Recognition Of Lower Limb Motor Imagery EEG Signals Based On Riemannian Geometry Framework

Motor imagery(MI)is a motor intention without specific body movements.Brain-computer interface(BCI)based on MI is capable of converting various MI electroencephalography(EEG)signals into corresponding control commands.As a result,MI-BCI receives widespread attention in rehabilitation medicine and other search fields.Lower limb motor function is a basic motor function for humans.Therefore,MI-BCI systems for lower limb rehabilitation have great clinical application value in the rehabilitation research of diseases like stroke.However,in existing researches,the classification of rest-state and lower limb MI EEG signals usually suffers from problems such as low accuracy and the need for great quantities of training samples,which limits the application of lower limb MI-BCI systems.Therefore,based on the Riemannian geometry framework,this research attempted to improve accuracy or reduce the number of training samples in recognition of lower limb MI EEG signals.Firstly,aiming at the problem of low accuracy in the recognition of lower limb MI EEG signals,this research proposed two clustering algorithms in the Riemannian manifold: margin based Riemannian clusters and statistics based Riemannian clusters.They could aggregate data of the same class in the Riemannian manifold into different subclusters to improve the accuracy of recognition of lower limb MI EEG signals.According to offline experiment results of 9 subjects,when the number of training samples was no less than 30,the average classification accuracies of two clustering algorithms in the Riemannian manifold were significantly higher than those of general comparison algorithms such as minimum distance to Riemannian mean,and the average classification accuracies of MBRC and SBRC could reach up to 98.74±1.39%and 98.89±2.51%.Secondly,aiming at the problem of the need for great quantities of training samples in the recognition of lower limb MI EEG signals,this research proposed an adaptive strategy in the Riemannian tangent space,to update the tangent space nearest centroid and tangent space linear discriminant analysis by mapping parameters between the Riemannian tangent space and the Riemannian manifold.According to offline experiment results of 9 subjects,when the number of training samples was 10,the classification accuracies of supervised-adaptive tangent space classifiers were significantly higher than those of non-adaptive tangent space classifiers(4.79%improvement).Finally,to test the online/pseudo-online performance of the clustering algorithms in the Riemannian manifold and the adaptive strategy in the tangent space,this research built a lower limb MI EEG signals recognition online system,completed 20person-time data collection,and designed two experiments for lower limb MI EEG signals online recognition.Online/pseudo-online results showed that when the number of training samples was 20,the classification accuracy of statistics based Riemannian clusters was 73.12±13.56%,significantly higher than 70.96±13.99% of minimum distance to Riemannian mean.And when the number of training samples was 10,the classification accuracies of supervised-adaptive tangent space classifiers were significantly higher than those of non-adaptive space classifiers(5.25% improvement).In summary,this research proposed cluster algorithms in the Riemannian manifold and adaptive strategy in the tangent space,achieved acceptable classification accuracy,or reduced the number of training samples in the recognition of lower limb MI EEG signals.The research contents are expected to provide technical supports for the optimization of lower limb MI-BCI systems and promote the development of lower limb rehabilitation.