Prediction of Epilepsy Seizure Based on Cepstrum Analysis and Deep Learning

There are specific changes in the frequency components of the electroencephalogram (EEG) signals before a epilepsy seizure. The Mel frequency can simulate the features of human auditory perception. If the EEG signal is converted into the Mel frequency domain, the frequency features associated with epilepsy seizure are highlighted, changes in these features are analyzed, and epilepsy seizure can be predicted. EEG is a non-stationary signal whose statistical properties change with time. Seizure patterns and associated Mel frequency signatures may differ at different seizure stages and time points. Therefore, it is necessary to develop analysis methods and prediction models that can adapt to non-stationary signals to improve the accuracy and stability of prediction. This paper proposes a prediction model for epilepsy seizure. The EEG signals are processed by Mel-frequency cepstral coefficients (MFCC) and linear predictive coding cepstral coefficients (LPCC). More comprehensive EEG features are extracted by integrating convolutional neural network (CNN) and long short-term memory (LSTM). The experimental data comes from publicly available CHB-MIT epilepsy EEG dataset. In order to verify the effectiveness of the proposed model, support vector machine, K-nearest neighbors, linear discriminant analysis, naive Bayes, decision tree, random forest, and logistic regression are selected for comparative experiments. Accuracy, sensitivity and specificity of the proposed model are 94 $\%$ , 96 $\%$ and 92 $\%$ , respectively. The performance of the proposed model is better than that of the contrasting methods. The proposed model is effective for the prediction of epilepsy seizure.