Abstract
Epilepsy is a prevalent neurological disorder. Effective seizure prediction methods can warn of impending seizures, enabling patients to take preemptive measures. In this study, we propose an efficient epilepsy prediction model, 3D-SERESNet (squeeze-and-excitation residual network). We use short-time fourier transform (STFT) to convert EEG segments into time-frequency representations. We construct 3D representations from these time-frequency graphs, which serve as input to a subsequent 3D residual network designed to learn spatio-spectral features hierarchically. To address data imbalance, we employ the focal loss function. We evaluated the model's performance on the CHB-MIT dataset under both patient-specific and patient-independent settings. The model achieved an average sensitivity of 90.77% with a false positive rate (FPR) of 0.090/h in patient-specific experiments, and 84.41% sensitivity with 0.232/h FPR in patient-independent experiments.