Robust deep network with maximum correntropy criterion for seizure detection.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Effective seizure detection from long-term electroencephalogram (EEG) is crucial for epilepsy diagnosis.
• Current methods often optimize features and classifiers independently, limiting overall system performance.
• EEG signals are susceptible to impulsive noise and outliers from electromyography (EMG) artifacts.

Purpose of the Study:

• To develop a deep learning network for simultaneous optimization of feature learning and classification in EEG seizure detection.
• To enhance the robustness of feature extraction against noise and outliers in EEG data.

Main Methods:

• A deep network architecture was designed to jointly learn features and a classifier.
• A robust stacked autoencoder (R-SAE) incorporating the maximum correntropy criterion (MCC) was utilized for noise-resilient feature extraction.
• The MCC was employed to mitigate the impact of impulsive noises and outliers by down-weighting their influence compared to mean square error (MSE).

Main Results:

• The proposed method achieved 100% sensitivity and 99% specificity on 33.6 hours of scalp EEG data from six patients.
• The R-SAE effectively reduced the influence of EMG artifacts and other impulsive noises.
• Simultaneous optimization of feature learning and classification led to improved seizure detection performance.

Conclusions:

• The proposed deep network with MCC-based robust feature learning offers a promising approach for accurate and reliable seizure detection.
• This method demonstrates significant potential for clinical application in epilepsy diagnosis and patient monitoring.
• Joint optimization of feature extraction and classification is a viable strategy for improving EEG signal analysis.