An Event-Driven Neural Signal Processor for Closed-Loop Seizure Prediction
Student Contest:
Yes
Affiliation Type:
Academia
Keywords:
event-driven, hardware reuse, epileptic seizure prediction, deep neural network, digital chip design
Abstract:
Epileptic seizure prediction which enables timely intervention, is critical for closed-loop seizure controls. However, real-time neural signal processing, one of the effective approaches to address seizure prediction, poses a significant challenge due to its high power consumption, especially for implantable medical devices. In this paper, we propose a neural signal processor featuring an event-driven processing manner, designed for low-power, high-precision real-time epileptic seizure prediction. Our system monitors neural signals using a binary neural network (BNN) to detect any possible abnormal events and reconfigures to a convolutional neural network (CNN) for high-precision prediction and to reject false alarms. Validated with real-world neural signals from patients with epilepsy, our system demonstrates an F1 score of 0.97 in predicting seizures, outperforming existing state-of-the-art approaches. Thanks to the event-driven design, the average power consumption of our system is less than 200 µW, making it suitable for implantable closed-loop medical devices.
Track ID:
18.8
Track Name:
AI-Assisted Closed-loop Neuromodulation Systems for Neuroscience Applications
