A Lightweight R peak Detection Algorithm For Noisy ECG Signals

The electrocardiogram (ECG) signal is used to monitor the electrical activity of the heart throughout each cardiac cycle. Cardiovascular disease (CVD) and arrhythmias can be diagnosed using an ECG. Meanwhile, as Artificial Intelligence progressed, deep learning is increasingly being utilized to detect and classify ECGs. Even though various algorithms for detecting and classifying ECG signals have been developed, the majority of them focus on signals with a high signal-to-noise ratio (SNR) obtained from hospital patients. Detecting R peaks in noisy signals collected by wearable dynamic ECG devices remains research value and practical significance. In this paper, an approach combined 8-layer U-net with depthwise separable convolution named 8-DS-Unet is proposed to locate R peaks, particularly during the low-quality signal episode. Additionally, a six-layer network (6-DS-Unet) is proposed to further improve inference speed. The algorithm's high accuracy and low computational complexity enable model porting to wearables with limited hardware resources. The network was trained using the CPSC 2019 dataset and then validated on four additional datasets to verify robustness. In the CPSC 2019 dataset, the proposed 8-DS-Unet network achieved a Precision of 0.9966 and a Recall of 0.9915.