基于切片滤波器组共空间模式的跨数据集运动想象分类方法研究  

Cross-dataset motor imagery classification based on slice filter bank common spatial pattern

作　　者：张金辉[1] 郑宇博 邹冰 申牧 罗莹莹[2] 李蕾[2] ZHANG Jin-hui;ZHENG Yu-bo;ZOU Bing;SHEN Mu;LUO Ying-ying;LI Lei(Equipment Support Section of Logistics Support Center,Chinese PLA General Hospital,Beijing 100853,China;School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]解放军总医院服务保障中心装备保障室,北京100853 [2]北京邮电大学人工智能学院,北京100876

出　　处：《医疗卫生装备》2022年第8期8-13,40,共7页Chinese Medical Equipment Journal

基　　金：军队装备综合研究项目(LB2020A010010)。

摘　　要：目的:提出一种基于切片滤波器组共空间模式(slice filter bank common spatial pattern,SFBCSP)的运动想象分类方法,降低脑电运动想象数据集之间的系统性差异对分类模型的影响,提升分类准确率。方法:基于SFBCSP进行运动想象分类,首先对原始脑电信号进行预处理,其次通过带通滤波器和滑动窗口将脑电信号切分为不同频带和时间片段,然后利用空间滤波器提取脑电信号不同电极通道之间的空域特征,最后对多维度特征进行筛选和分类。利用神经信息处理系统大会提供的竞赛数据集对SFBCSP方法与脑电神经网络EEGNet、卷积循环注意力模型(convolutional recurrent attention model,CRAM)、共空间模式(common spatial pattern,CSP)和滤波器组共空间模式(filter bank common spatial pattern,FBCSP)的分类准确率进行比较。结果:提出的SFBCSP方法平均分类准确率均高于EEGNet、CRAM、CSP和FBCSP,相较于FBCSP,SFBCSP在单个数据集、跨数据集和混合数据集上的平均分类准确率分别提升了7.26%、5.16%和1.96%。结论:所提出的SFBCSP方法能够降低跨数据集的系统性差异对分类模型的影响,对运动想象和脑机接口的研究具有重要意义。Objective To propose a motor imagery classification model based on slice filter bank common spatial pattern(SFBCSP)to reduce the influence of systematic differences between EEG motor imagery datasets on the classification model and improve the classification accuracy.Methods Motor imagery classification was carried out based on SFBCSP.The raw EEG signals were pre-processed and sliced into different frequency bands and time segments by band-pass filters and sliding windows,then the spatial filter was used to extract the spatial domain features between different electrode channels of the EEG signals,and finally the multidimensional features were filtered and classified.The classification accuracy of the SFBCSP-based method was compared with those of EEGNet,convolutional recurrent attention model(CRAM),common spatial pattern(CSP)and filter bank common spatial pattern(FBCSP)using the competition dataset provided by the Neural Information Processing Systems Conference.Results The average classification accuracy of the proposed SFBCSP-based method was higher than those of EEGNet,CRAM,CSP and FBCSP,which was improved by 7.26%,5.16%and 1.96%on single dataset,cross-dataset and mixed dataset respectively when compared with that of FBCSP.Conclusion The proposed SFBCSP-based method reduces the impact of systematic differences in different datasets on classification models and has important implications for the study of motor imagery and brain-computer interfaces.[Chinese Medical Equipment Journal,2022,43(8):8-13,40]

关 键 词：运动想象 脑电 SFBCSP FBCSP CSP 脑机接口 

分 类 号：R318[医药卫生—生物医学工程]