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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate prediction of preictal states from EEG is crucial for epilepsy management.
• The specific contribution of temporal modeling techniques to seizure prediction accuracy remains unclear.

Purpose of the Study:

• To conduct a controlled comparison of Convolutional Neural Network (CNN) and CNN-LSTM architectures for EEG-based seizure prediction.
• To evaluate the impact of temporal modeling on cross-patient performance using the CHB-MIT dataset.

Main Methods:

• Utilized the CHB-MIT EEG dataset for analysis.
• Employed leave-one-subject-out cross-validation to assess generalizability.
• Compared CNN and CNN-LSTM models under identical conditions.

Main Results:

• Temporal modeling significantly improved overall performance (Macro F1: 0.66 to 0.72, p < 0.05).
• Preictal recall saw the largest improvement (+11%) with temporal modeling.
• Performance gains saturated at moderate temporal context lengths (5-10 steps).
• Substantial cross-patient variability persisted, highlighting generalization as a bottleneck.

Conclusions:

• Temporal modeling offers bounded improvements in seizure prediction.
• Improving cross-subject robustness is more critical than increasing temporal complexity.
• The CNN-LSTM model presents a moderate computational overhead (∼50%) but remains suitable for real-time applications.