Research On Atrial Fibrillation Recognition Technology Based On Deep Learning

Atrial fibrillation is a common tachyarrhythmia.With an aging global population and increasing average life expectancy,the prevalence and incidence of atrial fibrillation is increasing and has become a global epidemic and a major public health problem.In the era of rapid development of artificial intelligence and big data technology,computer-assisted or even alternative to manual analysis of ECG signals is a task of great significance.With the rapid development of artificial intelligence and big data technology,deep learning is being more and more widely used in various fields.In order to solve the problem of difficult recognition of atrial fibrillation features faced by the existing technology in the noisy shorttime ECG signal recognition task and the problem of low detection rate of atrial fibrillation in the complex long-time ECG signal recognition task,this paper investigates the automatic recognition technology of atrial fibrillation based on deep learning,and proposes two atrial fibrillation recognition models for the above problems,which are important for early intervention of atrial fibrillation and avoiding further deterioration of the disease is of great significance.The main research of this paper is as follows:(1)An atrial fibrillation recognition model based on a multi-scale residual shrinkage network is proposed for the problem of difficult recognition of atrial fibrillation features faced in short-time ECG signal recognition tasks with high noise.ECG signals usually contain much noise,and neural networks are very sensitive to noise,and the presence of noise affects the recognition accuracy of the model.In order to reduce the influence of noise on the neural network and simplify the process of ECG signal preprocessing,a multi-scale threshold denoising module is constructed in this paper to implement the threshold denoising task in the neural network.First,the threshold is obtained by the global attention module;then,the multi-scale features are extracted by the multi-scale attention module composed of global attention and local attention;finally,the acquired threshold and Garrote threshold function are used to denoise the multi-scale features.In addition,an adaptive synthetic sampling algorithm is used to oversample the dataset and achieve category balancing by generating new samples to alleviate the overfitting problem of the neural network.The method was experimented and validated on the Physio Net 2017 short-time atrial fibrillation dataset,which improved the accuracy of short-time atrial fibrillation signal recognition.(2)To address the problem of low atrial fibrillation detection rate faced in long-time ECG signal recognition tasks with complex features,an atrial fibrillation recognition model based on a deep residual capsule network is proposed.ECG signal is a waveform with complex detailed features and high similarity between waveforms.Capsule neural network takes into account the relative spatial relationship between features and can discriminate the complex waveforms of ECG signal.First,this paper designs two structures of residual neural networks to extract features of ECG signals;then,the extracted features are reshaped into low-level vector capsules by capsule neural networks;finally,the dynamic routing algorithm between capsules is used to transform low-level vector capsules into high-level vector capsules and predict the occurrence of atrial fibrillation based on the length of the capsules.The method was experimented and validated on the MIT-BIH long-time AF dataset,and improved the accuracy of long-time AF signal recognition.This paper investigates the deep learning-based AF recognition method and proposes different AF recognition models for different application scenarios,which improves the accuracy of long-time and short-time AF signal recognition and provides a theoretical basis for the research of portable cardiac monitoring devices.