Video-based detection of tonic-clonic seizures using a three-dimensional convolutional neural network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Automated video analysis aids seizure detection in epilepsy monitoring units (EMUs).
• Machine learning reduces resources needed for diagnostic monitoring of drug-resistant epilepsy.

Purpose of the Study:

• Employ a 3D convolutional neural network (3D-CNN) for automated seizure detection.
• Identify seizures from EMU videos using a fine-tuned model.

Main Methods:

• Utilized a two-stream inflated 3D-ConvNet (I3D) architecture for video classification.
• Fine-tuned a pretrained action classifier on 11 hours of video data with 49 tonic-clonic seizures.
• Validated performance using leave-one-patient-out cross-validation and external site data.

Main Results:

• Achieved a leave-one-patient-out cross-validation F1-score of 0.960 ± 0.007 and AUC of 0.988 ± 0.004 at site A.
• Detected all seizures with a median latency of 0.0 seconds; average false alarm rate of 1.81 alarms/hour.
• Demonstrated generalizability at site B, though cross-site evaluation showed diminished performance.

Conclusions:

• The fine-tuned I3D model shows high performance in epileptic seizure detection from video.
• Outperforms existing models, providing a foundation for real-time EMU and potential at-home monitoring.
• Suggests a reliable and cost-effective approach for detecting tonic-clonic seizures.