Mixture of checkpoint experts for explainable seizure detection using wearable devices

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Neurology

Background:

• In-hospital video-Electroencephalography (vEEG) is the standard for epilepsy seizure detection but is resource-intensive.
• Wearable devices offer a feasible alternative for long-term, at-home seizure monitoring.
• Existing machine learning (ML) models often function as black boxes, hindering clinical trust and auditability.

Purpose of the Study:

• To introduce a novel, transparent ML algorithm, Mixture of Checkpoint Experts (MoCE), for seizure detection.
• To evaluate MoCE's performance and explainability compared to traditional black-box models.
• To demonstrate the clinical utility of transparent ML in epilepsy management.

Main Methods:

• Development of the Mixture of Checkpoint Experts (MoCE) ML algorithm.
• Deployment of wrist-worn devices for data collection in 14 epilepsy patients.
• Comparative analysis of MoCE against existing neural networks for seizure detection accuracy and false alarm rates.

Main Results:

• MoCE demonstrated statistically significant improvement in false alarm rate compared to existing neural networks.
• MoCE achieved equivalent recall performance while offering enhanced model transparency.
• The algorithm provided global and local insights into model behavior, aiding in prediction auditing.

Conclusions:

• MoCE offers a transparent and effective alternative to black-box ML models for seizure detection.
• The algorithm's explainability enhances clinician and data scientist trust and utility in real-world epilepsy monitoring.
• MoCE represents a significant advancement in leveraging wearable technology and ML for epilepsy care.