Ensemble Manifold Regularized Multi-Modal Graph Convolutional Network for Cognitive Ability Prediction.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning

Background:

• Multi-modal functional magnetic resonance imaging (fMRI) offers potential for predicting individual behavioral and cognitive traits via brain connectivity networks.
• Current data-fusion models face limitations in fully leveraging complementary information from multi-modal fMRI.

Purpose of the Study:

• To develop an interpretable multi-modal graph convolutional network (MGCN) model for enhanced cognition prediction.
• To integrate fMRI time series and functional connectivity (FC) data for a comprehensive analysis of brain networks.
• To identify significant cognition-related biomarkers through model interpretability techniques.

Main Methods:

• Proposed an interpretable multi-modal graph convolutional network (MGCN) model incorporating fMRI time series and functional connectivity (FC).
• Employed manifold-based regularization to capture subject relationships within and between modalities.
• Introduced gradient-weighted regression activation mapping (Grad-RAM) and edge mask learning for model interpretation and biomarker identification.

Main Results:

• Validated the MGCN model on the Philadelphia Neurodevelopmental Cohort for predicting Wide Range Achievement Test (WRAT) scores.
• Achieved superior predictive performance compared to single-modality GCN and other competing methods.
• Cross-validated identified biomarkers using multiple approaches, confirming their significance.

Conclusions:

• Developed a novel interpretable graph deep learning framework for cognition prediction, addressing limitations of existing data-fusion models.
• Demonstrated the efficacy of MGCN in analyzing multi-modal fMRI data.
• Highlighted the potential of MGCN for discovering significant biomarkers in human brain studies.