Learning dynamic graph embeddings for accurate detection of cognitive state changes in functional brain networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Brain functions and cognitive states are not static but dynamically change, even during rest.
• Detecting cognitive state changes from blood-oxygen-level-dependent (BOLD) signals is challenging due to subtle signal variations and reliance on experimental design.
• Existing methods struggle to accurately identify dynamic cognitive shifts without prior task knowledge.

Purpose of the Study:

• To develop a novel dynamic graph learning approach for improved detection of cognitive states from brain activity.
• To create subject-specific dynamic graph embeddings that capture evolving brain network patterns.
• To disentangle cognitive events more accurately than using raw BOLD signals.

Main Methods:

• A dynamic graph learning approach was employed to generate subject-specific dynamic graph embeddings.
• Representation learning projected BOLD signals into a latent vertex-temporal domain.
• The learned domain integrated harmonic waves for functional connectivity topology and Fourier bases for temporal dynamics.

Main Results:

• The dynamic graph embeddings accurately disentangled cognitive events.
• The method demonstrated higher accuracy in detecting multiple cognitive states compared to state-of-the-art approaches.
• Validation was performed on simulated data and working memory task-based fMRI datasets.

Conclusions:

• Dynamic graph embeddings offer a powerful new methodology for investigating brain activity and cognitive states.
• This approach provides a robust framework for analyzing self-organized functional fluctuation patterns in relation to cognitive status.
• The proposed method enhances the ability to detect and differentiate cognitive states using fMRI data.