CN117357079A - Human blood pressure measuring method based on individual correction - Google Patents

Abstract

The invention discloses a human blood pressure measurement method based on individual correction. The method utilizes correction waveforms and correction blood pressure to carry out individual correction prediction on the blood pressure of the waveform to be detected based on an individual blood pressure correction prediction network; the measuring method comprises the following steps: based on the idea of template matching, extracting a single-period pulse waveform from pulse wave signals of the correction waveform and the waveform to be detected for a period of time; and calibrating the blood pressure prediction of the waveform to be measured based on the individual correction blood pressure prediction network by using the correction waveform and the correction blood pressure to obtain a more accurate blood pressure prediction result. The individual blood pressure correction prediction network fuses the characteristic characterization of U-Net and the residual error structure of ResNet, and provides a better characteristic extraction module for carrying out characteristic characterization on the input waveform; the training mode of the twin network is adopted, so that the characteristic extractors of the correction waveform and the waveform to be detected share network parameters, thereby better extracting the characteristic difference of the homomodal data and improving the accuracy of blood pressure prediction.

Description

Human blood pressure measurement method based on individual correction

Technical field

The invention relates to an individually calibrated non-squeezing blood pressure measurement method, in particular to a blood pressure measurement method based on fingertip pulse waves, which is applied to the fields of human physiological status detection and physical health status monitoring.

Background technique

By monitoring human blood pressure, the operating status of the cardiovascular system can be effectively assessed to gain a general understanding of the human body's health status. Blood pressure is an important physiological indicator that is of great significance in daily life and medical testing. People can adjust their living habits and behaviors based on blood pressure to detect and treat high or low blood pressure early. Blood pressure monitoring provides an important reference for maintaining human health.

Due to the translucent nature of skin, when light irradiates the skin surface of a person's face, part of the light will penetrate the epidermis and enter the blood vessels. This part of the light will be partially absorbed by hemoglobin and other substances in the blood, and produce diffuse reflection by the camera. captured. As the blood flow rate changes, the light absorbed by the blood under the skin also changes. Therefore, by extracting the chromaticity information of the fingertips, the human body's pulse wave can be obtained, and the blood pressure information can be extracted from the detailed changes in the pulse wave.

With the development of machine learning and deep learning, it has become possible to establish a blood pressure prediction model based on manual features with the help of a huge amount of training parameters and model parameters. The usual approach is to extract manual features based on the analysis of the time domain and frequency domain of the pulse wave, and then use machine learning and deep learning to fit the blood pressure value. For example, in the paper "A NeuralNetwork-based Method for Continuous Blood Pressure Estimation from a PPGSignal" at the 2013 Instrumentation & Measurement Technology Conference, the researchers performed time-domain manual feature extraction on the fingertip PPG signal, and then introduced a three-layer feedforward neural network Model and train the blood pressure prediction problem to complete the blood pressure prediction of PPG signals. However, although neural networks can better learn the relationship between manual features and blood pressure, since blood pressure prediction mainly relies on detailed changes in waveforms, there is a performance bottleneck in blood pressure prediction relying on manual features.

With the successful application of the encoder-decoder structure in the field of deep learning, using the encoder structure to extract features from PPG signals has become an important means to break through the bottleneck of blood pressure prediction algorithms based on PPG signals. Compared with the neural network structure that directly inputs features, the encoding and decoding structure can convert the input signal into a fixed-length vector for feature characterization, thus avoiding the step of manual feature selection and allowing the network to extract features of changes in waveform details by itself. Better utilize the learning ability of neural networks. In the paper "PPG-based blood pressure estimation can benefit from scalable multi-scale fusionneural networks and mutli-task learning" in the 2022 Biomedical Signal Processing and Control journal, the researchers first extracted the first-order derivative and second-order derivative of the PPG waveform. It is used to enrich the input waveform information, and then builds an encoder module to extract the characteristics of the waveform and predict blood pressure through the feedforward network. This method shows a good fitting effect on the training set. However, the algorithm has a large difference in blood pressure prediction between different individuals, which affects the effectiveness of the blood pressure prediction algorithm in practical applications.

