Research On Blood Pressure Measurement Method From Pulse Wave Based On Deep Learning Algorithm

Cardiovascular disease(CVD)is one of the leading causes of death all over the world,and blood pressure(BP)is a key point for the prevention and treatment of CVD.Convenient,accurate and continuous BP measurement is of great importance both in clinical and in daily life.Common BP measurement methods require users to wear a cuff,which causes inconvenient measurement.In recent years,researchers have focused on using easy-to-collect pulse wave(PW)signal,which contains physiological information such as BP,to achieve cuff-less BP measurement.However,the relationship between PW and BP is difficult to fit.Fortunately,deep learning has become an effective solution,which can create an end-to-end mapping between PW and BP through network training.When designing deep learning algorithms,the following issues need to be considered.First,as one of the four diagnostic methods of traditional Chinese medicine,PW contains rich information.Accurate identification,extraction,and utilization of these characteristics play a decisive role in the accuracy of BP measurement.Besides,neural networks for BP measurement need to be trained individually for each user to make the measurement as accurate as possible.The training of neural networks consumes a lot of time and computation resources.It is necessary to design algorithms to achieve efficient training and ensure measurement accuracy in multi-user scenarios.Finally,the research and application of BP measurement algorithms often face the problems of low quality and insufficient quantity of data.Few-shot data limits the performance of neural networks,which hinders the application of deep learning algorithms.The main research contents of this manuscript are as follows:First,a hybrid neural network model is designed applying position encoding(PE),convolutional neural network(CNN)and long and shortterm memory(LSTM)network.The fingertip PW data is the input of the model,which end-to-end outputs BP measurement values.In the proposed model,PE converts the time information of PW signal into position encoding vectors in a high-dimensional space,then position encoding vectors and the original PW signal are mapped,fused,and input to the CNN.By carefully designing the number of layers and structure of the network,the receptive field of the feature map of CNN completely covers the length of the input,so that it can capture the long-term dependence and sequence information of the input,and realize the automatic extraction of features.Then LSTM is used to process features.The special input gates,forget gates,and output gates in the unit of LSTM can further exploit the long-term dependence and sequence information in the signal.Experimental results based on real-life dataset show that the proposed hybrid model realizes the full utilization of information in PW signal and improves the accuracy of BP measurement.The mean absolute error(MAE)in measuring systolic blood pressure(SBP)and diastolic blood pressure(DBP)are 3.26 mmHg and 1.47 mmHg,respectively.The accuracy of BP measurement achieves grade A of British Hypertension Society.The proposed model shows improvement compared to existing works.Besides,a transfer learning(TL)scheme based on discrete wavelet transform(DWT)and clustering algorithm is designed to improve the training efficiency of neural networks in multi-user scenarios.The scheme which is called TLDK is preformed based on the clustering results of users.TLDK fully considers the similarity between the source domain and the target domain.Since PW data contains many data points,it is difficult to measure the similarity between the PW data directly.DWT is firstly used to obtain approximate components containing fewer data points.K-means algorithm is then performed based on the approximate components to divide the users into several clusters.In each cluster,the user with the closest average distance to other users is used to train the base model,which is then transferred to other target users.Specifically,each target user uses its own data to train the base model.During the process of training,the parameters of the front part of CNN layers are frozen,and the parameters of other layers are fine-tuned,so as to retain the ability of extracting features of the base model and reduce the parameters that need to be trained.The influence of freezing and finetuning layers on measurement accuracy is obtained.The experimental results show that the TLDK scheme improves BP measurement accuracy while reducing the training time by 43.11%.The MAE of measuring SBP and DBP are 2.42 mmHg and 1.34 mmHg,respectively.Finally,this manuscript recruits volunteers to collect fingertip PW and BP data.Aiming at the problem of low quality of collected data,DWT is used to remove the power line interference.The quality index of pulse wave is introduced to identify and remove artifacts.The complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)is used to remove the baseline drift.As a result,highquality PW data windows are formed.Aiming at the problem of insufficient quantity of data,generative adversarial network(GAN)is used and designed based on PW to achieve data augmentation on the collected data.Based on the collected data,the generative model and the discriminative model in GAN play game with each other during the training process.The well-trained generative model can generate synthetic data with random noise as input.Using the synthetic data and the collected data together for the training of the neural network can break through the limitation of few-shot PW data on the performance of the neural network,thus improving the performance of measuring SBP and DBP by 3.64%and 1.33%,respectively.The BP measurement function is integrated as a module to achieve the transformation of proposed algorithm,which realizes the BP measurement based on the fingertip PW on mobile terminals.This manuscript focuses on BP measurement based on fingertip PW data,researches the design of neural networks,optimization of network training,data augmentation,transformation of algorithm and other issues in the full process.The proposed methods achieve high accuracy of BP measurement based on both datasets and collected data.This manuscript innovates the BP measurement method and makes positive contributions to Healthy China.