BCIwiki (bciwiki.com) — A hierarchical graph learning framework developed at Lanzhou Jiaotong University outperformed baseline models such as EEGNet on three-class motor imagery decoding in virtual reality (VR), researchers reported in Journal of Neural Engineering on September 23, 2026 (PMID 42777778; DOI 10.1088/1741-2552/aeabf5).
Pairing VR with action observation (AO) gives users richer visual guidance for motor imagery (MI), but electroencephalography (EEG) recorded in dynamic environments is non-stationary and has a low signal-to-noise ratio, which undermines decoding stability in motor imagery brain-computer interfaces (BCIs).
The framework uses multivariate variational mode decomposition (MVMD) to break EEG into rhythmic components for multi-domain feature analysis, applies hybrid feature selection to obtain a compact representation, and builds functional brain networks from phase locking values. Each trial is represented as a sparsified graph, and a hierarchical graph convolutional network captures multi-scale spatial dependencies and high-order interactions for classification. The team evaluated it on a lab-acquired VR motor imagery dataset.
The authors say the VR-based AO plus MI paradigm enhanced sensorimotor connectivity and promoted large-scale network integration, indicating more coordinated neural dynamics during the task, and that the framework supports potential advanced BCI applications. The abstract does not report specific accuracy figures or the number of participants.