WO2019100827A1

WO2019100827A1 - Method and apparatus for extracting blood pressure data from pulse wave signal

Info

Publication number: WO2019100827A1
Application number: PCT/CN2018/106181
Authority: WO
Inventors: 张跃; 王占宇; 张拓; 雷夏飞
Original assignee: 深圳市岩尚科技有限公司; 清华大学深圳研究生院
Priority date: 2017-11-24
Filing date: 2018-09-18
Publication date: 2019-05-31
Also published as: CN109833034A

Abstract

A method for extracting blood pressure data from a pulse wave signal, comprising: model training (S1) and blood pressure data extraction (S2), wherein the model training (S1) comprises: obtaining pulse waves and corresponding blood pressure data, and preprocessing same (S11); performing complete waveform detection and separation on the preprocessed pulse waves (S12); performing frequency-domain transformation on each complete pulse waveform, and extracting frequency-domain features (S13); and using the frequency-domain features as an input, and a corresponding blood pressure value as an output, and performing prediction model training in combination with a neural network to obtain a blood pressure data prediction model (S14); the blood pressure data extraction (S2) comprises: inputting frequency-domain features to be predicted to the blood pressure data prediction model, and outputting the blood pressure value (S24). The method can avoid the problem of difficulty in feature point detection in a time-domain feature extraction method, and provides a more precise sample for model training by using the neural network, thereby improving blood pressure measurement prediction precision, and reducing prediction errors. Also disclosed is an apparatus for extracting blood pressure data from a pulse wave signal.

Description

Method and device for extracting blood pressure data from pulse wave signal

Technical field

The invention relates to the field of intelligent medical treatment, and in particular to a method and a device for extracting blood pressure data from a pulse wave signal.

Background technique

This research work has been approved by the National Natural Science Foundation of China (NO.61571268) and the Guangdong Provincial Science and Technology Department major science and technology project--based remote sensing of human body physiology multi-parameter real-time monitoring and analysis of the Internet of Things platform and demonstration project, and Shenzhen Municipal Development and Reform Commission major science and technology project - based on smart phone-based remote human physiological multi-parameter real-time monitoring and analysis of the industrialization of the network platform.

At each beat of the heart, blood is ejected from the ventricles to the aorta. Each shot produces a force on the blood, which produces a pressure wave from the heart to the peripheral blood vessels. The pressure wave travels along the arterial tree and its flow depends on the nature of the artery, such as the elasticity, stiffness or thickness of the arterial wall. The correlation between blood pressure and arterial properties makes it a very good indicator of cardiovascular system status. If the blood pressure value remains outside the normal range for a long time (for example due to cardiovascular disease), it can have fatal consequences for the patient. Therefore, it is very important to conduct long-term continuous monitoring of human blood pressure.

At present, non-invasive blood pressure measurement mainly includes auscultation, oscillometric method, tension measurement method, volume compensation method, etc., but the auscultation method and oscillometric method cannot meet the needs of long-term continuous monitoring. Although the tension measurement method and the volume compensation method can be realized. Continuous monitoring, but these two methods require the patient's tested site to be squeezed by pressure for a long time, which is prone to discomfort and is not suitable for continuous monitoring of blood pressure for a long time.

In the prior art, pulse wave is usually used for non-invasive continuous blood pressure measurement, which can realize long-term continuous blood pressure monitoring without causing discomfort to the tester, and thus attracts many scholars to conduct research. However, there are still many shortcomings in the current method. For example, many existing studies establish a multiple regression equation between pulse wave and blood pressure. Because blood pressure and pulse wave are not simply linear, the generalization ability of the model is poor. The prediction error is large. There are also many studies that extract the time-domain characteristics of pulse wave signals by detecting the characteristic points of pulse waves, and use machine learning methods to train blood pressure prediction models. However, in the measurement of the actual pulse wave signal, there are various kinds of noise interferences, which makes it difficult to detect the feature points of the pulse wave signal. A non-invasive, continuous, convenient and accurate method for measuring dynamic blood pressure of the human body.