Contents of the invention

The purpose of the present invention is to solve the problem in the prior art that blood pressure prediction has poor universality in the face of individual differences, and to propose a blood pressure prediction method based on individual correction. In order to overcome the influence of individual differences, such as differences in skin color, physical condition and other factors, the present invention proposes an individual correction method, which performs feature correction on the waveform to be measured by providing a correction waveform, and performs blood pressure prediction based on the corrected blood pressure. At the same time, the present invention generates a robust single-cycle pulse waveform based on the idea of template matching, which is used to characterize the entire waveform signal to improve prediction robustness. The feature extraction module proposed by the present invention combines the advantages of U-Net and ResNet models and can extract pulse wave waveform features more robustly. The present invention uses a twin network training method to share the feature extraction module parameters of the correction waveform and the waveform to be measured during the training process, which not only reduces the amount of network parameters, but also better captures the characteristic differences of the waveforms and improves the accuracy of blood pressure prediction.

The present invention adopts the following technical solutions to achieve:

A human blood pressure measurement method based on individual correction, using correction waveforms and corrected blood pressure, and performing individual correction predictions on the blood pressure of the waveform to be measured based on an individual blood pressure correction prediction network. The training process of the individual blood pressure correction prediction network includes the following steps:

S1: Data set construction: Construct a data set containing the corrected waveform, corrected blood pressure, waveform to be measured and the true value of the blood pressure of the waveform to be measured.

S2: Perform single-cycle pulse wave acquisition based on template matching for the photoplethysm pulse waveform in the data set.

First, based on the Fourier transform, the approximate length of the single-cycle waveform template is calculated.

Then, based on the idea of template matching, the similarity of waveforms is used to select single-cycle waveform segments, and the selected single-cycle waveform segments are fused to obtain a single-cycle pulse wave for the overall representation of the waveform.

S3: Blood pressure prediction based on individual corrected blood pressure prediction network.

S31: Use the waveform feature extraction module Uresnet to extract the single-cycle pulse wave features of the corrected waveform and the waveform to be measured; then, compare the single-cycle pulse wave features of the corrected waveform and the waveform to be measured to obtain the difference in waveform features.

S32: Using the difference in waveform characteristics and corrected blood pressure, perform individual correction of the blood pressure of the waveform to be measured based on the correction prediction module.

S4: Using the true value of the blood pressure of the waveform to be measured as the supervision information and the mean square error as the loss function, train the individual corrected blood pressure prediction network to convergence, and obtain the optimal individual blood pressure correction prediction network.

In the above technical solution, further, the individual corrected blood pressure prediction network is based on the twin network training method and utilizes the same modal characteristics of the corrected waveform and the waveform to be measured to share the network parameters of the feature extractor, thereby extracting high-quality waveform features. Based on the difference in high-quality waveform characteristics, the blood pressure of the waveform to be measured is finally accurately corrected individually.

Further, in step S1, a data set is generated based on the existing MIMIC physiological characteristic data set for training and testing of the individual corrected blood pressure prediction network. details as follows:

The pulse wave and dynamic blood pressure wave in the MIMIC database are cut according to a certain length of time (30-90s), and the first pulse wave and dynamic blood pressure are used as the correction waveform and corrected blood pressure.

Take the maximum and minimum values of each section of ambulatory blood pressure as the true values of systolic blood pressure and diastolic blood pressure.

Match the corrected waveform and corrected blood pressure with the subsequent waveform to be measured and the true blood pressure value of the waveform to be measured one by one to generate a data set.

Further, step S2 performs template-matching single-cycle waveform extraction on the waveforms in the correction data set. details as follows:

Wavelet transform is used to detrend each section of the waveform to be measured and the correction waveform, so that only the AC component in the waveform is retained.

