Research On Multi-label ECG Anomalies Classification Based On Deep Learning

In recent years,the death rate of cardiovascular disease is still in the first place,and electrocardiogram（ECG）as a noninvasive detection means,can detect abnormal heart beat from the ECG signal types,cardiovascular disease is the most basic and simple,economic and efficient method of clinical examination,and the traditional interpretation of electrocardiogram（ECG）requires human has the rich professional knowledge and a lot of clinical experience,So it is necessary to recognize ECG abnormality accurately and quickly.With the development of computer technology,it is hot and meaningful to use deep learning technology to help solve medical problems.However,in the current clinical application,deep learning has some problems to be solved in the identification of ECG abnormalities.For example,most studies based on ECG in recent years define the classification of ECG abnormalities as multi-classification tasks,that is,one ECG signal corresponds to only one label,which is inconsistent with the actual clinical situation.Or the classification accuracy is not high because the problem is divided into multi-label classification tasks.In addition,most of the ECG databases used in current studies have a serious imbalance in the distribution of ECG abnormality categories and a large amount of data,which impeds the improvement of the accuracy of ECG abnormality identification.The research based on single lead or double lead electrocardiogram has some disadvantages in mining the deep characteristics of electrocardiogram.Therefore,it is of great significance to study the classification of multi-label ECG anomalies.In view of the above problems,the main purpose of this thesis is to improve the recognition performance of 55 types of ECG including normal ECG signals through deep learning and other technologies,in which the original 8-channel data set can be expanded to 12-lead ECG anomaly database with multiple labels.Through experimental comparison and analysis,the model designed in this thesis can further improve the performance of multi-label ECG anomaly recognition.The main work contents are as follows:（1）In this thesis,the original ECG signal is preprocessed.The noise source of the original ECG signal was analyzed and denoised according to the frequency band of the noise and effective ECG signal.Pan-tompkins algorithm was used to detect QRS wave groups and divide heart beats,and time domain features with clinical significance were extracted.The label of each ECG signal was one-HOT coded and the label correlation matrix was converted.（2）Design ECG anomaly recognition model.Based on the analysis of similarity,locality of anomalies and correlation between anomalies in ECG database,the deep learning technology and the design of network structure are selected to start from the characteristics of ECG signals.Convolutional neural network and short-and long-term memory network are combined to extract local and temporal features of ECG signals.For certain similarities among leads,SENet network,which can obtain channel importance,is used to combine with other deep learning networks to extract features.ResNet and SENet extracted ECG characteristics from network depth,Xception and SENet extracted ECG characteristics from network width.In addition,tag correlation features are extracted to improve the performance of the model.The effectiveness of the model is verified by designing comparative experiments.（3）The ECG anomaly recognition model was optimized from the perspective of data.In addition to the network structure of the model itself,which can improve the model recognition performance,for the deep learning algorithm driven by big data,this thesis verified that the classification accuracy can be further improved by reducing ECG data through compressed sensing and solving the problem of sample imbalance through GNN.