KR102154652B1 - Method for estimating continuous blood pressure using recurrent neural network and apparatus thereof - Google Patents

Abstract

In an embodiment of the present invention, feature information for receiving a user's biosignal measured in at least one or more methods in a past period, analyzing the received biosignal, and extracting feature information for each time point at which the biosignal was measured Extraction step; A parameter calculating step of calculating a blood pressure related parameter for the past period based on the extracted feature information; And a blood pressure estimation control step of inputting the calculated blood pressure related parameters into a recurrent neural network (RNN) in time series to output the estimated blood pressure of the user in a future period including the present time point. Provides a method for estimating section blood pressure using a neural network.

Description

A method for estimating section blood pressure using a circulatory neural network, and a device for estimating section blood pressure to implement the method.

The present invention relates to a method and apparatus for estimating section blood pressure using a circulatory neural network, and more specifically, in estimating section blood pressure using a circulatory neural network, by purifying and pre-processing input data of the circulatory neural network. The present invention relates to a method and an apparatus for accurately estimating continuous blood pressure at a future point based on the measured biometric signals.

Recently, hypertension and hypotension patients are increasing, and even if they are not patients with blood pressure-related diseases, sequelae disorders caused by a sudden increase or decrease in blood pressure have a permanent characteristic, so that changes in blood pressure are identified and blood pressure in the future There have been many studies on how to estimate.

According to statistics, cardiovascular disease affects 37% of the total population and is the most important cause of death. Among cardiovascular diseases, especially high blood pressure, early diagnosis and prevention are important. Therefore, in order to diagnose and treat various cardiovascular diseases including hypertension at an early stage, it is necessary for patients or potential patients to continuously monitor changes in their own blood pressure and cardiovascular diseases.

The conventionally known blood pressure estimation method estimates the blood pressure without taking into account the biological signal for blood pressure estimation, the change of parameters over time, and the personal characteristics of the blood pressure measurement object. Since the deviation of the parameters for blood pressure estimation varies from person to person, and the baseline value representing the normal range of blood pressure varies greatly from person to person, the future blood pressure is estimated using only the current value as in the known method. If you do, there is a limitation that you can only expect low accuracy.

Recurrent Neural Network (RNN) is known as a key tool among deep learning techniques, as it is introduced as an artificial neural network (ANN) with memory. Conventional ANNs receive a parameter at a specific point in time as an input and estimate an output value corresponding to that time, but an RNN has a characteristic of receiving time series parameters as input data and outputting time series data. Structurally, it can be seen as a technique devised in consideration of the fact that the input of one hour affects not only that time but also various points in the future.

1 is a diagram illustrating an example in which a circulatory neural network operates.

Referring to FIG. 1, it can be seen that the circulatory neural network is composed of an embedding layer 110, a hidden layer 130, and an output layer 150. More specifically, FIG. 1 shows that when a user completes a sentence up to'you will never believe' in the process of typing through a keyboard, an RNN is driven to estimate what word the next typed word will be. It is a diagram schematically showing the operation. The circulatory neural network used in FIG. 1 is a bi-directional RNN.

It can be seen that the embedding layer 110 is a layer that accepts time series input data, and in FIG. 1, the embedding layer 110 accepts four words you, will, never, believe, or more as time series input data. have. Time series data input from the embedding layer 110 is labeled and transmitted to the hidden layer 130.

The hidden layer 130 is a layer that receives time-series input data, analyzes it for each point of view, and learns about adjacent points of time, and you inputted at time t-4, which is the highest priority, will be input at time t-3, which is the next time point. It is learned to affect never entered at time t-2. The hidden layer 130 is composed of a forward layer and a backward layer and is designed to learn time series data, and is repeatedly learned to minimize the value of the internally set cost function. (training) and being tested.

The output layer 150 is a layer that outputs a final output value through a result value optimized by the hidden layer 130. In FIG. 1, the output layer 150 may output words such as what, how, and where in response to a time series parameter input to the embedding layer 110.

