CN115886763A - Blood pressure characteristic information extraction method and blood pressure value estimation system - Google Patents
Blood pressure characteristic information extraction method and blood pressure value estimation system Download PDFInfo
- Publication number
- CN115886763A CN115886763A CN202211370210.0A CN202211370210A CN115886763A CN 115886763 A CN115886763 A CN 115886763A CN 202211370210 A CN202211370210 A CN 202211370210A CN 115886763 A CN115886763 A CN 115886763A
- Authority
- CN
- China
- Prior art keywords
- blood pressure
- characteristic information
- signal
- pulse wave
- redundancy
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000036772 blood pressure Effects 0.000 title claims abstract description 458
- 238000000605 extraction Methods 0.000 title claims abstract description 35
- 238000012545 processing Methods 0.000 claims abstract description 23
- 230000002068 genetic effect Effects 0.000 claims abstract description 21
- 238000003062 neural network model Methods 0.000 claims abstract description 21
- 238000010219 correlation analysis Methods 0.000 claims abstract description 15
- 239000013598 vector Substances 0.000 claims description 151
- 238000012549 training Methods 0.000 claims description 49
- 238000012360 testing method Methods 0.000 claims description 39
- 238000000034 method Methods 0.000 claims description 27
- 230000000737 periodic effect Effects 0.000 claims description 21
- 230000001174 ascending effect Effects 0.000 claims description 13
- 230000009467 reduction Effects 0.000 claims description 11
- 230000000630 rising effect Effects 0.000 claims description 10
- 238000012163 sequencing technique Methods 0.000 claims description 10
- 238000012795 verification Methods 0.000 claims description 7
- 230000003205 diastolic effect Effects 0.000 claims description 6
- 230000001360 synchronised effect Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 4
- 238000011156 evaluation Methods 0.000 claims description 3
- 230000035772 mutation Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 238000009394 selective breeding Methods 0.000 claims description 2
- 238000010276 construction Methods 0.000 abstract description 10
- 238000012216 screening Methods 0.000 abstract description 6
- 230000009286 beneficial effect Effects 0.000 abstract description 5
- 230000000875 corresponding effect Effects 0.000 description 14
- 238000005259 measurement Methods 0.000 description 6
- 238000011160 research Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 239000000284 extract Substances 0.000 description 4
- 238000009530 blood pressure measurement Methods 0.000 description 3
- 230000002596 correlated effect Effects 0.000 description 3
- 238000010187 selection method Methods 0.000 description 3
- 230000035488 systolic blood pressure Effects 0.000 description 3
- 238000010989 Bland-Altman Methods 0.000 description 2
- 239000008280 blood Substances 0.000 description 2
- 210000004369 blood Anatomy 0.000 description 2
- 210000004204 blood vessel Anatomy 0.000 description 2
- 230000035487 diastolic blood pressure Effects 0.000 description 2
- 230000002526 effect on cardiovascular system Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 206010020772 Hypertension Diseases 0.000 description 1
- 238000010220 Pearson correlation analysis Methods 0.000 description 1
- 238000012952 Resampling Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035479 physiological effects, processes and functions Effects 0.000 description 1
- 230000008288 physiological mechanism Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 230000000284 resting effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
The invention provides a blood pressure characteristic information extraction method and a blood pressure value estimation system, wherein the blood pressure characteristic information extraction method is used for extracting waveform characteristics which have obvious influence on blood pressure by combining blood pressure influence factors and consideration factors of a Moens-Korteweg equation on the basis of denoising and characteristic point identification processing of extracted pulse wave (PPG) signals and Electrocardiogram (ECG) signals; furthermore, the invention screens the characteristic parameters for blood pressure estimation based on correlation analysis and weight operation analysis, selects different numbers of characteristic parameters to form a few-parameter characteristic set according to the screening result, and preferentially selects a few-parameter characteristic set combination through a back propagation neural network model (MPGA-BP model) optimized by a multi-population genetic algorithm, thereby providing a foundation for the construction of a continuous blood pressure value estimation system based on few characteristic parameters and high prediction precision, and being beneficial to the blood pressure value estimation system to realize high-accuracy noninvasive blood pressure prediction.
Description
Technical Field
The invention relates to the technical field of electronic sensing information extraction and processing, in particular to a blood pressure characteristic information extraction method and a blood pressure value estimation system.
Background
In order to improve management of hypertension and enable continuous blood pressure measurement to be applied to daily life, a noninvasive cuff-less blood pressure measurement method is widely studied. Among them, noninvasive continuous blood pressure monitoring based on pulse wave (PPG) is considered as a promising noninvasive measurement technique for ambulatory blood pressure. PPG is produced by the heart beat, formed by the blood propagating peripherally along the arterial vessel; when flowing blood presses the wall of a blood vessel, the blood vessel is deformed and vibrated to form pulse changes related to blood pressure changes, so that extremely rich cardiovascular information is contained in the pulse waves.
Considering the abundant cardiovascular information in PPG and the complex physiological mechanisms of blood pressure, researchers tend to introduce more PPG feature parameters to fit the physiological factors affecting blood pressure, further mapping the features to blood pressure values when non-invasively estimating the blood pressure model construction. The current literature shows that the number of features extracted from PPG by signal processing algorithm is concentrated on about 20, and some researches relate to more than 50 parameters; and in the research of automatically learning the characteristics of the signals by using the deep learning network to construct the model, the related characteristic parameters can be hundreds or thousands.
For the characteristic parameters extracted by the signal processing algorithm, whether the extraction is accurate or not is the most important factor influencing the output result of the noninvasive blood pressure prediction model. However, the accuracy problem of parameter extraction caused by the difference of waveforms of different populations is always a challenge for whether the noninvasive dynamic noninvasive blood pressure prediction model can be really applied, and the more parameters represent the more uncertainty and the greater error rate, which affects the stability and accuracy of the model estimation result. The number of the characteristic parameters which are automatically extracted by stacking the neural network in front of the non-invasive blood pressure prediction model is very large, and the redundancy among the characteristics can be directly used as the input of the non-invasive blood pressure prediction model without processing, so that the training efficiency of the model is seriously limited, the complexity of the model is increased, and the integration of the model on a portable device is influenced. Although the large number of parameters means that more physiological signal information is utilized and the blood pressure estimation accuracy is higher, the larger number of parameters also means that the redundancy among the parameters is more, the more points need to be located when the characteristic points are located, and the more factors influence the blood pressure estimation result. Therefore, the number of parameters needs to be further reduced regardless of the feature extraction method.
Therefore, how to design a new blood pressure estimation feature information extraction method to reduce the above negative effects and better help to improve the accuracy and processing efficiency of the blood pressure estimation, becomes an important technical problem in research in the field.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a blood pressure estimation feature information extraction method, which is used for reducing the quantity of extracted feature information as much as possible while ensuring the accuracy of the extracted blood pressure estimation feature information for executing blood pressure value estimation so as to reduce the redundancy among the feature information and the influence factors of a model estimation result.
In order to solve the technical problems, the invention adopts the following technical scheme:
a blood pressure characteristic information extraction method is used for applying the extracted blood pressure characteristic information vector to a blood pressure prediction model for non-invasive blood pressure prediction; the blood pressure characteristic information extraction method comprises the following steps:
s1: synchronously acquiring pulse wave signals and electrocardiosignals of an object to be detected, and performing noise reduction pretreatment;
s2: respectively extracting signal characteristic parameters from the pulse wave signals and the electrocardiosignals after noise reduction pretreatment to form a blood pressure characteristic information vector containing Nk signal characteristic parameters;
s3: performing pairwise correlation analysis on each signal characteristic parameter in the blood pressure characteristic information vector, and performing redundancy removal processing on the signal characteristic parameters with correlation redundancy to obtain redundancy-removed blood pressure characteristic information vectors;
s4: performing weight operation on each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector to obtain weight sequence of each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector;
s5: selecting a plurality of signal characteristic parameters from the redundancy-removed blood pressure characteristic information vector according to the sequence of weight sequencing from large to small to form a test blood pressure characteristic information vector, thereby constructing a plurality of test blood pressure characteristic information vectors with different numbers of signal characteristic parameters;
s6: inputting each constructed test blood pressure characteristic information vector into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation prediction, wherein the noninvasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm; and then, preferentially selecting a test blood pressure characteristic information vector as the extracted blood pressure characteristic information vector according to the blood pressure estimation prediction result and the prediction performance index.
In the above blood pressure feature information extraction method, preferably, in step S2, the signal feature parameters respectively extracted from the pulse wave signal and the electrocardiographic signal after the noise reduction preprocessing include an individual feature parameter, a pulse wave time feature parameter, a pulse wave curve area feature parameter, a curve slope feature parameter, and a signal amplitude feature parameter.
