Efficient and Accurate Epilepsy Seizure Prediction and Detection Based on Multi-Teacher Knowledge Distillation...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Neurology

Background:

• Epileptic seizures are unpredictable, posing challenges for continuous monitoring.
• Current deep learning models for seizure prediction/detection are computationally intensive, limiting wearable device application.
• High computational costs and latency hinder real-time seizure prediction on wearables.

Purpose of the Study:

• To develop a lightweight deep learning model for unified seizure prediction and detection.
• To enable real-time epilepsy monitoring on resource-constrained wearable devices.
• To address the computational and latency limitations of existing seizure prediction models.

Main Methods:

• Integration of Feature-wise Linear Modulation (FiLM) with Ring-Buffer Gated Recurrent Unit (Ring-GRU) for causal consistency.
• Utilizing a multi-teacher knowledge distillation strategy to transfer knowledge to a lightweight student model.
• Development of the RGF-Model, a unified causal framework for seizure prediction and detection.

Main Results:

• The RGF-Model demonstrates superior efficiency compared to state-of-the-art teacher models on CHB-MIT and Siena datasets.
• Achieved 99.54% AUC and 0.01 FPR/h for prediction, and 98.78% Accuracy for detection on CHB-MIT.
• The model has only 0.082 million parameters, significantly reducing complexity while maintaining accuracy.

Conclusions:

• The RGF-Model offers a highly efficient and accurate solution for real-time epilepsy monitoring.
• This lightweight network is suitable for deployment on wearable devices.
• The unified causal framework effectively addresses computational and latency challenges.