Exploring the Temporal Structure of Scalp EEG for Epilepsy Seizure Prediction

Accurate and reliable prediction of epileptic seizures has potential in epilepsy monitoring, diagnosis, and rehabilitation. Current mainstream deep learning models treat the epileptic seizure prediction problem as a binary classification task, where hard labels are assigned to all signal segments for modeling. However, since the signal classes are manually selected, this approach can lead to many false positives during classification. Moreover, using hard labels that do not vary over time results in the inability to maintain the accuracy’s monotonicity over time during prediction. This means that the system can only provide results after a certain period once the alarm has started, which may cause anxiety for patients. This paper proposes a new modeling approach that uses early event prediction to model the labels of signals preceding seizure onset. Additionally, an innovative introduction of a large EEG model into the seizure prediction task is presented. The proposed model is validated on the CHB-MIT dataset, achieving an overall sensitivity of 94.8% and reducing false positive alarms to 0.03. Comparison with state-of-the-art methods indicates that the proposed model exceeds previous prediction performance. Furthermore, our method demonstrates the potential for maintaining accuracy monotonicity over time for certain patients, allowing them to predict seizures based on the model’s time-based judgments, thereby alleviating anxiety.