CN116509357A

CN116509357A - Continuous blood pressure estimation method based on multi-scale convolution

Info

Publication number: CN116509357A
Application number: CN202310550097.2A
Authority: CN
Inventors: 宫玉琳; 王慧杰; 陈晓娟; 胡命嘉; 景治新; 张福君; 袁文豪
Original assignee: Changchun University of Science and Technology
Current assignee: Changchun University of Science and Technology
Priority date: 2023-05-16
Filing date: 2023-05-16
Publication date: 2023-08-01

Abstract

The invention belongs to the technical field of blood pressure monitoring, in particular to a continuous blood pressure estimation method based on multi-scale convolution, which comprises the following steps: and step A, training the model by utilizing blood pressure data in a blood pressure database. According to the multi-scale convolution-based neural network provided by the invention, the characteristic extraction on the time-frequency domain is carried out by utilizing three different-scale convolution check input signals (ECG and PPG) through the multi-scale convolution neural network so as to extract more accurate and rich characteristic vectors, the dimension unification is carried out on the three-dimensional characteristics through a transverse connection module, a characteristic pyramid is constructed, and the regression analysis module carries out regression analysis on the characteristic pyramid to obtain systolic pressure (SBP) and diastolic pressure (DBP).

Description

Continuous blood pressure estimation method based on multi-scale convolution

Technical Field

The invention relates to the technical field of blood pressure monitoring, in particular to a continuous blood pressure estimation method based on multi-scale convolution.

Background

With the increasing economic and social level, especially the increasing population aging problem, cardiovascular disease (CVD) has become a major public health problem, and statistics of reports published by the world health organization in 2017 indicate that CVD is the first leading cause of death and disability worldwide, and about 1770 ten thousand people die from CVD in 2016, and the number of deaths increases by 14.5% compared with 2006. This gives the medical and research staff of various countries a knock off the alarm, there is a need to seek to solve or reduce the destructive effects of the disease, hypertension is a direct factor leading to CVD, serious hypertension causes a rapid increase in the risk of suffering from CVD, according to the report 2018 of chinese cardiovascular diseases issued by the national health commission, the prevalence and mortality of chinese cardiovascular diseases are still in the rising stage, the number of first suffering from cardiovascular diseases in China is 2.9 million, the mortality is first, the rate of total causes of cardiovascular diseases in rural areas is 45.5%, wherein hypertension accounts for more than 80%, and the trend of younger is continuously maintained, so it is known that hypertension is already the first killer of human health, and if no measures are taken, the prevalence and mortality will continuously increase, so how to prevent and treat hypertension has become a significant topic of attention of the society researchers.

Blood pressure monitoring is mainly performed on the highest value of blood pressure, the lowest value of systolic pressure (Systolic Blood Pressure, SBP) and blood pressure and the estimation of diastolic pressure (Diastolic Blood Pressure, DBP) in one cardiac interval, and currently, the clinical detection means of blood pressure are mainly divided into two methods, namely a direct measurement method and an indirect measurement method, wherein the direct measurement method is also called invasive blood pressure monitoring, and the measurement method of invasive blood pressure monitoring is mainly that one end of a catheter is inserted into a human epidermal blood vessel, and the other end of the catheter is directly connected with a pressure sensor. Because the blood flows to strike the vessel wall to generate pressure, along with technological development and improvement of human requirements on life quality, the medical field is also revolutionized, the wearable medical technology is presented, more convenient, efficient, safe and comfortable life style is brought to people, the wearable medical equipment can be directly worn on the human body, the human body signals are recorded and analyzed in an auxiliary mode through the sensing device, and the portable medical health electronic equipment which is movable, wearable and sustainable is used for measuring continuously blood pressure through a noninvasive method.

However, the existing blood pressure monitoring device has some obvious disadvantages, and a great deal of researches show that the most outstanding problems existing in the blood pressure monitoring technology at present are as follows: the accuracy of the estimate of systolic pressure (SBP) is widespread and significantly higher than that of diastolic pressure (DBP), and the present technique proposes a solution mainly to this problem.

Disclosure of Invention

(one) solving the technical problems

Aiming at the defects of the prior art, the invention provides a continuous blood pressure estimation method based on multi-scale convolution, which solves the problems of injury to a patient, discomfort caused by a cuff type and obviously lower systolic pressure (SBP) precision than diastolic pressure (DBP) precision in the traditional learning method.

