Research And Realization Of Atrial Fibrillation Detection Based On Deep Neural Network

Atrial fibrillation is a growing health care burden in the whole world.It is often asymptomatic and even may be a very short duration.Although there are many results on atrial fibrillation detection until now,most of these studies are limited to the standard databases and rarely were applied to the clinical.For multi-lead dynamic ECG machine wearing inconvenience,it was not suitable for personal health monitoring.Development of atrial fibrillation automated detection in order to achieve the early diagnosis and treatment is a challenging task.Deep learning technology is developed rapidly in recent years and it is widely used in various fields most of which are achieved good results.We want to adopt deep learning technology and single lead ECG data of wearable holter monitoring,aiming at proposing a novel method to realize automatic detection of atrial fibrillation in this thesis which has high sensitivity and strong specificity.We also want to apply this method to clinical diagnosis,screening and prognosis of atrial fibrillation,as were as other auxiliary work.What the main work of this thesis includes are as follows:(1)A great number of training samples is one of characteristics in deep learning models.The acquisition costs of labeled data are too high for they require experienced experts to mark them one by one which wastes too much time.We have collected enough clinical ECG data as the main component of unsupervised pre-training samples in our deep learning network.Before our experiments,we need some preprocessing for our ECG data.Firstly,we analyzed the R-R interval and P wave characteristics during atrial fibrillation in dynamic electrocardiogram.Secondly,we split the waveform by time window and normalized each dataset.We detect the R waves using wavelet transform method.It is facilitated to extract R-R interval time sequence correlation features.At last,construct the input sample features combination by combining with P wave and R-R interval characteristic.(2)We adopt stacked auto-encoders as the deep learning models to realize the automatic detection of atrial fibrillation.The decoder of auto-encoder can restructure the input,and the encoder can learn all the useful information of the raw input,which is the self-taught learning of deep learning.We add some sparsity restrictions to the auto-encoder can effectively accelerate the convergence of the model.This process is simulating our brain,when in the information processing,only a small number of neurons participate in.Through the unsupervised pre-training we can accomplish the ECG feature extraction automatic,and then achieve the global optimal initialization weights of the network.We can get the best estimate model by fine-tuning the whole network with labeled data.(3)The MIT-BIH standard library are used to estimate our experiments performance.The sensitivity of our result is 96.57%,the specificity is 99.04 %,and the overall accuracy is 98.31%...