On the other hand, as a non-invasive method of measuring blood pressure, the method of measuring the cuff around the upper arm, calf, and thigh (sphygmogram) is the most widely used, but the user's compliance is low, so it is not suitable for frequent measurement. As such, it is time to need a method for estimating blood pressure in consideration of changes in blood pressure at the past and individual blood pressure changes while making frequent measurements easy.

Here, the meaning of low user compliance means that the frequency of measuring blood pressure by ordinary users is not very high, and even if the user measures blood pressure, it is necessary to accurately measure blood pressure by white coat hypertension. In addition to being in a difficult situation, a user's blood pressure changes variously over a day. Therefore, a blood pressure measurement and estimation technique that considers individual characteristics and temporal changes is required more than conventional techniques.

1. US Patent Publication No. 5339818 (registered on August 23, 1994) 2. US Patent Publication No. 9767557 (registered on September 19, 2017)

Innovative continuous non-invasive cufless blood pressure monitoring based on photoplethysmography technology. Intensive Care Med, 39:1618-1625. Ruiz_rodriguez, et. al. (2013).

The technical problem to be solved by the present invention is to accurately estimate a continuous change in blood pressure at a future point in consideration of both temporal changes in parameters used for estimating blood pressure and unique characteristics of an object for measuring blood pressure. It is to provide a method and a device for that.

The method according to an embodiment of the present invention for solving the above technical problem includes receiving a user's biosignal measured in at least one or more methods in the past period, analyzing the received biosignal, and measuring the biosignal. A feature information extraction step of extracting feature information for each time point; A parameter calculating step of calculating a blood pressure related parameter for the past period based on the extracted feature information; And a blood pressure estimation control step of inputting the calculated blood pressure related parameters into a recurrent neural network (RNN) in a time series to output the estimated blood pressure of the user in a future period including the present time point.

An apparatus according to another embodiment of the present invention for solving the above technical problem, receives a user's biosignal measured in at least one or more methods in a past period, analyzes the received biosignal, and generates the biosignal A feature information extracting end for extracting feature information for each measured time point; A parameter calculator that calculates a blood pressure related parameter for the past period based on the extracted feature information; And a blood pressure estimating control unit that inputs the calculated blood pressure related parameters into a recurrent neural network (RNN) in time series to output the estimated blood pressure of the user in a future period including the present time point.

An embodiment of the present invention provides a computer-readable recording medium storing a program for implementing a method for estimating section blood pressure using a circulatory neural network.

According to the present invention, it is possible to accurately estimate blood pressure through biometric signals collected through a non-invasive method.

1 is a diagram illustrating an example in which a circulatory neural network operates.
2 is a diagram schematically showing an example of a process of estimating section blood pressure through an RNN according to the present invention.
3 is a diagram schematically illustrating a method for estimating section blood pressure according to the present invention.
4 is a block diagram showing an example of an apparatus for estimating section blood pressure according to the present invention.
5 is a diagram for describing a process in which blood pressure related parameters are normalized.
6 is a diagram schematically showing a process in which a blood pressure related parameter calculated from a biological signal of a 90 cardiac cycle cardiac cycle is applied to an RNN.
7 is a diagram for explaining blood pressure related parameters calculated from biological signals through ECG and PPG.
8 is a diagram for explaining a function performed by the class classification unit.
FIG. 9 is a diagram showing additional waveforms of blood pressure estimated by the class classification method in FIG. 8.
10A is an example of a diagram schematically showing the accuracy of an estimation method according to the present invention.
10B and 10C are diagrams showing diagrams corresponding to the systolic phase and the diastolic phase as a result of applying the present invention.
11 is another example of a diagram schematically showing the accuracy of an estimation method according to the present invention.
12 is another example of a diagram schematically showing the accuracy of an estimation method according to the present invention.
13 is a flowchart illustrating an example of a method for estimating section blood pressure according to the present invention.

