A hierarchical graph learning framework for VR-based motor imagery decoding driven by MVMD and feature optimization

While virtual reality (VR) combined with action observation (AO) provides enriched visual guidance for motor imagery (MI), the decoding stability of MI Brain-Computer Interface (BCI) remains a challenge due to the inherent non-stationarity and low signal-to-noise ratio of Electroencephalography (EEG) signals in dynamic environments. Enhancing decoding accuracy and understanding neural representations in VR-based AO+MI tasks remain important challenges in neural engineering.

This study proposes a hierarchical graph learning framework tailored for VR-based MI decoding, leveraging multivariate variational mode decomposition (MVMD) to decouple complex EEG signals into rhythmic components for enhanced multi-domain feature analysis, hybrid feature selection for compact representation, and functional brain network modeling based on phase locking value. Each EEG trial is represented as a sparsified graph, and a hierarchical graph convolutional network is designed to capture multi-scale spatial dependencies and high-order interactions for classification.

Experimental evaluations on a lab-acquired VR-MI dataset demonstrate that the proposed framework achieves competitive three-class decoding accuracy. This performance surpasses baseline models such as EEGNet, as observed under VR-based motor imagery conditions. In addition to enhanced decoding performance, the study enables neurophysiological interpretation by characterizing VR-induced modulations in brain network organization.

VR-based AO+MI paradigm is found to enhance sensorimotor connectivity and promote large-scale network integration, indicating more coordinated neural dynamics during MI tasks. These findings suggest that the proposed framework effectively improves VR-based AO+MI decoding while providing interpretable insights into neural mechanisms, supporting its potential for advanced BCI applications.