Research And Application On Atrial Fibrillation Diagnosis Model Based On Machine Learning

With development of economy and the aging of the population,the incidence of cardiovascular disease in China has increased year by year.At present,there are about 290 million patients with cardiovascular disease,and the mortality rate ranks first among various diseases.Atrial Fibrillation(AF)is a very common type of arrhythmia and an important cause of stroke,heart failure,and sudden death.Therefore,the prevention,diagnosis and treatment of AF are essential.This paper analyzes electrocardiogram(ECG)from AF with digital signal processing technology,data mining and machine learning technology,and establishes early diagnosis models of AF disease.It is hoped that it can help early warning of AF disease and reduce family and medical institutions burden.This paper chooses MIT-BIH AF database(static database),PhysioNet/Computing in Cardiology Challenge 2017 database(dynamic database),China PhysioNet Signal Challenge 2018 database and the wearable data,and uses machine learning methods to train the diagnosis model of atrial fibrillation.The main research content is as follows:(1)Research on atrial fibrillation diagnosis model based on RR interval.Seven short-term features of AF signals were extracted:Entropy-based AF,Sample entropy,Coefficient of sample entropy,mean RR intrerval,minimum heart rate,maximum heart rate and median heart rate.RBF-SVM and a hyperparameter search algorithm using grid search was used to construct diagnostic model of AF.(2)Research on AF diagnosis model based on modified frequency slice wavelet transform(MFSWT)and convolutional neural network(CNN).MFSWT was used to transform 1-D ECG waveforms into 2-D Time-Frequency(T-F)images.CNN was trained on 2-D T-F MFSWT images.(3)Research on the diagnosis model of AF based on enhanced multiple features.The predicted probability from the CNN model is taken as a new feature,and it is used as feature vector together with seven short-term atrial fibrillation signal features.A support vector machine is also used to establish the AF diagnosis model.The results show that using the predicted probability output by the CNN as a new feature can improve the performance of the AF diagnosis model.On the static database,the result from the 5-fold cross-validation:with the 30s time window,a sensitivity of 97.91%,a specificity of 97.82%and an accuracy of 97.87%;with 10s time window,sensitivity of 96.14%,specificity of 96.02%and an accuracy of 96.09%.This result is better than most of the results currently published,which indicates that the predicted probability from the CNN as a new feature can provide more useful information for the diagnosis of atrial fibrillation,and further apply the method to other physiological signals.It provides reference for other disease diagnosis and analysis.