In the above blood pressure characteristic information extraction method, preferably, the individual characteristic parameters include Weight and heart rate HR of the individual to be detected;
the pulse wave time characteristic parameters comprise: time span t from starting point to peak value of pulse wave periodic signal up Time span t from the start of the pulse wave periodic signal to the start of the dicrotic wave bf The time span t from the maximum slope of the ascending branch to the minimum slope of the descending branch of the pulse wave ae The time span t from the peak to the trough of the pulse wave and its first order differential signal down And dt down Time span t from pulse wave start point to pulse wave end point fb Time span t from start to end of pulse wave period signal bb Propagation time PTT between the start points of electrocardio R wave and pulse wave synchronous periodic signals b Propagation time PTT between the maximum rising branch slope point of the electrocardio R wave and the pulse wave a Propagation time PTT between electrocardio R wave and pulse wave peak value point c And time spans tw1, tw2, tw3, tw4, tw5 from the start of the pulse wave period signal to 1/6, 2/6, 3/6, 4/6, 5/6 of the maximum amplitude, respectively;
the pulse wave curve area characteristic parameters comprise: area of systolic phase A of pulse wave bf Pulse wave diastolic area A fb The ascending branch area dAA of the pulse wave and the descending branch area dDA of the pulse wave; pulse wave characteristic parameter K value, K = (A) m -I c )/(I c )，A m Is the average pulse wave area of the period, I c Is the amplitude of the pulse peak point c; and the area under the pulse wave curve A1 between the periodic signal starting point b and the ascending branch slope maximum point a, the area under the pulse wave curve A2 between the ascending branch slope maximum point a and the peak point c, the area under the pulse wave curve A3 between the peak point c and the descending branch slope minimum point e, and the area under the pulse wave curve A4 between the descending branch slope minimum point e and the pulse wave next period signal starting point b' in the pulse wave;
the curve slope characteristic parameters comprise: slope S from pulse wave periodic signal starting point to dicrotic wave starting point cf The rising slope S of the pulse wave and its first order differential signal bc And dS bc The falling branch slope S of the pulse wave and its first order differential signal cb And dS cb ；
The signal amplitude characteristic parameters comprise: the ratio dPIR between the pulse wave peak amplitude and the periodic signal starting amplitude, and the ratio I of the intensity of the rising branch slope maximum point a, the falling branch slope minimum point e, the counterpulsation wave starting point f and the counterpulsation wave peak point g to the peak intensity ar 、I er 、I fr 、I gr 。
In the above blood pressure characteristic information extraction method, preferably, in step S3, a pearson correlation coefficient is used as an evaluation index for evaluating the correlation between two signal characteristic parameters, so as to perform pairwise correlation analysis on each signal characteristic parameter in a blood pressure characteristic information vector; any two signal characteristic parameters x in each signal characteristic parameter in the blood pressure characteristic information vector i And x j The pearson correlation coefficient between them is determined as follows:
wherein ρ i,j Is the signal characteristic parameter x of the ith dimension i And a signal characteristic parameter x of the j dimension j Pearson's correlation coefficient between, σ i 、σ j Respectively is the signal characteristic parameter x of the ith dimension in the blood pressure characteristic information vector acquired by multiple times of acquisition i Variance of (c) and signal characteristic parameter x of jth dimension j Variance of (E), E [ (x) i -μ i )(x j -μ j )]For a characteristic parameter x of the signal i And a signal characteristic parameter x j Standard deviation between, mu i Signal characteristic parameter x representing ith dimension in blood pressure characteristic information vector obtained by multiple times of acquisition i Mean value of (d) (. Mu.) j Representing the signal characteristic parameter x of the jth dimension in the blood pressure characteristic information vector acquired by multiple times of acquisition j Of the average value of (a). .
In the above method for extracting blood pressure characteristic information, preferably, in the step S3, a specific manner of performing redundancy removal processing on the signal characteristic parameter with correlation redundancy is as follows:
counting the Pearson correlation coefficient between every two signal characteristic parameters in each signal characteristic parameter of the blood pressure characteristic information vector, and if the Pearson correlation coefficient value is larger than a preset correlation coefficient threshold value rho 0 The two signal characteristic parameters are used for judging mutual redundant characteristic parameters, and for the signal characteristic parameters with the same redundant characteristic parameters, the mutual correlation redundant characteristic parameters are judged; then, selecting one of the signal characteristic parameters of the mutual redundancy characteristic parameters and the mutual correlation redundancy characteristic parameters for reservation, and removing other redundancy characteristic parameters and correlation redundancy characteristic parameters; therefore, redundancy removing processing is carried out on the signal characteristic parameters with the relevant redundancy from the blood pressure characteristic information vector, and a redundancy removing blood pressure characteristic information vector is obtained.
In the above blood pressure feature information extraction method, preferably, in step S4, for each signal feature parameter in the redundancy-removed blood pressure feature information vector, a weighted value of each signal feature parameter is calculated by using a ReliefF algorithm, and the specific flow is as follows:
s401: creating a weight coefficient w with an initial value of 0 for each signal characteristic parameter in the de-redundant blood pressure characteristic information vector i I.e. by
Representing the initial value of the weight of the ith signal characteristic parameter before iteration in the redundancy-removed blood pressure characteristic information vector, i is from {1,2, \8230;, N f }，N f The total number of the signal characteristic parameters contained in the redundancy-removed blood pressure characteristic information vector is calculated; thus, a characteristic weight vector corresponding to the redundancy-removed blood pressure characteristic information vector is formed by the set of weight coefficients; meanwhile, initializing iteration times m =1;
s402: randomly selecting a redundancy-removing blood pressure characteristic information vector sample X in a redundancy-removing characteristic training set; the redundancy-removing characteristic training set comprises a plurality of blood pressure sample data marked with blood pressure value labels, and each blood pressure sample data comprises a pulse wave signal which is synchronously acquired and corresponds to the blood pressure value label and a redundancy-removing blood pressure characteristic information vector extracted from the electrocardiosignal;
s403: selecting K nearest neighbor samples from the samples in the redundancy-removing characteristic training set, wherein the samples are similar to the redundancy-removing blood pressure characteristic information vector sample X, and marking the samples as the NearHit samples of the redundancy-removing blood pressure characteristic information vector sample X; selecting K nearest neighbor samples from other types of samples which are not in the same class as the redundancy-removing blood pressure characteristic information vector sample X in the redundancy-removing characteristic training set, and marking the K nearest neighbor samples as NearMiss samples of the redundancy-removing blood pressure characteristic information vector sample X;
s404: for the weight coefficient corresponding to the ith signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector, calculating the iterative weight value according to the following mode:
wherein the content of the first and second substances,
and &>
Respectively represent the i-th signal characteristic parameter x i With corresponding weight coefficients at the m-th iteration and m-1 th iterationA value; />
Representing the ith signal characteristic parameter x i And the kth NearHit sample->
M is iteration times, M is a preset upper limit value of the iteration times, K belongs to {1,2, \8230, K }, and K is the number of searched NearHit samples; c. C 1 Representing the class in which the sample X was taken, p (c) being the prior probability of the class c, p (c) 1 ) Is c 1 The prior probability of a class is determined,
representing the ith signal characteristic parameter x i And the kth NearMiss sample->
The distance of (d);
therefore, the values of the weight coefficients corresponding to the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector in the mth iteration are respectively calculated;
s405: adding 1 to the iteration number m, and then returning to the step S402;
s406: repeatedly executing the steps S402 to S405 until the value of the iteration number M reaches a preset iteration number upper limit value M, and executing a step S407;
s407: and sequencing the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector according to the descending order of the values of the weight coefficients corresponding to the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector, and taking the sequence as the weight sequencing of the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector.
Correspondingly, the invention also provides a blood pressure value estimation system combining the blood pressure characteristic information extraction method; therefore, the following technical scheme is adopted in the application:
a blood pressure value estimation system comprising:
the blood pressure characteristic information extraction module is used for synchronously acquiring the pulse wave signals and the electrocardiosignals of the object to be detected and extracting blood pressure characteristic information vectors from the pulse wave signals and the electrocardiosignals of the object to be detected by adopting the blood pressure characteristic information extraction method;
the blood pressure estimation and prediction module is used for inputting the extracted blood pressure characteristic information vector of the object to be detected into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation and prediction and outputting a blood pressure estimation result of the object to be detected; the noninvasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm.
