A 1.83 μ J High-Robust Cardiac Health Monitoring with Adaptive-Threshold QRS Detector and Hybrid Neural Network Arrhythmia Classifier

Arrhythmia is a crucial method for monitoring cardiac health. However, Electrocardiogram (ECG) signals obtained from wearable devices are often contaminated by noise (such as electrode motion artifacts, baseline wander, and muscle artifacts). Therefore, in this paper, a high-robust cardiac health monitoring with an adaptive-threshold QRS detector and hybrid neural network arrhythmia classifier is proposed. The QRS detector is based on event-driven sampling, combined with an adaptive duration threshold, adaptive RR interval threshold, and automatic error correction threshold. And arrhythmia classifier is based on LSTM and ANN hybrid neural networks (HNN) with Synthetic Minority Over-sampling Technique (SMOTE) technique for imbalance handling and three pathological feature fuses. Evaluated on the MIT-BIH Arrhythmia Database (MIT-ARDB) and MIT-bIH Noise Stress Test Database (MIT-NSTDB), the average R-peak performance Se/p+ are over 97.38%/97.08% in MIT-NSTDB, and the inter-patient classification performance accuracy is 93.47% on MIT-ARDB and 90.1% under 6 dB noise ECG signal. Finally, the abnormal heartbeat detection mode of cardiac health monitoring energy consumption is $1.51 \mu \mathrm{~J}$ and the arrhythmia detection mode is $1.83 \mu \mathrm{~J}$ on TSMC 65lp process. The overall standby time is estimated to be approximately 74.7 hours.