Research On Detection And Prediction Of Atrial Fibrillation Based On Machine Learning

Cardiovascular disease is the leading cause of death.Atrial fibrillation is a very common arrhythmia-type cardiovascular disease.At present,the clinical intervention rate of atrial fibrillation is very little,so it is particularly important to explore the intelligent detection and prediction of atrial fibrillation.In this thesis,different machine learning models were established to realize automatic detection and prediction of atrial fibrillation.In order to solve the problem of insufficient data in machine learning,this thesis proposed a data augmentation method based on multiple time scales.In order to solve the problem that the ECG signal changes are not obvious before the onset of atrial fibrillation,this thesis combined Convolutional Neural Network and Long and Short Term Memory network to study,and achieved good results.The main research contents are as follows:(1)Research on atrial fibrillation detection model based on multi-time scale data augmentation and support vector machine.The ECG data were preprocessed and R peaks were detected,then 12 features based on RR interval which are commonly used in AF detection were extracted.Then the data was enhanced 6 times by multi-time-scale data augmentation method,and the intelligent detection model of atrial fibrillation was established by support vector machine.The classification accuracy,sensitivity and specificity of the model on the MIT-BIH are 98.06%,95.40%and 99.39%,respectively,and then,the 24-hour dynamic ECG data collected by the dynamic ECG laboratory of Shandong Provincial Hospital were used for testing,and the accuracy is 93.18%.(2)Research on atrial fibrillation prediction model based on modiefied frequency slice wavelet transform and CNN-LSTM hybrid model.The modiefied frequency slice wavelet transform was applied to the 10s ECG signals before atrial fibrillation occurred at least 45min and normal ECG in database to obtain the two-dimensional time-frequency diagram.The twodimensional time-frequency diagram was used as the input of CNN-LSTM hybrid model to predict atrial fibrillation.The prediction accuracy,sensitivity and specificity of the model on the MIT-BIH AFPDB are 98.06%,95.40%and 99.39%,respectively.