In the blood pressure value estimation system, preferably, the specific way of optimizing the back propagation neural network model by the multi-population genetic algorithm is as follows:
step a01: constructing 5 different populations, wherein each population comprises 20 individuals, each individual represents a group of weight and threshold for initializing the BP network, and meanwhile, the control parameters of each population are different;
step a02: coding each individual of each population, and calculating the fitness of each individual in each population;
step a03: selecting each population by using a selection algorithm, and generating new individuals according to crossover and mutation operators;
step a04: b, calculating the fitness of the filial generation in the step a 03;
step a05: finding out individuals with highest fitness in the population to carry out immigration operation;
step a06: selecting individuals with highest fitness in each population by using a manual selection operator to form an essence population;
step a07: determining the individual with the highest fitness in the essence population, recording the algebra of the individual in the optimal state in the essence population, stopping the algorithm when the algebra exceeds the preset algebra, and decoding to obtain the initial weight and the threshold value of the back propagation neural network model optimized by the multi-population genetic algorithm to be used as a non-invasive blood pressure prediction model.
In the above blood pressure value estimation system, preferably, after the back propagation neural network model optimized by the multi-population genetic algorithm is obtained as the non-invasive blood pressure prediction model, the non-invasive blood pressure prediction model is required to be subjected to blood pressure prediction training to obtain a pre-trained non-invasive blood pressure prediction model; the specific mode of carrying out blood pressure prediction training on the noninvasive blood pressure prediction model is as follows:
step b01: acquiring a blood pressure sample data set from a blood pressure sample database, wherein the blood pressure sample data set comprises a plurality of blood pressure sample data marked with blood pressure value labels, and each blood pressure sample data comprises a pulse wave signal and an electrocardiosignal which are synchronously acquired and correspond to the blood pressure value label;
step b02: respectively extracting blood pressure characteristic information vectors of each blood pressure sample data in the blood pressure sample data set by adopting the blood pressure characteristic information extraction method;
step b03: selecting a training sample and a test sample from a blood pressure sample set to respectively form a training sample set and a test sample set;
step b04: the blood pressure characteristic information vector of each blood pressure sample data in the training sample set is used as the input of the non-invasive blood pressure prediction model, the blood pressure value label of each blood pressure sample data in the training sample set is used as the output verification label, and the non-invasive blood pressure prediction model is subjected to blood pressure prediction training to adjust the blood pressure prediction parameters of the non-invasive blood pressure prediction model;
step b05: inputting the blood pressure characteristic information vector of each blood pressure sample data in the test sample set into a non-invasive blood pressure prediction model for blood pressure prediction, adopting the blood pressure value label of each blood pressure sample data in the test sample set as an output verification label, comparing and verifying the blood pressure prediction result of the non-invasive blood pressure prediction model, and evaluating the blood pressure prediction performance of the non-invasive blood pressure prediction model;
step b06: if the blood pressure prediction performance of the noninvasive blood pressure prediction model does not reach the preset target, returning to execute the step b04; and if the blood pressure prediction performance of the non-invasive blood pressure prediction model reaches a preset target, finishing training to obtain a pre-trained non-invasive blood pressure prediction model.
In the above blood pressure value estimation system, preferably, in the step b03, the ratio of the number of the blood pressure samples of the training sample to the number of the blood pressure samples of the test sample is selected to be 8.
Compared with the prior art, the invention has the following beneficial effects:
1. the method for extracting the blood pressure characteristic information extracts the waveform characteristics which have obvious influence on the blood pressure by combining blood pressure influence factors and consideration factors of a Moens-Korteweg equation on the basis of denoising and characteristic point identification processing of extracted pulse wave (PPG) signals and Electrocardiogram (ECG) signals; furthermore, the invention screens the characteristic parameters for blood pressure estimation based on correlation analysis and weight operation analysis, selects different numbers of characteristic parameters to form a less-parameter characteristic set according to the screening result, and preferentially selects the less-parameter characteristic set based on a back propagation neural network model (MPGA-BP model) optimized by a multi-population genetic algorithm to obtain the less-parameter characteristic set, thereby providing a basis for the construction of a continuous blood pressure value estimation system based on less characteristic parameters and high prediction precision, and being beneficial to the blood pressure value estimation system to realize high-accuracy noninvasive blood pressure prediction.
2. Before the method uses the Relieff algorithm for the feature selection of the noninvasive estimation continuous blood pressure research, the redundancy among features is eliminated by utilizing the correlation analysis, the number of the highly correlated features is effectively reduced, and the influence degree of the selected features on a blood pressure prediction model keeps a certain gradient; furthermore, the Relieff algorithm gives weights which can reflect actual conditions to different characteristics, and the selected characteristic dimensions with the same number contain more pulse wave waveform information, so that the construction of a noninvasive blood pressure prediction model with few parameters is facilitated; the reduction of the number of parameters required by the model construction can further reduce the complexity of the model, reduce the influence factors of the output result of the model and improve the stability and the accuracy of the constructed noninvasive blood pressure prediction model.
3. The method effectively reduces the number of pulse wave characteristics used on the premise of ensuring the accuracy of the constructed noninvasive blood pressure prediction model, improves the intelligibility and interpretability of the blood pressure prediction model result, and creates conditions for realizing the high-precision and low-error continuous noninvasive blood pressure prediction model.
Drawings
For the purposes of promoting a better understanding of the objects, aspects and advantages of the invention, reference will now be made to the following detailed description taken in conjunction with the accompanying drawings in which:
FIG. 1 is a flow chart of a feature selection method for constructing a non-invasive blood pressure prediction model based on pulse wave waveform features according to the present invention.
Fig. 2 is a schematic diagram of synchronous acquisition states of pulse wave PPG, electrocardiogram ECG signals and cuff blood pressure data in data acquisition according to an embodiment of the invention.
Fig. 3 is an exemplary diagram of a pulse wave PPG, electrocardiograph ECG signal acquired synchronously in an embodiment example of the invention.
Fig. 4 is a signal example diagram of the pulse wave PPG of fig. 3 and its first order differential signal dPPG in an example of implementation of the invention.
Fig. 5 is a distribution range diagram of the measurement values of the blood pressure of the cuff obtained in the embodiment of the present invention.
FIG. 6 is a graph comparing the predicted and actual measured values of the systolic blood pressure SBP on the MPGA-BP model for different data sets in an example of implementation of the present invention.
FIG. 7 is a comparison between estimated and measured values of the diastolic DBP pressure on the MPGA-BP model for different data sets in an example of an embodiment of the present invention.
FIG. 8 is a Bland-Altman plot of the five parameter dataset estimate predictions SBP and DBP selected in an example embodiment of the present invention.
Detailed Description
The invention will be further elucidated and described with reference to the drawings and the detailed description. The technical features of the embodiments of the present invention can be combined correspondingly without mutual conflict.
The invention provides a blood pressure estimation characteristic information extraction method based on pulse wave signals and electrocardiosignals, which is used for applying extracted blood pressure characteristic information vectors to a non-invasive blood pressure prediction model to carry out non-invasive blood pressure prediction; the flow of the blood pressure characteristic information extraction method is shown in figure 1,
the method comprises the following steps:
s1: synchronously acquiring pulse wave signals and electrocardiosignals of an object to be detected, and performing noise reduction pretreatment;
s2: respectively extracting signal characteristic parameters from the pulse wave signals and the electrocardiosignals after noise reduction pretreatment to form a blood pressure characteristic information vector containing Nk signal characteristic parameters;
s3: performing pairwise correlation analysis on each signal characteristic parameter in the blood pressure characteristic information vector, and performing redundancy removal processing on the signal characteristic parameters with correlation redundancy to obtain redundancy-removed blood pressure characteristic information vectors;
s4: performing weight operation on each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector to obtain weight sequence of each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector;
s5: selecting a plurality of signal characteristic parameters from the redundancy-removed blood pressure characteristic information vector according to the sequence of weight sequencing from large to small to form a test blood pressure characteristic information vector, thereby constructing a plurality of test blood pressure characteristic information vectors with different numbers of signal characteristic parameters;
s6: inputting each constructed test blood pressure characteristic information vector into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation prediction, wherein the noninvasive blood pressure prediction model is a back propagation neural network model (MPGA-BP model) optimized by a multi-population genetic algorithm; and then, preferentially selecting a test blood pressure characteristic information vector as the extracted blood pressure characteristic information vector according to the blood pressure estimation prediction result and the prediction performance index.
The blood pressure characteristic information extraction method provided by the invention is used for extracting the waveform characteristics which have obvious influence on the blood pressure by combining blood pressure influence factors and consideration factors of a Moens-Korteweg equation on the basis of denoising and characteristic point identification processing of extracted pulse wave (PPG) signals and Electrocardiogram (ECG) signals. Furthermore, the invention screens the characteristic parameters for blood pressure estimation based on correlation analysis and weight operation analysis, selects different numbers of characteristic parameters to form a less-parameter characteristic set according to the screening result, and obtains the blood pressure estimation value with the less-parameter characteristic set based on the preference of a back propagation neural network model (MPGA-BP model) optimized by a multi-population genetic algorithm, thereby providing a foundation for the construction of a continuous blood pressure monitoring model based on less-characteristic parameters and high prediction precision.
In a specific implementation, the blood pressure characteristic information extraction method of the present invention can use the ECG and PPG signals shown in fig. 2 and the cuff blood pressure synchronous acquisition manner to synchronously acquire the pulse wave signal and the ECG signal of the object to be detected, so as to perform extraction processing on the blood pressure characteristic information vector of the object to be detected. As an example of experimental applications, the blood pressure measuring device may be an ohron electronic sphygmomanometer (HEM-7210), and the device for measuring ECG and PPG waveforms is an acquisition front end of ECG and PPG waveforms that is autonomously developed by a subject group. In the experimental process, a testee sits on an armchair, a pulse wave sensor is clamped on the index finger of the right hand of the testee, and an electrocardioelectrode plate is attached to the corresponding position (LA, RA and LL) of a human body in an ECG three-lead detection mode. The cuff of the electronic sphygmomanometer is worn on the left upper arm. The subject started to acquire PPG and ECG signals simultaneously 5 minutes after resting. After the waveforms of the two signals are not changed too much and are in a relatively stable state, the blood pressure is measured by an ohm dragon electronic sphygmomanometer every other minute, and the measured blood pressure value and the measuring time are recorded. The duration of each signal acquisition is around 20 minutes. After signal acquisition, the signal waveform data is divided into time windows of 20 seconds, and data with blood pressure record values corresponding to the time is stored. After removing waveforms without corresponding blood pressure values, this example uses a total of 552 sets of data, each set of data including the individual's height, weight, PPG and ECG waveforms having a duration of 20 seconds, and corresponding measurements of Systolic (SBP) and Diastolic (DBP) pressures.
Denoising the acquired ECG signals through wavelet threshold processing; after the PPG signal is denoised by wavelet threshold processing, cubic spline interpolation processing is carried out by locating a periodic signal starting point b of the pulse wave signal shown in figure 3 to remove the baseline drift of the signal.
After the denoising preprocessing is performed, the R wave of the electrocardiographic signal shown in fig. 3 and 4, the periodic signal starting point b, the maximum ascending branch slope point a, the peak point c, the minimum descending branch slope point e, the dicrotic wave starting point f, the dicrotic wave peak point g of the pulse wave signal, and the periodic signal starting point, the peak point and the trough of the first-order differential dPPG signal are located. By means of the located feature points, signal feature parameters can be respectively extracted from the pulse wave signals and the electrocardiosignals after the noise reduction pretreatment, and in the specific implementation of the step S2, the extracted signal feature parameters can include individual feature parameters, pulse wave time feature parameters, pulse wave curve area feature parameters, curve slope feature parameters and signal amplitude feature parameters.
The extracted individual characteristic parameters comprise Weight and heart rate HR of an individual to be detected;
the extracted pulse wave time characteristic parameters comprise: time span t from starting point to peak value of pulse wave periodic signal up Time span t from the start of the pulse wave periodic signal to the start of the dicrotic wave bf Time span t from maximum slope of ascending branch to minimum slope of descending branch of pulse wave ae The time span t from peak to trough of the pulse wave and its first order differential signal dPPG down And dt down Time span t from pulse wave start point to pulse wave end point fb Time span t from start to end of pulse wave period signal bb Propagation time PTT between the start points of electrocardio R wave and pulse wave synchronous periodic signals b Propagation time PTT between the maximum rising branch slope point of the electrocardio R wave and the pulse wave a And the propagation time PTT between the peak point of the electrocardio R wave and the pulse wave c And time spans tw1, tw2, tw3, tw4, tw5 from the start of the pulse wave period signal to 1/6, 2/6, 3/6, 4/6, 5/6 of the maximum amplitude, respectively;
the extracted pulse wave curve area characteristic parameters comprise: area of systolic phase A of pulse wave bf Area of diastolic phase of pulse wave A fb The ascending branch area dAA of the pulse wave and the descending branch area dDA of the pulse wave; pulse wave characteristic parameter K value, K = (A) m -I c )/(I c )，A m Is the average pulse wave area of the period, I c Is the amplitude of the pulse peak point c; and the area A1 under the pulse wave curve between the starting point b of the periodic signal and the maximum slope point a of the ascending branch, the area A2 under the pulse wave curve between the maximum slope point a of the ascending branch and the peak point c, and the area between the peak point c and the minimum slope point e of the descending branch in the pulse waveThe area A3 under the pulse wave curve and the area A4 under the pulse wave curve between the minimum point e of the slope of the descending branch and the starting point b' of the next period signal of the pulse wave;
the extracted curve slope characteristic parameters comprise: slope S from pulse wave periodic signal starting point to dicrotic wave starting point cf The rising slope S of the pulse wave bc And descending branch slope S cb Slope dS of rising branch of first order differential signal dPPG of pulse wave PPG bc And decreasing the slope dS of the branch cb
The signal amplitude characteristic parameters comprise: the ratio dPIR between the pulse wave peak amplitude and the periodic signal starting amplitude, the ratio I of the intensity of the rising branch slope maximum point a, the falling branch slope minimum point e, the dicrotic wave starting point f and the dicrotic wave peak value point g to the peak intensity ar 、I er 、I fr 、I gr 。
Therefore, nk =36 signal characteristic parameters are extracted for each set of data according to the located waveform characteristic point pairs, and a blood pressure characteristic information vector is formed, and the definition of each signal characteristic parameter is shown in table 1.
TABLE 1
Next, step S3 is performed to perform pairwise correlation analysis on each signal feature parameter in the blood pressure feature information vector. In the specific implementation, the invention proposes to adopt the Pearson correlation coefficient as an evaluation index for judging the correlation between two signal characteristic parameters, so as to carry out pairwise correlation analysis on each signal characteristic parameter in the blood pressure characteristic information vector; any two signal characteristic parameters x in each signal characteristic parameter in the blood pressure characteristic information vector i And x j The pearson correlation coefficient between them is determined as follows:
where ρ is i,j Is the signal characteristic parameter x of the ith dimension i And a signal characteristic parameter x of the j dimension j Pearson's correlation coefficient between, σ i 、σ j Respectively is the signal characteristic parameter x of the ith dimension in the blood pressure characteristic information vector acquired by multiple times of acquisition i And the signal characteristic parameter x of the j dimension j Variance of (E), E [ (x) i -μ i )(x j -μ j )]For a characteristic parameter x of the signal i And a signal characteristic parameter x j Standard deviation, μ between i Signal characteristic parameter x representing ith dimension in blood pressure characteristic information vector obtained by multiple times of acquisition i Mean value of (d) (. Mu.) j Representing the signal characteristic parameter x of j dimension in the blood pressure characteristic information vector obtained by multiple times of acquisition j Of the average value of (a). Wherein the signal characteristic parameter x i And a signal characteristic parameter x j The value range of the dimension angle mark i, j is determined according to the number of signal characteristic parameters contained in the blood pressure characteristic information vector; signal characteristic parameter x i And a signal characteristic parameter x j The variance and the standard deviation between the variance and the standard deviation need to be calculated and determined according to the sample size of the acquired blood pressure characteristic information vector.
In the experimental example of the present invention, 522 blood pressure characteristic information vectors are collected in total, and each vector contains the above-mentioned 36-dimensional signal characteristic parameters. For the 552 blood pressure feature information vector samples containing 36 dimensions, pearson correlation analysis is performed between every two dimensions according to the following formula, and pearson correlation coefficients are obtained, so that feature parameters with strong correlation and correlation coefficients thereof are obtained as shown in table 2.
TABLE 2
After the step S3, carrying out pairwise correlation analysis on each signal characteristic parameter in the blood pressure characteristic information vector, carrying out redundancy removal processing on the signal characteristic parameters with the correlation redundancy to obtain redundancy-removed blood pressure characteristic information vectors; in specific implementation, a specific way of performing redundancy removal processing on the signal characteristic parameters with the correlation redundancy is as follows: counting the Pearson correlation coefficient between every two signal characteristic parameters in each signal characteristic parameter of the blood pressure characteristic information vector, and judging that the two signal characteristic parameters are mutually redundant characteristic parameters when the Pearson correlation coefficient value is larger than a preset correlation coefficient threshold value; then, selecting and reserving one of the signal characteristic parameters which are mutually redundant characteristic parameters, and removing other redundant characteristic parameters and relevant redundant characteristic parameters; selecting and reserving a basic principle of screening, selecting and extracting simple and error-prone parameters from parameters with high correlation degree as far as possible, for example, in the positioning of characteristic points of pulse waves, the peak value, the starting point of periodic signals and the point with the maximum slope are most easily positioned, while the dicrotic waves are relatively difficult to position due to individual difference, if redundant characteristic parameters exist, selecting and reserving the parameters which are easy to position and not easy to position errors, so that the accuracy and the robustness of prediction estimation for a blood pressure value estimation system are improved more favorably; therefore, redundancy removing processing is carried out on the signal characteristic parameters with the relevant redundancy from the blood pressure characteristic information vector, and a redundancy removing blood pressure characteristic information vector is obtained.
