Spatio-Temporal Attention Network for Epileptic Seizure Prediction

TL;DR

Timely seizure prediction from multichannel EEG with variable preictal durations is addressed by a unified Spatio-Temporal Attention Network (STAN) that jointly models spatial and temporal patterns. The model is trained with an adversarial discriminator to distinguish preictal from interictal attention, and employs incomplete supervision to learn patient-specific preictal durations, yielding personalized warning windows. Key contributions include a cascaded STAN architecture with multiple attention heads, a gradient-penalty discriminator, and demonstrated state-of-the-art results on CHB-MIT and MSSM datasets, with high sensitivity and very low false alarm rates. The approach enables reliable, timely interventions for clinical and home monitoring, significantly advancing personalized epilepsy management.

Abstract

In this study, we present a deep learning framework that learns complex spatio-temporal correlation structures of EEG signals through a Spatio-Temporal Attention Network (STAN) for accurate predictions of onset of seizures for Epilepsy patients. Unlike existing methods, which rely on feature engineering and/or assume fixed preictal durations, our approach simultaneously models spatio-temporal correlations through STAN and employs an adversarial discriminator to distinguish preictal from interictal attention patterns, enabling patient-specific learning. Evaluation on CHB-MIT and MSSM datasets demonstrates 96.6\% sensitivity with 0.011/h false detection rate on CHB-MIT, and 94.2% sensitivity with 0.063/h FDR on MSSM, significantly outperforming state-of-the-art methods. The framework reliably detects preictal states at least 15 minutes before an onset, with patient-specific windows extending to 45 minutes, providing sufficient intervention time for clinical applications.

Figures (4)

• Figure 1: Architecture of STAN showing three cascaded attention networks processing raw EEG input. Each network contains spatial and temporal attention modules with H=4 attention heads. The resulting M=3 spatial and temporal attention maps are aggregated via MLP and passed to the discriminator for adversarial training.
• Figure 2: Detailed architecture of spatial and temporal attention modules. Both employ 1D CNN encoders followed by multi-head attention (H=4) and include residual connections with layer normalization for training stability.
• Figure 3: Discriminator architecture employing MLPs to process aggregated attention patterns. Trained adversarially with gradient penalty to distinguish preictal from interictal states, generating anomaly scores for real-time seizure prediction.
• Figure 4: Real-time seizure prediction showing discriminator scores 90 minutes before onset. Recording 03 (top) and Recording 26 (bottom) demonstrate consistent gradual shift from interictal to preictal states, providing at least 15 minutes of intervention time.