Since the present invention can apply various transformations and have various embodiments, specific embodiments are illustrated in the drawings and will be described in detail in the detailed description. Effects and features of the present invention, and a method of achieving them will be apparent with reference to the embodiments described later in detail together with the drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various forms.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, and when describing with reference to the drawings, the same or corresponding components are assigned the same reference numerals, and redundant descriptions thereof will be omitted. .

In the following embodiments, terms such as first and second are used for the purpose of distinguishing one component from other components, not for a limiting meaning.

In the following embodiments, a singular expression includes a plurality of expressions unless the context clearly indicates otherwise.

In the following embodiments, the terms include or have means that the features or elements described in the specification are present, and do not preclude the possibility of adding one or more other features or components in advance.

When a certain embodiment can be implemented differently, a specific process order may be performed differently from the described order. For example, two processes described in succession may be performed substantially simultaneously, or may be performed in an order opposite to the described order.

2 is a diagram schematically showing an example of a process of estimating section blood pressure through an RNN according to the present invention.

In FIG. 2, a result value is output based on the RNN described in FIG. 1, and therefore, a description overlapping with that described in FIG. 1 will be omitted below.

In FIG. 2, as time series data, a user's heart rate (HR) from time t-4 to time t-1 is input to the embedding layer. Here, the time point t is regarded as the present time point, and the time point t-4 to time t-1 is regarded as the past time point. Since the RNN receives time series data and outputs time series data, if the input time series data is data for heart rate at time t-4 to time t-1, it should be able to output data on heart rate at time t and after time t. , As described above, in order to estimate a person's blood pressure from a non-invasive bio-signal, the changes over time and the intrinsic characteristics of the target person must all be considered. Preprocessing is required to be able to print.

Although there have been attempts to estimate blood pressure through RNN in the past, there is a problem in that the accuracy is very low because data input to the RNN is not processed to accurately estimate blood pressure. According to the present invention, it is possible to accurately estimate the blood pressure of the future point including the present point by effectively pre-processing the data input to the RNN. Hereinafter, the interval blood pressure is regarded as collectively referring to a series of blood pressure values over a certain period of time, not a specific fixed time point.

Measurement target Blood pressure function BP(t) of A 0.3*HR(t-4)+0.4*HR(t-3)+0.5*HR(t-2)+0.7*HR(t-1)+80 BP(t) 0.1*HR(t-4)+0.8*HR(t-3)+1.33*HR(t-2)+2.1*HR(t-1)+40

Table 1 shows an example of the blood pressure function at time t, which is the current time point of user A and user B.

As shown in Table 1, the blood pressure at time t, which is the current time point, may be defined in a manner in which different slopes are multiplied by a blood pressure function at time t-4 to time t-1, and an offset is added. As above, the parameters constituting the blood pressure function at time t have positive slopes and offsets, but their magnitudes have different characteristics. As shown in Table 1, the blood pressure at the present time has a characteristic that is finally determined by being influenced by the blood pressure at various past time points before the present time.

As described above, since Table 1 is only an example of the blood pressure function, the blood pressure function at time t may be configured as a function according to other biological signals in addition to the blood pressure function before the time t.

3 is a diagram schematically illustrating a method for estimating section blood pressure according to the present invention.

The method for estimating section blood pressure according to the present invention aims to accurately estimate the section blood pressure of a user after the current point of time from non-invasive and continuous biosignals.

First, the primary information 310 includes a non-invasive and continuous raw biosignal. As a biosignal that can be included in the primary information 310, electrocardiogram test ECG: electrocardiogram, seismic beat curve test (seismocardiogram), impedance cardiogram, photoplethysmogram (PPG), cardiac ballistic test ( There are biological signals measured by BCG: ballistocardiogram), plethysmogram, and ultrasound. Although the above-described test is a non-invasive test that does not accompany an incision, it has the characteristic of providing meaningful information in measuring and estimating the blood pressure of a user to be measured.

