Efficient Nonlinear Autoregressive Neural Network Architecture for Real-Time Biomedical Applications

Medical devices, such as continuous glucose monitors (CGMs) and drug-delivery pumps, are often combined in closed-loop systems for treating chronic diseases. Generally, these systems consist of sensors and actuators whose operation is modulated based on sensed stimuli. Closed-loop systems may be susceptible to a number of different security and reliability issues which may result in incorrect operation which may endanger patients. Nonlinear autoregressive neural networks (NARNNs) may be used in such systems for error detection and correction due to their predictive capabilities; however, an efficient implementation is needed for use in wearables and biomedical implants. In this paper, we present an area-and energy-efficient, pipelined NARNN hardware architecture suitable for such constrained devices. The architecture was tested on FPGA to confirm functionality, then synthesized targeting the SAED 32nm EDK. This NARNN implementation requires an estimated area of 0.02 mm(2), 0.54 mu s and 0.76 nJ per inference.