Seizure Detection and Prediction by Parallel Memristive Convolutional Neural Networks

Area of Science:

• Neurology
• Computer Engineering
• Materials Science

Background:

• Epileptic seizure detection and prediction algorithms have advanced significantly over the last two decades.
• Hardware implementation of these algorithms faces challenges in power and area-constrained environments, especially with multiple recording channels.
• Conventional technologies like Complementary Metal-Oxide-Semiconductor (CMOS) present limitations for efficient deployment.

Purpose of the Study:

• To propose a novel low-latency parallel Convolutional Neural Network (CNN) architecture for epileptic seizure detection and prediction.
• To implement this CNN on analog crossbar arrays using Resistive Random-Access Memory (RRAM) devices.
• To provide a comprehensive hardware benchmark, evaluating latency, power, and area.

Main Methods:

• Developed a parallel CNN architecture with significantly fewer parameters than State-Of-The-Art (SOTA) models.
• Implemented the CNN on analog crossbar arrays of RRAM devices.
• Utilized Quantization Aware Training (QAT) and a stuck weight offsetting methodology to address non-idealities and device variations.
• Simulated, laid out, and determined hardware requirements for the CNN component.

Main Results:

• Achieved high accuracy: 99.84% for detection, 99.01% and 97.54% for prediction across multiple datasets (Bonn, CHB-MIT, SWEC-ETHZ).
• Reduced network parameters by 2-2,800x compared to SOTA CNNs.
• Achieved a 2-orders of magnitude reduction in latency through parallel execution on analog crossbars.
• Recovered up to 32% accuracy using the stuck weight offsetting methodology without retraining.
• Estimated power consumption of 2.791 W and area of 31.255 mm² in a 22 nm FDSOI CMOS process.

Conclusions:

• The proposed RRAM-based parallel CNN offers a highly efficient solution for real-time epileptic seizure detection and prediction.
• This hardware-efficient approach overcomes limitations of conventional CMOS technologies in power and area-constrained settings.
• The system demonstrates robustness against non-idealities and device variations, paving the way for practical wearable epilepsy monitoring devices.