(II) technical scheme

The invention adopts the following technical scheme for realizing the purposes:

a continuous blood pressure estimation method based on multi-scale convolution, the method comprising the steps of:

step A, training a model by utilizing blood pressure data in a blood pressure database;

step B, collecting photoelectric volume pulse signals and electrocardiosignals;

step C, noise reduction and filtering are carried out on the two paths of collected signals;

step D, segmenting the PPG and the ECG by using a window function;

and E, inputting the ECG and the PPG obtained in the step D into a multi-scale convolutional neural network model to estimate systolic pressure (SBP) and diastolic pressure (DBP).

Further, the step A is divided into:

step A1, obtaining sample data from a blood pressure test database, wherein the sample data comprises pulse wave (PPG), electrocardiosignal (ECG) and invasive arterial blood pressure signal (ABP), and the sampling frequency fs=125;

step A2, constructing a multi-scale convolution-based blood pressure estimation network, wherein the multi-scale convolution-based blood pressure estimation network comprises a multi-scale convolution neural network, a transverse link module and a regression analysis module, and the overall model architecture of the multi-scale convolution-based blood pressure estimation network is shown in figure 1;

and A3, carrying the sample into the blood pressure estimation network to train, obtaining weight parameters of the blood pressure estimation network, and obtaining the trained blood pressure estimation network.

Further, the step B is further divided into:

step B1, sticking an electrode patch of an ECG acquisition module to corresponding positions of the chest and the abdomen of a subject, and placing the PPG acquisition module at the position of the finger tip of the index finger of the subject;

step B2, the subject is in a sitting state, does not move violently within 30 minutes before being tested, and is measured in a calm state;

and step B3, synchronously acquiring ECG and PPG signals of the subject in a calm state, and obtaining signal output of real-time test data.

Further, the step C is further divided into:

step C1, removing baseline wander by PPG, wherein a cubic spline interpolation method is the most widely applied baseline wander removal algorithm, the cubic spline interpolation is actually a piecewise polynomial interpolation, a blood pressure signal data set N is given, the data set is divided into N-1 segments, and then a cubic polynomial is constructed between two adjacent data points for fitting;

step C2, removing myoelectric interference by the ECG signal, wherein the myoelectric signal interference is caused by trembling of muscle fibers, the duration time is short, the voltage range is small, the myoelectric interference can occur in a wider frequency band, but the myoelectric interference is mainly concentrated and distributed in a range of 30-300 Hz, and the main frequency of the ECG signal is concentrated between 0-45 Hz, so that the myoelectric interference can be in frequency spectrum aliasing with the ECG signal, and frequency components above 45Hz are filtered by a high-frequency filter;

step C3, the power frequency interference of the ECG signal is suppressed, the power frequency interference is a main interference source of the ECG signal, the domestic power frequency interference is mainly concentrated at about 50Hz, and the power frequency interference can be filtered by utilizing a wave trap;

and C4, removing baseline drift of the ECG signal, wherein the ECG signal is acquired by adopting a patch electrode, so that slight shake of a subject in the acquisition process can generate weak low-frequency interference on the ECG signal, the ECG signal fluctuates up and down, a low-frequency component generating the baseline drift is filtered out through a low-pass Butterworth filter, and the baseline drift can be removed by subtracting the low-frequency component from the original signal.

Further, the step E is further divided into:

step E1, the processed data is subjected to matrix transformation (such as a reshape function) to adjust the number of rows, the number of columns and the number of dimensions of the function;

e2, performing convolution calculation on an Electrocardiosignal (ECG) and a pulse signal (PPG) through three scales, wherein an input signal is subjected to first-layer convolution, the convolution kernel of the first-layer convolution is 256, and a W1 wave band is convolved to extract high-frequency information C1 of the input signal;

e3, after the first layer convolution operation, the output size of the signal becomes smaller, and at the moment, second layer convolution is carried out, the size of a second layer convolution kernel is 64, convolution is carried out on a W3 wave band, and intermediate frequency information C2 of an input signal is extracted;

e4, after the second layer convolution operation, the output size is reduced again, at the moment, the third layer convolution is carried out on the signal, the third layer convolution kernel size is 32, the convolution is carried out on the W2 wave band, and the low-frequency information C3 of the input signal is extracted;

in step E5, in order to ensure that the features in the feature pyramid can be normally input to the next module, it must be ensured that the dimensions of each feature are the same, and in order to obtain the feature pyramid, { C1, C2, C3} needs to be processed, where the processing procedure is as follows:

