A Seizure Prediction Method for Epilepsy Utilizing Lightweight Convolutional Neural Networks and Time-frequency Spectrograms of EEG Signals

Seizure prediction is of great significance for the treatment of patients with epilepsy. However, distinguishing between the ictal state, inter-ictal state and pre-ictal state, especially the latter two, is a challenging task. In this paper, we propose a seizure prediction method based on the time-frequency spectrograms of single-channel electroencephalogram (EEG) signals. The method mainly contains preprocessing algorithm of EEG signals and classification algorithm using neural networks. The timefrequency spectrograms of EEG signals are obtained by short-time Fourier transform (STFT) of single-channel EEG signals. The task of classifying ictal state, inter-ictal state and pre-ictal state is implemented using the time-frequency spectrograms of EEG signals as input to a convolutional neural network (CNN). The proposed method achieves the accuracy of 99% and 92.63% based on Bonn Dataset and CHB-MIT Dataset, respectively, which demonstrates its compatibility. Moreover, we present the hardware implementation of the proposed neural network based on a micro-controller and the performance validation is carried out. Compared with the existing works, our proposed method has higher prediction accuracy and smaller model size, which shows the advantages for embedded processor of wearable systems.