Automatic Detection and Classification of Epileptic Seizures from EEG Data: Finding Optimal Acquisition Settings and...

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Spontaneous seizures are often missed or misclassified.
• Deep learning (DL) offers a promising approach for automated seizure detection and differentiation.
• Existing methods require robust and accurate analysis of electroencephalography (EEG) data.

Purpose of the Study:

• To develop and evaluate a DL system for accurate seizure detection and classification.
• To investigate the impact of EEG sampling rate and electrode count on model performance.
• To enhance model interpretability using machine learning techniques.

Main Methods:

• A system architecture using top-performing DL models was developed.
• The non-overlapping window technique was applied to the TUSZ dataset.
• Model performance was analyzed with varying EEG sampling rates (50-250 Hz) and electrode numbers (8-21).
• Interpretable ML techniques, including activation maximization and cortical source projection, were used.

Main Results:

• The system achieved high accuracy in seizure detection (87.7% Sn, 91.16% Sp) and differentiation (95-100% Acc).
• Increasing EEG sampling rate significantly improved seizure detection and differentiation precision.
• Reducing electrodes impacted seizure differentiation more than detection, highlighting varied informative value.
• Interpretable ML confirmed the system recognizes biologically relevant epileptic features.

Conclusions:

• The proposed DL system demonstrates high efficacy in detecting and differentiating seizure types from EEG data.
• EEG sampling rate is a critical factor for reliable DL model training in epilepsy diagnosis.
• Electrode selection influences seizure differentiation accuracy, suggesting specific electrode relevance.
• Interpretable ML validates the clinical relevance of the DL model's findings in epilepsy.