Brain Networks and Intelligence: A Graph Neural Network Based Approach to Resting State fMRI Data

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Resting-state functional magnetic resonance imaging (rsfMRI) enables the study of intrinsic brain functional organization.
• Predicting cognitive abilities like intelligence from brain activity is a key challenge in neuroscience.

Purpose of the Study:

• To introduce BrainRGIN, a novel graph neural network architecture for intelligence prediction.
• To evaluate BrainRGIN's performance on predicting fluid, crystallized, and total intelligence using rsfMRI data.

Main Methods:

• Utilized graph neural networks (GNNs) on static functional network connectivity matrices derived from rsfMRI.
• Incorporated clustering-based embedding and graph isomorphism networks within the GNN architecture.
• Employed TopK pooling and attention-based readout functions for enhanced network analysis.

Main Results:

• BrainRGIN demonstrated superior performance in predicting intelligence compared to existing graph architectures and traditional machine learning models.
• Achieved lower mean squared errors and higher correlation scores across all intelligence prediction tasks.
• Identified significant contributions of the middle frontal gyrus to fluid and crystallized intelligence.

Conclusions:

• The BrainRGIN model effectively predicts individual differences in intelligence from rsfMRI data.
• Brain network connectivity patterns hold valuable information for understanding cognitive abilities.
• The findings highlight the complex neural underpinnings of total intelligence.