broad set of models grouped into three categories: (1) GDL models that explicitly treat the functional connectivity matrix as a graph and perform message‑passing (e.g., GCN, GAT, GIN, GraphSAGE, NeuroGraph, BrainNetTF); (2) non‑graph deep‑learning models such as multilayer perceptrons (MLPs) that ignore the graph structure; and (3) traditional ML models (logistic regression, ElasticNet, kernel ridge regression, SVM/SVR) that use the raw connectivity values as features.

The key empirical finding is that GDL models do not outperform, and often underperform, the simpler baselines. In binary classification tasks (ABIDE, PNC) logistic regression, ElasticNet, and MLPs achieve area‑under‑the‑curve (AUC) scores comparable to or higher than the best GDL methods (e.g., NeuroGraph, BrainNetTF). In the regression tasks (HCP, ABCD) classical models again match or exceed the performance of graph‑based networks, with Pearson correlation coefficients that are statistically indistinguishable from the top GDL results.

To probe why GDL underperforms, the authors manipulate graph density by retaining only the top K % of edges (positive correlations for most models, both signs for BrainGB and BrainNetTF). When K = 0, the graph contains no edges, effectively disabling any message aggregation. Across all GDL variants that rely on aggregation (GCN, GAT, GIN, GraphSAGE, BrainGB, BrainGNN), performance decreases monotonically as density increases, indicating that more aggregation harms prediction. Models that incorporate residual or skip connections (NeuroGraph, BrainNetTF) are more robust to density changes, but when the residual paths are removed, they exhibit the same degradation pattern. This suggests that the aggregation operation itself—not the overall architecture—is detrimental in this domain.

Further analysis quantifies the “smoothness” of the learned node embeddings by computing the mean cosine similarity of the final‑layer node vectors across density levels. Higher similarity (i.e., smoother, more homogeneous embeddings) correlates strongly and negatively with predictive accuracy (Pearson r < 0, p < 0.05) across all datasets and models. The authors interpret this as an over‑smoothing phenomenon: message passing drives node representations toward a common subspace, erasing discriminative information needed for downstream prediction.

The paper also examines the intrinsic dimensionality of the input node features (the rows of the functional connectivity matrix). Using an effective‑rank metric, they find that only 7–38 % of the total feature dimensions carry independent variance, confirming that the raw connectivity profiles are highly redundant and low‑rank. Because the input already encodes global information, additional graph‑based smoothing amplifies redundancy and further diminishes useful signal.

Motivated by these observations, the authors propose a dual‑pathway hybrid model that combines a linear pathway (global connectivity captured by a simple linear regression on the full matrix) with a graph‑attention pathway (local structure captured by a Graph Attention Network using BOLD time‑series as node features). The two pathways are fused before the final prediction layer. This design leverages the interpretability and global pattern detection of linear models while still allowing the GAT branch to highlight localized subnetworks and salient ROIs. Empirically, the hybrid model achieves performance on par with or slightly better than the best baseline on all four datasets, and it provides richer interpretability: the linear branch emphasizes whole‑brain connectivity efficiency, whereas the GAT branch reveals modular subnetworks that align with known neurocognitive systems.

In conclusion, the study challenges the prevailing assumption that graph‑based deep learning automatically yields superior predictive power for functional connectome data. The message‑aggregation mechanism, a hallmark of GDL, can be counterproductive when node features are already globally informative and low‑rank. The authors advocate for rigorous benchmarking against strong classical baselines, careful consideration of feature redundancy, and a renewed focus on model interpretability rather than raw predictive accuracy. Future work should explore alternative node representations (e.g., frequency‑domain features, structural connectivity) and graph constructions that avoid over‑smoothing, as well as hybrid architectures that balance global and local information without unnecessary complexity.