Özet:

Detection of epileptic activities requires detailed analysis of the electroencephalogram (EEG) data. Scoring manual epileptic activities is both difficult and inconsistent. Machine learning techniques are faster and more consistent than manual scoring. Therefore, there is a need for an effective machine learning technique to classify EEG data. Because of the success of modeling nonlinear data, Support Vector Machines (SVM), which is a supervised learning algorithm, is preferred. This success is achieved only when the appropriate kernel function is selected. Commonly used kernel functions are linear, polynomial and radial based (RBF). Since the nature of the data is not known in advance, it is difficult to make appropriate selection from the kernel functions. For this reason, when creating the model, it should be selected using multiple kernel functions to give the best performance among them. In this study, EEG data from Bonn University and 9 different classification problems are discussed. EEG signals were analyzed in 5 different frequency bands and feature vectors were generated from the standard deviation values of each frequency band. The generalization capabilities of linear, polynomial, radial based and Pearson VII(PUK) kernel functions are compared. The effect of PUK kernel functions parameter values on success rates is also investigated. With the model proposed in the study, processing load was reduced, dimensionality reduction algorithms were eliminated, and less processing load was created. It was concluded that PUK kernel function has better generalization performance than other functions. 100% success rate was achieved in two-class problems.

Anahtar Kelimeler: