Research On Myocardial Ischemia ECG Signal Recognition Based On Deep Neural Network

With the development of artificial intelligence,ECG-assisted diagnosis has become an important research topic in the field of artificial intelligence.ECG signal is the most common physiological signal for the diagnosis of heart disease.Using deep learning technology to achieve ECG-assisted diagnosis will greatly improve the efficiency of heart disease diagnosis.Since the amount of patients with myocardial ischemia increases gradually,early prevention of myocardial ischemia become much more important to the patient health.Based on the above background,this paper studies on the identification of myocardial ischemic ECG signals based on deep neural networks.The main research work and innovations are as below:1.Use the wavelet threshold denoising algorithm to denoise ECG signals based on the clinical data set.This algorithm has no obvious effect in denoising clinical ECG signals,while has obvious effect in denoising ECG signals on the MIT-BIH dataset.In order to send the ECG signal to the convolutional neural network for training,a method for converting the 8-lead ECG signal into a two-dimensional image has been designed.In addition,due to the low proportion of myocardial ischemic ECG signals in large data sets,sample equalization was performed by downsampling to obtain a sample balanced data set.Data set 1 consists of 2109 myocardial ischemic ECG signals and 7891 non-myocardial ischemic ECG signals.Data set 2 consisted of 2109 myocardial ischemic ECG signals and 7891 normal ECG signals.2.A simplified version of single convolutional neural network VggNet,a bidirectional long short-term memory network Bi-L STM,and a fusion network of the two were used to perform a two-class experiment on myocardial ischemia ECG signals on dataset 1.Compared with Bi-LSTM,the simplified version of VggNet is 3.5%,15.08%,and 1.46%higher in recognition accuracy,sensitivity,and specificity.Compared with the simplified version of VggNet,the fusion network has a 3.07%improvement in recognition sensitivity.It shows that the fusion network can improve the sensitivity of the model.3.A 28-layer deep residual network model has been designed to identify myocardial ischemic ECG signals.The accuracy,sensitivity,and specificity of recognition on dataset 1 were 95.10%,89.15%,and 96.70%,respectively.The network fusion is applied to deep residual networks.Resnet+Bi-LSTM,a fusion network of deep residual networks and bidirectional long short-term memory networks,has 93.95%,92.22%,and 94.42%recognition accuracy,sensitivity,and specificity.Although the accuracy and specificity are slightly lower than those of a single deep residual network,the sensitivity has been improved by 3.07%,which is effective and less misdiagnosed,indicating that the method of network fusion can further improve the sensitivity of deep residual network identification.4.In order to compare the recognition effect of the deep residual networks algorithm and the recognition algorithm based on waveform features,the waveform feature-based algorithm is used to identify myocardial ischemia ECG signals on dataset 2.The deep residual network is 2.77%,4.95%,and 0.16%higher in recognition accuracy,sensitivity,and specificity than the recognition algorithm based on waveform features.At the same time,complex feature extraction processes are avoided.