An end-to-end seizure prediction approach using long short-term memory network

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy affects many patients, with seizure onsets causing significant pain and reduced quality of life.
• Current epileptic seizure prediction methods often rely on complex, subjective feature extraction, limiting their effectiveness.
• Developing sensitive and reliable seizure prediction tools is crucial for patient management.

Purpose of the Study:

• To propose an end-to-end deep learning approach for epileptic seizure prediction.
• To overcome the limitations of subjective and complex feature engineering in existing methods.
• To improve the sensitivity and reduce false predictions for epileptic seizures.

Main Methods:

• Utilized a long short-term memory (LSTM) network for an end-to-end seizure prediction model.
• Inputted only the gamma band of raw electroencephalography (EEG) signals directly into the LSTM network.
• Incorporated the postictal stage, in addition to preictal and interictal stages, for enhanced classification.

Main Results:

• Achieved a mean sensitivity of 91.76% and a false prediction rate (FPR) of 0.29/h using the CHB-MIT database.
• Further improved performance by including the postictal stage, reaching a sensitivity of 92.17% and an FPR of 0.27/h.
• Provided a mean warning time of 44.46 minutes, allowing for timely intervention.

Conclusions:

• The proposed LSTM-based method offers a simplified and effective approach to epileptic seizure prediction.
• Directly using gamma band EEG signals with LSTM avoids complex feature design and subjectivity.
• The inclusion of the postictal stage enhances prediction accuracy and provides clinically relevant warning times.