Automatic Classification Method Of Arrhythmia Based On Discriminative Deep Belief Networks

Cardiovascular disease is the main killer of human health,its morbidity and mortality are increasing year by year,which poses a serious threat to human life.As an important non-invasive clinical tool for cardiologists,electrocardiogram(ECG)reflects the rhythm of cardiac activity and is widely used in the diagnosis of cardiovascular disease.Arrhythmia is one of the common cardiovascular diseases,the ECG signals arrhythmia automatic classification method based on deep learning can effectively solve problems such as the difficulty in screening caused by the artificial feature design of traditional classification methods.Conducting ECG signals automatic classification and diagnosis based on deep learning is of great value for timely diagnosis and treatment of cardiovascular diseases.The main research contents of this thesis include:(1)Pretreatment of ECG signals: ECG signals denoising process using improved wavelet threshold.According to the ECG signals and various noise distribution characteristics,the bior3.7 wavelet basis function is used to perform 9-layer wavelet decomposition on the ECG signals,the 9th layer scale coefficient is set to zero,the 1 to 4 layers of wavelet coefficients are processed and reconstructed using an improved threshold and threshold function to effectively suppress noise such as baseline drift and high frequency interference in the ECG signals.(2)Feature extraction and classification of ECG signals: A heartbeat feature extraction and classification network model for arrhythmia based on discriminative deep belief network(DDBNs).The DDBNs is stacked by a three-layer Restricted Boltzmann machine(RBM).Firstly,the morphological characteristics of heartbeat signal are extracted by the constructed two-layer generated restricted boltzmann machine(GRBM).Then,a discriminative restricted boltzmann machine(DRBM)with the ability of feature learning and classification is introduced to carry out pre-training according to the extracted morphological features and RR interval features,and DDBNs is initialized with the weight parameters obtained from pre-training.(3)Feature extraction and classification model adjustment of ECG signals: In order to improve the classification performance of DDBNs,DDBNs is converted into a deep neural network(DNN)which uses Softmax regression layer for supervised classification,and the network is fine-tuned by back propagation(BP).Using the fine-tuned DDBNs model,the arrhythmia heartbeat automatic classification is completed.The massachusetts institute of technology and beth israel hospital arrhythmia database(MIT-BIH AR)is used to evaluate the performance of DDBNs with different network structures,RR interval features integration and before and after fine-tuning.The experimental results show that the DDNBs-based arrhythmia automatic classification method proposed in this thesis has better ability of deep feature mining and classification,which verifies its effectiveness in arrhythmia automatic feature extraction and classification.