Prediction of Epileptic Seizure onset based on EEG signals and learning approaches

Abstract

Epileptic seizures happen due to sudden bursts of electrical activity in the brain. This uncontrolled outburst may produce physical problems, abnormal behavior. Before the beginning of the seizure, a prediction is very useful to prevent the seizure by medication. This can be done by applying machine learning techniques and computational methods on EEG signals. However, EEG signals, in raw form, are hard to process. Feature measurement and noise cancellation can be done. Therefore, we come up with a model that presents the predictable methods of both preprocessing and feature extraction. We applied statistical methods for preprocessing and extracted time and frequency phase from the EEG signals. Our model detects the interictal state, which is the time frame between two seizures, preictal state, which is the time frame before Epileptic seizure, and ictal state, which is onset to the end of an epileptic seizure. We considered 1 hour and 30 minutes for every seizure duration to create this model. We have used the Savitzky-Golay filter for data smoothing and we used the energy of the signal, mean amplitude, skewness, and kurtosis of the signal as the features to classify seizure and non-seizure period. For classification, we have used two classifiers such as support vector machines and naive Bayes classifiers. The model is applied on the scalp EEG Children Hospital of Boston(CHB)-MIT dataset of 17 subjects and we obtained accuracy of more than 75 percent for predicting with a high true positive rate. In the proposed method, derived sensitivity is 42 percent, specifity is 80 percent, precision is 47 precent and negative predictive value is 32 percent.