Research On Arrhythmia Detection Model Based On Machine Learning

Cardiovascular disease is a serious disease faced by middle-aged and elderly people at present.The initial stage of the disease is often accompanied by symptoms of arrhythmia.Therefore,it is of great significance for the prevention and early treatment of cardiovascular diseases to detect patients’ arrhythmia symptoms quickly and accurately with information and intelligent technology.However,traditional detection methods for arrhythmia based on ECG signal have limitations,some of these methods rely heavily on expert experience to form reasoning strategies,some of these methods need to select and extract cardiac rhythm characteristics artificially.As a result,the accuracy of the detection model is not high,and missed detections and false detections occur sometimes.Therefore,based on the characteristics of the ECG signals,this thesis conducts comparative experiments on various machine learning algorithms to select and optimize our algorithm,in order to build a better performance arrhythmia detection model.The main work and contributions of this thesis are as follows:(1)Firstly,based on the in-depth analysis of the MIT-BIH arrhythmia data characteristics,this thesis constructs the arrhythmia detection model based on multi-classification support vector machine and random forest respectively,and determines the appropriate key parameters of the model to optimize the performance index of the model through the comparison experiment and principle.At the same time,on the basis of more in-depth analysis of the detection results of the two models,the specific ideas for the further improvement of the model are determined.(2)Secondly,in order to solve the problems of mapping space classification and over-fitting inherent in the above two traditional machine learning algorithms,one dimensional convolutional neural networks for arrhythmia detection model is constructed in this thesis,and the basis and method for determining the hyperparameters of the model are described.On this basis,aiming at the timing characteristics of cardiac signals,this thesis improved the 1D-CNN model and added the long-short term memory module.The experimental results show that the detection accuracy of the improved 1D-CNN model can reach 98.87%.(3)Finally,in order to reduce the missed diagnosis rate of the case category of the model,traditional up-sampling is used to balance the original arrhythmia data set,but the effect is not good.Therefore,this thesis uses Auxiliary Classifier Generative Adversarial Networks for data balancing.Experimental results show that after data balancing,the missed diagnosis rate of arrhythmia cases of the improved 1D-CNN model is reduced from 5.09% to 3.42%.In order to enhance ACGAN’s ability to simulate data timing characteristics,this thesis adds an LSTM module to ACGAN,and proposes Auxiliary Memory Generation Adversarial Networks for data balancing.Experiments show that AMGAN significantly reduces the rate of missed detection for the four machine learning models discussed in this thesis.Compared with ACGAN,AMGAN further reduces the missed detection rate of the improved 1D-CNN model by 1.11%,which further enhances the practicability of the detection model.