Modeling multiscale time-frequency complex networks on Riemannian manifolds for motor imagery BCI classification with graph convolutional networks

Abstract

Motor imagery brain-computer interface (MI-BCI) classification faces challenges such as low decoding accuracy and difficulty in capturing the spatiotemporal dynamics of EEG signals. The use of Riemannian geometry classifiers for this task has become one of the most popular classification methods. However, current Riemannian geometry classifiers typically compute the covariance matrix over a period of time to capture spatial features, neglecting the multiscale characteristics of EEG signals in both time and frequency, which limits their classification performance. To address these issues, this study proposes a novel framework. Specifically, we introduce graph convolutional network (GCN) on Riemannian geometry (GR) to process multiscale networks, using virtual nodes to capture global topological features and integrating spatial features across time and frequency domains. This method significantly enhances the feature extraction capability of Riemannian geometry classifiers. The proposed method was validated on three public datasets, with average classification accuracies of 91.87 % ± 7.33 %, 87.96 % ± 7.6 %, and 82.50 % ± 7.74 %, respectively. Ablation experiments show that, compared to traditional single-scale methods, the average classification accuracy improved by 9.85 %, highlighting the effectiveness and versatility of the proposed method. This research provides a new perspective for multiscale EEG signal analysis and advances the development of motor imagery BCI classification technology.

Keywords: Geometric deep learning; Graph convolutional networks; Motor imagery BCI classification; Multiscale time-frequency complex networks on riemannian manifolds.