Research On Decoding Algorithm For Motor Imagery EEG Induced By Virtual Rowing Game

Brain-computer interfaces(BCIs)can establish a new channel for humans to directly communicate with the external environment by mind but not relying on the conventional peripheral nervous and musculoskeletal system.In particular,motor imagery(MI)based BCI has great significance and application value in motor rehabilitation,which has attracted broad attention from researchers.In order to make users immerse themselves in the MI-BCI training,virtual games have been adopted to enhance their enjoyment.However,the decoding algorithms for virtual game induced electroencephalogram(EEG)are poorly studied.Therefore,this study carried out in-depth research on the virtual game-induced EEG response and its decoding algorithm.Firstly,to investigate the MI induced EEG patterns,this study designed a virtual rowing game based MI-BCI paradigm including left and right hand asks.Fourteen healthy subjects were required to participated in the experiment for three times on different days.Then,we analyzed the corresponding EEG features from views of the temporal,spectral and spatial domains under both within-time and cross-time conditions.The results indicated that event-related desynchronization(ERD)and its contralateral dominance in spatial distribution were occurred for the most subjects.However,these phenomena intermittently existed in some local time and frequency intervals of the EEG,and changed greatly across time for the same subjects.The above results provided a theoretical basis for the follow-up EEG decoding algorithm research.Secondly,to improve the decoding of MI-EEG patterns,this paper proposed an ensemble learning based common spatial pattern(EL-CSP)algorithm.This novel method could comprehensively use the local EEG information in time,frequency,and space domains,and dynamically optimize the weights of all weak classification models.The results showed that the average classification accuracy of EL-CSP was 81.32%,which was 10.98% higher than the traditional CSP algorithm(p<0.001).Moreover,the proposed method showed superior stability with a standard deviation of only 0.17%when the number of base models was greater than 100.Finally,to enhance the generalization ability of the cross-time classification model,this study proposed an ensemble learning and Riemannian geometric alignment based minimum distance to mean(EL-RAMDRM)algorithm.Specifically,the whole MI period was first segmented into multiple subseries using a sliding window approach.Then,adaptive boosting(Adaboost)method was adopted to extract features and train multiple weak classifiers.At last,the voting results of these classifiers determines the final output of the BCI.The results showed an average cross-time classification accuracy of70.22% by the EL-RAMDRM method,which was 7.34% higher than the traditional CSP method(p<0.05).In addition,EL-RAMDRM also had a good performance in within-time recognition of MI-EEG patterns,with the average accuracy of 78.39%.In summary,this paper focused on the decoding of virtual rowing game induced EEG patterns,and developed two efficient methods,i.e.,EL-CSP and EL-RAMDRM,which could provide technical supports for the design of high-performance active BCI system and its application in rehabilitation.