NSF-NSERC: AI Chip for Seizure Monitoring with an Ear-based Wearable Device

About This Grant

Epilepsy care primarily involves periodic visits to a neurology clinic for assessment and refinement of medications. A key gap in care is the inability to robustly monitor seizure activity outside of the clinic. Currently available wearable devices include wrist-based seizure detection devices that cannot record brain activity and suffer from poor sensitivity. At-home brain recordings with bulky full-scalp electroencephalography (EEG) caps are not practical for chronic use. To meet these challenges, this project is developing a wireless ear-based wearable, or “earable,” for monitoring neural activity using EEG electrodes behind each ear. The familiar form factor, similar to commercial earphones, will be more compatible with use during activities of daily living and can reduce the stigma that often accompanies wearable medical devices. The earable will feature a state-of-the-art custom integrated circuit with programmable artificial intelligence (AI) to infer seizure activity from the ear EEG signals. It will provide low-power wireless integration with a mobile app to alert the user and caregivers when seizure events are occurring or likely to occur. Access to remote seizure monitoring could improve refinement of medication while decreasing health care costs by requiring fewer in-person visits to the neurology clinic. A promising recent approach to epilepsy monitoring is to use ear-based wearables due to their access to EEG signals and unobtrusive form factor. It has been demonstrated that a limited behind-the-ear EEG device can have the same seizure detection sensitivity and false detection rate as a full EEG cap. However, robust real-time seizure monitoring with earables has yet to be achieved. Engineering challenges hamper practical, everyday use. Key issues include miniaturization, power efficiency, and real-time edge AI capability to detect critical features in noisy ear EEG signals. This project aims to: (1) collect an unprecedented dataset in treatment-refractory epilepsy patients comprising ear-EEG and intracranial EEG, the latter being the gold standard for seizure localization, (2) develop a novel two-stage edge deep learning model for ear-EEG-based seizure detection, achieving high sensitivity and low false positives by leveraging transfer learning, data augmentation, and knowledge distillation; and (3) design a programmable AI chip featuring dynamic quantization, mixed-precision computation, specialized processing element arrays, and optimized data flow scheduling, enabling ultra-low power inference with minimal latency. The resulting dataset, algorithm, and hardware can individually advance current knowledge in their respective domains. Together, they could establish a path toward improved remote care and quality-of-life for individuals with epilepsy. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.