CNN-based classification of epileptic states for seizure prediction using combined temporal and spectral features

Reliable prediction of epileptic seizures is of paramount importance in reducing the serious consequences of seizures by detecting their onset and warning patients early enough to take prompt and effective intervention measures, thereby ensuring the safety of patients who cannot be treated with pharmaceutical treatments or surgery. Indeed, the classification of epileptic states by deep learning methods based on electroencephalography (EEG) signals has attracted much attention in recent years. Nevertheless, the performance of classification of these states is strongly related to the preprocessing phase. The study of stability and detection of transitions between epileptic states is paramount to improve prediction algorithms. In this work, a stability index (SI) based on multivariate autoregressive modeling, capable of quantifying the phenomena observed during transitions between epileptic states and indicating the stability state of the epileptic neural system, is computed and fed to a convolutional neural network model among other known features in order to improve the learning performance of high-level features of the EEG signal and thus the classification of epileptic states. The experimental results highlight that the integration of the SI can stabilize the implemented learning model, satisfactorily improve the classification of epileptic states and permit our model to be competitive, based on many performance measures, to the state-of-the-art studies. Regarding the distinction between preictal and interictal states, our proposed model achieved an average accuracy of 90.1% to 94.5% and an average sensitivity of 88.6% to 92.8% for preictal interval durations of 30 and 60 min, respectively, on the CHB-MIT data set.