Towards Clinically Interpretable Early Seizure Prediction in Pediatric Epilepsy Using Hybrid Deep Neural Networks and Multi-Resolution EEG Analysis

Early seizure prediction in Pediatric epilepsy remains a significant clinical challenge due to complex neurophysiological dynamics, high inter-patient variability, and the non-stationary nature of electroencephalogram (EEG) signals. Although deep learning models have demonstrated promising performance in seizure detection, their translation into clinical practice is limited by insufficient interpretability and inadequate modelling of multi-scale temporal patterns preceding seizure onset. Addressing these limitations, this study proposes a clinically interpretable hybrid deep neural network framework for early seizure prediction using multi-resolution EEG analysis in paediatric patients. Multi-channel scalp EEG recordings obtained from the CHB-MIT Scalp EEG Database were pre-processed using adaptive bandpass filtering and artifact suppression techniques to enhance signal quality. A discrete wavelet transform–based multiresolution decomposition strategy was employed to capture time–frequency characteristics across delta, theta, alpha, beta, and gamma bands, enabling robust representation of preictal dynamics. The extracted feature maps were integrated into a hybrid CNN–BiLSTM architecture designed to simultaneously learn spatial correlations across EEG channels and long-range temporal dependencies. To enhance clinical transparency, gradient-based attribution mapping and SHAP-driven feature importance analysis were incorporated to identify dominant frequency bands and channel contributions associated with preictal transitions. The proposed framework was evaluated using patient-wise cross-validation to ensure generalization across unseen subjects. Performance was assessed using sensitivity, specificity, F1-score, area under the receiver operating characteristic curve (AUC), and early prediction horizon metrics. Experimental results demonstrate that the hybrid architecture significantly outperforms conventional standalone CNN and LSTM models, achieving improved predictive accuracy while maintaining interpretability of model decisions. Importantly, the explainability analysis revealed clinically meaningful neurophysiological patterns consistent with established pediatric seizure biomarkers. Overall, the integration of multi-resolution EEG modelling, hybrid deep learning, and explainable artificial intelligence provides a robust and clinically aligned approach for early seizure prediction in pediatric epilepsy, supporting real-time monitoring systems and advancing personalized therapeutic interventions.