Epilepsy is a critical and widely seen neurological disorder for people and electroencephalogram (EEG) signals are used to diagnose epilepsy. To accurately diagnose epilepsy, distinctive features of the EEG signals should be extracted. Therefore, a novel texture descriptor is presented for distinctive feature extraction in this study and an EEG recognition method is proposed. The proposed method consists of four main phases. These are feature extraction, feature concatenation, feature reduction and classification. Firstly, the EEG signal is divided into 1 × 25 size of overlapping blocks and these blocks are converted to 2 dimensional blocks with size of 5 × 5. Because, the proposed novel local senary pattern (LSP) uses 5 × 5 size of blocks for feature extraction. 1536 Features are extracted using the proposed LSP. The proposed LSP is used ternary function to extract features and as we know that the main problem of the ternary function is to find optimal threshold value. Therefore, we used 10 threshold values by using standard deviation function and 1536 × 10 = 15,360 features are extracted from an EEG signal. In the feature combining phase, these features are concatenated. In order to reduce these features, a neighborhood component analysis based feature reduction method is used. In the classification phase support vector machine, k nearest neighborhood, quadratic discriminant analysis and linear discriminant analysis are utilized as the classifiers. To test success of the proposed method, the widely used EEG signals dataset which is Bonn University EEG database is used and 7 cases are defined for testing using this database and the proposed method achieved 93.0% classification accuracy for 5 classes case. The obtained results and comparisons clearly indicated success of the proposed LSP based method.
Keywords: EEG; Epilepsy diagnosis; Feature extraction; Local senary pattern.