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

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Graph neural networks (GNNs) are increasingly used for brain fMRI analysis.
• Traditional Functional Connectome (FC) methods struggle with noisy fMRI data and neglect structural/causal information.
• Existing GNN approaches often oversimplify brain connectivity by using sparse connections.

Purpose of the Study:

• To address the challenges in defining robust ROI connectivity in fMRI data for GNNs.
• To propose a novel GNN model that combines predefined and data-driven connectivity.
• To enable flexible and spatio-temporal learning of ROI interaction patterns.

Main Methods:

• Development of the Spatio-Temporal Node Attention Graph Neural Network (STNAGNN).
• Integration of sparse predefined Functional Connectome (FC) with dense, data-driven spatio-temporal connections.
• Utilizing attention mechanisms for flexible learning of ROI interactions.

Main Results:

• The proposed STNAGNN offers a data-driven alternative to traditional connectivity measures.
• It allows for a more comprehensive and flexible representation of brain ROI interactions.
• The model facilitates improved learning of complex spatio-temporal patterns in fMRI data.

Conclusions:

• STNAGNN effectively combines the strengths of predefined and data-driven connectivity for fMRI analysis.
• This approach overcomes limitations of FC and sparse GNN edge selection.
• The model provides a promising direction for advanced GNN applications in neuroscience.