Combining Discriminative Deep Belief Networks And Active Learning ECG Classification Method

Electrocardiogram(ECG)is a non-invasive detection method widely used to reflect the condition of the heart.ECG can further understand the heart condition and diagnose various heart diseases.Careful inspection of its behavior is essential for detecting arrhythmia.It is based on deep learning The ECG classification method can provide effective solutions for cardiologists,and improve diagnostic accuracy while saving diagnostic time.The main research contents of this article include:(1)ECG signal preprocessing: The ECG signal is denoised by the wavelet threshold method.The wavelet basis function is symmetric tightly supported biorthogonal wavelet bior3.7,and the number of decomposition layers is determined to be 9 layers.Adapt the threshold,select the threshold function to improve the threshold function,do not perform the threshold operation on layers 5-8,set the wavelet coefficients of layer 9 to zero,perform heartbeat detection and segmentation on the ECG signal after denoising,identify the heartbeat feature points,and start the current heartbeat P wave.The current heartbeat R peak,the current heartbeat T wave ends,and the beginning of the P wave to the end of the current heartbeat T wave are divided as a heartbeat segmentation interval to complete the heartbeat division.(2)Feature extraction and classification of ECG signals: Following the American Medical Instrument Promotion Association(AAMI)standard,a heartbeat classification model based on discriminative deep confidence networks(DDBNs)is constructed.DDBNs are composed of three-layer restricted Boltzmann machine(RBM)),The first two layers use the generated restricted Boltzmann machine(GRBM)to extract the morphological features of the heartbeat,and then combine the RR interval characteristics to classify the heartbeat by discriminative restricted Boltzmann machine(DRBM),and then A DNN classifier is added to the top layer of the DDBNs network model by adding a Softmax regression layer that complies with the four types of heartbeat(N,S,V,and F)output of the AAMI standard,and then fine-tuning the DNN parameters by back propagation(BP)to optimize the classification performance of the DDBNs network model.(3)Fine-tuning with active learning: Design a classification model combining active learning(AL)and DDBNs,use probability-based break band(BT)algorithm as the sampling strategy,use the maximum rule and iterate the DNN to fuse,then reclassify the heartbeat,To reduce the training set while still improving the classification effect.This article selects the Massachusetts Institute of Technology and Beth Israel Hospital Arrhythmia Database(MIT-BIH)as the ECG signal data source,strictly follows the AAMI standard,and conducts experimental comparison and performance evaluation on different DDBNs structures,before and after BP fine-tuning,and before and after combining with AL The experimental results show that the proposed ECG classification method combined with DDBNs and active learning is superior to other algorithms.