Research On Algorithm Of Heart-rate Detection Using BCG Based On Lightweight Convolutional Neural Network

Since the 1980 s,heart rate detection has been the research target of biomedical practice.The methods of recording heart rate include ECG analysis,echocardiography,etc.These monitoring schemes are accurate and reliable,but they require close contact between the electrode or heart probe and the human body,so they need to be operated by medical experts,which will cause inconvenience and psychological pressure on the subjects.Ballistocardiogram（BCG）records the signal waveform of body movement generated by blood flow in each heartbeat cycle.The heartbeat detection method based on BCG does not require direct contact between the human body and the sensor,and can be operated without an expert.It is an ideal method for long-term home monitoring.However,compared with other clinical cardiac diagnostic methods,factors such as interference and noise will affect BCG signals,which makes it challenging to extract heartbeat from BCG.At present,a deep learning-based method for heart rate detection has emerged,which enables the deep learning algorithm to autonomously learn the characteristics of BCG signals.However,the parameters of the deep learning methods currently designed are very large and cannot be applied to devices with limited computing power and limited memory.To solve the above problems,this paper uses BCG data collected in different scenes for experiments,proposes an IJK-wave transformation method based on U-Net,and designs a lightweight network model for heart rate detection.This paper is mainly introduced in the following three parts:Firstly,to address the issue of a missing BCG public dataset,short-term BCG data from 24 healthy subjects were collected in the laboratory,and long-term data from elderly people collected in the hospital were used.Finally,BCG with different signal-to-noise ratios were produced by simulating the method of BCG generation.Secondly,this paper presents a BCG-based IJK-wave conversion method to simplify the problem of heart rate detection.Firstly,the original BCG waveform is converted into a J-peakcentered IJK waveform by U-Net,and the major peaks in BCG are restored.After the IJK-wave is obtained,the J-peak is located by using the traditional maximum value method in the search area.This method has high accuracy for heartbeat detection of BCG waveforms with no obvious peak compared with the method of direct binary classification of J-peak.Finally,in order to accommodate devices with limited computing resources,a lightweight network model is designed to solve the problem of too many parameters and computations in the network model.This model compresses the convolution layer structure of U-Net in a bottom-up manner,using the shallow symmetric network as the basic framework.A lightweight network module is designed based on inverted residual and group convolution combined with lightweight network design criteria,which accelerates the network speed.In order to further improve the accuracy,a multiscale dense residual module is designed,and the weights are redistributed to the channel characteristics using an adaptive weight allocation method.The IJK-wave output by the lightweight network is post-processed,and the method of searching for the maximum value in the region is used to locate the heartbeat efficiently and estimate the heart rate.The proposed new model greatly improves the speed of operation and compresses the network structure while losing a small amount of precision.It can still show better performance on data with different signal-tonoise ratios.To sum up,this paper proposes a conversion algorithm of IJK-wave based on BCG,which can simplify the problem of heart rate detection in BCG.Based on the conversion algorithm,a lightweight network model is proposed to efficiently convert BCG waveform into IJK-wave to further realize heart rate detection.The proposed lightweight convolutional neural network can be applied to devices with limited computing resources,and provides a theoretical basis for extracting heart rate information from hardware with high configuration requirements.