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

TL;DR

The paper questions the assumed superiority of graph deep learning for functional brain connectome analysis by re-evaluating four large-scale datasets and showing that message aggregation often harms predictive performance. It introduces a dual-pathway LM-GAT framework that combines a linear path on flattened connectivity with a graph attention path built on BOLD time-series embeddings, achieving robust accuracy and richer interpretability. The results challenge the default preference for aggregation-heavy GDLs and highlight the importance of rigorous baselining and domain-informed interpretability for neuroscience applications. The work advocates prioritizing interpretability and network-level insights over mere prediction accuracy, and points to future directions including integration of structural connectivity and end-to-end graph construction.

Abstract

Functional brain connectome is crucial for deciphering the neural mechanisms underlying cognitive functions and neurological disorders. Graph deep learning models have recently gained tremendous popularity in this field. However, their actual effectiveness in modeling the brain connectome remains unclear. In this study, we re-examine graph deep learning models based on four large-scale neuroimaging studies encompassing diverse cognitive and clinical outcomes. 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.

Figures (6)

• Figure 1: Performance comparison across different baselines on four datasets. For the ABIDE and PNC datasets, the prediction tasks are binary classification for autism disease and biological gender, respectively, evaluated using AUROC scores. For the HCP and ABCD datasets, the prediction task is fluid intelligence score prediction, a regression task evaluated using the Pearson correlation between predicted and actual values. The baselines are grouped into three categories, distinguished by color: green for graph deep learning, blue for non-graph deep learning, and red for classical machine learning models. Models highlighted with bold borders represent those with the best predictive performance among all the baselines. There is no significant difference ($p > 0.05$) in performance among these highlighted models, showing that simpler models such as Logistic Regression, MLP, ElasticNet, and Kernel Ridge Regression perform comparably to the most advanced graph deep learning models. The significance bars marked with asterisks ($*$) indicate that the performance of non-graph deep learning models is significantly better than that of the graph deep learning models displayed to the left ($p < 0.05$). Hollow circles and numbers in each box indicate the average results from 10 independent runs.
• Figure 2: Performance of graph deep learning models with varying graph densities. To investigate the effect of aggregation in graph deep learning models on brain connectome analysis, we adjust graph density by retaining the top $K\%$ edges, thereby regulating the extent of aggregation between ROIs. Lower densities imply reduced aggregation. Specifically, when $K = 0$, no edges exist between ROIs, meaning node features are transformed independently without any aggregation. Each data point here represents the average result from 10 independent runs with the corresponding graph density. The left panel of each subfigure illustrates prediction accuracy trends of GCN, GAT, GIN, GraphSage, BrainGB, and BrainGNN, showing a consistent decline in performance as graph density increases. This indicates that message aggregation does not improve prediction outcomes but rather degrades performance. In contrast, the right panel of each subfigure highlights the performance of BrainNetTF and NeuroGraph, which remain stable across different graph densities due to the inclusion of residual connections. When residual connections are removed in their respective variants (BrainNetTF w/o Res and NeuroGraph w/o Res), performance declines similarly to models on the left, restating that aggregation alone harms prediction accuracy in brain connectome analysis.
• Figure 3: Important connections identified by two pathways in the ABCD dataset.(a)-(b) Heatmap of important connections for GAT and LM pathways. The attention scores from the GAT pathway are rescaled to [0,1] using min-max normalization, while weights from the LM pathway are scaled to [-1,1] by normalizing with the maximum absolute value of weights. The heatmaps show that the GAT pathway highlights within-system connections, particularly in the Default Mode (DMN), Somatomotor (SM) and Central Executive (CE) networks, aligning with their roles in cognitive processes. In contrast, the LM pathway reveals more uniformly distributed connections, which reflects the long-range network integration essential for cognitive flexibility and large-scale communication. (c)-(d) Chordplots of the top $0.1\%$ most significant connections (65 in total) for GAT and LM pathways. The GAT pathway reveals symmetrical connections between homologous ROIs across hemispheres, highlighting its focus on both local and bilateral brain organization. In contrast, the LM pathway shows a more diffusive pattern with both positive and negative weights distributed across networks, accounting for both facilitative and inhibitory interactions across ROIs, necessary for maintaining cognitive flexibility.
• Figure 4: Visualization of the top 20 salient ROIs identified by two pathways in the ABCD dataset. Color intensity represents the importance of each ROI, with darker shades indicating lower importance and brighter shades indicating higher importance. For the GAT pathway, node importance is calculated by summing the attention scores across all its connections, corresponding to each row in the attention heatmap. The identified salient ROIs are primarily concentrated in the Central Executive Network (CE) and Default Mode Network (DMN), with key regions located in the frontal and temporal lobe. For the LM pathway, node importance is calculated separately for positive and negative weights. Salient ROIs identified from the LM pathway’s positive weights are concentrated in the Ventral Salience Network (VS) and Somatomotor Network (SM), predominantly in the temporal lobe, while negative weights highlight regions within the DMN, particularly in the frontal and parietal lobes.
• Figure 5: Graph properties of highlighted subgraphs derived from the top 5% of edges from two pathways in the ABCD dataset.(a)-(b) Node degree distribution of subgraphs for GAT and LM pathways. The GAT-induced subgraph exhibits a skewed distribution with a few highly connected hubs, indicative of a scale-free network, often associated with the brain's robustness and resilience to localized failures. The LM-induced subgraph exhibits a broader distribution with a higher concentration of nodes having intermediate degrees, reflecting the distributed processing across multiple brain regions. (c) Comparison between other numerical graph properties. The GAT-induced subgraph demonstrates higher clustering coefficients and modularity, indicating tightly-knit communities and clear modular divisions. Additionally, the small-worldness metric is significantly higher, suggesting an optimal balance between local specialization and global integration. In contrast, the LM-induced subgraph features shorter average path lengths and higher global efficiency, reflecting a network structure optimized for rapid information transfer. Besides, the higher assortativity coefficient in the LM subgraph points to stronger connections between functionally similar brain regions.