Seizure prediction based on classification of EEG synchronization patterns with on-line retraining and...

Area of Science:

• Neuroscience
• Machine Learning
• Medical Technology

Background:

• Epilepsy is a common neurological disorder impacting daily life due to unpredictable seizures.
• Electroencephalography (EEG) signal variations (e.g., wakefulness, sleep) present a major challenge for seizure prediction.
• Traditional offline training methods struggle with EEG signal variability, leading to performance degradation over time.

Purpose of the Study:

• To address the variation issue in EEG-based seizure prediction.
• To propose and evaluate an online retraining method for seizure prediction models.
• To reduce false alarms in seizure prediction through a post-processing scheme.

Main Methods:

• Developed an online retraining approach that gradually enlarges the training set with recent data.
• Incorporated a post-processing scheme to minimize false seizure alarms.
• Utilized a machine learning-based classification of bivariate patterns.

Main Results:

• Achieved 74.2% sensitivity on Electrocorticogram (ECoG) data and 52.2% sensitivity on scalp EEG data.
• Improved sensitivity by 29.0% (ECoG) and 17.4% (scalp EEG) compared to offline training methods.
• Demonstrated significant performance gains through online retraining.

Conclusions:

• Online retraining effectively leverages recent data to enhance seizure prediction reliability.
• The proposed method shows promise for developing more robust and dependable seizure prediction systems.
• Continuous model adaptation is crucial for overcoming EEG signal variability in epilepsy management.