EEG-GSoPNet: A Convolutional Network with Global Second-Order Pooling for Multiclass Classification of Upper Limb Motor Imagery in BCI

The application of brain-computer interface (BCI) in motor imagery (MI) training has been validated as a beneficial method for promoting upper limb rehabilitation in stroke survivors. However, the traditional MI paradigm is difficult to meet the diverse needs of patients for different joint movements during upper limb rehabilitation. To address this limitation, we proposed a new model called EEG-GSoPNet, which aims to classify six different categories of upper limb motor imagery. The model employs a multi-scale convolutional neural network to separately capture both temporal and spatial features, and combines the global second-order pooling (GSoP) module to effectively capture the dependencies between features, thereby enhancing the model's ability to express complex relationships in EEG signal.