EEG signal classification and forecasting for epileptic seizure prediction

Abstract

EEG signal recordings are increasingly replacing the old methods of diagnosis in medical field of many neurological disorders. Our contribution in this project is the study and development of EEG signal classification and forecasting algorithms for epilepsy diagnosis using machine learningusing one rhythm; for classification, an optimum classifier is proposed with only one used rhythm so that both execution time and number of features are reduced; for forecasting, the value of RMSE is minimized when using LSTM where the best hyperparameters are found. Firstly, we used wavelet packet decomposition (WPD) to extract the five rhythms of brain activity from the public Epilepsy-EEG recordings in order to represent each signal with features vector; then we applied on it the well-known classification methods. Secondly, we implemented forecasting methods for predicting seizures states on the signals using statistics methods and LSTM. A statistical study is done to validate the different algorithms.