Epilepsy Prediction via Time-Frequency Features and Multi-Scale Hybrid Neural Networks.

The prediction of epileptic seizures heavily depends on the precise embedding and classification of complex, multi-dimensional electroencephalogram (EEG) signals. Due to individual variability and the dynamic non-linear nature of EEG signals, extracting highly discriminative spatiotemporal features is a core challenge in this field. In this study, to address this issue, we proposed a novel architecture based on the Epilepsy Prediction using Multi-Scale Hybrid Neural Network (EPM-HNN), which integrates adaptive channel weighting, multi-scale spatial feature extraction, and bidirectional temporal dependency modeling. Specifically, we incorporated a sliding window mechanism with spatiotemporal resolution into the feature extraction process, enhancing the model's sensitivity to neural dynamics across frequency bands and improving its ability to capture micro-patterns. We used the Res2Net-50 multi-scale feature extractor to enhance the convolutional neural network's capacity to process complex local micro-features, such as polyspike-and-slow-wave complexes. Additionally, we introduced Squeeze-and-Excitation Networks (SENet) to adaptively capture potential effective features between different EEG channels. This dynamic weighting mechanism based on adaptive attention demonstrates strong robustness and high generalization across individual subject data. Furthermore, we proposed a non-single-subject, non-specific cross-subject training and testing method, demonstrating its ability to combat overfitting when addressing differences in data distribution. Experiments on the CHB-MIT scalp EEG dataset achieved an overall prediction accuracy of 97.7%, validating the effectiveness of the proposed EPM-HNN architecture.