Enhanced Non-EEG Multimodal Seizure Detection: A Real-World Model for Identifying Generalised Seizures Across the...

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Non-electroencephalogram (non-EEG) seizure detection offers potential for early identification of generalized onset seizures.
• Existing non-EEG methods struggle with high false alarm rates and differentiating normal movements from seizures.

Purpose of the Study:

• To develop and evaluate AMBER (Attention-guided Multi-Branching pipeline with Enhanced Residual Fusion), a novel multimodal seizure detection model.
• To improve the accuracy of Ictal-Phase Detection using non-EEG data, specifically acceleration and heart rate.

Main Methods:

• Utilized the Open Seizure Database, analyzing 94 events (approx. 5.5 hours) with clinician-annotated 5-second timesteps (Normal, Pre-Ictal, Ictal).
• Developed AMBER, a multimodal model with independent feature extraction branches for each sensing modality (acceleration, heart rate).
• Employed a Residual Fusion layer to combine features, followed by densely connected blocks for classification.

Main Results:

• AMBER achieved an accuracy of 0.9027 and an F1-score of 0.9035 on unseen test data.
• Demonstrated high True Positive Rates: 0.8342 (Normal), 0.9485 (Pre-Ictal), and 0.9118 (Ictal).
• Achieved a low average False Positive Rate of 0.0502.

Conclusions:

• The proposed Ictal Phase Detection technique enhances seizure phase classification accuracy.
• AMBER demonstrates reduced false alarms, paving the way for advanced non-EEG seizure detection systems.
• This research contributes to improved seizure monitoring and management through innovative machine learning approaches.