Accurate electroencephalogram classification by variational Bayesian learning and graph convolutional networks

Non-invasive electroencephalogram (EEG) classification is crucial for motor-based brain–computer interfaces (BCIs), as it avoids surgical implantation and can be applied in a wider range of scenarios. However, EEG studies often encounter limited sample sizes, making it difficult for existing deep learning models to generalize effectively to EEG signals susceptible to noise. To address the multi-class EEG classification (MEC) problem, we formulate a model for MEC and introduce an integrated algorithmic framework. The proposed method constructs a neurophysiologically grounded spatiotemporal graph to represent functional brain connectivity, employs graph convolutional networks (GCNs) to capture spatial dependencies, and integrates a multi-head self-attention mechanism to model temporal dynamics. Furthermore, by embedding variational Bayesian (VB) inference with Monte Carlo sampling into the network weights, our approach effectively captures parameter uncertainty and mitigates overfitting under limited data conditions. This probabilistic architecture supports adaptive optimization and substantially enhances cross-subject and cross-scenario generalization. Experimental results on both public BCI competition datasets and a self-collected handwritten-character EEG dataset demonstrate that ESP consistently outperforms existing state-of-the-art methods in classification accuracy, stability, and robustness, confirming its effectiveness for EEG signal analysis.