Deep-learning-based seizure detection and prediction from electroencephalography signals.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Manual electroencephalography (EEG) analysis for epilepsy diagnosis is time-consuming and has low inter-rater agreement.
• Automated systems are needed for faster diagnosis, reduced errors, and timely seizure prediction.
• Existing methods often focus on binary classification, limiting comprehensive EEG signal analysis.

Purpose of the Study:

• To develop and evaluate effective approaches for classifying EEG signals into normal, pre-ictal, and ictal activities.
• To introduce patient-specific and patient-non-specific models for seizure detection and prediction.
• To present a generalized three-class classification framework for comprehensive EEG analysis.

Main Methods:

• Three models were developed: two Convolutional Neural Network (CNN) models using spectrograms (13-layer and 3-layer), and a third model employing Phase Space Reconstruction (PSR) with a 5-layer CNN.
• The first two models performed binary classification for seizure prediction (normal vs. pre-ictal) and detection (normal vs. ictal).
• The third model utilized PSR for direct time-domain projection, addressing spectrogram limitations, and was tested for a three-class classification (normal, pre-ictal, ictal) on the CHB-MIT dataset.

Main Results:

• The PSR-based CNN model demonstrated superior performance compared to the spectrogram-based CNN models and existing state-of-the-art methods.
• The patient-non-specific PSR model proved effective for general EEG classification tasks.
• The study successfully classified normal, pre-ictal, and ictal EEG activities, paving the way for improved epilepsy management.

Conclusions:

• The proposed Phase Space Reconstruction (PSR) method combined with a Convolutional Neural Network (CNN) offers a robust and superior approach for EEG signal classification in epilepsy.
• This generalized three-class classification framework enhances the potential for accurate seizure detection and prediction, improving patient care.
• The patient-non-specific nature of the best-performing model suggests broad applicability in clinical settings for epilepsy diagnosis and management.