Fourier transform is performed on each waveform to be measured and the correction waveform to obtain the approximate heart rate value, thereby determining the template length of the single-cycle pulse wave.

The idea of template matching is reflected in the similarity of waveforms. Calculate the first derivative of each waveform to be measured and the corrected waveform after wavelet transformation, and obtain the zero point coordinates. Calculate the area of the image formed by the adjacent zero points and the x-axis, and set the threshold α to float up and down based on the median of all areas. α is used as the area selection interval to select the starting zero points of all waveforms.

With the starting zero point as the center and the template length of the single-cycle pulse wave as the length, select the lowest point of the original single-cycle pulse wave as the starting point of the waveform, and cut out the single-cycle pulse wave of the template length.

The final single-cycle pulse wave is obtained by averaging all the cut single-cycle pulse waves.

Further, in step S3, the individual corrected blood pressure prediction network is constructed as follows: constructing the waveform feature extraction module and the correction prediction module of twin towers, and the twin network training method is used in the training of the individual corrected blood pressure prediction network. details as follows:

Based on the twin-tower structure and combined with U-Net and ResNet, the waveform feature extraction module Uresnet is constructed to extract features of the correction waveform and the waveform to be measured respectively, and calculate the waveform difference features accordingly. details as follows:

Input the correction waveform PPG _adj and the waveform to be measured PPG _mea into the feature extraction modules Uresnet _adj and Uresnet _mea of the correction waveform and the waveform to be measured respectively, and extract the single-cycle pulse wave features F _adj and F _mea of the correction waveform and the waveform to be measured. Secondly, in order to extract the feature difference, the single-cycle pulse wave features F _adj and F _mea of the correction waveform and the waveform to be measured are subtracted to obtain the waveform feature difference F _diff .

The correction prediction module mainly introduces the correction of blood pressure, performs individual correction on the blood pressure of the waveform to be measured, and then outputs the blood pressure prediction result. details as follows:

First, the difference in waveform characteristics is used to predict the blood pressure coefficient: Coef _bp = f (F _diff ), where f represents the full connection operation and Coef _bp represents the predicted blood pressure coefficient; then the corrected blood pressure is multiplied by the blood pressure prediction coefficient to achieve accurate blood pressure prediction. : BP _est = BP _adj *Coef _bp , where BP _adj is the corrected blood pressure and BP _est is the predicted blood pressure.

The individual corrected blood pressure prediction network is trained using the twin network training method, sharing the parameters of two waveform feature extraction modules instead of the double-ended feature extraction module. Based on the fact that the correction waveform and the waveform to be measured belong to the same modal data, sharing the parameters of the waveform feature extraction module can better guide the network to extract the difference between the correction waveform and the waveform to be measured.

The feature extraction modules Uresnet _adj and Uresnet _mea of the correction waveform and the waveform to be measured share network parameters. The implementation method is: only one feature processing module Uresnet _share is defined globally. The single-cycle pulse wave features F _adj and F _mea of the correction waveform and the waveform to be measured only use the Uresnet _share for feature extraction, and the back propagation only updates the parameters of the Uresnet _share . , reduce the amount of model parameters, and better extract the characteristic differences between the waveform to be measured and the corrected waveform.

The beneficial effects of the present invention are: based on the individual correction mechanism, the difference factors in different individual measurements are eliminated. The correction waveform and the guidance information for correcting blood pressure give the waveform to be measured an additional reference value, thereby obtaining a more accurate blood pressure prediction; at the same time, through the waveform The design of the feature extraction module and the selection of the individual corrected blood pressure prediction network training method can obtain more obvious feature differences between the corrected waveform and the waveform to be measured, and predict the blood pressure results more accurately. For individual non-squeezing blood pressure measurement, the present invention has strong application value.

Description of the drawings

Figure 1 is a flow chart of the method of the present invention.

