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 brain functional organization without task-based paradigms.
• Understanding the neural basis of intelligence is crucial for cognitive neuroscience.
• Existing models for intelligence prediction from neuroimaging data have limitations.

Purpose of the Study:

• To introduce BrainRGIN, a novel graph neural network architecture for predicting intelligence.
• To leverage rsfMRI-derived static functional network connectivity (sFNC) for intelligence prediction.
• To evaluate BrainRGIN's performance against established methods on a large dataset.

Main Methods:

• Utilized graph neural networks with a novel architecture (BrainRGIN) incorporating clustering-based embedding and graph isomorphism networks.
• Applied TopK pooling and attention-based readout functions for efficient network representation.
• Trained and validated the model on the Adolescent Brain Cognitive Development (ABCD) Dataset using sFNC matrices.

Main Results:

• BrainRGIN demonstrated superior performance in predicting fluid, crystallized, and total intelligence, achieving lower mean squared errors and higher correlation coefficients.
• The model outperformed existing graph architectures and traditional machine learning approaches.
• Identified significant contributions of the middle frontal gyrus to fluid and crystallized intelligence, and a diverse set of regions for total intelligence.

Conclusions:

• BrainRGIN is an effective tool for predicting individual differences in intelligence from rsfMRI data.
• The findings highlight the role of specific brain regions and networks in different facets of intelligence.
• The study provides a novel computational approach for advancing intelligence research using neuroimaging data.