MGMAP-Net: A Multi-Graph Modality-Aware Network for Enhanced Fluid Intelligence Prediction Using Multimodal Brain...

Area of Science:

• Neuroscience
• Cognitive Science
• Data Science

Background:

• Fluid intelligence prediction using neuroimaging is vital for understanding cognition.
• Multimodal brain connectivity (functional connectivity - FC, structural connectivity - SC) offers rich data.
• Existing methods struggle with feature redundancy when fusing FC and SC.

Purpose of the Study:

• To develop a novel network, MGMAPNet, for enhanced fluid intelligence prediction.
• To effectively disentangle modality-specific and shared features from multimodal brain connectivity data.
• To address information redundancy in fused FC and SC data.

Main Methods:

• Proposed Multi-Graph Modality-Aware Predictive Network (MGMAPNet).
• Incorporated private and shared feature extraction modules for FC and SC data.
• Designed specialized loss functions for feature separation.

Main Results:

• MGMAPNet outperformed unimodal and existing multimodal approaches on the UK Biobank Dataset.
• t-SNE visualization confirmed effective isolation of private and shared features.
• Demonstrated improved prediction of fluid intelligence by leveraging complementary connectivity information.

Conclusions:

• MGMAPNet successfully disentangles modality-specific and shared information from FC and SC.
• The proposed method enhances fluid intelligence prediction by reducing redundancy and utilizing complementary connectivity data.
• This approach offers a promising direction for multimodal neuroimaging data analysis in cognitive neuroscience.