Feature selection applied to wavelet packet transform for an efficient EEG signal classification

This paper deals with feature selection for electroencephalogram (EEG) signal classification. The wavelet packet transform (WPT) technique is used to decompose EEG signals. Then, features are computed and extracted from different wavelet packet coefficients for each frequency sub-band. Then, principal components analysis (PCA) is used for feature dimension reduction. The features are used as inputs for support vector machine (SVM) classifier. The SVM classify EEG signals into two categories: Epileptic or Healthy. To select the best feature basis, a feature ranking method based on SVM recursive feature elimination technique (SVM-RFE) is used. Best results are obtained for a combination of three features, namely: range, maximum and energy values of wavelet coefficients in the sub-bands. The classifier was tested against data obtained from three different datasets. Data length is approximately 120 hours. Obtained results were compared to those obtained in state of the art papers. The classification accuracy reached 99.91% for the first dataset (around 1 hour 20 minutes data) and reached 100% for the second dataset (around 20 hours data) and reached 99.78% for the third dataset (around 98 hours data). Those results are among the best classification results reported so far.