In addition, even if not included in the types of tests described above, the present invention is a non-invasive and continuous signal, as long as it is a test that can provide meaningful information for measuring and estimating the blood pressure of a user to be measured. It is obvious that it can be applied.

The secondary information 330 refers to information that can be derived by processing the primary information 310. Since the primary information 310 is information on the waveform of the bio-signal itself, the amount of information is excessively large and noise is quite mixed in estimating blood pressure, so the secondary information 330 accurately estimates the blood pressure. And, in order to minimize the amount of computation through a computer, it can be calculated by extracting only information that can be viewed as a characteristic from raw data. As information included in the secondary information 330, a pulse transmission time (PTT), a pulse arrival time (PAT), a large artery stiffness index (LASI), and the like may be included.

When the secondary information 330 is obtained from the primary information 310, the RNN receives the secondary information 330 as input data and finally outputs a blood pressure. At this time, the blood pressure displayed as the output value of the RNN is the blood pressure at the future time including the present time point, and the input data and output data become the section blood pressure due to the characteristics of the RNN, which is displayed as time series data

4 is a block diagram showing an example of an apparatus for estimating section blood pressure according to the present invention.

Referring to FIG. 4, the section blood pressure estimation apparatus 400 according to the present invention includes a feature information extracting unit 410, a parameter calculating unit 430, a class classifying unit 450, and a blood pressure estimation control unit 470. Able to know. Hereinafter, for convenience of explanation, it will be described with reference to FIGS. 2 and 3.

The feature information extracting unit 410 receives the user's bio-signals measured in at least one or more methods in the past period, analyzes the received bio-signals, and extracts feature information for each time point at which the bio-signals are measured. Here, the term “one or more” means an electrocardiogram (ECG), an optical volume pulse wave test (PPG), and the like described in the primary information 310 of FIG. 3. In addition, the feature information is information obtained from a bio-signal and means information necessary to estimate a user's blood pressure, and the secondary information 330 of FIG. 3 may be the feature information. The past period refers to a period in the past based on the present time t, like the time t-4 to t-1 described in FIG. 2.

In the present invention, the interval between the time points t-2 and t-1 may be one cardiac cycle. For example, if the user's heart goes through four systolic diastoles at time t-4, the time point is considered to be time t. Depending on the embodiment, unlike the above, the interval for each viewpoint may simply be in units of 1 second or 5 seconds.

As an optional embodiment, the feature information extractor 410 may extract feature information by applying an ensemble average method to a waveform of a bio-signal. Since the user's bio-signals include various slopes, as the gender and age of the population (to which multiple users belong) are diversified, the feature information tends to not be properly extracted from the bio-signals, so in the present invention, the ensemble average Through the method, the blood pressure characteristic of the user included in the biometric signal is extracted. Through the ensemble averaging method, information that meaningfully reflects the magnitude of the waveform of the bio-signal and the feature of the waveform can be extracted.

The feature information extracting unit 410 may additionally apply a support vector machine (SVM) to the feature information extracted through the above process to remove noise or detect a section in which noise is greater than or equal to a preset value. Here, the preset value is a predefined value of the support vector machine and may be changed to increase the accuracy of blood pressure estimation.

The ensemble averaging and support vector machine methods are known techniques that are already well known, and the specific application method of the technique is not described further in the present specification, and noise is generated like a denoising auto-encoder. It is obvious to those skilled in the art that the above-described process of the present invention can be implemented through programs or algorithms other than the ensemble averaging method and the support vector machine as long as it is a method capable of well extracting features from many bio signals. Do.

The parameter calculation unit 430 calculates a blood pressure-related parameter for each time point in the past period based on the feature information extracted by the feature information extraction unit 410. The parameter calculation unit 430 stores in advance a parameter calculation criterion for calculating a blood pressure related parameter.

