Epileptic Seizure Detection on an Ultra-Low-Power Embedded RISC-V Processor Using a Convolutional Neural Network

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Medical Devices

Background:

• Refractory epilepsy treatment demands implantable devices with low power consumption and efficient processing.
• Convolutional Neural Networks (CNNs) offer a promising approach for analyzing brain signals for epileptic seizure detection.

Purpose of the Study:

• To develop and implement an ultra-low-power CNN for epileptic seizure detection on a microprocessor.
• To evaluate the CNN's performance using the CHB-MIT dataset and real-world recordings.

Main Methods:

• Implementation and optimization of a CNN in MATLAB for epileptic seizure detection.
• Deployment of the CNN on a GAP8 microprocessor with RISC-V architecture.
• Validation using the CHB-MIT EEG dataset and recordings from epileptic rats.

Main Results:

• The CNN achieved a median sensitivity of 90% and specificity over 99% on the CHB-MIT dataset.
• On the microcontroller, the CNN achieved 85% sensitivity, classifying 1s of EEG data in 35ms with ~140μW power consumption.
• The detector demonstrated a 6x improvement in power efficiency compared to related methods.

Conclusions:

• The developed ultra-low-power CNN is suitable for real-time epileptic seizure detection on implantable devices.
• This technology facilitates the creation of next-generation, energy-efficient medical devices for epilepsy management.
• The CNN's effectiveness was confirmed in both human and animal models, highlighting its broad applicability.