Fi＝gi(Ci,θg,i)

wherein g (-) represents the dimension reduction process of the feature vector, ci represents the i-th layer output of the convolutional neural network, thetag, i are trainable parameters of the transverse connection layer, fi is the output of the transverse connection module, after passing through the transverse connection module, the feature vector is unified into the same feature dimension to form a feature pyramid, and in the network model, the feature dimension is finally the same as 64;

after the feature dimensions are unified, the feature pyramid is input into a regression analysis module, the regression analysis module carries out regression analysis on the feature pyramid and outputs a finally estimated blood pressure waveform, and the module can analyze the relation between contexts in a time sequence in the feature pyramid and estimate the blood pressure according to the time relation and the time-frequency characteristics of the waveform;

step E6, inputting the feature vectors F1, F2 and F3 into a BiLSTM layer, wherein the BiLSTM module can analyze the context information related to the time sequence in the feature pyramid at the same time, and the BiLSTM layer can better capture the bidirectional semantic dependence, and the output dimension of the BiLSTM layer is 128;

e7, integrating 128-dimensional output characteristics of BiLSTM through a full connection layer (FC) and outputting the output characteristics as a value, namely Arterial Blood Pressure (ABP), wherein the influence of characteristic positions on regression can be greatly reduced through the full connection layer;

in step E8, after the output of FC, the ABP signal contains information of Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP), where the specific information is that the peak value of ABP is the Systolic Blood Pressure (SBP) of the current blood pressure, the valley value of ABP is the Diastolic Blood Pressure (DBP) of the current blood pressure, and the peak-valley value extraction is performed to extract the (systolic blood pressure) SBP and the (diastolic blood pressure) DBP of the current blood pressure.

(III) beneficial effects

Compared with the prior art, the invention provides a continuous blood pressure estimation method based on multi-scale convolution, which has the following beneficial effects:

according to the multi-scale convolution-based neural network provided by the invention, the characteristic extraction on the time-frequency domain is carried out by utilizing three different-scale convolution check input signals (ECG and PPG) through the multi-scale convolution neural network so as to extract more accurate and rich characteristic vectors, the dimension unification is carried out on the three-dimensional characteristics through a transverse connection module, a characteristic pyramid is constructed, and the regression analysis module carries out regression analysis on the characteristic pyramid to obtain systolic pressure (SBP) and diastolic pressure (DBP).

Drawings

FIG. 1 is a schematic diagram of a multi-scale convolutional neural network based structure of the present invention;

FIG. 2 is a schematic diagram of time-frequency signals of different wave bands of an input signal according to the present invention;

FIG. 3 is a schematic diagram of a regression analysis module according to the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Examples

As shown in fig. 1-3, a continuous blood pressure estimation method based on multi-scale convolution according to an embodiment of the present invention includes the following steps:

step A3, carrying the sample into the blood pressure estimation network for training to obtain weight parameters of the blood pressure estimation network, and obtaining a trained blood pressure estimation network;

step B, collecting photoelectric volume pulse signals and electrocardiosignals;

the step B signal acquisition comprises the following steps:

step B3, ECG and PPG signals of the subject in a calm state are synchronously collected, and signal output of real-time test data is obtained;

in the step C, the collected original PPG and ECG are preprocessed, both belong to micro-electric signals, noise interference exists, but PPG signals have a low frequency range, 99% of energy is concentrated and distributed between 0Hz and 10Hz, the main frequency is generally smaller than 3Hz, and ECG signals mainly have the following two characteristics: the ECG signal amplitude is little, and the frequency is low, and the frequency range is usually between 0.05 ~ 45Hz, and dominant wave frequency is around 1Hz, and secondly ECG signal belongs to nonstationary quasi-periodic signal, has the characteristics of randomness, because of the random noise intensity, so mix into the noise very easily when gathering, its noise source mainly has three types: the first type of myoelectric disturbance, the electric activity due to tremor and movement of the human body's muscles, produces a noise signal up to 10% of the amplitude of a normal electrocardiogram, at a frequency of 10kHz, and lasting about 50 ms; the second kind of power frequency interference is an interference signal with the frequency of 60Hz or 50+/-0.2 Hz generated by a power supply; the third baseline drift, the noise of 0-0.5 Hz is generated due to respiratory motion, the amplitude is 15% of the total electrocardiogram amplitude;

step C4, removing baseline drift of the ECG signal, wherein the ECG signal is acquired by adopting a patch electrode, so that slight shake of a subject in the acquisition process can generate weak low-frequency interference on the ECG signal, the ECG signal fluctuates up and down, a low-frequency component generating baseline drift is filtered out by a low-pass Butterworth filter, and the baseline drift can be removed by subtracting the low-frequency component from an original signal;

step D, segmenting the PPG and the ECG by using a window function;

step D, setting a calculation window with a reasonable window function size (e.g. size=400), and segmenting the processed ECG and PPG;

