Research On Wearable Ecg Data Labeling And Signal Quality Assessment Methods

Wearable ECG monitoring is an important technical means to realize portable telemedicine monitoring.Dynamic ECG monitoring plays an important role in early detection of heart disease risk.With the advent of the health care big data era,wearable ECG real-time,dynamic,continuous and intelligent monitoring has become a new trend in ECG clinical applications and family health monitoring.Establishing a standard and effective wearable ECG database and carrying out dynamic ECG intelligent analysis research are very important.However,the ECG signal is a weak physiological signal that is susceptible to interference,especially in dynamic monitoring scenarios.Due to the complex and variable motion conditions and the influence of noise factors such as equipment and environment,the recorded signal quality is uneven.If the long-recorded ECG big data is directly pushed to the doctor without any quality assessment,it will greatly increase the workload of the medical staff,the noise information dimension disaster,and even serious misjudgment.Therefore,effective signal quality assessment is necessary before wearable ECG risk testing.In this paper,the research focuses on the aspects of wearable ECG system design,data online annotation system,dynamic ECG quality manual labeling and machine learning based signal quality classification.Firstly,using the 3-lead ECG monitoring equipment jointly developed with the cooperation unit,the 24-hour long-range dynamic wearable ECG data of more than 200 patients with cardiovascular disease(mainly arrhythmia)were collected,and an online ECG data annotation platform was built.Then we use the data annotation platform to classify the wearable ECG data into three categories including clinically useful and good signal quality,clinically useful but poor signal quality and clinical uselessness signal,further dividing each signal quality into subclasses.Based on that,16 signal quality subtypes had been established.According to the data annotation above,three small-scale signal annotation databases of wearable ECG signal quality classification,signal quality refinement sub-category and ECG QRS position annotation were established.Then,the features of the wearable dynamic ECG signal and its noise interference type were analyzed.A total of 12 ECG signal quality features were extracted from three domains including time,frequency and nonlinear as the evaluation indicators of signal quality classification,combined with support vector machine(SVM)algorithm for automatic evaluation of signal quality of wearable ECG.At the same time,the grid search algorithm was used to optimize the SVM parameters and the model for evaluating wearable ECG signals was trained.The accuracy was 92.3%.Meanwhile,this paper also studied the ECG quality assessment method that integrates ECG time-frequency analysis with deep learning methods.The ECG signal segment was converted into a time-frequency two-dimensional image,which would be input into the convolutional neural network(CNN).The signal quality was classified into three categories through the CNN model.The classification result of the CNN model output was 79.5%.Finally,the differences between the SVM and CNN models in the evaluation of wearable ECG signals were analyzed and compared,and the future work was prospected.