SecNet: A second order neural network for MI-EEG

Abstract

Motor imagery-based brain-computer interfaces (MI-BCIs) hold significant potential for individuals with severe paralysis who are aware and alert by but unable to reliably control their muscles. MI-BCIs have garnered increasing attention from researchers in the field of motor function rehabilitation as a healthcare technology. However, the variability of inter-session and the difficulty in extracting energy features bring huge challenges to the information processing of such systems. To overcome this, we propose a novel architecture called SecNet, designed to capture relationships between convolutional features. SecNet utilizes multiple branches to learn spatio-temporal features of EEG signals and then pools them into a covariance. We integrate regularization and attention mechanisms to enhance the learning efficiency of features, followed by the adoption of a second order pooling method. Lastly, we employ Riemannian geometry learning to map features derived from symmetric positive definite covariance matrices. Evaluation experiments are conducted on a dataset from stroke patients and further compared its performance on two public datasets. Experimental results show that the SecNet is superior to the benchmark methods and achieves accuracy rates of 72.90%, 87.08% and 74.28% on the Stroke dataset, BCI IV 2a dataset and OpenBMI dataset, respectively. These results demonstrate its efficacy and robustness in inter-session decoding for MI-BCI, showing its practical utility for application. Our code is available at https://github.com/SecNet-mi/SecNet.