Figure 2 is a flow chart of single-cycle pulse wave acquisition according to the present invention.

Figure 3 is a template matching flow chart.

Figure 4 is a network structure diagram of the present invention.

Figure 5 is a schematic structural diagram of the feature extraction module of the present invention.

Detailed ways

The present invention will be further described below with reference to specific embodiments and drawings.

Under the condition of keeping as still as possible, the human body to be measured first uses a fingertip pulse meter with a sampling frequency fs of 60Hz to record the fingertip pulse wave, and at the same time uses a cuff sphygmomanometer to measure the blood pressure. The fingertip pulse wave collection duration is not less than 10 seconds. , record the fingertip pulse wave and blood pressure values as corrected waveform PPG _adj and corrected blood pressure BP _adj .

The preprocessing steps for correcting the waveform are shown in Figure 2.

Use wavelet transform to detrend the corrected waveform PPG _adj to obtain PPG _adj-d . Perform Fourier transform on PPG _adj-d to calculate the heart rate HR, and thereby calculate the approximate single-cycle pulse wave template waveform length.

Perform template matching on the detrended waveform to obtain a more robust waveform representation. The steps are shown in Figure 3, specifically as follows:

Calculate the first-order derivative waveform APG _adj-d for the corrected waveform PPG _adj-d after detrending transformation, and calculate the zero point coordinate Zeros _n , where n represents the number of zero points.

Calculate the image area Area _n-1 surrounded by adjacent zero points and x-axis coordinates, and take its median Area _middle . The zero point Zeros _m within the upper and lower offset thresholds is considered to be the selectable starting zero point of the waveform.

In this example, the threshold is set to 0.1.

Use the corresponding zero point in Zeros _m before and after in PPG _adj-d Find the lowest point of the waveform and cut out a single-cycle waveform PPG _i of length L, where i represents the i-th cut out single-cycle waveform.

Average the cut single-cycle waveforms to obtain the single-cycle pulse wave of the corrected waveform.

Single-cycle pulse wave that will correct the waveform and corrected blood pressure BP _adj serve as correction inputs to the individual corrected blood pressure prediction network.

Under the condition of keeping as still as possible, the human being tested uses a fingertip pulse meter to collect pulse waves. The collection duration is not less than 10 seconds, and the waveform to be measured PPG _mea is recorded.

Repeat steps [0044]-[0050] for PPG _mea to obtain the single-cycle pulse wave of the waveform to be measured. Follow-up blood pressure measurements were taken.

The network architecture of the individual corrected blood pressure prediction network proposed by the method of the present invention is shown in Figure 4, which is mainly divided into two modules: waveform feature extraction and correction prediction.

The waveform feature extraction module is shown in Figure 5. It is mainly based on the Encoder structure of U-net for feature extraction. At the same time, in order to improve the feature representation ability and alleviate the problem of gradient disappearance, combined with the idea of ResNet, each layer of U-net convolution is Add a layer of short wires to the module for identity mapping.

The waveform feature extraction process based on Uresnet is described in detail as follows: the waveform to be measured and correction waveform It is processed in two ways: first, through the first convolutional layer, the dimension is expanded to obtain the initial waveform characteristics to be measured/> and correction waveform characteristics/>

Initial waveform characteristics to be measured and initial correction waveform characteristics/> Both tensor dimensions are C×D, where C and D represent the channel number and length of the feature tensor respectively. Then continue to pass the initial features through the convolution layer with a convolution kernel of 5, and then through the ReLU activation function, and then Repeat the convolution step twice and upsample to obtain/> and/> The tensor latitude becomes At the same time, the features also pass through short lines and generate residual features R _mea and R _adj through the convolution layer with a convolution kernel of 1 (i.e., the waveform feature extraction module), and the tensor latitude is/> Then, the residual feature and the upsampled waveform feature are added to generate the final waveform feature F _mea and the corrected waveform feature F _{adj of} the waveform to be measured.

