A hybrid Local Binary Pattern and wavelets based approach for EEG classification for diagnosing epilepsy

Abstract

Epilepsy is one of the grave neurological ailments affecting approximately 70 million people globally. Detection of epileptic attack is commonly carried out by viewing and analysing long-duration multi-channel EEG records. To counter this time-consuming process, a hybrid Local Binary Pattern-Wavelet based approach, classifying EEG in epileptic patients, is adopted in this research. Epilepsy is characterized by multiple ictal patterns in the form of synchronous epileptiform discharge transients. This work attempts to classify seizure from normal EEG recordings using the low-frequency activity. In order to perform this classification, the EEG signal is filtered and then transformed using Local Binary Pattern (LBP) into a new signal. Discrete Wavelet Transform (DWT) is employed to decompose the obtained signal. Wavelet coefficients are calculated to 5 levels of decomposition. A combination of univariate and bivariate features forms the feature set for seizure detection. This feature set extracted from low-frequency band coefficients helps in bringing out the dispersion, symmetry, and peakedness present in the EEG signal. A novel LBP based Spatio-temporal analysis of the continuous EEG signal for epilepsy detection is carried out on 105 seizures from 14 randomly selected subjects of CHB-MIT EEG database. A sensitivity of 100% is achieved on the CHB-MIT database while long term EEG is being tested with Linear Discriminant Analysis (LDA) classifier. The algorithm works well to obtain a false detection rate (FP/Hour) of 0.59. The specificity of 99.8% is attained with a mean accuracy of 99.6% when tested on 498.9 h of EEG data.