A novel neural network model with SAGPooling graph decodes changes in cognitive trajectories.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Predicting neurodegenerative disease progression and identifying key brain regions is crucial.
• Self-Attention Graph Pooling (SAGPooling) can improve prediction accuracy.
• Existing methods may not fully capture complex brain network dynamics.

Purpose of the Study:

• To propose an atlas-constrained spatial-informational SAGPooling graph neural network (ACSI-SGNN) model.
• To enable precise prediction of cognitive resilience.
• To identify brain regions associated with resilience in neuroimaging.

Main Methods:

• Utilized frequency-domain Granger causality to calculate functional states and information flow in brain networks.
• Integrated functional data with spatial encoding to construct a brain network model.
• Employed SAGPooling within a graph neural network (GNN) framework with ROI selection for feature aggregation.

Main Results:

• The ACSI-SGNN model achieved high accuracy (e.g., 87.50% ACC on ADNI dataset) for cognitive resilience classification.
• The model outperformed other methods across multiple evaluation metrics.
• Identified key brain regions critical for cognitive resilience, consistent with prior neuroimaging studies.

Conclusions:

• The ACSI-SGNN model is effective for identifying brain aging rates and neurodegenerative disease progression.
• The proposed method accurately predicts cognitive resilience and highlights functionally relevant brain regions.
• This approach advances neuroimaging analysis for understanding brain health and disease.