Seizure detection using ultra-long-term subcutaneous electroencephalography: A deep learning CNN-BiLSTM approach.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Technology

Background:

• Ultra-long-term monitoring in epilepsy can yield novel insights but generates massive electroencephalographic (EEG) data.
• Manual review of extensive EEG data is impractical for identifying seizure patterns.
• Automated seizure detection is crucial for analyzing large datasets and enabling new therapeutic strategies.

Purpose of the Study:

• To develop and validate a deep learning algorithm for automated seizure detection.
• To utilize two-channel subscalp electroencephalographic (EEG) recordings for enhanced epilepsy monitoring.
• To improve the efficiency and accuracy of seizure detection compared to conventional methods.

Main Methods:

• A nine-layer convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) hybrid algorithm was developed.
• The algorithm processed 5-minute spectrograms with 50% overlap from retrospective subscalp EEG data of 16 patients.
• Performance was evaluated against a spectral power classifier and compared using intrapatient training-testing approaches.

Main Results:

• The CNN-BiLSTM algorithm trained on scalp and subscalp EEG achieved an AUROC of 0.98 and AUPRC of 0.50 (94% sensitivity, 1.11 false detections/day).
• Models trained solely on intracranial EEG (iEEG) or using only the first 45% of subscalp data showed lower performance.
• The CNN-BiLSTM detector significantly outperformed a conventional spectral band power detector.

Conclusions:

• A CNN-BiLSTM framework enables highly sensitive and specific automated seizure detection in two-channel subscalp EEG.
• Training with subcutaneous EEG data yields superior performance compared to intracranial EEG data.
• Augmenting training data with scalp EEG seizures proved beneficial for algorithm performance.