A Deep Learning Method for Classification of Interictal and Preictal EEG Signals in Epileptic Seizure Prediction

Epilepsy is one of the most common nervous system diseases in the world, which will cause serious consequences to the patient's nerves, cognition and psychology, and affect the patient's physical and mental health and quality of life. Early detection of seizures can help prevent seizures and reduce harm through medical intervention. In this study, a deep learning (DL) framework was proposed that employs two branches to extract time-domain and frequency-domain features of EEG signals, which are then fused together to improve feature extraction and classification abilities. The proposed model demonstrates satisfactory performance in epileptic seizure prediction (ESP). To ascertain the effectiveness of the model, four experiments were conducted. Initially, five different segmentation lengths were tested in order to determine the optimal one. Subsequently, ablation experiments were executed to assess the effectiveness of the model’s components. Furthermore, the model was tested on specific patient data, resulting in an average accuracy of 0.9636 and an Fl-score of 0.9638 on the CHB-MIT dataset, outperforming other advanced models. Finally, when compared to other advanced models, the proposed framework yielded the highest accuracy in both the CHB-MIT dataset and the newborns with epilepsy annotation EEG recording dataset (DNERSA), with accuracies of 0.9943 and 0.9745, respectively. The proposed model exhibits promising potential in the field of ESP.