Research On The Quality Evaluation Method Of ECG Signal Based On Deep Learning

With the development of 5G technology and healthy Internet of Things technology,portable dynamic ECG devices have become a new trend,which is crucial for the early detection and prevention of heart disease.However,because the detection environment of dynamic ECG devices is complex and changeable,and ECG signal is a kind of bioelectrical signal that is vulnerable to interference,the quality of the collected wearable ECG signals is uneven.Therefore,before the diagnosis and treatment of heart disease,it is very necessary to eliminate the ECG data without diagnostic value through the quality assessment of ECG signal.Existing methods of ECG signal quality assessment are not very accurate due to the single feature extraction method or the small number of features used.As a result,quality assessment has been unable to be really put into the application of wearable ECG devices.Aiming at the above problems,this thesis mainly studies the database establishment of dynamic ECG signal quality assessment,the multi-feature fusion of ECG signal manually extracted features,the time-frequency representation features of ECG signals and the fusion of two kinds of features through convolutional neural network.It can be divided into the following four points.Firstly,the annotation rules of the existing ECG signal quality evaluation database are summarized and the new idea of the annotation rules is put forward in this thesis.Then,the PhysioNet/Computing in Cardiology Challenge 2011(PICC)used in this thesis is emphatic introduced.Second,nine typical manually extracted features of ECG signal are extracted,and the quality of ECG signal is evaluated by SVM classifier after feature fusion.Thirdly,the characteristics of the ECG signal are studied from a single time domain or frequency domain by time-frequency transformation and extended time-frequency domain.Fourth,a CNN network with two input layers is designed to fuse the two types of features,one of which is input the layer first sends the 12-lead ECG signal to the S transform layer to obtain its time-frequency representation,and uses a CNN network with 3 convolutional blocks to perform automatic feature extraction.At the same time,the manual features of the other input layer including lead shedding,R peak and baseline drift are fused with features automatically extracted by the convolutional network for training.The methods proposed in this thesis all use the PICC database for performance evaluation.The multi-feature fusion method of manually extracting features and the multi-feature fusion method of two types of features have achieved the effect of improving the accuracy.At the same time,this thesis applies the S transform to the field of ECG signal quality evaluation for the first time,and proposes a new method that combines the features extracted by the convolutional neural network and the manually extracted features,and finally achieved good results.In future work,the method in this thesis can be verified in more databases for ECG signal quality evaluation.