Blood Pressure Waveform Measurement By Deep Neural Network Based On Pulse Visualization

Nowadays,the fast-paced life style and great work pressure have brought a variety of health problems to people,especially cardiovascular diseases have caused great damage to people’s heart and brain.Once cardiovascular diseases occur,serious cerebral hemorrhage or cerebral ischemia and other problems are very likely to occur,thus seriously threatening the health of patients.Blood pressure is one of the vital signs showing the basic health information of patients,and is also the main basis for clinical diagnosis and treatment.Keeping blood pressure under control on a daily basis is the primary goal of a healthy life.Therefore,how to measure blood pressure effectively and predict the trend of blood pressure has become the key to prevent cardiovascular diseases.Among them,the method based on pulse wave has attracted wide attention because of its easy acquisition and accurate measurement.Pulse wave is the regular pulsation of arteries,which contains important information about blood pressure.Therefore,this thesis studies blood pressure measurement based on pulse wave.Specific research methods are as follows:Firstly,pulse blood pressure data set is constructed.Based on the formation mechanism of pulse and blood pressure,a human pulse simulation platform is developed in this thesis to provide a reference platform for non-invasive measurement of blood pressure waveform.In order to obtain the pulse waveform data generated by the human pulse simulation platform(blood pressure waveform data is obtained by the internal tube pressure transmitter of the human pulse simulation platform),binocular vision acquisition platform is developed for pulse image data acquisition,and then pulse image is processed to obtain pulse wave.Pulse and blood pressure data sets are generated after simple preprocessing of pulse and blood pressure waves,such as filtering,baseline drift removal,normalization and interpolation/extraction.Secondly,the blood pressure waveform measurement model was constructed through the back propagation neural network,the back propagation neural network improved by genetic algorithm,the back propagation neural network improved by particle swarm optimization algorithm and the back propagation neural network improved by Sparrow algorithm to explore the relationship between the threedimensional radial artery wall displacement and blood pressure waveform.Correlation coefficient and mean square error are used as evaluation criteria of the model.The results show that the back propagation neural network improved by sparrow swarm algorithm has high accuracy.However,the overall accuracy of the blood pressure waveform measurement model constructed by the back propagation neural network and its optimization model is not high enough to be used as a real blood pressure waveform measurement model.It is necessary to explore a new blood pressure waveform measurement model.Thirdly,because the data set shows a complex temporal correlation,a recurrent neural network module is added to the new blood pressure waveform measurement model.However,when the input pulse and blood pressure sequence is very long,the traditional recurrent neural network will have the problem of gradient explosion or disappearance.In order to solve the long term dependence problem of recurrent neural networks,this thesis uses two other representative models based on recurrent neural networks.One is the long short term memory network,and the other is the gated recurrent unit.Experiments show that both models have better performance than the traditional recurrent neural network model.Then,this thesis constructs a mixed model of convolutional neural network and long short term memory network,and makes comparison with network models such as convolutional neural network,long short term memory network,bidirectional long short term memory neural network,gated recurrent unit network and recurrent neural network.Experimental results show that the mixed model of convolutional neural network and long short term memory network has the highest accuracy.It can be used as a good model for measuring blood pressure waveform.Finally,attention mechanism is introduced to optimize the hybrid model of convolutional neural network and long short term memory network.The main principle of attention mechanisms is that models learn by focusing on areas of interest,and attention mechanisms have also recently been shown to be effective in signal processing.Attention mechanisms can assign more weight to important information,and various approaches have emerged to apply them.Since in the proposed model,the sequential feature vectors of the long short term memory networks may have different effects on the value of blood pressure measurement,this thesis adds the importance of an attention mechanism model to automatically train the feature vectors in each time step.In this thesis,a blood pressure waveform measurement model was constructed based on the Informer model,and a mixed model of convolutional neural network and long short term memory network with attention mechanism module was constructed as a comparison model.The experimental results showed that the models with the added attention mechanism showed better results than the other models.This thesis also proves this conclusion by comparing with other literatures at home and abroad.