For example, in the experimental example, t is shown in the data of Table 2 down 、dt down 、t bb HR and t fb These five parameters are highly correlated. Selecting one of the heart rate HR and PPG time characteristics t of one of the blood pressure physiology influencing factors down As a parameter for the blood pressure estimation of this study, dt down 、t fb 、t bb Is excluded; PTT a 、PTT b And PTT c High degree of correlation, excluding PTT a And PTT b (ii) a tw4 and tw5 are excluded between tw3, tw4 and tw5; s bc 、dS bc Exclusion of dS between dAA and dAA bc And dAA; s cb And dS cb Do between cb ；A bf And A fb Between exclude A fb . The basic principle of screening is to select and extract parameters which are simple and difficult to make mistakes as far as possible from the parameters with high degree of correlation, thereby being beneficial to improving the robustness of the whole system in the subsequent engineering realization research. Accordingly, a total of 11 parameters are filtered out, and the remaining 25 signal characteristic parameters constitute a redundancy-free blood pressure characteristic information vector.
After the redundancy among the characteristics is removed by utilizing the correlation analysis, the method of the invention goes to step S4, and then the weight operation is carried out on each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector to obtain the weight sequence of each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector; in specific implementation, for each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector, a weight value of each signal characteristic parameter is calculated by using a ReliefF algorithm, and the specific flow is as follows:
s401: creating a weight coefficient w with an initial value of 0 for each signal characteristic parameter in the de-redundant blood pressure characteristic information vector i I.e. by
Representing the initial value of the weight of the ith signal characteristic parameter before iteration in the redundancy-removed blood pressure characteristic information vector, i is from {1,2, \8230;, N f }，N f The total number of the signal characteristic parameters contained in the redundancy-removed blood pressure characteristic information vector is calculated; thus, a characteristic weight vector corresponding to the redundancy-removed blood pressure characteristic information vector is formed by the set of weight coefficients; meanwhile, the number of initialization iterations m =1;
s402: randomly selecting a redundancy-removing blood pressure characteristic information vector sample X in a redundancy-removing characteristic training set; the redundancy-removing characteristic training set comprises a plurality of blood pressure sample data marked with blood pressure value labels, and each blood pressure sample data comprises a pulse wave signal which is synchronously acquired and corresponds to the blood pressure value label and a redundancy-removing blood pressure characteristic information vector extracted from the electrocardiosignal;
s403: selecting K nearest neighbor samples from the samples in the redundancy-removing characteristic training set, wherein the samples are similar to the redundancy-removing blood pressure characteristic information vector sample X, and marking the samples as the NearHit samples of the redundancy-removing blood pressure characteristic information vector sample X; selecting K nearest neighbor samples from other types of samples which are not classified into the redundancy-removing characteristic training set and the redundancy-removing blood pressure characteristic information vector sample X, and marking the K nearest neighbor samples as NearMiss samples of the redundancy-removing blood pressure characteristic information vector sample X;
s404: for the weight coefficient corresponding to the ith signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector, calculating the iterative weight value according to the following mode:
wherein the content of the first and second substances,
and &>
Respectively represent the i-th signal characteristic parameter x i The values of the corresponding weight coefficients at the mth iteration and the (m-1) th iteration; />
Representing the i-th signal characteristic parameter x i And the kth NearHit sample +>
M is iteration times, K belongs to {1,2, \8230;, K }, and K is the number of sought NearHit samples; c. C 1 Representing the class in which the sample X is extracted, p (c) being the prior probability of the class c, p (c) 1 ) Is c 1 A priori probability of a category, <' > based on>
Representing the i-th signal characteristic parameter x i And the kth NearMiss sample +>
The distance of (d);
therefore, the values of the weight coefficients corresponding to the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector in the mth iteration are respectively calculated;
s405: enabling the iteration number m to be added by 1, and then returning to the step S402;
s406: repeating the steps S402 to S405 until the value of the iteration number m reaches the preset iteration number upper limit value, and executing the step S407;
s407: and sequencing the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector according to the descending order of the values of the weight coefficients corresponding to the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector, and taking the sequence as the weight sequencing of the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector.
As can be seen from the above flow, the significance of the weighting is to subtract the difference of the features of the same classification and add the difference of the features of different classifications. That is, if the feature is associated with a class, the value of the feature for the same class should be similar and the values for different classes should not be similar.
Table 3 shows the weight ranking calculated by using the ReliefF algorithm on the 522 collected redundancy-removed blood pressure feature information vectors in the above embodiment of the present invention, which is taken as the top ten features.
TABLE 3
Then, step S5 of the method selects a plurality of signal characteristic parameters from the redundancy-free blood pressure characteristic information vector according to the sequence of weight sorting from large to small to form a test blood pressure characteristic information vector, thereby constructing a plurality of test blood pressure characteristic information vectors with different numbers of signal characteristic parameters.
In order to further verify the effectiveness of the feature selection method constructed based on the PPG waveform feature estimation blood pressure value model, the implementation example of the invention selects the first three, five and seven most important weight ranking and all 25 signal feature parameters in the redundancy-removed blood pressure feature information vector, and constructs four test blood pressure feature information vectors with different numbers of signal feature parameters to form an experimental data set. And (4) segmenting the training set and the test set by adopting an autonomous segmentation method for the data set. The principle is that the training set is formed by the returned randomly selected samples with the same number as the data set samples, and the unselected samples are formed into the testing set. Therefore, the method fully utilizes the given observation information, does not need other assumptions of the model, and has high robustness and efficiency. Meanwhile, the method avoids the problem of sample reduction caused by cross validation by resampling, increases the randomness of data, and enables the training result of the model to be validated at the data level.
Furthermore, a non-invasive blood pressure prediction model is set up for blood pressure estimation and prediction, and the non-invasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm, namely an MPGA-BP model for short; the specific mode for optimizing the back propagation neural network model by the multi-population genetic algorithm is as follows:
step a01: constructing 5 different populations, wherein each population has 20 individuals, each individual represents a group of weight and threshold values for initializing the BP network, and simultaneously, the control parameters of each population are different;
step a02: coding each individual of each population, and calculating the fitness of each individual in each population;
step a03: selecting each population by using a selection algorithm, and generating new individuals according to crossover and mutation operators;
step a04: b, calculating the fitness of the filial generation in the step a 03;
step a05: finding out the individual with the highest fitness in the population to carry out immigration operation;
step a06: selecting individuals with highest fitness in each population by using an artificial selection operator to form an essence population;
step a07: determining the individual with the highest fitness in the essence population, recording the algebra of the individual in the optimal state in the essence population, stopping the algorithm when the algebra exceeds the preset algebra, and decoding to obtain the initial weight and the threshold value of the back propagation neural network model optimized by the multi-population genetic algorithm to be used as a non-invasive blood pressure prediction model.
In step S6 of the method, each constructed test blood pressure characteristic information vector is input into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation prediction, wherein the noninvasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm; and then, preferentially selecting a test blood pressure characteristic information vector as the extracted blood pressure characteristic information vector according to the blood pressure estimation prediction result and the prediction performance index.
When one test blood pressure characteristic information vector is selected preferentially, the performance of blood pressure estimation prediction can be performed according to the fact that each test blood pressure characteristic information vector is input to a pre-trained noninvasive blood pressure prediction model, and the complexity of model construction, the accuracy of model output and the stability and reliability of model operation are balanced to perform final determination.
In the implementation example of the present invention, the above-mentioned four test blood pressure feature information vectors of the three-parameter, five-parameter, seven-parameter, and full-parameter data sets are respectively input into the MPGA-BP model to train the model, and after the training is completed, the test set of each data set is used to observe the blood pressure estimation effect of the model, and the results are shown in fig. 6 and 7. Fig. 6 is a comparison graph of the predicted value of the systolic blood pressure SBP and the actual measurement value of the three-parameter dataset (fig. 6 (a)), the five-parameter dataset (fig. 6 (b)), the seven-parameter dataset (fig. 6 (c)), and the full-parameter dataset (fig. 6 (d)), and fig. 7 is a comparison graph of the predicted value of the diastolic blood pressure DBP and the actual measurement value of the three-parameter dataset (fig. 6 (a)), the five-parameter dataset (fig. 6 (b)), the seven-parameter dataset (fig. 6 (c)), and the full-parameter dataset (fig. 6 (d)).
