EPILEPTIC Seizure Classification Using Gradient Tree Boosting Classifier

Analysis of electroencephalography (EEG) is widely used for the diagnosis of epilepsy in which relevant information extraction from EEG signals poses great challenge due to noise and interference with various environmental factors. This paper proposes a binary classification system through which EEG signals are analyzed to distinguish between ictal and normal signals. For this purpose discrete wavelet transform (DWT), along with gradient boosting is used for classification. Two level, Daubechies order 4 wavelet are used to decompose the signal into three sub-bands after which Hjorth mobility and Hjorth complexity are calculated from these sub-bands resulting in a 6-dimensional feature vector. We use two benchmark datasets in our experimentation i.e., the Bonn's dataset and CHB-MIT dataset. We establish our classifier using training samples from the Bonn's dataset. Classification accuracy of 99.4% is achieved when tested on same dataset using different samples. To validate the effectiveness and better generalization of our system, we cross-test on CHB-MIT dataset which yielded accuracy of 96.8%. Achieved performance surpasses previous state of the art technologies, giving better classification results than other well-known techniques used for seizure classification. Considering low feature dimension and hence decreasing complexity, coupled with the high performance on both datasets prove the given method to be favourable for distinguishing between epileptic and non-epileptic EEG signals.