Research On Myocardial Infarction Electrocardiogram Feature Extraction And Classification Algorithms

Myocardial infarction is one of the cardiovascular diseases that threaten people's life and health.It is characterized by rapid onset and high mortality rate.Moreover,the morbidity of myocardial infarction increases year by year as living tempo speeds up,so timely diagnosis of myocardial infarction has significant and essential research value.Electrocardiogram(ECG)is a regular tool to manifest and detect myocardial infarction by reflecting the signal of human bioelectrical potential changes.However,the ECG system for clinical use always produces vast amounts of data.To diagnose and analyze different changes and types of ECG signal by professional doctors is time-consuming and inefficient.The implementation of automatic detection and analysis of ECG for myocardial infarction can solve these problems to some extent.Furthermore,with the prevalence of portable ECG devices,to apply and transplant these methods to portable devices make it possible to monitor abnormal ECG in individuals and families,which can further decrease the mortality of myocardial infarction and have high clinical significance and social value.Aiming at the automatic detection and analysis of myocardial infarction ECG,this thesis proposes two different schemes,which include traditional feature extraction approach and deep learning method.The main contents of this thesis are as follows:By studying and analyzing the interference and noise of the ECG signal,this thesis utilizes wavelet transform methods to filter three different kinds of noise.After removing the noise,the QRS wave is positioned through detecting the extremum pairs zero-crossing point that is produced by biorthogonal quadratic B-spline wavelet transform.Using the QRS wave location results,the detection and calculation of P-wave,T-wave and other characteristics are completed.Based on the above work and the theoretical knowledge of myocardial infarction ECG,this thesis figures out the 16 time-domain and non-linear features for the diagnosis and classification of myocardial infarction.Automatic classification of myocardial infarction is completed by traditional feature extraction method firstly.16 different ECG signal features are calculated and feature selection is accomplished by PCA.Then this thesis uses support machine vector,BP neural network,and Adaboost algorithms to classify and detect myocardial infarction.Meanwhile,this thesis discusses the influence and optimization of the kernel function,soft margin parameter,the number of neural nodes and weak learners,etc.The PTB database is used to validate and assess the performance of three different algorithms,and the experiment results show that Adaboost has a better performance than SVM and BP neural network.To solve the drawbacks of traditional feature detection and extraction methods,such as cumbersome processing steps and heavily relying on prior knowledge,a deep learning model based on convolutional neural network and long short-term memory network is established.The model consists of five channels and includes 16 layers.Fixed long heartbeat is used as input to automatically extract its spatial and time features through convolution neural network and LSTM network to diagnose myocardial infarction.Finally,the relative parameters of the model are studied and discussed,such as the number of channels and kernel size,etc.The results indicate that the deep learning network model can avoid the steps of complex handcrafted features selection and have a good classification performance.