Heart Rate Estimation Via Deep Regression Method From BCG Signal

With the development of our country’s economy and the improvement of people’s living standards,especially the changes in people’s diet structure and lifestyle.According to the survey of the national health department,the prevalence of heart disease in our country is gradually increasing every year,which has become the greatest threat to people’s health.For early detection and treatment of heart disease,long-term monitoring of the heart in daily life is of great importance for the detection,prevention and diagnosis of heart disease.Research on more comfortable,convenient and accurate heart rate monitoring technology is also becoming more and more popular.Ballistocardiograms(BCG),as a non-invasive and non-contact method for monitoring cardiac function,can obtain BCG signal containing rich information on human physiological characteristics,in a comfortable and convenient way,which meets the needs of our daily life cardiac monitoring.Among the existing heart rate extraction methods from BCG signal,heart rate estimation methods based on signal processing and heartbeat detection are popular.However,due to unavoidable noise interference,individual differences and other factors,there is much great room for improvement in heart rate estimation accuracy,robustness and individual adaptability.Therefore,based on the existing research,this paper proposes two heart rate estimation methods for BCG signal based on the idea of deep regression.The main contents are divided into the following parts:(1)In order to improve the heart rate estimation accuracy of BCG signal while reducing the labor cost and enhancing the stability of the heart rate estimation algorithm,a heart rate estimation method from BCG signal based on self-attention memory network is proposed.First,this method enhances the heartbeat pulse characteristics of the original BCG signal by fusing the information of each channel,Fast Fourier variation,and introduces the heart rate prior information to reconstruct the enhanced signal.Then,based on the idea of deep regression,the Bi-directional Long Short Term Memory(Bi-LSTM)network is used to initially extract the temporal features such as BCG signal amplitude and periodicity,and the self-attention temporal features mechanism is introduced to weight the fusion of the extracted features at each moment.Finally,the fully connected regression network is used to obtain the heart rate estimation.The experimental results show that the end-to-end heart rate estimation method improves the accuracy of heart rate estimation while effectively reducing the labor cost and has good stability.(2)In the first method,the heart rate estimation is directly obtained from BCG signals,without the heartbeat detection step.In order to obtain the heartbeat detection result at the same time as the heart rate estimation,and further reduce the dependence on the label,a weakly supervised heart rate estimation method from BCG signals based on temporal signal reconstruction is proposed.In this method,the idea of signal reconstruction is innovatively presented.It’s only by introducing the heart rate and other heartbeat prior information that completes the reconstruction of the BCG signal and optimize through the interpolation method.Then,the deep convolutional neural network is used for feature extraction and fusion from BCG signal,which maps the BCG signal into a regression signal easier to detect heartbeat.Finally,the heartbeat detection is performed by the peak search algorithm to complete the heart rate estimation.The main contribution of this method is that the introduction of finger pulse signal is avoided by the idea of signal reconstruction,and the mapped signal labels required by the regression model are generated,thus solving the mismatch problem between BCG signal and reference signal,and greatly saving the cost of reference signal acquisition;In addition,this method based on the regression idea instead of the traditional heartbeat classification method for heartbeat detection,does not require precise tagging of heartbeats,which greatly reduces the cost of manual tagging in the data pre-processing process.The experimental results show that the weakly supervised method can effectively improve the accuracy of heart rate estimation while reducing the labor cost.