Four-class EEG Classification for Seizure Prediction and Detection Using a Lightweight CNN-LSTM

Epilepsy, a neurological disorder, necessitates prediction and detection of seizures to safeguard patient health. Previous studies have employed 2-class or 3-class classification neural network models for seizure monitoring. However, these models overlooked electroencephalogram (EEG) characteristics in postictal stage, leading to reduced specificity in real-world scenarios. In this work, we introduce a lightweight four-class classification method that combines seizure prediction and detection. By applying the short-time Fourier transform (STFT), raw EEG signals are transformed into time-frequency spectra. The employment of the one-dimensional convolutional neural network (1D-CNN) paired with long short-term memory (LSTM) ensures meticulous extraction and classification of complex EEG patterns. Experimental results demonstrate that the average accuracy of the proposed four-class classification method is $\mathbf{9 8. 4 4 \%}$. In addition, when postictal samples are included in the test set, the specificity of seizure prediction and detection is 99.28% and 100%, respectively. The proposed method effectively addresses the reduction in specificity caused by neglecting postictal stages. The parameters of the proposed method are only 3.7 K, whereas other models’ parameters range from 1.7 to 60.7 times more, making it particularly suitable for wearable devices in resource-constrained edge computing scenarios. The project code is openly available at https://github.com/0HENGMENG0/network-for-seizures.