In order to better compare the results of the blood pressure predicted by the blood pressure prediction models with different parameter numbers, the embodiment uses the Mean Absolute Error (MAE), the Root Mean Square Error (RMSE) and the standard deviation (STD) as the measurement indexes of the accuracy of the model for estimating the blood pressure. Table 4 shows the blood pressure prediction ability of different number of parameters expressed on the MPGA-BP model.
TABLE 4
The three-parameter, five-parameter, seven-parameter, and full-parameter feature sets shown in table 4 all met the measurement standards of the institute of advancement of medical instrumentation (AAMI) blood pressure (MAE < =5mmhg, std < =8 mmHg) when estimating SBP and DBP. When estimating the SBP, the five-parameter, seven-parameter and full-parameter feature sets have ideal effects. From the MAE point of view, five and seven parameters perform optimally. This shows that the full-parameter estimation SBP belongs to a state of feature redundancy, and too much less-valuable training data influences the complexity and accuracy of the model, further explaining the importance of feature selection. From the STD and RMSE, the five-parameter and full-parameter effects are optimal. When the DBP is estimated, the effects of the three-parameter, five-parameter, seven-parameter and full-parameter feature sets are ideal. Wherein, the five-parameter and full-parameter characteristics are superior to the three-parameter and seven-parameter characteristics. In conclusion, the correlation analysis-Relieff feature selection method well extracts feature information required by blood pressure estimation, blood pressure prediction with the same precision is realized on the MPGA-BP model constructed by the invention by preferentially selecting five parameters, and the five parameters show higher accuracy on SBP and DBP estimation prediction as can be seen from fig. 6 and 7.
The technical object of the blood pressure characteristic information extraction method provided by the invention is to apply the extracted blood pressure characteristic information vector to a non-invasive blood pressure prediction model to carry out non-invasive blood pressure prediction. Therefore, the invention also provides a blood pressure value estimation system, which comprises:
the blood pressure characteristic information extraction module is used for synchronously acquiring the pulse wave signals and the electrocardiosignals of the object to be detected, and extracting the blood pressure characteristic information vectors of the pulse wave signals and the electrocardiosignals of the object to be detected by adopting the blood pressure characteristic information extraction method;
the blood pressure estimation and prediction module is used for inputting the extracted blood pressure characteristic information vector of the object to be detected into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation and prediction and outputting a blood pressure estimation result of the object to be detected; the non-invasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm.
However, after the back propagation neural network model optimized by the multi-population genetic algorithm is obtained as the non-invasive blood pressure prediction model, the pre-trained non-invasive blood pressure prediction model is obtained after the blood pressure prediction training of the non-invasive blood pressure prediction model is required; the specific mode of carrying out blood pressure prediction training on the noninvasive blood pressure prediction model is as follows:
step b01: acquiring a blood pressure sample data set from a blood pressure sample database, wherein the blood pressure sample data set comprises a plurality of blood pressure sample data marked with blood pressure value labels, and each blood pressure sample data comprises a pulse wave signal and an electrocardiosignal which are synchronously acquired and correspond to the blood pressure value label;
step b02: respectively extracting the blood pressure characteristic information vector of each blood pressure sample data in the blood pressure sample data set by adopting the blood pressure characteristic information extraction method;
step b03: selecting a training sample and a test sample from a blood pressure sample set to respectively form a training sample set and a test sample set;
step b04: the blood pressure characteristic information vector of each blood pressure sample data in the training sample set is used as the input of the non-invasive blood pressure prediction model, the blood pressure value label of each blood pressure sample data in the training sample set is used as the output verification label, and the non-invasive blood pressure prediction model is subjected to blood pressure prediction training to adjust the blood pressure prediction parameters of the non-invasive blood pressure prediction model;
step b05: inputting the blood pressure characteristic information vector of each blood pressure sample data in the test sample set into a non-invasive blood pressure prediction model for blood pressure prediction, adopting the blood pressure value label of each blood pressure sample data in the test sample set as an output verification label, comparing and verifying the blood pressure prediction result of the non-invasive blood pressure prediction model, and evaluating the blood pressure prediction performance of the non-invasive blood pressure prediction model;
step b06: if the blood pressure prediction performance of the noninvasive blood pressure prediction model does not reach the preset target, returning to execute the step b04; and if the blood pressure prediction performance of the non-invasive blood pressure prediction model reaches a preset target, finishing training to obtain a pre-trained non-invasive blood pressure prediction model.
In the specific application implementation, in the training process of the non-invasive blood pressure prediction model, the ratio of the number of the blood pressure samples of the training samples and the test samples selected in the step b03 can be designed as 8; then, performing blood pressure prediction training on the noninvasive blood pressure prediction model by using a training sample set, selecting an optimal model with the lowest loss of an optimal verification set, predicting a test sample set, and evaluating whether the performance reaches a preset service target based on indexes; if the preset business target is not reached, the parameters in the process can be adjusted, and the training step is repeatedly executed until the performance reaches the preset target.
FIG. 5 illustrates a distribution range diagram of the systolic blood pressure SBP (FIG. 5 (a)) and the diastolic blood pressure DBP (FIG. 5 (b)) of cuff blood pressure measurements obtained in an example of implementation; FIG. 8 shows a Bland-Altman plot of the five parameter estimated systolic SBP (FIG. 8 (a)) and diastolic DBP (FIG. 8 (b)) values for an experimental example of the present invention, with the middle solid line indicating the mean error between measured and predicted values, and it can be seen that the mean difference between SBP and DBP measured and predicted values is 0.21 and 0.56, respectively; the upper and lower dashed lines in the graph represent the upper and lower bounds of the confidence interval, and it can be seen that most of the errors for SBP and DBP lie between the upper and lower lines of the 95% confidence interval. It is shown that the five characteristic parameters selected by the method of the present invention contain sufficient physiological information required for blood pressure estimation, and high-precision and low-error blood pressure estimation is achieved within the blood pressure distribution range of fig. 5.
In summary, the method for extracting blood pressure estimation feature information provided by the invention has the following technical advantages:
1. the method for extracting the blood pressure characteristic information extracts the waveform characteristics which have obvious influence on the blood pressure by combining blood pressure influence factors and consideration factors of a Moens-Korteweg equation on the basis of denoising and characteristic point identification processing of extracted pulse wave (PPG) signals and Electrocardiogram (ECG) signals; furthermore, the invention screens the characteristic parameters for blood pressure estimation based on correlation analysis and weight operation analysis, selects different numbers of characteristic parameters to form a small parameter characteristic set according to the screening result, and obtains the blood pressure estimation value with the small parameter characteristic set based on the preference of a back propagation neural network model (MPGA-BP model) optimized by a multi-population genetic algorithm, thereby providing a foundation for the construction of a continuous blood pressure value estimation system based on small characteristic parameters and high prediction precision, and being beneficial to the blood pressure value estimation system to realize high-accuracy non-invasive blood pressure prediction.
2. Before the method uses the Relieff algorithm for the feature selection of the noninvasive estimation continuous blood pressure research, the redundancy among features is eliminated by utilizing the correlation analysis, the number of the highly correlated features is effectively reduced, and the influence degree of the selected features on a blood pressure prediction model keeps a certain gradient; furthermore, the Relieff algorithm gives weights which can reflect actual conditions to different characteristics, and the selected characteristic dimensions with the same number contain more pulse wave waveform information, so that the construction of a noninvasive blood pressure prediction model with few parameters is facilitated; the number of parameters required by model construction is reduced, so that the complexity of the model can be further reduced, the influence factors of the output result of the model are reduced, and the stability and the accuracy of the constructed noninvasive blood pressure prediction model are improved.
3. The method effectively reduces the number of pulse wave characteristics used on the premise of ensuring the accuracy of the constructed noninvasive blood pressure prediction model, improves the intelligibility and interpretability of the blood pressure prediction model result, and creates conditions for realizing the high-precision and low-error continuous noninvasive blood pressure prediction model.
Finally, it should be noted that the above embodiments are only used for illustrating the technical solutions of the present invention and not for limiting the technical solutions, and those skilled in the art should understand that modifications or equivalent substitutions can be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions, and all that should be covered by the claims of the present invention.
