Algorithm Research On Preprocessing And Classification Of ECG Data

With the rapid development of computer-aided diagnosis and treatment system in recent years,the analysis of electrocardiogram can not only relieve the pressure of doctors,but also provide diagnosis and treatment for patients timely,furthermore it can improve the accuracy to a certain extent.ECG signal is a weak physiological signal,vulnerable to be interfered in the acquisition process,so it is important to eliminate noise from ECG signal.The ultimate goal of ECG is heartbeat disease diagnosis,most methods focus on feature extraction mainly at present,with deep learning in the analysis of biological and medical data getting success,researchers applied it to the ECG gradually,but it still can be improved.This paper studies the two aspects of the preprocessing and classification of electrocardiogram.The main contents are as follows:(1)In order to solve the problems that the IMF components is difficult to select and the noise components are always eliminated directly when removing the ECG data noise by the ensemble empirical mode decomposition(EEMD)method,an adaptive threshold algorithm based on EEMD is proposed.Firstly,the noisy ECG signal is decomposed to intrinsic modal functions(IMFs)by the EEMD method,and then to judge the IMFs which are noised IMFs and which are signal IMFs according to the Mahalanobis distance.After that,the threshold of the noised IMF is determined using the fruit fly optimization algorithm.The denoised ECG signals are reconstructed by the new IMFs and the rest of IMFs after the threshold denoising.Finally,the method is applied to ECG data from MIT-BIH database.The experiment results indicate that the method can preserve the signal details while denoising.(2)In order to improve the accuracy of ECG heartbeat classification,a multi-scale fuzzy entropy feature extraction strategy based on EEMD is used.Firstly,the ECG signal is decomposed by EEMD into a series of IMF components.Then,select the effective IMF components to calculate fuzzy entropy and constitute the feature vector.Finally,the feature vector is sent to the RBF neural network for training and recognition.Using the data from MIT-BIH Arrthythmia Database to do simulation experiment,the experimental results show that the proposed algorithm is superior to EMD multi-scale fuzzy entropy methods.(3)In order to solve this problem that the parameters need repeated adjustments byhuman using the traditional deep learning network to extract the feature,this paper proposes a ECG learning approach algorithm based on adaptive parameters marginalised Stacked Denoising Autoencoder(mSDA).Firstly,input the damaged training set into mSDA structure whose parameters are optimized by FOA algorithm.To find the minimum value of the reconstruction error in the search space,the best number of hidden layers,best learning rate,and noise level are determined.Then,the original ECG signal with 340 sampling points is input to the m SDA.Finally,adding the Softmax layer at the top of the mSDA to form a deep learning network for testing on training set DS1.The total accuracy reaches high level under test set DS2.The results show that mSDA can obtain more robust features with less artificial participation.