Research On LSTM Algorithm For Prediction Of Atrial Fibrillation

Cardiovascular disease is the leading cause of death worldwide and has become increasingly young in recent years.Atrial fibrillation,one of the arrhythmias,is more urgent because it is closely associated with non-ischemic stroke.Among the main means of cardiovascular research,electrocardiogram(ECG)is one of the non-invasive and widely used in the diagnosis of cardiovascular disease detection tools.Usually,the diagnosis and treatment of ECG requires professional doctors to rely on observation experience and manual measurement,which is susceptible to many objective factors and inefficient.With the advancement of machine learning and deep learning in many fields,some scholars have used ECG data to carry out related researches on automatic detection algorithm of atrial fibrillation.However,in the prediction of atrial fibrillation,there are few studies on the real-time prediction of patients with paroxysmal atrial fibrillation,and there is a large gap in related engineering studies.In this paper,a prediction model of atrial fibrillation will be built based on the LongShort Term Memory(LSTM)network in deep learning.By analyzing ECG before and after the onset of atrial fibrillation and normal sinus rhythm of ordinary people,depth characteristics were extracted and trained.Multiple models are combined to predict the occurrence of paroxysmal atrial fibrillation and to enable patients to detect and eliminate the potentially fatal risk in time.Based on the existing basic medical theories,this study found that the predictability is mainly shown within 20 to 30 minutes before the occurrence.Based on this,target data sets are obtained from multiple open source databases.In this paper,an automatic extraction algorithm is designed for three types of label data,and atrial fibrillation-related data from three public databases is obtained using the label type of the target database while controlling a single variable as much as possible.For the differences among non-homologous databases,a detailed normalization method is proposed to eliminate the influence on the later model prediction.In this paper,Convolutional Neural Network is added for sequence reconstruction and feature extraction to reconstruct the input multi-lead data into the input form and feature suitable for the next layer LSTM network.The reconstruction method eliminates the dependence on the traditional expert characteristics and saves the cost of manpower and time.ECG belongs to time series data and has good agreement with LSTM network.In this paper,the Bidirectional LSTM network is introduced,which can extract the deep atrial fibrillation features from both the forward and reverse timing directions,and is more advantageous for the extraction of related rhythmic change information.In order to further optimize the performance of the model,the Attention Mechanism is added in this paper,which adaptively focuses on the more predictive parts from the large feature map.In this paper,we compare the current research advances,which show that the proposed model is superior in feature selection method,model optimization and model performance compared to the traditional machine learning model and the single network structure model.The accuracy rate of the model in the validation set reached 94.5%,and the accuracy rate reached 94.2%.In the final clinical dataset test of self-research database,the accuracy rate reached 92.7%,which proved the excellent generalization ability of the model.