Secondly, during the network training process, the twin network training method is used to share parameters between the two feature extraction modules, that is, the parameters W _mea and W _adj of the feature extraction modules of the waveform to be measured and the correction waveform are the same. In implementation, a single feature extraction module Uresnet _share is defined, and only this module is calculated during calculation and gradient descent. The training method of Siamese network has two important effects on waveform feature difference extraction. By sharing parameters, the training parameters of the model are effectively reduced and the model is prevented from overfitting. At the same time, since the waveform to be measured and the corrected waveform belong to the same modal data, using the same feature extraction module can better extract the difference in waveform features.

The correction prediction module uses the waveform difference feature F _diff and the corrected blood pressure BP _adj respectively, fuses the two information through the product, and finally outputs the predicted blood pressure. The specific operation is as follows: F _diff generates the blood pressure coefficient Coef _bp through two fully connected layers. The dimension of the blood pressure coefficient is 1×2. Then the idea of individual correction is introduced, and BP _adj is multiplied by Coef _bp to obtain BP _est . Using the true value of the blood pressure to be measured as the supervision information and the mean square error as the loss function, the individual corrected blood pressure prediction network is trained to convergence. The important role of introducing corrected blood pressure is to improve the use value of the model and improve the poor generalization problem of the original neural network blood pressure prediction algorithm.

Based on the existing MIMIC correction data set, this invention is compared with "ANeural Network-based Method for Continuous Blood Pressure Estimation from a PPG Signal" (marked as Method 1 in the table) and "PPG-based blood pressure estimation can benefit from scalable multi The blood pressure prediction algorithm proposed by "-scalefusion neural networks and mutli-task learning" (marked as method 2 in the table) has been significantly improved in the test sets of different individuals, which reflects the effectiveness of the present invention in improving the generalization ability of the blood pressure prediction method. Comparison The results are shown in Table 1.

Table 1 Comparison of results of different algorithms

The above embodiments are used to illustrate the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modifications and changes made to the present invention fall within the protection scope of the present invention.

Claims (8)

1. The human blood pressure measurement method based on individual correction is characterized in that the individual correction prediction is carried out on the blood pressure of the waveform to be measured based on an individual blood pressure correction prediction network by using a correction waveform and correction blood pressure, and the training process of the individual blood pressure correction prediction network comprises the following steps:

s1: data set construction: constructing a data set containing correction waveforms, correction blood pressure, waveforms to be measured and blood pressure true values of the waveforms to be measured;

s2: carrying out single-period pulse wave acquisition based on template matching on waveforms in the data set;

firstly, calculating the approximate length of a monocycle waveform template based on Fourier transformation; then, based on the idea of template matching, selecting segments of the monocycle waveform by utilizing the similarity of the waveforms, and fusing the selected segments of the monocycle waveform to obtain monocycle pulse waves for overall characterization of the waveforms;

s3: performing blood pressure prediction based on the individual correction blood pressure prediction network;

s31: extracting single-period pulse wave characteristics of the correction waveform and the waveform to be detected by utilizing a waveform characteristic extraction module Uresnet; then, comparing the single-period pulse wave characteristics of the correction waveform and the waveform to be detected to obtain waveform characteristic differences;

s32: the method comprises the steps of performing individual correction on the blood pressure of a waveform to be detected based on a correction prediction module by utilizing waveform characteristic differences and correction blood pressure to obtain a blood pressure prediction result;

s4: and training the individual correction blood pressure prediction network to converge by taking the true value of the waveform blood pressure to be detected as supervision information and the mean square error as a loss function to obtain the optimal individual blood pressure correction prediction network.

