Research On Unsupervised Domain Adaptive Of Myocardial Infarction Location Method

Myocardial infarction is one of the most widespread cardiovascular diseases affecting human health.Because of its high mortality rate and permanent damage to myocardial cells,it needs to be diagnosed as early as possible.At present,ECG is the most common and effective diagnostic tool for detection.At present,12 lead ECG has been used to diagnose cardiovascular diseases such as myocardial infarction in an endless stream.However,the existing intelligent diagnosis methods for myocardial infarction have some limitations,mainly because they do not consider the impact of differences between patients on the model,which makes it difficult for the model to play a role in the actual diagnosis process.In this thesis,we use the depth learning method to build the myocardial infarction location network and use the unsupervised domain adaptive method under migration learning to fine tune the model,and finally achieve the detection and location of myocardial infarction.The specific research contents are as follows:(1)In view of the problem that the ECG signal is easy to mix with noise in the acquisition process,which is not conducive to the training of diagnostic model,wavelet denoising method is used to achieve the denoising of ECG signal.Daubechies 6(db6)wavelet function is used to decompose the original ecg signal to remove the noise components contained in the wavelet coefficients of each scale,and then the wavelet coefficients of each scale are reconstructed to obtain the signal after removing the noise.(2)In order to effectively extract the deep features of ECG signals,an automatic classification network of ECG signals for myocardial infarction location was designed.The multi-scale features of 12 lead ECG signals are extracted by combining the ideas of the Space Pool Pyramid Pooling(ASPP)module,attention mechanism and residual block.Finally,the location of myocardial infarction is classified and verified on the Physikalisch Technische Bundesanstalt(PTB)ECG database.(3)In view of the problem that the trained diagnostic model can decline in practical application due to the difference of ECG among patients,the unsupervised domain adaptive method in transfer learning is used to divide the source domain and target domain for patients according to the paradigm requirements between patients,and the heartbeat data of patients belonging to different domains are used for training to reduce the domain drift caused by the large difference of ECG individuals,and then the domain invariant features are extracted.At the same time,according to the idea of clustering,two loss function regularization terms are designed for sub domain adaptation.Finally,experiments are carried out on PTB dataset and PTB-XL dataset,and the experimental results show that the proposed algorithm has good classification performance improvement effect on target domain data.