Summary of the invention

In order to solve the problem that the pulse wave feature point detection is difficult and the model generalization ability is poor, the present invention provides a method and a device for extracting blood pressure data from a pulse wave signal, which can measure the dynamic blood pressure of the human body in a non-invasive, continuous, convenient and accurate manner. .

The method for extracting blood pressure data in the pulse wave signal provided by the present invention comprises: model training and blood pressure data extraction; the model training comprises: S11. acquiring a pulse wave and corresponding blood pressure data, and performing preprocessing; S12. The pulse wave performs complete waveform detection and separation; S13. Performs frequency domain transformation on each complete pulse wave waveform and extracts frequency domain features; S14. takes the frequency domain feature as an input, and the corresponding blood pressure value as an output, combined with the nerve The network performs prediction model training to obtain a blood pressure data prediction model; the blood pressure data extraction includes: S21. acquiring a pulse wave and corresponding blood pressure data, and performing preprocessing; S22. performing a complete waveform detection and separating the preprocessed pulse wave; S23. Perform frequency domain transformation on each complete pulse wave waveform and extract frequency domain features; S24. Input the frequency domain feature into the blood pressure data prediction model in S14, and output the blood pressure value.

The invention also provides a computer storage medium storing a computer program for electronic data exchange, the computer program causing a computer to perform a method as described above.

The invention also provides the application of the method for extracting blood pressure data in the pulse wave signal described above in a blood pressure measuring device or a physiological multi-parameter monitoring device.

The present invention also provides an apparatus for extracting blood pressure data from a pulse wave signal, comprising: an acquisition unit that acquires a pulse wave and corresponding blood pressure data and performs preprocessing; and a separation unit that performs a complete waveform detection and separation of the preprocessed pulse wave; The conversion unit performs frequency domain transformation on each complete pulse wave waveform and extracts frequency domain features; the training unit takes the frequency domain feature as an input, and combines the neural network to perform prediction model training to obtain a blood pressure data prediction model; The frequency domain feature to be predicted is input to the blood pressure data prediction model, and the blood pressure data is output.

The invention has the beneficial effects that after the complete waveform detection and separation of the pulse wave is performed, the separated complete pulse wave waveform is obtained, and then the frequency domain transform is performed to extract the frequency domain feature, which can effectively avoid the previous time domain feature extraction method. The problem of feature point detection is difficult, and feature redundancy and dimensional disaster are effectively avoided, and more accurate samples are provided for subsequent model training using neural network, thereby improving the prediction accuracy of blood pressure measurement and reducing the prediction error.

DRAWINGS

FIG. 1 is a schematic flow chart of a method for extracting blood pressure data from a pulse wave signal according to an embodiment of the present invention.

2 is a schematic flow chart of a method for detecting and separating a complete waveform peak according to an embodiment of the present invention.

FIG. 3 is a schematic flow chart of a method for detecting and separating a complete waveform valley according to an embodiment of the present invention.

4 is a schematic diagram of a discrete-cosine transform of a single-heart beat pulse wave according to an embodiment of the present invention.

FIG. 5 is a network structure diagram of a BP neural network according to an embodiment of the present invention.

Detailed ways

The present invention will be further described in detail with reference to the accompanying drawings and the accompanying drawings.

The embodiment provides a system for extracting blood pressure data from a pulse wave signal, comprising: an acquiring unit, acquiring a pulse wave and corresponding blood pressure data and performing preprocessing; and a separating unit, performing a complete waveform detection and separating the preprocessed pulse wave; The conversion unit performs frequency domain transformation on each complete pulse wave waveform and extracts frequency domain features; the training part takes the frequency domain feature as an input, the corresponding blood pressure value as an output, and combines the neural network to perform prediction model training to obtain blood pressure data. The prediction model; the output unit inputs the frequency domain feature to be predicted into the blood pressure data prediction model, and outputs the blood pressure value.

In this embodiment, the method for extracting blood pressure data from the pulse wave signal is as shown in FIG. 1 and includes two parts: model training S1 and blood pressure data extraction S2.

