Automatic detection of epilepsy from EEGs using a temporal convolutional network with a self-attention layer.

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Epilepsy affects over 60% of children globally, necessitating early diagnosis and treatment.
• Automated epilepsy detection from EEG is crucial but challenged by unpredictable seizure occurrences during exams.
• Existing models may show inflated performance due to reliance on seizure-specific data, highlighting the need for robust, universally applicable solutions.

Purpose of the Study:

• To develop a universally applicable model for automated epilepsy detection in real-world clinical scenarios.
• To address limitations of existing algorithms that risk artificially enhanced performance metrics.
• To improve the accuracy and reliability of epilepsy diagnosis in pediatric patients.

Main Methods:

• A novel Temporal Convolutional Neural Network with Self-Attention (TCN-SA) model was developed.
• The TCN component extracts time-variant features from EEG signals.
• A self-attention (SA) layer prioritizes critical features for enhanced classification accuracy.

Main Results:

• The TCN-SA model achieved high accuracies: 95.50% on a pediatric dataset and 97.37% (A v. E) / 93.50% (B vs E) on the Bonn dataset.
• Demonstrated superior performance compared to other deep learning architectures, including standard TCN, SA networks, and CNNs.
• Validated efficacy on both a custom pediatric epilepsy dataset and the established Bonn dataset.

Conclusions:

• The TCN-SA approach is a highly effective tool for automated epilepsy detection.
• Its proven performance offers significant benefits in diverse and complex real-world clinical settings.
• The model holds potential for improving epilepsy diagnosis and patient outcomes.