Energy-Efficient Neural Network for Epileptic Seizure Prediction.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Neurology

Background:

• Epilepsy affects millions globally, with drug-refractory cases posing significant challenges.
• Current deep learning seizure prediction models are resource-intensive, limiting their use in real-time, low-power devices.
• Wearable or implantable devices are needed for continuous, real-time seizure monitoring and prediction.

Purpose of the Study:

• To develop energy-efficient and hardware-friendly methods for predicting epileptic seizures.
• To create compact seizure prediction models suitable for resource-constrained wearable and implantable devices.
• To evaluate model performance across multiple public epilepsy datasets.

Main Methods:

• Utilized neural architecture search to develop a compact seizure prediction model (45 kB).
• Applied model compression techniques to further reduce model size for scalp and intracranial EEG data.
• Evaluated model sensitivity, false prediction rate, and AUC across three public datasets.
• Estimated energy consumption per inference for scalp and intracranial EEG data.

Main Results:

• Achieved high sensitivity (up to 99.81%) and low false prediction rates (as low as 0.005/h) across datasets.
• Developed models <50 kB for scalp EEG and <10 kB for intracranial EEG, outperforming similar-sized models.
• Demonstrated low energy consumption: <10 mJ/inference for scalp EEG and <0.5 mJ/inference for intracranial EEG.
• Obtained AUC values up to 1 for scalp EEG and 0.977 for intracranial EEG.

Conclusions:

• Developed highly efficient, hardware-friendly seizure prediction models suitable for low-power devices.
• The compact models meet the stringent energy requirements for wearable and implantable epilepsy monitoring.
• These advancements promise improved quality of life and reduced anxiety for epilepsy patients through real-time prediction.