2. The method for measuring human blood pressure based on individual correction according to claim 1, wherein in the step S1, a data set is generated based on the existing MIMIC physiological characteristic data set, and the specific method is as follows: s11, cutting pulse waves and dynamic blood pressure waves in a MIMIMIIC database according to a certain time length, and taking the first pulse wave and the dynamic blood pressure as correction waveforms and correction blood pressure;

s12, taking the maximum value and the minimum value of each section of dynamic blood pressure as systolic pressure and diastolic pressure values;

s13, matching the corrected waveform and the corrected blood pressure with the follow-up waveform to be measured, wherein the blood pressure true values of the waveform to be measured are matched one by one, and a data set is generated.

3. The method for measuring human blood pressure based on individual correction according to claim 1, wherein the step S2 is specifically as follows:

s21, carrying out trending treatment on each section of waveform to be detected and correction waveform by using wavelet transformation, so that only alternating current components in the waveform are reserved;

s22, carrying out Fourier transformation on each section of waveform to be detected and the correction waveform, and solving an approximate heart rate value so as to determine the template length of the single-period pulse wave;

s23, obtaining a first derivative of each section of waveform to be detected and correction waveform after wavelet transformation, obtaining zero coordinates, calculating the image area surrounded by adjacent zero and an x axis, setting a threshold value alpha, taking the floating alpha of the median of all areas as an area selection interval, and selecting the initial zero of all waveforms;

s24, taking a starting zero point as a center, taking the template length of the single-period pulse wave as a length, selecting the lowest point of the original single-period pulse wave as a starting point of a waveform, and cutting out the single-period pulse wave with the template length;

s25, carrying out average calculation on all the cut monocycle pulse waves to obtain a final monocycle pulse wave.

4. The method for measuring human blood pressure based on individual correction according to claim 1, wherein in step S3, the individual correction blood pressure prediction network is trained based on a twin network training mode, and uses the homomodal characteristics of the correction waveform and the waveform to be measured to share the network parameters of the waveform feature extraction module urenet.

5. The method according to claim 1, wherein in the step S31, the waveform feature extraction module urenet specifically adopts an Encoder module of U-Net to perform feature extraction, and adds a layer of shorting wire to perform identity mapping in each layer of convolution module of U-Net, so as to improve feature representation capability and alleviate gradient disappearance problem.

6. The method for measuring human blood pressure based on individual correction according to claim 1, wherein the step S31 specifically comprises: first, correct waveform PPG _adj And waveform to be measured PPG _mea Feature extraction module Uresnet for respectively inputting correction waveform and waveform to be detected _adj And Uresnet _mea Extracting single-period pulse wave characteristic F of correction waveform _adj And single-period pulse wave characteristics F of the waveform to be measured _mea The method comprises the steps of carrying out a first treatment on the surface of the Then, the single-period pulse wave of the correction waveform is correctedFeature F _adj And single-period pulse wave characteristics F of the waveform to be measured _mea Performing subtraction operation to obtain waveform characteristic difference F _diff 。

7. The method for measuring human blood pressure based on individual correction according to claim 6, wherein the step S32 is specifically: firstly, predicting a blood pressure coefficient by utilizing waveform characteristic differences: coef _bp ＝f(F _diff ) Wherein f represents a full connection operation, coef _bp Representing predicted blood pressure coefficients; then, the predicted blood pressure is obtained by multiplying the corrected blood pressure by the blood pressure coefficient: BP (BP) _est ＝BP _adj *Coef _bp Wherein BP is _adj To correct blood pressure, BP _est In order to predict blood pressure.

8. The method for measuring human blood pressure based on individual correction according to claim 6, wherein the feature extraction module urenet of the correction waveform and the waveform to be measured _adj And Uresnet _mea The specific method for sharing the training parameters comprises the following steps: globally defining only one feature processing module Uresnet _share Single period pulse wave feature F of correction waveform _adj And single-period pulse wave characteristics F of the waveform to be measured _mea All using solely Uresnet _share Feature extraction is carried out, and backward propagation only updates the Uresnet _share Is a parameter of (a).