Research On Intelligent Prediction And Accurate Diagnosis Algorithm Of Cardiovascular Disease Based On Deep Learning

With the changes in population structure and social environment in China,cardiovascular disease has become the leading cause of death,surpassed tumors.More than 40% of the total number of deaths per year of residents is caused by cardiovascular disease,and the number is increasing year by year.More seriously,with the increase of air pollution and the spread of bad living habits,the patient population of cardiovascular disease appears to be younger and younger.The monitoring and prevention of cardiovascular diseases consume a large amount of funds and resources from the state,which has become a major problem in improving the national health level and accelerating the development of national health undertakings.The main cause of high mortality of cardiovascular disease is its insidious and sudden onset.As the most effective tool for clinical diagnosis of cardiovascular disease,dynamic electrocardiogram(ECG)and medical imaging have their own characteristics.The portability of ECG makes it possible to monitor and predict sudden cardiovascular disease in real time,and has a prominent role in predicting cardiovascular disease.However,it is a weak electrical signal in vitro which limits its use in peeping internal mechanism.The high precision of medical imaging makes it possible to explore the underlying causes of cardiovascular disease and plays a prominent role in the precise diagnosis and treatment of cardiovascular diseases,but it is not a real-time performance.Therefore,combining artificial intelligence technology with them,giving full play to the advantages of ECG and medical imaging,can effectively improve the efficiency of cardiovascular disease prevention and treatment,and reduce the mortality rate of cardiovascular disease.In view of the above problems and challenges,this paper studies the intelligent prediction and accurate diagnosis algorithms of cardiovascular diseases from the two data levels of ECG and medical imaging.At the data level of ECG,the intelligent prediction and real-time warning algorithm for high-risk cardiovascular disease wasstudied.At the data level of medical imaging,coronary vascular morphology assessment and automatic identification of intravascular plaques and vulnerable plaques that are closely related to high-risk cardiovascular disease were studied.The innovative research work of this thesis mainly includes the following aspects:1.A sparse auto-encoded deep neural network with a four-layer stack structure was designed to automatically extract the deep features of the heartbeat in ECG.Through stratified training and optimization,normal,atrial premature,ventricular premature,left bundle branch blockage,and right bundle branch blockage beats were accurately recognized.The average recognition accuracy was 99.5%.It provided the technical assistance for intelligent prediction of high-risk cardiovascular disease in ventricular fibrillation2.An intelligent prediction algorithm for sudden cardiac death(SCD)based on the echo state network is proposed.By designing an echo state network with a multilayer structure,the intelligent distinction between SCD signals and non-SCD signals is realized.The average prediction accuracy is 94.32% testing on 5 minutes signal before the occurrence of SCD.It provides new ideas for the prediction of SCD and it provides a guarantee for intelligent monitoring and real-time early warning of high-risk cardiovascular diseases.3.An automatic extraction algorithm for the intima of the coronary OCT image based on the linear label maximum flow algorithm was proposed.The extraction problem of the intima contour was transformed into the segmentation problem of the linear label region.The radial gray distribution of OCT image is used to set the grayscale label value in the linear label maximum flow algorithm,thus achieving accurate extraction of the intima contour of different characteristic OCT images.The algorithm has a good effect on the OCT images with boundary blur,stents,plaques or thrombus.The average Dice coefficient is 0.972.It provides data support for accurate assessment of coronary vessel morphology.4.An automatic plaque and vulnerable plaque recognition algorithm based on A-line depth modeling was proposed.Using the ability of self-learning of stacked sparse auto-encoding network,a large number of unlabeled data were designed to train the sparse auto-encoding neural networks to automatically extract the deep representation of plaque features.Then introduce a small number of labeled data to micro-train the entire network weights.Fibrous,fibrous-calcific,and fibrous-lipid plaques were accurately identified with limited labeled data.The average regional coincidence of the three types of plaques was 87%,87%,and 85%,respectively.Through the automatic analysis of the thickness of the fibrous cap in the plaque,the vulnerable plaque with thin fibrous cap was automatic identified.The average coincidence of the vulnerable region was 87%.The overall time consumption for plaque and vulnerable plaque recognition is 0.54 seconds.It provides theoretical support for accurate diagnosis and inner cause analysis of high-risk cardiovascular diseases.