特征选择算法在ECoG分类中的应用

研究了基于运动想象的皮层脑电信号ECoG的特点,针对BCI2005竞赛数据集I中的ECoG信号,通过提取频带能量获得了想象左手小指及舌头运动时的特征,结合Fisher,SVM-RFE及L0算法对特征进行选择,采用10段交叉验证的方法得到训练数据集在各维特征数下的识别正确率并选出最佳特征组合.结果表明:三种特征选择方法中SVM-RFE算法所选出的特征组合可以获得最低的识别错误率以及最低的特征维数,针对所选出的特征组合,使用训练数据集的特征对线性支持向量机进行训练,使用训练好的模型对测试数据集进行分类,识别正确率可以达到94%.

Application of Feature Selection and SVM for ECoG Classification

The motor imagery ECoG(electrocorticongraph) was investigated,specifically for classifying different imagined movements of the left little finger and tongue through ECoG,with BP(band power) of the ECoG signal extracted as the feature of BCI2005 competitive dataset I.Then Fisher,L0,and SVM-RFE were each used to select the best features.After a 10-fold cross validation of the training dataset,the features selected by SVM-RFE were the best of the three feature selection methods because it provided the lowest classification error rate and the least feature dimensionality.In addition,selected features of the training dataset were used to train a linear SVM model while selected features of the testing dataset were used to predict the labels by the model.Final classification accuracy was 94%.