Auditory Spatial Attention Detection Based on Feature Disentanglement and Brain Connectivity-Informed Graph Neural Networks

Auditory spatial attention detection (ASAD) aims to determine which speaker in a surround sound field a listener is focusing on from a single-trial electroencephalogram (EEG). Latest studies have represented non-Euclidean structured EEGs by graph-based modeling, but how to better incorporate brain connectivity into EEG graphs remains a great challenge. Moreover, due to inter-subject distribution shifts in EEGs, most existing models perform well only on specific subjects. To address these issues, we propose a new ASAD model. EEG graphs are constructed based on brain effective connectivity, and then mapped into embedding spaces by a graph neural network architecture. Meanwhile, feature disentanglement combined with correlation alignment is utilized to learn subject-invariant EEG patterns relevant to ASAD tasks. Experiments on open datasets demonstrate that in cross-subject scenarios, the proposed model outperforms state-of-the-art models, and the algorithmic complexity is relatively low.