STNAGNN: Data-driven Spatio-temporal Brain Connectivity beyond FC

TL;DR

This work tackles the problem of learning brain connectivity from noisy task-based fMRI by introducing STNAGNN, a data-driven spatio-temporal GNN that integrates sparse predefined FC with dense, node-level spatio-temporal connections. The model employs two graph convolution layers, a 2D absolute positional encoding across space and time, and a global nodewise attention mechanism across all graph snapshots to capture multi-scale ROI interactions, formalized as a mapping $f:\{{G}_{i}\}_{i=1}^{T}\rightarrow Z$ for $K$-class tasks. Empirical results on Biopoint ASD and HCP task datasets show superior classification performance (e.g., Biopoint Acc up to 79.2% with STNAGNN-GAT; HCP Acc up to 98.1% with STNAGNN-SAGE/GT) compared to SVM and various GNN baselines, with AUC reaching up to 0.999. Interpretability analyses using GNNExplainer reveal dynamic ROI importance patterns, highlighting regions such as the left parietal lobe and right thalamus as potential ASD biomarkers and guiding stimulus design for fMRI experiments.

Abstract

In recent years, graph neural networks (GNNs) have been widely applied in the analysis of brain fMRI, yet defining the connectivity between ROIs remains a challenge in noisy fMRI data. Among all approaches, Functional Connectome (FC) is the most popular method. Computed by the correlation coefficients between ROI time series, FC is a powerful and computationally efficient way to estimate ROI connectivity. However, it is well known for neglecting structural connections and causality in ROI interactions. Also, FC becomes much more noisy in the short spatio-temporal sliding-window subsequences of fMRI. Effective Connectome (EC) is proposed as a directional alternative, but is difficult to accurately estimate. Furthermore, for optimal GNN performance, usually only a small percentage of the strongest connections are selected as sparse edges, resulting in oversimplification of complex brain connections. To tackle these challenges, we propose the Spatio-Temporal Node Attention Graph Neural Network (STNAGNN) as a data-driven alternative that combines sparse predefined FC with dense data-driven spatio-temporal connections, allowing for flexible and spatio-temporal learning of ROI interaction patterns.

Figures (8)

• Figure 1: STNAGNN architecture
• Figure 2: Illustration of connectivity types: a) spatial connectivity between neighboring nodes in one graph, usually computed by GNN convolution; b) temporal connectivity between different time points of one node; c) spatio-temporal connectivity (magenta), allowing information flow between nodes in different graphs. Existing architectures usually consider only spatial connectivity BrainGNN or temporal connectivity nicha-lstm. Some spatio-temporal designs consider both spatial and temporal perspectives gcgrulstmgclstmlrgcnevolve but use a two-step spatial-then-temporal approach. Our proposed STNAGNN jointly considers all spatio-temporal connectivity.
• Figure 3: Interpreted ROI importance from $T=3, 4, 7, 8, 11, 12$. Temporal indices of graph snapshots are marked on the top of each plot. Time under biological motion stimuli are marked by the blue dashed square. Darker regions indicate higher importance. Blue and green circles mark left parietal lobe and right thalamus. The complete plots of 12 sliding window snapshots are shown in \ref{['fig_all_heatmaps']}.
• Figure 4: Preprocessing and graph construction pipeline on the biopoint data
• Figure 5: Visualization of calculated brain connectome using different approaches