A Deep Convolutional Neural Network Method to Detect Seizures and Characteristic Frequencies Using Epileptic...

Area of Science:

• Computational Neuroscience
• Medical Technology

Background:

• Deep learning for electroencephalogram (EEG)-based seizure detection shows promise but is limited by classifier design and signal representation.
• Existing methods struggle to effectively harness deep learning for accurate seizure diagnosis.

Purpose of the Study:

• To design and evaluate deep convolutional neural network (CNN) classifiers for improved seizure detection.
• To develop effective signal-to-image conversion methods for EEG data input into classifiers.
• To identify the most accurate classification approach for seizure detection using EEG.

Main Methods:

• Proposed signal-to-image conversion to transform time-domain EEG signals into time-frequency images.
• Developed and evaluated three classification methods with five distinct classifiers.
• Utilized Shapley Additive exPlanations (SHAP) to analyze feature frequency contributions.

Main Results:

• The proposed FT-VGG16 classifier achieved a leading average accuracy of 99.21%, outperforming previous studies.
• The study identified specific EEG frequency components crucial for accurate seizure classification.
• This is the first study to quantify the contribution of frequency components to seizure classification.

Conclusions:

• Developed deep CNN models effectively detect seizures and characteristic frequencies from EEG data.
• The models demonstrate clinical applicability for automated seizure detection.
• Identified seizure-related EEG frequencies offer insights into seizure mechanisms.