Study On Short Term ECG Classification Based On Residual-Concatenate Network And Metric Learning

With the development of biological science and technology,electrocardiogram has become one of the main methods for diagnosing cardiovascular diseases.According to the World Health Organization,cardiovascular disease is the disease with the highest lethality compared to other diseases.Therefore,accurate and rapid analysis of ECG is of great significance for the diagnosis of cardiovascular disease.In recent years,the analysis of ECG signals has become a hot spot in biomedical research.Machine learning methods have been applied to various research fields and achieved good experimental results due to their continuous improvement of their own performance based on existing knowledge systems.Therefore,the use of machine learning methods to classify ECG signals has also aroused the interest of many relevant experts.Although the recognition and research of ECG signals based on machine learning methods have made some progress,there are still some difficulties in the recognition of ECG signals.Because the shape and time characteristics of the ECG signal are different for different patients in different physical environments,the ECG signal of each patient may be different,and different patients may have different ECG shapes for the same disease.In addition,various kinds of inevitable noises will appear when collecting ECG signals.These problems have caused unavoidable problems to the classification research of ECG signals,and also brought many difficulties to the actual research.The actual clinical work requires analysis and research on long-term ECG signal recording,such as bedside monitoring or wearable online medical care.However,this may be very troublesome and time-consuming for a person.Due to the shortage of medical resources,some personal errors will occur in the entire ECG analysis.Aiming at the problem that it is difficult to distinguish different ECG abnormal categories in ECG signals,in order to improve the performance of ECG signal recognition and classification,this thesis classifies ECG signals from the perspective of feature representation.It is an effective computer-aided method to use deep learning methods in machine learning to process ECG signals,and to improve the residual network into a residual connection network in order to better learn the characteristics of ECG signals.Compared with traditional machine learning methods,the network processing of ECG signal data in deep learning can combine data preprocessing,feature extraction,classification and other steps to directly map the input data to a certain type of ECG abnormality.Moreover,the efficiency of deep learning methods in processing data is generally higher,and it can better serve ECG analysis devices that require higher immediacy.Using deep learning methods and starting from the expression of ECG characteristics can comprehensively improve the various indicators of the classification.For the identification of ECG signals and the study of arrhythmia,the most used database is the public database MIT-BIH Arrhythmia Database.This database has the problem of sample imbalance.Although the deep learning algorithm can achieve the same or better effect as the traditional algorithm,it is difficult to achieve various accuracy rates or balance the experimental evaluation indicators and improve the overall accuracy rate at the same time.Achieving this has also become a challenge for the study of ECG signals.In this regard,in this thesis,a 5-second ECG signal is used as the input of the entire network to be closer to the actual clinical diagnosis.From the perspective of strengthening the expression of the input data,the structure of the residual network is changed to better identify ECG Signal to achieve better algorithm performance.At the same time,this thesis also uses metric learning as a post-processing method to classify the ECG feature representation,further improving the overall classification accuracy.Experimental results show that,compared with other methods,the method proposed in this thesis improves the classification accuracy of abnormal heart rhythm,and for each experimental evaluation index,the method proposed in this thesis achieves a balance between these standards.