A 900-nW Wearable Interpatient Cardiac Arrhythmia Monitoring Processor With a Feature Engine-Based Artificial Neural Network

Real-time wearable arrhythmia monitoring is an important means to avoid sudden cardiac death and treat cardiovascular disease effectively. However, because of the limited computing resources and power consumption of wearable devices, it is difficult for existing on-chip arrhythmia monitoring processors to achieve the good generalization performance required by inter-patient applications. Therefore, this paper proposes several innovative methods to balance the trade-off between algorithm generalization performance and limited resource and power consumption. Using multi-level event-driven architecture, the standby power consumption of the processor is reduced through event wake-up. The abnormal heartbeat detection algorithm uses reconstructed multi-cycle heartbeat segments as the classification object and highlights the correlation information between heartbeats to improve generalization performance. From the bioelectrical signal mechanism, a set of morphological and rhythmic features rich in electrocardiogram pathological information is proposed to minimize the scale of the classification neural network while ensuring the accuracy of the arrhythmia classification algorithm. The processor is implemented using FPGA and TSMC65lp CMOS process. The measurement results demonstrated that the power consumption is only 810 nW in abnormal detection mode, with an algorithmically abnormal segment sensitivity of 96.77%, and 900 nW in arrhythmia classification mode, with an algorithmically inter-patient accuracy of 93.3