基于EEG脑网络下肢动作视觉想象识别研究

Identification of visual imagery of movements involving the lower limbs based on EEG network

基于想象的脑机接口(Brain‐Computer Interface,BCI)在运动障碍康复中有潜在的应用.传统的想象任务是运动想象(Motor Imagery,MI),但MI不易习得和控制,且存在“BCI(Brain Computer Interface)盲”现象,使得该类BCI的实用化受限.为寻找下肢运动障碍的康复方法,采用一种较少被研究且易完成的心理想象,即“视觉想象(Visual Imagery,VI)”来构建BCI,但该类BCI的分类难度较大,需要探索有效的特征提取方法.招募18名被试参加两种动态图片的视觉想象任务并采集脑电(Electroencephalogram,EEG)数据;采用EEG互信息构建功能网络,利用图论分析方法计算脑网络的网络属性特征,分别以网络属性特征、不同维度邻接矩阵空间特征与网络属性与邻接矩阵组合特征构建特征向量;最后采用支持向量机(Support Vector Machine,SVM)对两类视觉想象任务进行分类.结果显示,采用八维互信息邻接矩阵构建的空间特征集具有较好的可分性,平均分类精度为90.12%±5.43%,表明基于EEG互信息邻接矩阵空间特征是识别所设计的VI任务的有效特征,可望为构建新型的在线视觉想象脑机接口用于下肢运动障碍康复提供思路.

Imagery‐based brain‐computer interface(BCI)has potential applications in the rehabilitation of movement disorders.The traditional imagery task for BCI consists of motor imagery(MI).MI tasks are difficult for subjects,and there are even subjects who are incapable of MI.This makes MI‐BCI difficult to implement in practice.This study aimed at finding a rehabilitation method for lower limb movement disorders.We used a less studied and more easily performed mental imagery(Visual Imagery,VI)to implement a BCI.However,classification of VI tasks is challenging.Therefore,effective feature extraction methods require further exploration for VI‐BCI.In this study,18 subjects were recruited to participate in two kinds of dynamic pictures of visual imagery,during which EEG (Electroencephalogram) data were collected. Next,the mutualinformation based on EEG were used to construct a functional network,and graph theory analysis was used to calculate thenetwork attribute features of the constructed brain network. The feature vectors were constructed by the network attributefeatures:the spatial features of adjacency matrix of different dimensions and the combination features of network‐attribute andadjacency matrix. A support vector machine (SVM) was used to classify the two kinds of VI tasks. The results showed thatthe spatial characteristics constructed by an eight ‐ dimensional (8 ‐ D) mutual information adjacency matrix has goodseparability,and the average classification accuracy was 90.12%±5.43%,which showed that spatial characteristicsconstructed by an adjacency matrix of EEG mutual information was effective features for identifying VI tasks. Hence,ourfindings may provide new ideas for the construction of a novel on‐line VI‐BCI for the rehabilitation of lower limb movementdisorders.