Research On Multiple-label Aggregation Of Heart Rate Estimation For Dynamic ECG Signals

In China,cardiovascular deaths have accounted for more than 40% of deaths in residents currently,so the monitoring and analysis of ECG signals has received great attention.In order to realize the real-time monitoring and analysis of ECG signals,it is essential to label the QRS waves and heart rate automatically by algorithms rather than manual labeling.Because QRS waves locations and heart rate are not only the basis of intelligent diagnosis of cardiovascular diseases but also the important parameters of human body provided by mobile monitoring devices.At present,the annotating algorithm of QRS waves based on various theories has achieved more than 95% accuracy on one single database with good signal quality.However,the robustness and practicability of these algorithms are poor due to the weak and noisy ECG signals.Therefore,aiming at improving the accuracy of heart rate estimation,this dissertation proposes an unsupervised method to aggregate multiple annotations for estimating the heart rate of each ECG signal to be annotated.In this dissertation,modeling the multi-algorithm annotating process to improve heart rate estimation accuracy is implemented by studying the relationships between the underlying ground truth and multiple annotations,and the accuracy of each single annotator,and signal features,and signal quality.The main work of this dissertation is summarized as below.(1)The model for aggregation of continuous-valued multiple annotations is used to estimate the heart rate.The components of ECG feature vector is determined to estimate the unknown ground truth of heart rate in the linear regression model.The relationship between the unknown true label and the multiple annotations is described by a probabilistic model,and errors between them decided the precision of each single annotator.The contributions of single annotator for final aggregated heart rate results depend on thier precisions.The experiment results prove that this model has the highest accuracy improvement of 17.46%,23.12% and 42.23% respectively compared with the optimal single annatator and other continuous-valued label fusion method such as the mean strategy and the median strategy.In addition,the results on validation data set show the effect of the model to increase the annotating accuracy and the robustness of the model for unknown data.Furthermore,the estimated precision of the single annotator from the previous experiment served as the weight of the validation dataset.In validation test,the heart rate is computed by weighted average of multiple annotations,which is still superior to other heart rate estimation methods,demonstrating the effectiveness and robustness of the proposed model.(2)Based on the existing model,an improved model with the annotating bias and Bayesian prior distribution of parameters is proposed.The annotating bias of each single annotator is used to correct the differences between the multiple annotations and the potential true annotation.And the prior distribution of the annotator precision and annotator bias added to the model can reduce the dependence of the model on the observations of various annotations and improve the robustness of the model.Furthermore,a factor to adjust the annotator precision is proposed to connect the annotator precision with the signal quality.Finally,the Expectation Maximization algorithm is used to compute the maximum a posteriori estimate of the annotator bias,annotator accuracy,and the underlying ground truth.In this stage,according to the prior distribution of the annotator precision,the Generalized Maximum Value Distribution is modelled to find the threshold of precison.The threshold can guarantee the quick convergence of the algorithm.Compared with the previous model,the mean stragety and the optimal single annotator,the improved model further optimizes the heart rate estimation accuracy and the error reduction is up to 19.94%,26.38% and 25.69% respectively.(3)The two proposed models are tested on 2000 dynamic ECG signals.Compared with the mean strategy,the accuracy improvement of final aggregated heart rate annotations produced by two proposed models are up to 11.82% and 14.84% respectively.Compared with the optimal single annotator,the two models achieve 12.8% and 18.2% accuracy improvement respectively.The experimental results demonstrate that the two proposed models have different degrees of improvement on the annotation accuracy,and verify the robustness of proposed models for heart rate estimation of dynamic ECG signals.