TY - JOUR
T1 - RISC-V CNN Coprocessor for Real-Time Epilepsy Detection in Wearable Application
AU - Lee, Shuenn Yuh
AU - Hung, Yi Wen
AU - Chang, Yao Tse
AU - Lin, Chou Ching
AU - Shieh, Gia Shing
N1 - Funding Information:
Manuscript received March 21, 2021; revised May 16, 2021 and June 15, 2021; accepted June 17, 2021. Date of publication June 28, 2021; date of current version September 13, 2021. This work was supported in part by the Taiwan Semiconductor Research Institute, and the Ministry of Science and Technology (MOST), Taiwan, R.O.C., under Grant MOST 109-2218-E-006-022. (Corresponding author: Shuenn-Yuh Lee.) Shuenn-Yuh Lee, Yi-Wen Hung, and Yao-Tse Chang are with the Electrical Engineering Department of National Cheng Kung University, Tainan 70101, Taiwan (e-mail: [email protected]; [email protected]; [email protected]).
Publisher Copyright:
© 2007-2012 IEEE.
PY - 2021/8
Y1 - 2021/8
N2 - Epilepsy is a common clinical disease. Severe epilepsy can be life-threatening in certain unexpected conditions, so it is important to detect seizures instantly with a wearable device and to provide treatment within the golden window. The observation of the electroencephalography (EEG) signal is an imperative method to assist correct epilepsy diagnosis. To detect and classify EEG signals, a convolutional neural network (CNN) is an intuitive and appropriate method that borrows expertise from neurologists. However, the computational cost of training and inference on artificial intelligence (AI)-based solutions make software-only and hardware-only solutions incompetent for real-time monitoring on embedded devices. Hence, this study proposes three key contributions for the challenge, namely, an algorithm framework to provide real-time epilepsy detection, a dedicated coprocessor chip implementing this framework to enable real time epilepsy detection to offload and accelerate detection algorithm, and a custom interface with the coprocessor and reduced instruction set computer-V (RISC-V) instructions to reconfigure the coprocessor and transfer data. The epilepsy detection framework is implemented in 11-layer CNN. The proposed epilepsy detection algorithm performs 97.8% accuracy for floating-point and 93.5% for fixed-point operations through animal experiments with lab rats. The RISC-V CNN coprocessor is fabricated in the TSMC 0.18-μm CMOS process. For each classification, the coprocessor consumes 51 nJ/class. and 0.9 μJ/class. energy on data transfer and inference, respectively. The detection latency on the chip is 0.012 s. With the integration of the hardware coprocessor, AI algorithms can be applied to epilepsy detection for real-time monitoring.
AB - Epilepsy is a common clinical disease. Severe epilepsy can be life-threatening in certain unexpected conditions, so it is important to detect seizures instantly with a wearable device and to provide treatment within the golden window. The observation of the electroencephalography (EEG) signal is an imperative method to assist correct epilepsy diagnosis. To detect and classify EEG signals, a convolutional neural network (CNN) is an intuitive and appropriate method that borrows expertise from neurologists. However, the computational cost of training and inference on artificial intelligence (AI)-based solutions make software-only and hardware-only solutions incompetent for real-time monitoring on embedded devices. Hence, this study proposes three key contributions for the challenge, namely, an algorithm framework to provide real-time epilepsy detection, a dedicated coprocessor chip implementing this framework to enable real time epilepsy detection to offload and accelerate detection algorithm, and a custom interface with the coprocessor and reduced instruction set computer-V (RISC-V) instructions to reconfigure the coprocessor and transfer data. The epilepsy detection framework is implemented in 11-layer CNN. The proposed epilepsy detection algorithm performs 97.8% accuracy for floating-point and 93.5% for fixed-point operations through animal experiments with lab rats. The RISC-V CNN coprocessor is fabricated in the TSMC 0.18-μm CMOS process. For each classification, the coprocessor consumes 51 nJ/class. and 0.9 μJ/class. energy on data transfer and inference, respectively. The detection latency on the chip is 0.012 s. With the integration of the hardware coprocessor, AI algorithms can be applied to epilepsy detection for real-time monitoring.
UR - http://www.scopus.com/inward/record.url?scp=85112244944&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85112244944&partnerID=8YFLogxK
U2 - 10.1109/TBCAS.2021.3092744
DO - 10.1109/TBCAS.2021.3092744
M3 - Article
C2 - 34181550
AN - SCOPUS:85112244944
SN - 1932-4545
VL - 15
SP - 679
EP - 691
JO - IEEE transactions on biomedical circuits and systems
JF - IEEE transactions on biomedical circuits and systems
IS - 4
ER -