Continuous And Noninvasive Blood Pressure Measurement Based On Deep Neural Network And Its Applications

Hypertension has become a major problem that seriously endangers global public health.According to statistics,the current death of cardiovascular disease in China accounted for 41% of the total death,annual death of 3.5 million people,in which 70% of stroke and 50% of myocardial infarction are related with hypertension.Moreover,high blood pressure has become a "silent killer",not only because of high prevalence of hypertension and later mortality,more importantly,most hypertensive patients in their daily lives is not aware about their blood pressure.The awareness and control rates of hypertension are very unsatisfactory.Therefore,blood pressure management and regular monitoring are of great significance for early diagnosis of hypertension and prevention of hypertension.Blood pressure has a certain degree of fluctuation.Moreover,rapid changes of blood pressure may be closely related to the occurrence of cardiovascular diseases.However,current studies mostly focus on the accurate estimation of continuous blood pressure,even though the accuracy is not very satisfactory.Meanwhile,blood pressure variability is often overlooked.Therefore,this thesis has proposed blood pressure variability analysis through the time domain,frequency domain and nonlinear method,in order to find the key blood pressure parameters,which have predictive value of cardiovascular diseases.According to this research topic,this thesis has carried out the following aspects of the study:(1)Research on signal quality assessment based on constraint estimation.The accurate calculation of pulse transmit time is a prerequisite for accurate blood pressure estimation.The proposed method for signal quality assessment includes signal feature extraction module,constraint estimation modeling,timing prediction and signal quality analysis decision module,and an off-line parameter training and tuning module.For each signal segment,the extracted features are combined into a feature vector,and the constrained model is constructed according to the prior knowledge on the physiological signal.Finally three-tier signal quality assessment system is established in order to reject with serious interference.It is proven that this method can not only improve the efficiency of signal processing,but also improve the accuracy of feature extraction.(2)Research on continuous and noninvasive blood pressure estimation based on error back propagation neural network and deep learning.Considering the complexity of blood pressure,this thesis focuses on the continuous and noninvasive blood pressure estimation based on neural network architecture.In the preliminary study,the pulse transit time was calculated by electrocardiogram and photoplethysmogram.Combined with multiple individual information(heart rate,height,body weight,body mass index,sex,age)as the input of neural network,blood pressure can be estimated via the three-layer feedforward neural network.In the later experiment,we propose a novel continuous blood pressure estimation method based on combined information including waveform information,artificial features and personal features with a help of deep learning approach.It is proven that the algorithm is robust and has high accuracy in estimating blood pressure compared with the previous published studies,especially in the diastolic blood pressure.(3)Analysis of time domain,frequency domain and nonlinear blood pressure variability parameters.In this thesis,the evaluating method of blood pressure variability is proposed from the time domain,frequency domain and nonlinear theory respectively.In this thesis,we use the approximate entropy and sample entropy methods to assess blood pressure variability.(4)Evaluating the clinical value of blood pressure variability.Mainly from two aspects,the first aspect is for early identification of hypertension.This thesis collected the offsprings of hypertension patients and non-hypertensive people as subjects.We design the cold pressor test to explore the differences in the cardiovascular response between the two groups.The results showed that compared with the blood pressure level,blood pressure variabilities in the frequency domain were significantly different between the two groups,and thus the better predictors of early hypertension.The second aspect is about the target organ damage.In this thesis,the left ventricular mass index,carotid intima-media thickness,arterial wall distension,and plaques were measured.Through the statistical methods such as sample test,correlation analysis and multiple linear regression analysis,the relationship between blood pressure variability and these dependent variables was established,and moreover,the factors such as age and sex were taken into account in the model.Finally,this thesis summarizes the research work and the future work.