Deep learning models of cognitive processes constrained by human brain connectomes.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Brain decoding aims to understand cognitive processes from neural activity.
• Traditional methods use region-specific analysis and trial averaging.
• Graph neural networks (GNNs) show promise for whole-brain decoding at fine temporal scales.

Purpose of the Study:

• Investigate the impact of graph properties on GNN-based brain decoding.
• Explore the role of path lengths, node homogeneity, and edge types.
• Understand the inductive bias of connectome priors in deep learning models.

Main Methods:

• Evaluated GNN decoding models on 1200 participants from the Human Connectome Project.
• Tested models under 21 experimental conditions.
• Analyzed the effects of high-order graph convolutions, node homogeneity, and interaction types.

Main Results:

• Optimal decoding achieved by propagating neural dynamics within empirical functional connectomes using high-order graph convolutions.
• The model demonstrated high accuracy and robustness against adversarial graph attacks.
• Learned features were biologically meaningful, and representations resembled task-evoked activation maps.

Conclusions:

• A full-brain integrative model leveraging connectome constraints is crucial for large-scale cognitive decoding.
• Deep GNNs can effectively utilize human connectome information for studying cognition.
• This work provides principles for enhancing brain decoding models and exploring neural substrates of cognition.