Interpretable Cognitive Ability Prediction: A Comprehensive Gated Graph Transformer Framework for Analyzing...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Graph convolutional deep learning is a powerful tool for analyzing brain functional organization.
• Predicting cognitive ability from neuroimaging data is crucial for understanding brain function.

Purpose of the Study:

• To develop a novel framework using a gated graph transformer (GGT) for predicting cognitive ability from functional connectivity (FC) derived from fMRI.
• To enhance the interpretability of findings by identifying significant biomarkers from functional brain networks.

Main Methods:

• Utilized a gated graph transformer (GGT) model incorporating spatial knowledge and random-walk diffusion.
• Employed learnable structural and positional encodings (LSPE) with a gating mechanism for efficient disentanglement of positional encoding (PE) and graph embeddings.
• Applied an attention mechanism for multi-view node feature embeddings and dynamic weight distribution to identify significant FC biomarkers.

Main Results:

• The proposed GGT framework achieved superior prediction accuracy for cognitive ability compared to existing methods on the PNC and HCP datasets.
• The model demonstrated enhanced explainability, effectively identifying important functional connectivities (FCs) associated with cognitive behaviors.
• The framework successfully integrated spatial priors and diffusion strategies to capture complex brain network relationships.

Conclusions:

• The GGT-based framework offers a promising approach for accurate and interpretable cognitive ability prediction from fMRI data.
• The method advances the application of deep learning in neuroscience for biomarker discovery and understanding brain-behavior relationships.
• This work highlights the potential of graph-based deep learning models in unraveling the neural underpinnings of human cognition.