Motor imagery classification method based on Riemannian Manifold and CNN-LSTM for advanced brain computer interfaces

Motor imagery classification method based on Riemannian Manifold and CNN-LSTM for advanced brain computer interfaces

연구 분야: Infrastructure

학회: CCF Transactions on Pervasive Computing and Interaction

Motor imagery, a crucial brain–computer interface (BCI) paradigm, provides a new approach to facilitating human–computer interactions. However, its performance is considerably impaired by variances in the operating band and noise in the electroencephalogram (EEG) signal, thus resulting in weak performance. This study proposes a feature extraction method combining a filter bank with Riemannian manifolds for advanced BCIs. The filter bank is used to overcome frequency band variations, and the Riemannian metric is used to reduce noise interference. Covariance matrices are used to construct a high-dimensional Riemannian manifold; they are eventually mapped into feature vectors of the tangent space. Subsequently, a convolutional neural network-long short-term memory (CNN-LSTM) hybrid network is employed to fully utilize the extracted spatial and spectral features (as well as temporal information) for classification. Experiments on the public dataset BCI IV 2a (four classes) and our in-house TMI dataset (three classes) reveal that the proposed method outperforms all existing methods, with classification accuracies of 88.1% and 91.2% for the two datasets, respectively. The proposed method provides a new option for classifying multiclass motor imagery and promises application potential for real-time systems.