The model training S1 part is used to generate a blood pressure data prediction model through a large amount of data training, and the specific steps include:

S11. Obtain pulse wave (PPG) and corresponding blood pressure data, and perform pre-processing, including: denoising, removing baseline drift.

S12. The pre-processed pulse wave is subjected to complete waveform detection and separated.

There are two ways to detect and separate a complete waveform, one of which is the peak (maximum) detection method and the other is the trough (minimum value) detection method.

The peak (maximum value) detection method is shown in FIG. 2, and specifically includes the following steps:

A1. Perform peak (R wave) detection on the pulse wave signal after preprocessing, that is, detect the maximum value of the signal, and record the coordinates of the maximum value;

A2. Some pulse waves have a heavy beat wave. In order to prevent the beat wave peak from being detected, a threshold value T ₁ is set to the maximum value point, and it is judged whether the detected maximum value point is smaller than the threshold value, and only the point larger than the threshold value T ₁ Is considered to be a peak, deleting points that are less than the threshold;

T ₁ =P _max *c ₁

Where P _max is the maximum value in the pulse wave, and c ₁ is a constant between 0-1, which can be adjusted according to the actual situation.

A3. After step A2, the maximum value less than the threshold T ₁ can be screened out, but in the actual signal acquisition process, there are many noise interferences, and many noise spurs are falsely detected as maximum values, in order to solve the noise. For the interference of the maximum point, set the threshold T ₂ to determine whether the distance between the two maximum points is greater than the set threshold T ₂ . If the distance between the two extreme points is too close, delete the second Extreme point

T ₂ = D ₁ * c ₂

Where D ₁ is the distance between two R waves of a normal pulse wave, and c ₂ is a constant between 0-1, which can be adjusted according to actual conditions.

A4. After the steps A2 and A3, the R wave of the PPG signal can be detected. According to the waveform characteristics of the PPG signal, the point on the left of the maximum point is 1/3 of the total point, and the point on the right side of the maximum point is 2/. The point of 3 constitutes a complete pulse wave waveform. For example, with a sampling frequency of 125 Hz, 34 points to the left of the maximum point and 65 points to the right are selected, and a total of 100 data points constitute a complete pulse wave waveform.

The trough (minimum value) detection method is shown in FIG. 3, and specifically includes the following steps:

B1. Performing a valley detection on the pre-processed pulse wave signal, that is, detecting a minimum value of the signal, and recording coordinates of the minimum value;

B2. Some pulse waves have a heavy beat wave. In order to prevent the heavy wave wave trough from being detected, a threshold value T ₃ is set for the minimum value point, and it is judged whether the detected minimum value point is smaller than the threshold value, and only the point smaller than the threshold value T ₃ Is considered to be a trough, deleting points that are greater than the threshold;

T ₃ =P _min *c ₃

Where P _min is the minimum value in the pulse wave, and c ₃ is a constant between 0-1, which can be adjusted according to the actual situation.

B3. After step B2, the minimum value larger than the threshold T ₃ can be screened out, but in the actual signal acquisition process, there are many noise interferences, and many noise spurs are falsely detected as minimum values, in order to solve the noise. For the interference of the minimum point, set the threshold T ₄ to determine whether the distance between the two minimum points is greater than the set threshold T ₄ . If the distance between the two extreme points is too close, delete the second Extreme point

T ₄ = D ₂ * c ₄

Where D ₂ is the distance between the two troughs of the normal pulse wave, and c ₄ is a constant between 0-1, which can be adjusted according to the actual situation.

B4. After the steps B2 and B3, the valleys of the PPG signal can be detected. It is assumed that a total of n troughs are detected, which are respectively recorded as 1, 2, ..., n, and the coordinates corresponding to the trough are C _t , where _{1 ≤ t} ≤ n. The data point between the t-1th trough and the tth trough, the data point between the coordinates [C _t-1 , C _t ], constitutes a complete pulse wave waveform; between each two troughs The data is truncated to obtain n-1 separate pulse waveforms.

