Reconfigurable and Scalable Artificial Intelligence Acceleration Hardware Architecture With RISC-V CNN Coprocessor for Real-Time Seizure Detection

Epilepsy is a neurological disorder characterized by recurrent seizures. These seizures are caused by abnormal electrical activity in the brain. Seizures are often accompanied by involuntary partial or whole-body convulsions, frothing at the mouth, and possible loss of consciousness, putting a patient at high risk. Electroencephalogram (EEG) can be used to diagnose epilepsy. This study proposes a seizure detection algorithm to identify a seizure attack with EEG. This algorithm includes a simplified signal preprocessor and a nearly optimized convolutional neural network (CNN). This study also proposes an artificial intelligence acceleration (AIA) hardware architecture, including a deep learning accelerator (DLA) and a two-stage reduced instruction set computer-V (RISC-V) central control unit (CPU), to implement the detection algorithm in real-time operation. The accelerator is implemented in System-Verilog and validated on the Xilinx PYNQ-Z2 Field Programmable Gate Array (FPGA) board. The implementation consumes 3535 lookup tables, 2283 flip-flops, 28 KB of block random-access memory, six digital signal processors, and seven input/output (I/O). The total power consumption is 0.108 W in 1-MHz operation frequency. The detection algorithm provides 99.06% accuracy on fixed-point operations with a detection latency of 128 ms/class. The application-specific integrated circuit (ASIC) performance of the AIA hardware architecture is also tested with a 180 nm 1P6M process. The total power of the AIA is 1.29 mW. The core circuit of the RISC-V CPU and DLA consumes $80~\mu $ W and $84.5~\mu $ W, respectively. Moreover, the AIA can be reconfigurable. Thus, the accelerator can execute different deep-learning models to fit various wearable applications for biomedical acquisition systems.