Claims (10)
1. A blood pressure characteristic information extraction method is characterized in that the method is used for applying the extracted blood pressure characteristic information vector to a blood pressure prediction model for non-invasive blood pressure prediction; the blood pressure characteristic information extraction method comprises the following steps:
s1: synchronously acquiring pulse wave signals and electrocardiosignals of an object to be detected, and performing noise reduction pretreatment;
s2: respectively extracting signal characteristic parameters from the pulse wave signals and the electrocardiosignals after noise reduction pretreatment to form a blood pressure characteristic information vector containing Nk signal characteristic parameters;
s3: performing pairwise correlation analysis on each signal characteristic parameter in the blood pressure characteristic information vector, and performing redundancy removal processing on the signal characteristic parameters with correlation redundancy to obtain redundancy-removed blood pressure characteristic information vectors;
s4: performing weight operation on each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector to obtain weight sequence of each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector;
s5: selecting a plurality of signal characteristic parameters from the redundancy-removed blood pressure characteristic information vector according to the sequence of weight sequencing from large to small to form a test blood pressure characteristic information vector, thereby constructing a plurality of test blood pressure characteristic information vectors with different numbers of signal characteristic parameters;
s6: inputting each constructed test blood pressure characteristic information vector into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation prediction, wherein the noninvasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm; and then, preferentially selecting a test blood pressure characteristic information vector as the extracted blood pressure characteristic information vector according to the blood pressure estimation prediction result and the prediction performance index.
2. The method for extracting blood pressure characteristic information according to claim 1, wherein in the step S2, the signal characteristic parameters extracted from the pulse wave signal and the electrocardiographic signal after the noise reduction preprocessing include an individual characteristic parameter, a pulse wave time characteristic parameter, a pulse wave curve area characteristic parameter, a curve slope characteristic parameter, and a signal amplitude characteristic parameter.
3. The method for extracting blood pressure characteristic information according to claim 2, wherein the individual characteristic parameters include Weight and heart rate HR of the subject individual;
the pulse wave time characteristic parameters comprise: time span t from starting point to peak value of pulse wave periodic signal up Time span t from the start of the pulse wave periodic signal to the start of the dicrotic wave bf Time span t from maximum slope of ascending branch to minimum slope of descending branch of pulse wave ae The time span t from the peak to the trough of the pulse wave and its first order differential signal down And dt down Time span t from pulse wave start point to pulse wave end point fb Time span t from start to end of pulse wave period signal bb Propagation time PTT between the start points of the electrocardio R wave and the pulse wave synchronous period signals b Propagation time PTT between the maximum rising branch slope point of the electrocardio R wave and the pulse wave a Propagation time PTT between electrocardio R wave and pulse wave peak value point c And time spans tw1, tw2, tw3, tw4, tw5 from the start of the pulse wave period signal to 1/6, 2/6, 3/6, 4/6, 5/6 of the maximum amplitude, respectively;
the pulse wave curve area characteristic parameters comprise: area of systolic phase A of pulse wave bf Area of diastolic phase of pulse wave A fb The ascending branch area dAA of the pulse wave and the descending branch area dDA of the pulse wave; pulse wave characteristic parameter K value, K = (A) m -I c )/(I c )，A m Is the average pulse wave area of the period, I c Is the amplitude of the pulse peak point c; and the area under the pulse wave curve A1 between the periodic signal starting point b and the ascending branch slope maximum point a, the area under the pulse wave curve A2 between the ascending branch slope maximum point a and the peak point c, the area under the pulse wave curve A3 between the peak point c and the descending branch slope minimum point e, and the area under the pulse wave curve A4 between the descending branch slope minimum point e and the pulse wave next period signal starting point b' in the pulse wave;
the curve slope characteristic parameters comprise: slope S from pulse wave periodic signal starting point to dicrotic wave starting point cf The rising slope S of the pulse wave and its first order differential signal bc And dS bc The pulse wave reachesFalling branch slope S of first order differential signal cb And dS cb ；
The signal amplitude characteristic parameters comprise: the ratio dPIR between the pulse wave peak amplitude and the periodic signal starting amplitude, and the ratio I of the intensity of the rising branch slope maximum point a, the falling branch slope minimum point e, the counterpulsation wave starting point f and the counterpulsation wave peak point g to the peak intensity ar 、I er 、I fr 、I gr 。
4. The method for extracting blood pressure characteristic information according to claim 1, wherein in the step S3, a pearson correlation coefficient is used as an evaluation index for evaluating the correlation between two signal characteristic parameters, so as to perform pairwise correlation analysis on each signal characteristic parameter in the blood pressure characteristic information vector; any two signal characteristic parameters x in each signal characteristic parameter in the blood pressure characteristic information vector i And x j The pearson correlation coefficient between them is determined as follows:
wherein ρ i,j Is the signal characteristic parameter x of the ith dimension i And a signal characteristic parameter x of the j dimension j Pearson's correlation coefficient between, σ i 、σ j Respectively is the signal characteristic parameter x of the ith dimension in the blood pressure characteristic information vector acquired by multiple times of acquisition i And the signal characteristic parameter x of the j dimension j Variance of E [ (x) i -μ i )(x j -μ j )]For a characteristic parameter x of the signal i And a signal characteristic parameter x j Standard deviation between, mu i Signal characteristic parameter x representing ith dimension in blood pressure characteristic information vector obtained by multiple times of acquisition i Mean value of (d) (. Mu.) j Representing the signal characteristic parameter x of j dimension in the blood pressure characteristic information vector obtained by multiple times of acquisition j Is measured.
5. The method for extracting blood pressure feature information according to claim 4, wherein in the step S3, the specific manner of performing the redundancy removal processing on the signal feature parameters having the correlation redundancy is as follows:
counting the Pearson correlation coefficient between every two signal characteristic parameters in each signal characteristic parameter of the blood pressure characteristic information vector, and if the Pearson correlation coefficient value is larger than a preset correlation coefficient threshold value rho 0 Judging mutual redundant characteristic parameters for the two signal characteristic parameters, and judging mutual correlation redundant characteristic parameters for the signal characteristic parameters with the same redundant characteristic parameters; then, selecting one of the signal characteristic parameters of the mutual redundant characteristic parameters and the mutual correlation redundant characteristic parameters for reservation, and removing other redundant characteristic parameters and correlation redundant characteristic parameters; therefore, redundancy removing processing is carried out on the signal characteristic parameters with the relevant redundancy from the blood pressure characteristic information vector, and a redundancy removing blood pressure characteristic information vector is obtained.
6. The method according to claim 1, wherein in step S4, for each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector, a ReliefF algorithm is used to calculate a weight value of each signal characteristic parameter, and the specific process is as follows:
s401: creating a weight coefficient w with an initial value of 0 for each signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector i I.e. by
Representing the initial value of the weight of the ith signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector before iteration, i belongs to {1,2, \8230;, N f }，N f The total number of the signal characteristic parameters contained in the redundancy-removed blood pressure characteristic information vector is calculated; thus, the de-redundant blood pressure signature is formed by a set of weight coefficientsFeature weight vectors corresponding to the information vectors; meanwhile, the number of initialization iterations m =1;
s402: randomly selecting a redundancy-removing blood pressure characteristic information vector sample X in a redundancy-removing characteristic training set; the redundancy-removing characteristic training set comprises a plurality of blood pressure sample data marked with blood pressure value labels, and each blood pressure sample data comprises a pulse wave signal which is synchronously acquired and corresponds to the blood pressure value label and a redundancy-removing blood pressure characteristic information vector extracted from the electrocardiosignal;
s403: selecting K nearest neighbor samples from samples in the same class as the redundancy-removing blood pressure characteristic information vector sample X in the redundancy-removing characteristic training set, and marking the K nearest neighbor samples as NearHit samples of the redundancy-removing blood pressure characteristic information vector sample X; selecting K nearest neighbor samples from other types of samples which are not in the same class as the redundancy-removing blood pressure characteristic information vector sample X in the redundancy-removing characteristic training set, and marking the K nearest neighbor samples as NearMiss samples of the redundancy-removing blood pressure characteristic information vector sample X;
s404: for the weight coefficient corresponding to the ith signal characteristic parameter in the redundancy-removed blood pressure characteristic information vector, calculating the iterative weight value according to the following mode:
wherein the content of the first and second substances,
and &>
Respectively representing the ith signal characteristic parameter x i The values of the corresponding weight coefficients at the mth iteration and the (m-1) th iteration; />
Representing the i-th signal characteristic parameter x i And the kth NearHit sample->
M is iteration frequency, M is a preset upper limit value of the iteration frequency, K belongs to {1,2, \ 8230;, K }, and K is the number of sought NearHit samples; c. C 1 Representing the class in which the sample X was taken, p (c) being the prior probability of the class c, p (c) 1 ) Is c 1 The prior probability of a class of the object,
representing the ith signal characteristic parameter x i And the kth NearMiss sample->
The distance of (a);
therefore, the values of the weighting coefficients corresponding to the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector in the mth iteration are respectively calculated;
s405: enabling the iteration number m to be added by 1, and then returning to the step S402;
s406: repeatedly executing the steps S402 to S405 until the value of the iteration number M reaches a preset iteration number upper limit value M, and executing a step S407;
s407: and sequencing the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector according to the descending order of the values of the weight coefficients corresponding to the signal characteristic parameters at present, and taking the sequencing as the weight sequencing of the signal characteristic parameters in the redundancy-removed blood pressure characteristic information vector.
