Research On ECG Classification Based On Transfer Learning

Electrocardiogram(ECG)is a record of the electrical activity of the heart,a non-stable,nonlinear,highly random weak physiological signal.The ECG classification algorithm represented by deep learning techniques is of great significance for the clinical diagnosis of Cardiovascular Disease(CVD).Deep learning requires a large number of labeled ECG samples,which require a significant amount of time for a specialist physician.On the other hand,deep learning models are designed for specific ECG classification tasks.Training a neural network requires a lot of computational resources and time.These manually tuned models can only handle specified ECG classification tasks and cannot be applied directly to new ECG classification tasks.Utilizing common characteristics between data and reusing existing classification models is a technical way to deal with the shortage of ECG label samples and further improve the accuracy of classification.The thesis takes transfer learning as the entry point and deep learning as the technical basis,and explores ways to improve the recognition rate of ECG classification.The amount of data in the ECG classification field is limited.First,the image domain is used as the source domain,and the large pre-training network in the image classification field is reused.By fine-tuning the pre-training model,the ECG classification is realized;and then the ECG timing characteristics are used to build an ECG pre-training classification model,and use the labeled data to retrain it to achieve the target classification task;finally,when the labeled data is not available,the domain adaptation method based on the Maximum Mean Difference(MMD)is used to reduce the difference in the feature distribution of the source and target domains to achieve ECG Classified domain adaptation.The main research work includes:(1)In view of the limited sample size in the ECG open data set,the models trained by the large picture data set are modified for ECG classification.First,the original one-dimensional ECGs are converted into scalograms based on Continuous Wavelet Transform(CWT).These scalograms' size are modified to the input of the fine-tuned networks accordingly,including Alex Net,Goog Le Net and Squeeze Net.These networks are trained on Image Net and fine-tuned for the three-class output.The classification accuracy of the models were 93.75%,90.63% and 96.88% on the datasets of arrhythmias,sinus heart rhythm and heart failure.(2)In order to achieve model migration in the field of ECG classification,fine-tuning adaptation is performed based on Bidirectional Gate Recurrent Unit(Bi GRU).A neural network based on Bidirectional Gate Recurrent Unit(Bi GRU)is designed.First,the ECG is denoised and the baseline drift is processed.The pre-training model is established using the ECG source data,the target domain model is initialized with the weight of the pre-trained model.Then,according to the label space,two sets of datasets are designed,and different numbers of training samples are selected to build the model training set during the model retraining.One-way transfer experiment was conducted on the MIT-BIH Arrhythmia Database and the PTB myocardial infarction database,with a model accuracy of 95.87%.Two-way transfer experiments were conducted on Chinese Cardiovascular Disease Database(CCDD)fibrillation data and MIT-BIH atrial fibrillation database(AFDB).The accuracy of the model reached 98.30% and 89.19%.compared with unused adaptation methods increased by 0.13% and 0.94%,respectively.(3)For the unsupervised domain adaptation issue,the MMD-Bi GRU-Net based on Maximum Mean Discrepancy(MMD)is built.First,the source and target domains are mapped to the Regenerating Kernel Hilbert space(RKHS).Then the model learns to classify the public space by minimizing the distance between MMD measures,and using the label information of the source domain makes the model learn the classification of the public space,so as to achieve the purpose of classifying the target domain.Two-way unsupervised domain adaptation experiments were conducted on CCDD and AFDB,and the average accuracy reached 73.34% and 73.34%,respectively,which was 11.70% and 17.74% higher than those without domain adaptation.