Automated Classification of Epileptiform Discharges in EEG Signals Using the Wavelet Transform

This study analyzes the use of the wavelet transform for the improvement of automatic detection algorithms of epileptic discharges. The novel algorithm uses an eight-level wavelet decomposition on both normal and epileptic discharge EEG data. The obtained coefficients are characterized by several features: maximum amplitude, standard deviation, variance, skewness, kurtosis, Shannon and log energy entropy. These are fed into two classification algorithms: a support vector machine (SVM) and random forest (RF). The algorithm is applied on the CHB-MIT database, containing EEG recordings from 24 pediatric patients. Out of the proposed features, the amplitude, variance, standard deviation and the two entropy measures proved themselves to have the best discrimination power. The RF algorithm outperformed SVM for most of the patients. The use of multiresolution analysis combined with statistical features for coefficient characterization proved to be beneficial for the performance of the classification algorithms. The maximum obtained accuracy was of 100%, representing an improvement of 10% as compared to our previous study, for the same patient.