7. A blood pressure value estimation system, comprising:
a blood pressure characteristic information extraction module for synchronously acquiring the pulse wave signals and the electrocardiosignals of the object to be detected, and extracting the blood pressure characteristic information vectors of the pulse wave signals and the electrocardiosignals of the object to be detected by adopting the blood pressure characteristic information extraction method of any one of claims 1 to 6;
the blood pressure estimation and prediction module is used for inputting the extracted blood pressure characteristic information vector of the object to be detected into a pre-trained noninvasive blood pressure prediction model for blood pressure estimation and prediction and outputting a blood pressure estimation result of the object to be detected; the non-invasive blood pressure prediction model is a back propagation neural network model optimized by a multi-population genetic algorithm.
8. The blood pressure value estimation system according to claim 7, wherein the back propagation neural network model is optimized by a multi-population genetic algorithm in a manner that:
step a01: constructing 5 different populations, wherein each population has 20 individuals, each individual represents a group of weight and threshold values for initializing the BP network, and simultaneously, the control parameters of each population are different;
step a02: coding each individual of each population, and calculating the fitness of each individual in each population;
step a03: selecting each population by using a selection algorithm, and generating new individuals according to crossover and mutation operators;
step a04: b, calculating the fitness of the filial generation in the step a 03;
step a05: finding out the individual with the highest fitness in the population to carry out immigration operation;
step a06: selecting individuals with highest fitness in each population by using an artificial selection operator to form an essence population;
step a07: determining the individual with the highest fitness in the essence population, recording the algebra of the individual in the optimal state in the essence population, stopping the algorithm when the algebra exceeds the preset algebra, and decoding to obtain the initial weight and the threshold value of the back propagation neural network model optimized by the multi-population genetic algorithm to be used as a non-invasive blood pressure prediction model.
9. The blood pressure value estimation system according to claim 7, wherein after the back propagation neural network model optimized by the multi-population genetic algorithm is obtained as the non-invasive blood pressure prediction model, the non-invasive blood pressure prediction model is required to be subjected to blood pressure prediction training to obtain a pre-trained non-invasive blood pressure prediction model; the specific mode of carrying out blood pressure prediction training on the noninvasive blood pressure prediction model is as follows:
step b01: acquiring a blood pressure sample data set from a blood pressure sample database, wherein the blood pressure sample data set comprises a plurality of blood pressure sample data marked with blood pressure value labels, and each blood pressure sample data comprises a pulse wave signal and an electrocardiosignal which are synchronously acquired and correspond to the blood pressure value label;
step b02: respectively extracting the blood pressure characteristic information vector of each blood pressure sample data in the blood pressure sample data set by adopting the blood pressure characteristic information extraction method of any one of claims 1 to 6;
step b03: selecting a training sample and a test sample from a blood pressure sample set to respectively form a training sample set and a test sample set;
step b04: the blood pressure characteristic information vector of each blood pressure sample data in the training sample set is used as the input of the non-invasive blood pressure prediction model, the blood pressure value label of each blood pressure sample data in the training sample set is used as the output verification label, and the non-invasive blood pressure prediction model is subjected to blood pressure prediction training to adjust the blood pressure prediction parameters of the non-invasive blood pressure prediction model;
step b05: inputting the blood pressure characteristic information vector of each blood pressure sample data in the test sample set into a non-invasive blood pressure prediction model for blood pressure prediction, adopting the blood pressure value label of each blood pressure sample data in the test sample set as an output verification label, comparing and verifying the blood pressure prediction result of the non-invasive blood pressure prediction model, and evaluating the blood pressure prediction performance of the non-invasive blood pressure prediction model;
step b06: if the blood pressure prediction performance of the non-invasive blood pressure prediction model does not reach the preset target, returning to execute the step b04; and if the blood pressure prediction performance of the non-invasive blood pressure prediction model reaches a preset target, finishing training to obtain a pre-trained non-invasive blood pressure prediction model.
10. The blood pressure value estimation system according to claim 9, wherein in the step b03, the ratio of the number of the blood pressure samples of the training samples to the number of the blood pressure samples of the test samples is selected to be 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211370210.0A CN115886763A (en) | 2022-11-03 | 2022-11-03 | Blood pressure characteristic information extraction method and blood pressure value estimation system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211370210.0A CN115886763A (en) | 2022-11-03 | 2022-11-03 | Blood pressure characteristic information extraction method and blood pressure value estimation system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115886763A true CN115886763A (en) | 2023-04-04 |
Family
ID=86484927
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211370210.0A Pending CN115886763A (en) | 2022-11-03 | 2022-11-03 | Blood pressure characteristic information extraction method and blood pressure value estimation system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115886763A (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112488180A (en) * | 2020-11-27 | 2021-03-12 | 哈尔滨工程大学 | Feature selection method based on RealifF |
| CN114569096A (en) * | 2022-02-28 | 2022-06-03 | 西安交通大学 | Non-contact continuous blood pressure measuring method and system based on video stream |
-
2022
- 2022-11-03 CN CN202211370210.0A patent/CN115886763A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112488180A (en) * | 2020-11-27 | 2021-03-12 | 哈尔滨工程大学 | Feature selection method based on RealifF |
| CN114569096A (en) * | 2022-02-28 | 2022-06-03 | 西安交通大学 | Non-contact continuous blood pressure measuring method and system based on video stream |
Non-Patent Citations (1)
| Title |
|---|
| 季忠: "基于脉搏波的无创连续血压监测模型簇研究", 《仪器仪表学报》, 31 July 2020 (2020-07-31), pages 224 - 234 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107837082A (en) | Electrocardiogram automatic analysis method and device based on artificial intelligence self study | |
| CN106725426A (en) | A kind of method and system of electrocardiosignal classification | |
| CN110522444B (en) | Electrocardiosignal identification and classification method based on Kernel-CNN | |
| WO2006113697A1 (en) | Trainable diagnotic system and method of use | |
| CN112806977B (en) | Physiological parameter measuring method based on multi-scale fusion network | |
| CN113855038B (en) | Electrocardiosignal critical value prediction method and device based on multi-model integration | |
| CN110269605B (en) | Electrocardiosignal noise identification method based on deep neural network | |
| CN110558960A (en) | continuous blood pressure non-invasive monitoring method based on PTT and MIV-GA-SVR | |
| CN101835421B (en) | Feature value candidate generating device and feature value candidate generating method | |
| CN116451110A (en) | Blood glucose prediction model construction method based on signal energy characteristics and pulse period | |
| CN113116321A (en) | Non-invasive continuous blood pressure measuring system based on PSO-GRNN neural network | |
| CN111248881B (en) | Method and device for predicting blood pressure | |
| CN115281688A (en) | Cardiac hypertrophy multi-label detection system based on multi-mode deep learning | |
| CN116369877A (en) | Noninvasive blood pressure estimation method based on photoelectric volume pulse wave | |
| CN113729648B (en) | Wearable pulse-taking bracelet system based on multiple pulse sensors | |
| CN114431873A (en) | Method for constructing electrocardiogram heart rate multi-type prediction model based on graph convolution | |
| CN109033990B (en) | CNN model heartbeat classification method based on intra-class and inter-class distances | |
| CN114469041A (en) | Heart rate change data characteristic analysis method in exercise process | |
| CN114145725B (en) | PPG sampling rate estimation method based on noninvasive continuous blood pressure measurement | |
| CN115886763A (en) | Blood pressure characteristic information extraction method and blood pressure value estimation system | |
| CN114862865B (en) | Vessel segmentation method and system based on multi-view coronary angiography sequence image | |
| CN111325202A (en) | Ultrasonic Doppler-based color spectrogram frequency spectrum blood flow detection method | |
| CN116172573A (en) | Arrhythmia image classification method based on improved acceptance-ResNet-v 2 | |
| Hughes et al. | Novel signal shape descriptors through wavelet transforms and dimensionality reduction | |
| CN115089145A (en) | Intelligent blood pressure prediction method based on multi-scale residual error network and PPG signal |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |