Multiscale Feature Fusion Convolution Neural Network And Application In ECG Arrhythmia Recognition

Cardiovascular diseases pose a serious threat to human health and can be called the "number one.killer." At the same time,it has brought a huge burden to the medical and health system and is showing a trend of increasing year by year.Electrocardiogram(electrocardiogram,ECG)can display heart condition information,reflect the health status of human heart,and is widely used in clinical diagnosis of cardiovascular disease.In recent years,thanks to the development of hardware equipment,deep learning technology has made great progress,and some traditional machine learning algorithms have been replaced to become the mainstream of the industry.It has achieved fruitful results in computer vision and speech recognition.Timely detection and prevention of heart disease can effectively reduce mortality,and the application of deep learning technology to ECG detection and classification recognition is of great significance for the prevention and diagnosis of heart disease.This topic takes arrhythmia recognition as the core,and gives a basic introduction to the process of arrhythmia recognition,electrocardiogram theory,and arrhythmia diseases.This article takes CCDD arrhythmia database as the research object,and applies deep learning technology to multi-lead ECG arrhythmia recognition.The innovation and research work of this thesis mainly has the following two aspects:For traditional ECG automatic identification and diagnosis technology algorithms rely heavily on manual feature extraction,single-scale neural network extraction feature information is not rich enough and multi-scale convolutional network fusion does not consider the problems of different scales of different values and importance.In this paper,the convolution kernel of the two-dimensional convolutional network is changed to a one-dimensional convolution kernel.First,a multi-scale feature fusion convolution network model MSCNN suitable for multi-lead ECG is designed,and the model is optimized and improved to attract attention Mechanism,a MSACNN model for multi-lead ECG is proposed.Experiments show that the MSACNN model achieves an accuracy rate of 94.41%in the multi-lead ECG arrhythmia four-class classification task,which is superior to the single-scale model and the MSCNN model.For MSACNN and other models,multi-lead ECG data is treated as two-dimensional single-channel data.The same convolution kernel is used in different leads.Relatively speaking,the focus is on the common characteristics of the leads.This does not take into account the correlation between the channels.Nor does it take into account the fusion features between leads.To this end,two other multi-lead recognition models are designed in this paper,which are models that focus on the features of lead fusion and the connection between channels.In order to explore which combination fusion method is better,this paper uses two feature fusion methods,a total of 8 combination models,and constructs a multi-lead ECG multi-feature fusion convolutional neural network based on fully connected and attention mechanisms.Through comparative experiments,the results of the fusion model have further improved the classification effect on the four classifications and the positive and abnormal two classifications,and the accuracy rates have reached 95.24%and 98.26%,respectively.This verifies the effectiveness of the scheme in this paper and provides a new Ideas.