Research On Abnormal ECG Segment Detection Method Based On Deep Learning

ECG is an important basis of detecting cardiovascular diseases,and ECG analysis is also a classic scenario of computer technology integration in medical field.The efficiency improvement brought by computer automation helps doctors to make better diagnosis and treatment,and ECG analysis technology can monitor patients’ physical condition in real time.With the development of deep learning technology,a wave of technological revolution has been set off in the medical field,and more and more machine learning methods are applied to medical scenarios.Deep learning plays an increasing role in feature extraction,classification detection and clinical diagnosis in ECG.At present,the application of deep learning methods of the field of ECG is still in the process of exploration,which face problems such as inaccurate localization in dynamic ECG,insufficient accuracy of arrhythmia classification,and lack of ECG data labels,etc.In this thesis,the deep learning method is applied to the above problem with the tasks of ECG localization and ECG classification in the context of ECG detection,and a pre-training scheme is designed according to the characteristics of ECG.The specific work is as follows:(1)Aiming at the problem of QRS complex(an electrocardiogram band)inaccurate localization in ECG,a QRS complex location algorithm of Bi-LSTM(Bidirectional Long Short Term Memory)network with window structure is proposed.Through the preprocessing method of window structure,the feature of ECG is effectively extracted,and the position in ECG is marked through time sequence model,which has achieved good results in MITBIH data set;(2)Confronting the lack of performance in the classification of arrhythmias in cardiovascular diseases,an arrhythmia detection algorithm based on deep residual network and transformer structure is proposed.The potential features are extracted through the residual network.According to the characteristics of ECG,the transformer module is customized as a classifier,which effectively improves the effect of the algorithm in CINC data set;(3)Aiming at the lack of data labels in the field of ECG,an ECG pretraining model based on contrastive predictive coding is proposed,a large number of unlabeled data can be used through unsupervised pretraining,effectively improve the efficiency of data and promote the circulation of ECG data.In this thesis,we start from the simplest QRS detection to ECG classification,finally enhance the ECG classification algorithm by pretraining method,and verify the effectiveness of the method of the experiment.