Parmak hareketlerine ilişkin ECoG örüntülerin AR tabanlı öznitelikler ile sınıflandırılması

Classification of electrocorticography (ECoG) records related to finger movement is the main purpose of this study. Data set IV presented in BCI Competition IV was used in this paper. This data set contains brain signals from three epileptic subjects and the data records consist of both ECoG and electronic glove data. ECoG segments related finger movements were extracted by means of finger movement records generated by electronic glove. Features of segments having different data points were extracted using autoregressive (AR) model. The AR coefficients were classified with Support Vector Machine (SVM) and K nearest neighbors (kNN) classifiers. AR coefficients were calculated using the least squares method. To get the acceptable result, the AR coefficients were calculated in the range of [3-10] order. At the end of the analysis, the third-order AR coefficients produces acceptable results were observed. k nearest neighbor (kNN) algorithm was introduced by Dasarathy in1991. It is a machine learning algorithm that reads a set of labeled training set, and then it is used to classify an unlabeled testing set. In order to classify a testing pattern, it computes the distance between testing pattern and all the training patterns. Then, the k training patterns of closest distance to the testing pattern are used to determine the class of testing pattern. To get the acceptable results in our classification problem, the distance of patterns were calculated with the k parameter in the range of [1-10] order. At the end of the analysis, generally the highest classification performances were observed with small k values. Support vector machine (SVM) is a discriminative classifier formally defined by a separating hyperplane. In other words, SVMs are based on the concept of decision planes that define decision boundaries. A decision plane is one that separates between a set of objects having different class memberships. In our study, SVM was used to classify the ECoG patterns with the kernel of radial base function. At the end of the SVM classification, the optimal values were obtained for regularization and sigma parameters of SVM. To find the finger which is hardest to classify among all fingers, different applications were achieved. In the first application, it consists of two classes for P1 and P5. In the second application, it consists of three classes for P1, P2 and P5. In the third application, it consists of four classes for P1, P2, P3 and P5. In the last application, it consists of five classes for P1, P2, P3, P4 and P5.The results showed that performances were decreased with increasing of classes, which was also an expected result. We listed the performances of SVM and kNN methods. We found that the SVM method yielded the best performance in classifying the ECoG pattern related to finger movements. For the SVM classifier, the mean performances of three subjects were obtained as follows; classification rate 87.35% for two fingers, classification rate 66.97% for three fingers, classification rate 50.06% for four fingers and classification rate 34.41% for five fingers. For the kNN classifier, the mean performances of three subjects were obtained as follows; classification rate 75.35% for two fingers, classification rate 55.50% for three fingers, classification rate 39.00% for four fingers and classification rate 31.90% for five fingers. The results showed that we need different approaches to impove the performances. To get the higher performance, more effective methods should be improved. In the future work, we focused on this problem.