MQVAE: Capturing Metastable Dynamics from EEG for Brain-computer Interfaces

Abstract

Research on neural dynamics indicates that cognitive processes are driven by metastable state transitions in the brain, which highlights the temporally discontinuous nature of brain activity. However, many existing methods fail to account for this discontinuity, limiting their effectiveness in modeling the brain’s temporal dynamics. To address this, we propose the Metastability Quantized Variational Autoencoder (MQVAE), a neurologically inspired neural network designed to capture metastable dynamics from electroencephalography (EEG) for brain-computer interface (BCI). In MQVAE, EEG signals are quantized into a time series of discrete latent variables using vector quantization. This quantization facilitates the identification of metastable states and the construction of a latent state space. To model state transition patterns, we incorporate a transformer module to capture global temporal dependencies among the discrete latent variables. Furthermore, a multi-scale spatial convolution layer is embedded within the encoder to extract hierarchical spatial patterns. Experiments on two public datasets demonstrate that MQVAE outperforms state-of-the-art methods in both emotion recognition and motor imagery classification tasks. Interpretability analyses further confirm that the features learned by MQVAE are neurophysiologically meaningful.