B5. Since the pulse wave waveform is not completely equal in period, the number of data points included in the pulse wave waveform separated in step B4 is not completely the same. In order to facilitate subsequent transformation of the pulse wave waveform, the separated pulse wave is separated. The waveform is subjected to cubic spline interpolation processing and processed into a complete pulse waveform with the same waveform length.

A complete pulse waveform, the default is that the starting point and the end point are both troughs, and the peak is in the middle position, so the trough (minimum value) detection method is superior to the peak (maximum value) detection method. The trough (minimum value) detection method utilizes the inherent properties of the pulse wave waveform, which is more conducive to detecting the starting point and the end point, thereby forming a complete pulse wave waveform, avoiding the peak (maximum value) detection method. The number of left and right points of the large value is improperly selected to cause a defect of the non-complete pulse waveform. After obtaining a plurality of complete pulse wave waveforms, subsequent frequency domain transforms are facilitated to extract features.

S13. Perform frequency domain transformation on each complete pulse wave waveform and extract frequency domain features.

Among them, the transformation mainly transforms the complete pulse wave waveform separated in steps S11 and S12 by frequency domain transformation with energy concentration characteristics, and extracts a portion of the energy concentration as a feature. Frequency domain transforms with energy concentration characteristics include discrete cosine transform, K_L transform, power spectrum density estimation of periodic graph method and power spectral density estimation of multiple windows. The frequency domain transform with energy concentration characteristics can effectively avoid the defects of feature point detection in the time domain feature extraction method, and effectively avoids feature redundancy and dimensional disaster.

KL transform is a kind of transformation based on statistical properties. The eigenvalues and eigenvectors of the covariance matrix of the signal source must be obtained first, and then sorted according to the size of the eigenvalues. The larger the eigenvalues, the corresponding eigenvectors. The greater the contribution to the signal. According to the size of the feature value, the feature vector with smaller feature value is rounded off to achieve the purpose of dimension reduction.

The power spectrum density estimation of the periodogram method is a method of directly performing the Fourier transform of the sampled data x(n) of the signal to obtain the power spectral density estimation. The finite-length random signal sequence is assumed to be x(n). Its Fourier transform and power spectral density estimation have the following relationship:

Where N is the length of the random signal sequence x(n), x(f) is the signal after the Fourier transform of x(n), and S _x (f) is the periodic power spectral density estimate of the signal.

Multi-window power spectral density estimation: Obtain independent power spectrum estimates using multiple orthogonal windows, and then combine these estimates to obtain a power spectrum estimate of a sequence. Compared with the ordinary periodic graph method, this power spectrum estimation has greater degrees of freedom and has better effects in estimating accuracy and estimating fluctuations.

Discrete cosine transform (DCT) is used as the feature extraction method. If x[m] is a one-dimensional signal of length L, the discrete cosine transform can be performed according to the following formula:

among them,

X ^c [k] is a sequence after discrete cosine transform.

The energy concentration characteristic of the discrete cosine transform is as shown in FIG. 4, wherein the waveform a is a complete pulse wave waveform separated by the valley detecting method in step S12, and the waveform b is a waveform obtained by discrete cosine transforming the waveform. As can be seen from Fig. 4, the waveform energy after discrete cosine transform of the pulse wave is concentrated in front of the signal signal, and the energy of the waveform transformation of 100 data points is mainly concentrated on the first 15 data points. In the subsequent steps, these 15 data points were used as input to the neural network to perform blood pressure data prediction model training.

S14. Taking the frequency domain feature as an input, and corresponding blood pressure value as an output, and using a neural network to perform prediction model training, and obtaining a blood pressure data prediction model.

This embodiment combines BP neural network for predictive model training: multiple BP neural networks are used as weak classifiers, and BP_Adaboost strong classifier is constructed by Adaboost algorithm to generate final blood pressure data prediction model. FIG. 5 is a network structure diagram of a BP neural network, which adopts a dual hidden layer network structure, which can better map the complex nonlinear relationship between pulse wave characteristics and blood pressure. After experimentation, the optimal prediction results can be obtained when the network structure is [25 15].

