Arrhythmia Recognition Based On Deep Learning Hybrid Networks

With the development of people’s economic conditions,the rhythm and pressure of life is gradually increasing.Arrhythmia has become an important source of threat to people’s health.Arrhythmia is an abnormal heartbeat rhythm and a precursor to diseases of the human heart.It has a certain concealment in the early stages of the disease,and its risk increases with age.Electrocardiogram,as a standard tool for detecting and recording heart activity,faithfully reflects the health status of the human heart.By analyzing and interpreting it,potential irregularities can be diagnosed.However,due to its complexity,non-linearity and weak amplitude,the ECG signal is difficult to use artificial methods for rapid and accurate analysis.Therefore,it is necessary to automatically identify different abnormal heartbeats from a large amount of ECG data,and this is a necessary task in the clinical medical field.The work of this paper was to construct an automatic arrhythmia recognition model,identifying and classifying abnormal heartbeats.The main research contents and results are as follows:(1)Based on the deep learning method,this paper proposes a hybrid model of convolutional neural network and long-term short-term memory network according to the current research status.It targets six different types of ECG signals in four public ECG standard databases,namely normal sinus rhythm,atrial fibrillation,ventricular bigeminy,pacing rhythm,atrial flutter and sinus bradycardia for automatically classifying and identifying.In this work,two different hybrid model structures are proposed to optimize the classification results,and the feasibility of the optimized model and the generalization of new data are verified.(2)This paper uses a 12-layer convolutional neural network followed by 2 layers of short-and long-term memory networks to model the 10-second ECG signal segments from the MIT-BIH arrhythmia database.After that,this paper uses MIT-BIH normal sinus database and MIT-BIH atrial fibrillation database as independent databases for generalization analysis of the proposed model.In order to increase the diversity of participants in the training data,Challenge2017 database is added to the training data to train the model,and then the trained model is independently verified.According to the verification results,an improved hybrid multi-input network is proposed,that is,the corresponding RR interval of the input ECG signal is added,and the same method as before is used for testing and verification.(3)This paper uses a five-fold cross-validation strategy for the proposed single-input hybrid network to achieve an overall accuracy of 99.01%,a positive predictive rate of 96.02%,a sensitivity of 96.69%,and a specificity of 99.00% on the MIT-BIH arrhythmia database.For the improved multi-input hybrid network,a 50% cross-validation strategy is used on the database with increased training set to achieve an overall accuracy of 99.32%,a positive predictive rate of 97.66%,a sensitivity of 97.75%,and a specificity of 99.51%.Using an independent database to verify the multi-input network,we obtain an overall accuracy and sensitivity of 96.94% and 96.87%.Among them,94.30% of the independent ECG signals in the atrial fibrillation class are correctly identified,which is 17.71% higher than the single-input network.The above results show that the deep learning hybrid network constructed in this paper can well recognize several clinically abnormal heartbeats and achieve automatic diagnosis of patients with arrhythmia.Moreover,the improved multi-input hybrid network is ideal for independent ECG data recognition results,and the generalization of the model can be guaranteed.This method can be used as an auxiliary tool to help clinicians diagnose arrhythmia,and it is significant for practical applications.