Temporal Modelling for EEG Seizure Prediction: Saturation and Cross-Patient Variability

Accurate prediction of preictal states from EEG is critical for epilepsy management, yet the contribution of temporal modelling remains unclear. We present a controlled comparison of CNN and CNN-LSTM architectures under identical conditions using the CHB-MIT dataset, evaluating cross-patient performance under leave-one-subject-out validation. Temporal modelling improves overall performance (Macro F1: 0.66 → 0.72, p < 0.05), with the largest gain observed in preictal recall (+11%). Performance increases with temporal context but saturates at moderate lengths (5-10 steps), indicating diminishing returns. Despite these improvements, substantial variability across patients persists, identifying generalisation as the primary bottleneck. The CNN-LSTM model incurs a moderate computational overhead (∼50%), while remaining suitable for real-time use. These results show that temporal modelling provides bounded gains, and that improving cross-subject robustness is more critical than increasing temporal complexity.