Research On Bayesian Fusion Of Heart Rate Estimation For Dynamic ECG Signals

Continuous monitoring of dynamic electrocardiogram(ECG)is an effective means to achieve early detection of cardiovascular diseases.Heart rate(HR)is an important physiological indicator of ECG that reflects the heartbeat of the human body.It can also assess and evaluate cardiovascular disease effectively.However,dynamic ECG pollution is very common,and the labeling algorithms of HR are susceptible to signal noise,individual variations and other factors,thus decreasing the reliability of HR values greatly.The erroneous estimation of HR will affect the patients' treatment timing or cause alarm fatigue of the guardians.In this paper,with the help of signal processing technology and information mining of big data technology,an annotation fusion model is established,aimming to achieve a robust and accurate estimation of HR in dynamic ECG,so as to reduce the burden on families and medical staffs.Deriving from unsupervised ensemble learning,this paper establishes a fusion process of annotations based on the annotators' capacity.The main contributions of this paper are summarized as below.(1)An unsupervised fusion model of multiple HR annotations is proposed,and it estimates the potential ground truth of HR based on the annotators' capacity.The fusion model is added Bayesian prior probability and solved by the maximum posterior estimation(MAP)and expectation maximum(EM)algorithm,achieving a more reliable and accurate estimation of HR.It provides an unsupervised estimation method based on continuous-valued and time-series labels.Besides,it proposes a unified framework for fusing multiple labels to infer the ground truth while the deviation and accuracy of the annotator is modeled.Experiments on the Physio Net/ Computing in Cardiology(Cin C)in 2014 training data set show that our model surpasses the best annotator by 29.582%,indicating the superiority of the model.(2)On the base of the most common application scenarios of single-lead ECG and multilead ECG,this fusion model is modified to a single-lead-multi-annotator model(SLMA)and a multi-lead-single-annotator model for HR estimation(MLSA).It is difficult to achieve a stable,reliable and high-precision estimation of HR relied merely on a certain algorithm or single lead.Thus,SLMA or MLSA model have the potential to obtain the more reliable estimation of HR without the ground truth.The dynamic data-set of the China Physiological Signal Challenge(CPSC)2018 is used to validate the models,the SLMA model is 19.623% better than the best annotator in lead-I,and the MLSA model is 11.932% better than the best lead in standard 12-lead data-set.(3)This paper,from the view point of users,establishes an automatic HR annotation system which is quailified in application of ECG monitors.This system is suit for annotating on cross-database,cross-lead and cross-algorithm,as well as providing appropriate fusion process for varified databases(single-lead database or multi-lead database),running to the benefit of improving the accuracy of long-term dynamic estimation of HR and donating to the diagnosis of clinical cardiovascular diseases.