Through the Adaboost iterative algorithm, different predictors (weak predictors) are trained for the same training set, and then these weak predictors are combined to form a stronger final predictor (strong predictor).

By separating 10,000 PPG waveforms, and using the power spectrum density estimation of the periodogram method, the power spectral density estimation of the multi-window method, the K_L transform and the discrete cosine transform are used for frequency domain transform and the frequency domain features are extracted. Among them, 7000 data are used as training sets, 3000 data are used as test sets, and different features are used. The BP_adaboost model is used for training. The results are shown in the following table:

Among them, μ represents the average error and δ represents the standard deviation.

The average error represents the accuracy of the model prediction and is calculated using the following formula:

Where n is the number of samples, e _i is the prediction error of the ith sample, and the prediction error is the difference between the expected value and the model prediction value.

The standard error represents the degree of stability of the model prediction and is calculated using the following formula:

among them,

From the results in the table, the feature extraction method proposed by the present invention includes the power spectrum density estimation of the periodogram method, the power spectral density estimation of the multi-window method, the K_L transform, and the discrete cosine transform, which can effectively perform blood pressure prediction. The best predictive effect can be achieved with discrete cosine transform.

However, compared with other frequency domain transforms, the discrete cosine transform has the following outstanding advantages:

The computational complexity is low and the amount of computation is small, especially compared to the power spectral density method. The discrete cosine transform is reduced by more than half of the calculation.

To avoid noise interference, it is proved by experiments that after the signal added with noise is subjected to discrete cosine transform, the energy concentration part and the signal without noise added are transformed to have a small difference, indicating that the signal is multi-discrete cosine transformed regardless of whether the source signal contains noise or not. The characteristic of the energy concentration is small. In the actual signal acquisition process, it is inevitable that there will be noise, so the denoising feature of the discrete cosine transform makes it very suitable for the feature extraction method of the present invention.

The discrete cosine transform is also used to extract the feature parts, and BP_adaboost is compared with other machine learning, such as support vector machine and BP neural network. The results are as follows:

It can be seen from the above results that the BP_adaboost algorithm in this embodiment can achieve the best effect in blood pressure prediction. BP neural network has strong nonlinear mapping ability. Mathematical theory proves that the three-layer neural network can approximate any nonlinear continuous function with arbitrary precision, which makes it particularly suitable for solving complex problems of internal mechanism, but in this embodiment The solved blood pressure prediction model has many input characteristics and is a complex nonlinear problem, so the neural network model is particularly suitable for the present embodiment. However, BP neural network is also insufficient. The traditional BP neural network is a local search optimization method. The weight of the network is gradually adjusted by the direction of local improvement, which will cause the algorithm to fall into the local extremum and the weight convergence. To the local minimum point, the present embodiment combines the adaboost algorithm to effectively solve the defect that a single BP neural network is easy to fall into local optimum, and thus obtains the best prediction effect.

The specific steps of the blood pressure data extraction S2 part include: S21. acquiring the pulse wave and the corresponding blood pressure data, and performing preprocessing; S22. performing the complete waveform detection and separating the preprocessed pulse wave; S23. For each complete pulse wave waveform The frequency domain transform is performed, and the frequency domain feature is extracted; S24. The frequency domain feature is input to the blood pressure data prediction model in S14, and the blood pressure data is output.

Among them, steps S21-S23 are the same as steps S11-S13 in the model training S1 portion.

By predicting the blood pressure value by the above method, the present invention has stronger generalization ability and prediction accuracy than the conventional method, the systolic pressure prediction error is within 3 mmHg, and the diastolic pressure prediction error is within 2 mmHg.

The above is a further detailed description of the present invention in combination with specific/preferred embodiments, and it is not intended that the specific embodiments of the invention are limited to the description. It will be apparent to those skilled in the art that <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; It belongs to the scope of protection of the present invention.

