Finer-Grained Dynamic Functional Graph Structure Learning for EEG Sequence Modeling

Electroencephalography (EEG) serves as a critical neuroimaging observation instrument to understand brain dynamics, yet accurately modeling the dynamic functional connectivity inherent in EEG signals remains an open challenge. The existing statistically based coarse-grained associations based on fixed physical distances or static graph structures cannot reflect the spatio-temporal specificity of dynamic functional interactions between brain regions. The secondary computation based on threshold or attention mechanisms to eliminate redundant connections cannot dynamically obtain global structural changes at a fine-grained level, ignoring the rapid dynamic reorganization of brain region functions and lacking a fast response to input correlations. More recently, state space models have been shown to excel at processing long data sequences. However, directly applying such models to EEG data is far from satisfactory due to the lack of functional connectivity modeling between brain regions. In this paper, we propose the Dynamic Functional Graph Structure Learning framework (DFGSL) to capture the dynamic functional connectivity with state space models at a finer-grained level. The proposed DFGSL first constructs dynamic similarity probability maps to reveal information exchange between brain regions. Then, we simulate the entire dynamic evolution of dynamic functional connectivity at a finer-grained level through a selective state space model. By calculating the dynamic similarity probability between selected states, we obtain a compact state representation to describe the context of the dynamic evolution of brain state and reveal potential neural mechanisms. Empirical experiments on three benchmark datasets with different populations, electrode numbers, and brain states show that the proposed DFGSL consistently outperforms state-of-the-art methods, demonstrating strong functional modeling capabilities.