Causality-Driven Convolutional Manifold Attention Network for Electroencephalogram Signal Decoding

Deep learning-based methods have achieved remarkable success in brain-computer interfaces (BCIs). However, its inherent assumption of independent and identically distributed (i.i.d.) data renders it vulnerable to out-of-distribution (OOD) scenarios. To address this limitation, the present study proposed a causality-driven convolutional manifold attention network (CD-CMAN) that learned invariant representations from electroencephalogram (EEG) signals to enhance OOD generalization. The framework began with a spatiotemporal convolution module to extract rich temporal and spatial features. Guided by the defined structural causal model and leveraging the strengths of Riemannian geometry and deep learning, dual latent encoders with manifold attention units were crafted to explicitly separate spatiotemporal feature maps into semantic and variation latent factors. A reconstruction module with a dedicated loss was implemented to ensure these factors retaining informative, while the Hilbert-Schmidt independence criterion (HSIC) was introduced to enforce their statistical independence. Further, a variational information bottleneck and gradient reversal layer were incorporated to compress and disentangle the semantic and variation factors. Evaluations on two public datasets under both subject-dependent and subject-independent settings demonstrated that CD-CMAN consistently outperforms comparative baselines. These findings suggest that the proposed model could provide a new solution for the practical application of BCI technology.