Adaptive Temporal–Spatial Attention Framework for Early Epileptic Seizure Prediction Using Multichannel EEG Signals

Epileptic seizure prediction from electroencephalogram (EEG) signals plays a critical role in improving the safety and quality of life of individuals living with epilepsy. This study proposes a robust multi-stage seizure prediction framework that integrates noise reduction, data enrichment, spectral decomposition, dynamic connectivity learning, optimal feature selection, and attention-driven classification. Initially, Adaptive Neuro-Clean Filtering (ANCF) removes noise components while preserving seizure-relevant neural dynamics. To mitigate class imbalance and enhance model generalization, a Temporal-Physiological Mixup (TPMix) strategy generates physiologically consistent synthetic EEG samples. The refined signals are further processed using the Multivariate Adaptive Synchro-Spectral Transform (MASST) to extract high-quality time-frequency representations. Dynamic Graph Spectral Connectivity (DGSC) features are derived to capture evolving inter-channel relationships, and Hybrid Quantum-Optimized Fisher Selection (HQFS) chooses compact yet discriminative feature subsets. Finally, a TemporalSpatial Attention Graph Neural Network (TSAGNN) jointly models long-range temporal dependencies and spatial brainregion interactions to classify EEG into pre-ictal, ictal, post-ictal, and normal states. Experimental evaluation on the CHB-MIT dataset achieved an overall accuracy of 95.9%, demonstrating the model's strong predictive capability and clinical relevance.