Anchoring temporal convolutional networks for epileptic seizure prediction.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Accurate epilepsy seizure prediction is vital for patient management and preventing seizures.
• Current short-term prediction models are insufficient as prediction times are shorter than antiepileptic drug onset.
• Longer-term prediction is challenging due to the similarity between preictal and interictal periods.

Purpose of the Study:

• To develop a patient-specific, long-term epilepsy seizure prediction model.
• To improve the accuracy and timeliness of seizure prediction alarms.
• To enhance the quality of life for epilepsy patients through better seizure management.

Main Methods:

• Feature extraction from electroencephalography (EEG) signals using sample entropy.
• Development of an anchoring temporal convolutional networks (ATCN) model for prediction.
• Utilizing dilated causal convolutional networks and anchoring data for enhanced performance.
• Implementation of a multilayer sliding window prediction algorithm for seizure alarms.

Main Results:

• Achieved 100% sensitivity and 98.92 min average prediction time (APT) on the Freiburg intracranial EEG dataset.
• Achieved 97.44% sensitivity and 93.54 min APT on the CHB-MIT scalp EEG dataset.
• Demonstrated a low false prediction rate (0.09-0.12 per hour).

Conclusions:

• The proposed approach enables adequate long-term seizure prediction using both intracranial and scalp EEG data.
• The extended APT surpasses typical antiepileptic drug onset times, allowing for proactive intervention.
• This method can potentially prevent seizures and improve patient quality of life, particularly for those on timed medication.