the multi-scale convolution network model is characterized in that features of different time frequency domains are extracted by utilizing different convolution kernel sizes, so that targeted features are extracted for blood pressure change conditions of different stages, errors are reduced, and model accuracy is improved;

e, inputting the ECG and the PPG obtained in the step D into a multi-scale convolutional neural network model to estimate systolic pressure (SBP) and diastolic pressure (DBP);

firstly, through a multi-scale convolution neural network, the neural network comprises three convolution modules with different scales, the three convolution modules adopt three convolution kernels with different scales, and characteristic information on different time-frequency scales in input is respectively extracted;

taking an electrocardiograph signal (ECG) as an example, as shown in fig. 2, carrying out feature extraction on different wave band information in an input waveform by adopting a maximum convolution check for a wave crest W1 with the largest frequency change in fig. 2, carrying out feature extraction on a wave crest W2 with the most gentle frequency change by adopting a minimum convolution check, carrying out feature extraction on a wave trough with the frequency change between W1 and W2 by adopting a medium-scale convolution check, extracting waveform information of the input waveform by adopting convolution with different scales, and more accurately extracting the features of the input waveform at the wave crest and the wave trough;

step E2, performing convolution calculation on an Electrocardiosignal (ECG) and a pulse signal (PPG) through three scales, wherein an input signal is subjected to first-layer convolution, the convolution kernel of the first-layer convolution is 256, and a W1 wave band shown in FIG. 2 is convolved to extract high-frequency information C1 of the input signal;

step E3, after the first layer convolution operation, the output size of the signal becomes smaller, and at the moment, a second layer convolution is performed, the size of a second layer convolution kernel is 64, and the convolution is performed on the W3 wave band shown in FIG. 2, so that intermediate frequency information C2 of the input signal is extracted;

step E4, after the second layer convolution operation, the output size is reduced again, at this time, the third layer convolution is performed on the signal, the third layer convolution kernel size is 32, the convolution is performed on the W2 wave band shown in FIG. 2, and the low-frequency information C3 of the input signal is extracted;

after a convolution operation of three dimensions, the following set of feature sequences can be obtained:

{C1,C2,C3}＝f(X；θf)

wherein Ci represents the output of the ith convolution block of the multi-scale convolution neural network, f (·) represents the calculation process of the convolution neural network, X represents the input of the neural network, and θf represents the trainable parameter;

outputting the convolved feature matrixes { C1, C2 and C3} to a transverse connection module, and connecting the tail ends of the convolved output C1, C2 and C3 modules with transverse connection respectively, wherein the transverse connection module is used for processing the convolved feature matrixes into features with the same feature size, as shown by F1, F2 and F3 in fig. 1, and the features are also called pyramid feature sequences (namely feature pyramids);

Fi＝gi(Ci,θg,i)

wherein g (-) represents the dimension reduction process of the feature vector, ci represents the i-th layer output of the convolutional neural network, thetag, i are trainable parameters of the transverse connection layer, fi is the output of the transverse connection module, after passing through the transverse connection module, the feature vector is unified into the same feature dimension to form a feature pyramid, and in the network model, the feature dimension is finally unified into 64;

after the feature dimensions are unified, the feature pyramid is input into a regression analysis module, the regression analysis module carries out regression analysis on the feature pyramid and outputs a finally estimated blood pressure waveform, the module can analyze the relation between contexts in a time sequence in the feature pyramid, and estimate the blood pressure according to the time relation and the time-frequency characteristics of the waveform, wherein the structure of h () is shown in figure 3;

Finally, it should be noted that: the foregoing description is only a preferred embodiment of the present invention, and the present invention is not limited thereto, but it is to be understood that modifications and equivalents of some of the technical features described in the foregoing embodiments may be made by those skilled in the art, although the present invention has been described in detail with reference to the foregoing embodiments. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A continuous blood pressure estimation method based on multi-scale convolution is characterized in that: the method comprises the following steps:

step B, collecting photoelectric volume pulse signals and electrocardiosignals;

step D, segmenting the PPG and the ECG by using a window function;

2. A continuous blood pressure estimation method based on multi-scale convolution according to claim 1, characterized in that: the step A is divided into:

3. A continuous blood pressure estimation method based on multi-scale convolution according to claim 1, characterized in that: the step B is also divided into:

4. A continuous blood pressure estimation method based on multi-scale convolution according to claim 1, characterized in that: the step C is also divided into:

5. A continuous blood pressure estimation method based on multi-scale convolution according to claim 1, characterized in that: the step E is also divided into:

Fi＝gi(Ci,θg,i)