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

The characterization of directed interactions among brain regions provides important insights into both normal brain function and the pathological mechanisms of neurological disorders. However, existing methods often do not explicitly integrate electroencephalographic (EEG) features across spatial, spectral, and temporal domains, which may limit their ability to model directed brain network representations. Furthermore, subject-independent shared network structures have not yet been systematically modeled in existing directed connectivity frameworks, despite their importance for improving generalizability and interpretability. In this work, we propose a task-driven framework for learning discriminative directed inter-channel dependency patterns. First, a multi-scale temporal-spectral feature extraction network is employed to learn discriminative EEG representations for each subject. Then, a learnable graph convolutional network (L-GCN) is introduced to adaptively aggregate neighborhood information and refine spatial representations of EEG channels, where directed dependencies are modeled via a learnable asymmetric adjacency matrix. Finally, a joint loss function is designed to guide shared directed network learning by integrating prediction loss, classification loss, group sparsity regularization, and a class-link constraint to encourage consistent connectivity patterns across subjects. As a unified task-driven framework, SS-DBNet was validated on a simulated dataset and two real EEG datasets covering distinct neurological conditions.