Video-Based Detection of Generalized Tonic-Clonic Seizures Using Deep Learning.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Timely seizure detection is critical for epilepsy management and reducing risks like sudden unexpected death in epilepsy (SUDEP).
• Traditional automated seizure detection relies on hand-designed features, which may not capture all seizure complexities.
• Deep learning approaches offer automated feature extraction, showing promise in various classification tasks.

Purpose of the Study:

• To assess the feasibility and efficacy of deep learning for automated detection of generalized tonic-clonic seizures (GTCSs) from video data.
• To compare the performance of frame-based (CNN) and sequence-based (CNN+LSTM) deep learning models for GTCS detection.
• To evaluate the potential of deep learning as an alternative to traditional feature-based methods.

Main Methods:

• Retrospective analysis of 76 GTCS videos and seizure-free recordings from 37 participants undergoing long-term video-EEG monitoring.
• Utilized convolutional neural networks (CNNs) for frame-level analysis and CNN+long short-term memory (LSTM) networks for video sequence analysis.
• Employed a leave-one-subject-out cross-validation (LOSO-CV) approach to evaluate model generalizability.

Main Results:

• CNN+LSTM networks analyzing video sequences achieved superior performance compared to frame-based CNNs.
• The CNN+LSTM model demonstrated a mean sensitivity of 88% and a mean specificity of 92% across patients.
• Average detection latency was 22 seconds, with performance improving with larger training datasets.

Conclusions:

• Automated video-based GTCS detection using deep learning is feasible and effective.
• Deep learning methods overcome limitations of hand-crafted features and can serve as a benchmark for future research.
• Further improvements in detection accuracy are expected with the availability of larger datasets.