RMKnet: A novel and parameter-efficient architecture for robust r-peak detection

Abstract

The Electrocardiogram (ECG) has become the most important diagnostic criteria in the non-invasive diagnosis of heart disease. The automatic detection of r-peaks in an electrocardiogram signal is also crucial in a all kinds of applications about heart disease. Recently, lots of algorithms have been proposed to automatically detect r-peaks in ECG signals. Traditional detection methods have high detection accuracy on high quality ECG signals which don’t have so much noise, but they perform poorly on low quality ECG signal. Most of the deep learning methods are more robust and have higher accuracy than the traditional methods, but they are usually computationally expensive and difficult to implement and customize. In this paper, we proposed a novel and robust model architecture based on Unet[17] to detect r-peaks. After training, we use re-parameterization greatly accelerating the inference process and reducing the inference memory utilization. The proposed network is evaluated on two open-access ECG databases: CPSC-DB, MIT-BIH and achieves high F1 scores which showing good generalization on different databases comparing to all competing methods.