Epileptic Signal Classification with Deep Transfer Learning Feature on Mean Amplitude Spectrum

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy affects 6% of the global population, necessitating improved classification methods.
• Current epileptic state classification (preictal, ictal, interictal) lacks granularity, particularly in the preictal phase.
• The preictal stage, typically defined as one hour before seizure onset, requires finer temporal resolution for practical applications.

Purpose of the Study:

• To develop a novel deep learning method for granular classification of the preictal stage in epilepsy.
• To enhance epileptic seizure detection by analyzing Electroencephalogram (EEG) signals at a finer time scale.
• To improve the accuracy of epileptic seizure prediction through advanced feature extraction and transfer learning.

Main Methods:

• Utilized multichannel Electroencephalogram (EEG) data from the CHI-MIT epilepsy EEG database.
• Computed subband mean amplitude spectrum maps (MAS) for EEG signal representation.
• Employed transfer learning with three popular deep convolutional neural networks (CNNs) for feature extraction.

Main Results:

• The proposed algorithm achieved a highest overall accuracy of 92.77%.
• Optimal performance was observed when the one-hour preictal stage was divided into 20-minute segments.
• Demonstrated the effectiveness of CNN-based transfer learning for EEG feature extraction in epilepsy.

Conclusions:

• Finer temporal resolution in preictal stage analysis significantly improves epilepsy classification accuracy.
• The developed deep EEG feature extraction method using CNN transfer learning is effective for granular seizure detection.
• This approach offers a promising advancement for practical epilepsy management and seizure prediction systems.