SS-DBNet: A Task-Driven Framework for Learning Cross-Subject Shared Sparse Directed Brain Networks from EEG

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Characterizing directed brain interactions is crucial for understanding brain function and neurological disorders.
• Existing methods often fail to integrate multi-domain EEG features (spatial, spectral, temporal), limiting directed network modeling.
• Subject-independent network structures are vital for generalizability but are under-explored in directed connectivity.

Purpose of the Study:

• To develop a task-driven framework for learning discriminative directed inter-channel dependency patterns from EEG.
• To address limitations in existing methods by integrating multi-domain EEG features and modeling shared network structures.
• To improve the generalizability and interpretability of directed brain network representations.

Main Methods:

• A multi-scale temporal-spectral feature extraction network was used for subject-specific EEG representation learning.
• A learnable graph convolutional network (L-GCN) refined spatial representations by aggregating neighborhood information.
• Directed dependencies were modeled using a learnable asymmetric adjacency matrix, with a joint loss function for shared network learning.

Main Results:

• The proposed SS-DBNet framework successfully learned discriminative directed inter-channel dependency patterns.
• The framework demonstrated effective modeling of subject-independent shared network structures.
• Validation on simulated and real EEG datasets confirmed the framework's utility across different neurological conditions.

Conclusions:

• SS-DBNet offers a unified, task-driven approach for analyzing directed brain connectivity from EEG.
• The framework enhances the modeling of directed interactions by integrating multi-domain features and shared network structures.
• This approach holds promise for advancing the study of neurological disorders through improved EEG network analysis.