Research On Noninvasive Continuous Blood Pressure Measurement Based On Self-extraction Of Array Signal Features By Deep Neural Network

Continuous blood pressure measurement can detect abnormal blood pressure in the human body in time for early detection,prevention and treatment of cardiovascular diseases,thereby avoiding the occurrence of serious complications.In the non-invasive continuous blood pressure measurement method,the measurement method based on pulse wave analysis only needs to collect one pulse wave signal.By analyzing the correlation between pulse wave characteristic parameters and blood pressure,a model is established to realize blood pressure estimation.The measurement is simple and has obvious advantages.Most of the existing measurement methods based on pulse wave analysis require manual extraction of features.Due to the large individual differences,the feature points are difficult to identify,and the scope of application is limited.Deep neural networks can automatically extract features and mine potential information in the data.Its powerful characterization capabilities are expected to achieve accurate estimation of blood pressure for a wider range of people.However,most of the existing deep learning methods use relatively shallow simple networks and relatively single features,which makes it difficult to achieve the ideal blood pressure measurement accuracy.In response to the above problems,this thesis uses array sensors to collect pulse waves to improve the quality of the collected signals,and studies the blood pressure estimation method based on the attention mechanism of time-frequency domain joint modeling,and establishes blood pressure estimation models on the MIMIC III database and self-built dataset.The main content includes the following aspects:(1)The self-built dataset in this thesis is composed of multi-channel pulse wave signals(array signals for short)collected by array sensors.In the signal preprocessing part,the array signals are enhanced by channel fusion.Since the direction of light reflection is not fixed,a single-channel acquisition signal will lose part of the effective signal components.Collecting multi-channel signals through the array sensor and performing channel fusion can make full use of the effective signal components.In addition,it is proposed to use permutation entropy and autocorrelation coefficient for signal screening,which is more efficient than manual screening.(2)In term of model,a method based on attention mechanism and time-frequency dualdomain joint modeling is proposed.Firstly,the convolutional network is used to extract the time-domain features of the pulse wave,and then the fully connected network and self-attention module are used to extract the frequency-domain features of the pulse wave,and then the crossattention mechanism is used to jointly model the time-frequency features of the pulse wave self-extracted by the deep neural network.The cross-attention mechanism can automatically activate the effective information in the features and improve the accuracy of model prediction.This article randomly selected 2526 patients' pulse waves and blood pressure from the MIMIC III public database to establish a model.The average absolute error of the model for systolic and diastolic blood pressure are 1.0181 mm Hg and 0.5187 mm Hg,respectively,and the error meat the AAMI standard.Based on a self-built data set,including the pulse waves and corresponding blood pressures of 36 volunteers,the errors of the model for systolic and diastolic blood pressure are 1.2473 mm Hg and 0.6039 mm Hg,respectively.(3)Based on the above method,a blood pressure prediction system is built,and ambulatory blood pressure can be output in real time only by collecting pulse wave signals.