Claims

A method for extracting blood pressure data from a pulse wave signal, comprising: model training and blood pressure data extraction;

The model training includes:

S11. Obtain pulse wave and corresponding blood pressure data, and perform pretreatment;

S12. Performing complete waveform detection and separation on the pulse wave after preprocessing;

S13. Perform frequency domain transformation on each complete pulse wave waveform and extract frequency domain features;

S14. Taking the frequency domain feature as an input, and corresponding blood pressure value as an output, combining with a neural network to perform prediction model training, and obtaining a blood pressure data prediction model;

The blood pressure data extraction includes:

S21. Obtaining a pulse wave and corresponding blood pressure data, and performing preprocessing;

S22. Performing a complete waveform detection and separating the pulse wave after preprocessing;

S23. Perform frequency domain transformation on each complete pulse wave waveform, and extract frequency domain features;

S24. Input the frequency domain feature to the blood pressure data prediction model in S14, and output the blood pressure value.
The method according to claim 1, wherein said pre-processing in said step S11 and step S21 comprises: denoising, removing baseline drift.
The method of claim 1 wherein said detecting and separating the complete waveform in said step S12 and step S22 comprises:

Performing maximum value detection on the pulse wave after pretreatment;

Deleting a maximum value point smaller than the threshold value T 1 according to the set threshold value T 1 , and retaining a maximum value point larger than the threshold value T 1 ;

In the maximum value point larger than the threshold T 1 , if the distance between the two maximum value points is less than the set threshold value T 2 , the second maximum value point is deleted, and the retained maximum value point is the R wave. ;

According to the sampling frequency, a point occupying 1/3 of the total number of points is selected on the left side of the R wave, and a point occupying 2/3 of the total number of points is selected on the right side of the R wave to form a complete pulse wave waveform.
The method of claim 1 wherein said detecting and separating the complete waveform in said step S12 and step S22 comprises:

Performing a minimum value detection on the pulse wave after pretreatment;

Deleting a minimum value point larger than the threshold value T 3 according to the set threshold value T 3 , and retaining a minimum value point smaller than the threshold value T 3 ;

In the minimum value point smaller than the threshold value T 3 , if the distance between the two minimum value points is less than the set threshold value T 4 , the second minimum value point is deleted, and the retained minimum value point is the trough;

Separating data between two adjacent troughs respectively to obtain a separated pulse wave waveform;

Cubic spline interpolation is performed on the separated pulse waveform to form a complete pulse waveform with the same waveform length.
The method according to claim 1, wherein said frequency domain transform in said step S13 and said step S23 comprises a frequency domain transform having an energy concentration characteristic.
The method of claim 5 wherein said frequency domain transform having energy concentration characteristics comprises: discrete cosine transform, K_L transform, periodic map power spectral density estimation, multi-window power spectral density estimation.
The method according to claim 1, wherein said step S14 is combined with BP neural network for predictive model training; and a plurality of BP neural networks are used as weak classifiers, and a BP_Adaboost strong classifier is constructed by Adaboost algorithm to generate final blood pressure. Data prediction model.
A computer storage medium storing a computer program for electronic data exchange, the computer program causing a computer to perform the method of any of claims 1-7.
A method of extracting blood pressure data from a pulse wave signal according to any one of claims 1 to 7 for use in a blood pressure measuring device or a physiological multi-parameter monitoring device.
An apparatus for extracting blood pressure data from a pulse wave signal, comprising:

The acquisition unit acquires pulse wave and corresponding blood pressure data and performs preprocessing;

a separating unit that performs a complete waveform detection and separation of the pulse wave after pretreatment;

a conversion unit that performs frequency domain transformation on each complete pulse wave waveform and extracts frequency domain features;

The training unit takes the frequency domain feature as an input, and the corresponding blood pressure value as an output, and performs a prediction model training in combination with the neural network to obtain a blood pressure data prediction model;

The output unit inputs the frequency domain feature to be predicted to the blood pressure data prediction model, and outputs the blood pressure data.