Intelligent Diagnosis Algorithm Of Arrhythmia ECG Signal Based On Convolutional Neural Network

Electrocardiogram(ECG)is a widely used non-invasive detection technology that can help doctors diagnose a variety of heart diseases.Arrhythmia,as a high incidence of heart disease,usually requires timely and effective treatment.The dynamic ECG acquisition device can perform ECG detection at any time and in any place.However,during the use process,problems such as low recognition efficiency and noise interference can be encountered.Convolutional Neural Network(CNN)is powerful and shows great advantages in image and signal processing.Based on CNN,we propose three algorithms to solve three types of problems encountered in the use of dynamic ECG acquisition equipment: classification of common arrhythmia,detection of ECG signal noise,and classification of severe arrhythmia-atrial fibrillation(AF).The main research work of this thesis is as follows:(1)Research on one-dimensional convolutional neural network(1D-CNN)algorithm for intelligent diagnosis of arrhythmiaBy analyzing the actual problems in the diagnosis of arrhythmia,the data set is collected by Holter and annotated after preprocessing.Based on the classic classification model of CNN,for the 12 types of arrhythmias with high incidence during diagnosis,we propose four types of 1D-CNN networks: 1D-VGG16,1D-Res Net18,1D-Res Net34 and1D-Goog Le Net.The influence of different convolution kernel sizes on the classification effect is analyzed to determine the optimal convolution kernel size.Finally,by comparing with the classical algorithms SVM and RBF,it is verified that the proposed algorithm has a good classification effect on 12 kinds of arrhythmia and can be used to assist doctors in diagnosing the disease.(2)ECG signal noise object detection algorithm based on CNNThe ECG signal is affected by the surrounding environment during the acquisition process due to the mixing of different types of noise.In this thesis,We study two types of noise,electromyographic noise and electrode motion artifacts,which are difficult to remove by the filter.Based on the YOLO target detection algorithm,we detect the noise contained in the ECG signal and propose the YOLO-ECG noise object detection algorithm.First,noise-free ECG data and noise data are obtained in the public data set,and the two are superimposed to obtain noise ECG data with different signal-to-noise ratios.Through two sets of experiments with different noise levels and different noise types,it is concluded that the YOLO-ECG noise object detection algorithm has a good detection effect on noise.Compared with the classification accuracy and algorithm complexity of YOLOv3 and YOLOv3-Tiny target detection algorithms,it shows that YOLO-ECG can effectively reduce the complexity of the algorithm,accurately screen out the noise contained in ECG signals,and provide convenience for subsequent diagnosis of arrhythmia.(3)Research on AF detection algorithm for dynamic electrocardiographyAtrial fibrillation(AF)is one of the causes of stroke in patients.Because of its complexity and suddenness,Diagnosis of AF takes longer time to monitor the ECG than other arrhythmia.AF detection has become a research hotspot for many scholars.Based on Holter’s use environment,we use the Physio Net / Cardiology Challenge data held in 2017 to improve the Mobile Netv3 network structure which widely used in embedded devices,and propose a Mobile AF algorithm for AF detection.we also select the average F1 value as an index to measure the accuracy of the classification,and compare the top five algorithms in the open challenge.Considering the classification accuracy of the algorithm,the complexity of the model,and the training time,it is concluded that the classification accuracy of the Mobile AF algorithm ranks second among the six algorithms,which is only about 2% lower than the first algorithm.But complexity and training time have absolute advantages,which are about 90% less than other comparison algorithms.Finally,through experiments on clinically collected data,it is verified that the Mobile AF algorithm has a good detection effect on AF and is suitable for use in Holter.