EEG-Based Epileptic Seizure Prediction Using Attention-Based Temporal Convolutional Recurrent Neural Networks for Cross-Patient Generalization

Le résumé fourni par la source

Epilepsy affects approximately 1% of the global population, with timely seizure prediction being critical for mitigating life-threatening risks. This paper introduces a novel deep learning framework that integrates attention mechanisms with temporal convolutional and recurrent neural networks (ATC-RNN) for cross-patient seizure prediction using scalp EEG signals. We synergistically combine handcrafted feature extraction capturing time-frequency characteristics through maximum cross-correlation, phase-locking synchrony (SPLV), and statistical moments with automatic high-dimensional pattern mining via dilated causal convolutions. The architecture employs layer normalization and parametric ReLU activations to model long-range dependencies in EEG streams while an attention gate dynamically filters redundancies and amplifies discriminative preictal signatures. Evaluated comprehensively on the full CHB-MIT dataset (24 cases, ≈975 hours), our model achieves 0.988 AUC, 96.24% sensitivity, and 0.0053/h false prediction rate, surpassing state-of-the-art methods by >1.9% AUC. Implementation of a "5-of-10" post-processing rule further reduces false alarms by 68%, as validated in challenging cases like chb06 where false prediction rates decreased from 0.22/h to 0.07/h. These results demonstrate unprecedented cross-patient generalization capabilities essential for clinical deployment, though architectural simplification remains necessary for real-time implementation at bedside monitoring settings.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.