基于ECoG的运动想象分类  

Classifying ECoG-Based Motor Imagery Tasks

作　　者：安滨[1] 江朝晖[1] 宁艳[1] 陈强[2] 冯焕清[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230026 [2]合肥工业大学生物医学工程系,合肥230009

出　　处：《中国生物医学工程学报》2007年第1期64-68,共5页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金资助项目(60422201);中国科学技术大学研究生创新基金(200510)

摘　　要：目的以两种运动想象任务下采集的64导ECoG信号为训练样本,识别几天后重复进行的运动想象任务。方法以动作感知皮层区脑电图(ECoG)的μ节律(8Hz^13Hz频段)功率谱为特征。通过手工比较功率谱的差异显著性,从64导中粗选出11导最明显的信号。再用共同空间特征法(CSP)滤波提高信噪比,使信号从11维降到8维。采用K近邻分类器进行分类识别,其中依据交叉验证法得到最佳的近邻值。结果测试样本的预测精度达到94%。结论利用动作感知皮层区脑电μ节律能较好识别对应的特定(想象)运动;共同空间特征法滤波可以有效提高信噪比;只要预处理、特征抽取及分类得当,时间间隔和实验误差等因素对运动想象识别的影响不大。Objective The 64-channel ECoG signals recorded during two sorts motor imagery tasks performed were regard as train sample to classify the ECoG signals recorded under the same tasks performed a few days later. Methods The power spectral density of μ-rhythm （between 8Hz to 13Hz） extracted from ECoG in motor cortex was selected as feature. Total 11 channels of distinctive ECoG signal were selected after comparing the power spectra of all 64 channels of ECoG signals. And then the algorithm of common spatial patterns （CSP） was used in preprocessing to improve the signal-to-noise ratio, which made the dimension down from 11 to 8. A k-nearest neighbor classifier, the optimal k was chosen using the method of cross-validation, was applied for the final classification. Results The predictive accuracy was 94% for the test samples. Conclusion Specific motor imagery tasks can be recognized precisely by μ-rhythm extracted in corresponding motor cortex area. The preprocessing with CSP makes a notable improvement in signal-to-noise ratio. With the suitable pretreatment, features extraction and classifier design, the influence, which arose from time interval, experiment error and so on, can be nearly ignored.

关 键 词：运动想象 皮层脑电图 μ节律 共同空间特征法滤波 K近邻 

分 类 号：R318[医药卫生—生物医学工程]