A High Accuracy and Ultra-Energy-Efficient Zero-Shot-Retraining Seizure Detection Processor

Hits:

Indexed by:Journal paper

First Author:J. Liu

Correspondence Author:C. Wang, and J. Zhou

Co-author:X. Liu, X. Wang, Z. Xie, C. Guo, Z. Zhong, J. Fan, H. Qiu, Y. Xu, H. Qin, Y. Long, L. Zhou, Z. Shen, L. Chang, S. Liu, S. Lin,

Journal:IEEE Solid-State Circuits Journal (JSSC 2024)

Included Journals:SCI

Discipline:Engineering

First-Level Discipline:Electronic Science And Technology

Document Type:J

Date of Publication:2024-09-20

Impact Factor:4.6

Abstract:Recently, wearable devices integrating seizure detection processors have been developed to detect seizures in real time for alerting, recording, or in-device treatment purposes. High accuracy and low energy consumption are paramount for seizure detection processors. Many existing seizure detection processors are able to achieve high accuracy when large amounts of seizure data from the test patient is available for the training, which requires the test patient to undergo time-consuming and costly hospitalization due to the low occurrence of seizure data and therefore is impractical. This work proposes a high-accuracy and ultra-energy-efficient zero-shot-retraining seizure detection processor that requires no seizure data from the test patient for retraining. Two novel techniques have been proposed to improve the accuracy and reduce energy consumption, including a hybrid-feature-driven adaptive processing architecture with on-chip learning for improving the accuracy against inter-patient variation and reduce the classification energy consumption and a learning-based adaptive channel selection technique to identify the redundant electroencephalogram (EEG) channels for further energy saving while maintaining high accuracy. The proposed seizure detection processor has been implemented and fabricated in 55-nm CMOS technology. It demonstrates high sensitivity (i.e., 100% event-based sensitivity) and specificity (i.e., 94%) with extremely low energy consumption (i.e., 0.07 μ J/classification and 0.1 μ J/learning) while requiring no seizure data from the test patient for retraining, outperforming the state-of-the-art designs.