Research On Arrhythmia Classification Algorithm Based On Deep Learning And Hardware Design

As the economy and living standards rise,people begin to pay more attention to their physical health.In China,cardiovascular disease is the primary cause of death among residents,so the prevention,diagnosis and treatment of cardiovascular disease have received extensive attention.Among the methods of diagnosing cardiovascular diseases,the most important and widely used method is to rely on the electrocardiogram（ECG）for analyzation.Patients need to go to the hospital to use professional acquisition equipment to obtain ECG signals,and then doctors diagnose cardiovascular diseases.In recent years,the development of smart wearable devices has made it easier and more convenient to obtain ECG signals,but a large number of ECG signals generated by these portable devices still need to be diagnosed by doctors,thus increasing the workload of doctors.This problem can be solved using artificial intelligence-assisted diagnosis technology.Through machine learning or deep learning technology,the diagnosis of ECG signals can be made without the participation of professional doctors;by deploying algorithms to hardware devices,the automatic diagnosis system can be made portable,which improves the real-time performance of ECG signal diagnosis and is convenient for patients to use.In this paper,using deep learning technology,a detection and classification algorithm for two common cardiovascular diseases,atrial fibrillation and arrhythmia,is proposed,and the arrhythmia classification algorithm is deployed to hardware divices.The main content can be divided into the following aspects.First,a supervised learning algorithm based on one-dimensional Convolutional Neural Network（CNN）is established in the classification task of arrhythmia.Based on the classical Residual Neural Network（ResNet）,the algorithm uses dilated convolution of different scales to establish multiple feature extraction channels.Compared with ordinary multi-scale convolution operations,the introduction of dilated convolution significantly reduces the amount of parameters without affecting the classification effect,which is beneficial to the lightweight of the model and makes it easier to deploy to the hardware side.In order to verify its classification effect,this paper trains and validates the model on the actual database from a hospital in Shanghai.The results show that the model achieved an F1 value of 0.89 in the four-classification task of arrhythmia,which proves that the model is effective in the arrhythmia diagnosis task of measured ECG data.Second,in view of the large amount of raw ECG signal data and the small amount of ECG signal data with cardiovascular disease labels,this paper proposes an unsupervised learning method based on ECG compression technology,and applies it to the detection task of atrial fibrillation.At present,one of the problems existing in the application of artificial intelligence in the field of ECG signal classification is that the open source databases containing cardiovascular disease labels that can be used for supervised learning and the number of patients included are small,and the neural network models trained based on these databases overfit unavoidably,bringing poor performances when applied to actual classification tasks.One of the ways to solve this problem is transfer learning,which first uses a large amount of unlabeled ECG signal data for pre-training phase,and then fine-tunes the pre-trained model by the classic open source database with labels for classification task learning.This paper designs a pre-training task of ECG compression using Convolutional Auto Encoder on the actual data of a hospital in Shanghai,and the feature extraction ability of the front-layer convolutional neural network is trained by the procedure of convolutional coding and deconvolutional decoding.This paper designs and conducts a comparative experiment to compare the performance of atrial fibrillation detection on the PhysioNet CinC 2017 database using the same neural network model with and without unsupervised learning.Experimental results show that this pretrain task brought an increase of F1 score by4.83%.The F1 value is the harmonic mean of the recall and the precision,which is an indicator that comprehensively considers false detection and missed detection.Finally,for the lightweight model of the arrhythmia classification algorithm proposed in this paper,the portable hardware implementation of the model is carried out,and the trained model is deployed on a miniaturized hardware device,which can realize the four-class classification task of arrhythmia with ECG signal inputs.If the front-end acquisition module is added,a portable hardware device covering the whole process from ECG signal acquisition to arrhythmia classification can be realized.The work of this paper includes the research on intelligent diagnosis algorithms from single-category detection of cardiovascular diseases to multi-category classification.Aiming at transfering to hardware,a lightweight neural network that can be used for arrhythmia classification of measured data is designed;in order to enhance the practical scope of the ECG intelligent diagnosis algorithm,an unsupervised learning algorithm and network with ECG compression task as the upstream task are designed;the proposed neural network for the four-classification task of arrhythmia is deployed to the hardware side.Combining the work of this paper with smart wearable acquisition equipment and remote transmission technology is expected to play a certain role in the field of smart medical care for cardiovascular diseases.