Research On ECG-Related Diseases Detection Based On Machine Learning

With the rapid development of the economy and changes in the lifestyles of residents,the prevalence and mortality of chronic diseases in China have risen rapidly,and they have become the main factor in the death of urban and rural residents.However,the detection process of chronic diseases is extremely complicated and requires a lot of medical resources,resulting in the inability of a large number of patients to receive timely diagnosis and treatment.The General Office of the State Council released the ”New Generation Artificial Intelligence Development Plan” in 2017,proposing to build a new intelligent medical system and develop a wearable and biocompatible intelligent monitoring system.From point monitoring to continuous monitoring,the transformation from short-term health management to long-term health management can be realized,so as to achieve early detection,early prevention and early treatment of diseases,among which intelligent detection methods for diseases are the most important.The paper focuses on the detection methods of two ECG-related chronic diseases(sleep apnea and arrhythmia).The innovations are as follows:(1)A sleep apnea detection method(ECG-TD-MLP)based on time-dependent multilayer perceptron(MLP)is proposed.There is a certain time dependence between adjacent ECG signals,and the existing methods mainly perform sleep apnea detection based on the information of the current ECG signal,without considering the effect of time dependence on sleep apnea detection.To solve the problem,the research constructs a time feature based on the sliding window method,and uses the learning ability of the artificial neural network multilayer perceptron to extract the time dependence between adjacent ECG signals.Considering the impact of the number of hidden layer nodes in MLP on network performance,too few hidden layer nodes will result in the network not having the necessary learning and information processing capabilities;and too many nodes will not only greatly increase the complexity of the network structure,but also make the network fall into a local minimum.In the study,the hidden layer nodes are set based on the Kolmogorov superposition theorem.To make the network parameters more stable,Adam is used in the algorithm optimization stage to replace Stochastic Gradient Descent(SGD)to adjust the learning rate.The proposed ECG-TD-MLP method obtains an 87.3% accuracy of per-segment sleep apnea detection on the Apnea-ECG database.(2)An end-to-end sleep apnea detection method(ECG-Derived-CNN)based on Convolutional Neural Network(CNN)is proposed.Considering that traditional sleep apnea detection needs to utilize human-crafted features,there are certain limitations and complexities.The research introduces the convolutional neural network in deep learning into the sleep apnea detection research to realize automatic feature extraction.To avoid over-fitting caused by the direct use of electrocardiogram(ECG or EKG)signals,through the analysis of the physiological mechanism of sleep apnea,the derivative signals of electrocardiogram are used to replace the original electrocardiogram as the input of the convolutional neural network.Considering that there are too many parameters in the fully-connected layer,to reduce the risk of network overfitting,the Dropout technology is also introduced between the convolution layer and the fully-connected layer.Experimental results show that compared with traditional machine learning methods,the proposed ECG-Derived-CNN method can achieve end-to-end detection and improve sleep apnea detection performance.(3)An arrhythmia detection method(CNN-Focal-Loss)based on convolution neural network and focal loss is proposed.In the detection of arrhythmia,the number of normal heartbeats is much greater than the number of abnormal heartbeats,and the number of heartbeats between different types of arrhythmias is also very different.Existing arrhythmia detection methods mainly aim at the pursuit of high accuracy,without considering the serious imbalance between different types of heartbeats,resulting in a large number of minority arrhythmias being mistakenly classified into majority arrhythmias or normal heartbeats.To solve the problem,the research introduces a focus loss function to reduce the weight of simple class during training,so that the model can focus on hard-to-separate class,that is,the minority class.Considering the incomplete RR interval information caused by the heartbeat segmentation,the human-crafted RR interval features are also merged into the proposed CNN-Focal-Loss method.In order to speed up the training speed of the network and avoid the disappearance of the gradient in the optimization process,a Batch Normalization(BN)operation is added after each convolutional layer.Tested by the MIT-BIH database,the proposed CNN-Focal-Loss model not only obtains higher accuracy,but also effectively improves the detection performance of minority arrhythmias.(4)An arrhythmia detection method(CWT-2D-CNN)based on wavelet transform and 2D convolutional neural network(2D-CNN)is proposed.The aliasing of different frequency signals in the ECG signal and the interference of noise increase the difficulty of convolutional neural network representation.To solve the problem,the research uses wavelet transform’s excellent analysis ability in the time and frequency domain to decompose the ECG signal,so as to avoid the influence of different frequency signal aliasing on convolutional neural network representation.In order to take advantage of the powerful feature extraction capabilities of CNN on images,the coefficients of wavelet decomposition are constructed into a time-frequency scalogram,and 2D-CNN is used to extract features of heartbeats.Considering the incomplete RR interval information caused by the heartbeat segmentation,the human-crafted RR interval features are also merged with the CNN network.Experimental results show that the proposed CWT-2D-CNN method further improves the arrhythmia detection performance on the basis of previous work.