As an optional embodiment of the blood pressure-related parameter calculated by the parameter calculating unit 430, the parameter calculating unit 430 defines and calculates a difference in characteristic information of the biosignal measured by ECG and PPG as a blood pressure-related parameter. can do.

As another optional embodiment, the parameter calculation unit 430 may define and calculate a difference in pulse wave delivery time (PTT) calculated from characteristic information of bio-signals measured by ECG and PPG as a blood pressure related parameter.

As another optional embodiment, the parameter calculation unit 430 may define and calculate a value obtained by multiplying the difference in pulse wave transmission time calculated from the characteristic information of the bio-signals measured by ECG and PPG by a random integer as a blood pressure related parameter. have. Here, an arbitrary integer may be 4.

As another optional embodiment from the above-described example, the parameter calculation unit 430 may calculate a blood pressure related parameter from eight types of characteristic information of the bio-signal measured by PPG. Since PPG is a result of the waveform of arterial blood pressure (ABP), the eight characteristic information of PPG is very similar to the characteristic information of the ABP waveform. In this optional embodiment, the blood pressure related parameters include heart rate (HR), LASI, AI, a1, a2, PPGarea, t1, and t2. LASI (large artery stiffness index) is a parameter representing blood vessel elasticity, like PAT, and refers to the temporal gap between the forward wave that advances from the point of time and the backward wave that is reflected from the end of the blood vessel. do. LASI will be described later in FIG. 7. In particular, in the present invention, HR and LASI can be used as both characteristic information and blood pressure-related parameters, and when used as blood pressure-related parameters, labeling or numbering to distinguish them from other blood pressure-related parameters can be applied. have.

For convenience of explanation, the remaining six blood pressure related parameters excluding HR and LASI will be described with reference to FIG. 7 to be described later.

The blood pressure related parameters calculated through the above-described method may be processed through a normalization process.

5 is a diagram for describing a process in which blood pressure related parameters are normalized.

As shown in FIG. 5, the blood pressure-related parameters for various time points in the past period can be projected onto a normalized distribution curve through a normalization process, and the parameters that have undergone normalization have a more efficient feature in estimating the normalized interval blood pressure. More specifically, by going through the normalization process, the subject-dependent slope/offset effect as described in Table 1 no longer has an adverse effect on estimating the user's blood pressure. .

As an optional embodiment, the temporal length of the biosignal for calculating the blood pressure related parameter may be a fixed value related to the user's heart rate. For example, the parameter calculation unit 430 may calculate a blood pressure related parameter based on the characteristic information of the biosignal for 90 cardiac cycles. A biosignal of an appropriate temporal length allows the RNN to accurately estimate blood pressure, and can significantly reduce the time it takes for the computing device to process the operation. In the case of calculating the blood pressure related parameters from too short a biological signal, the section blood pressure output due to insufficient information is inevitably inaccurate. In addition, when the blood pressure-related parameters are calculated from a biological signal of an excessively long length, information that is not necessary for estimating blood pressure is accumulated and applied, so that not only the incorrect section blood pressure is estimated, but also the amount of computation that the computing device has to process is exponential. There is a disadvantage that it takes a lot of time because it is increased.

According to the present invention, a time length of a biological signal required to calculate a blood pressure-related parameter is defined as 90 cardiac cycles, so that both a fast calculation speed and accuracy in estimating interval blood pressure can be expected. The cardiac cycle, which is added each time the heart beats, is a value related to the user's heart rate, and may be measured in advance as a time length unique to each user. In addition to the above, the temporal length of the biological signal for calculating the blood pressure-related parameter may be 60 cardiac cycles, 60 cardiac cycles or more and 90 cardiac cycles, or 150 cardiac cycles. If the cardiac cycle is lengthened, it is possible to estimate the blood pressure without deteriorating the accuracy even if the number of types of feature information is reduced more than that in the case of shortening the cardiac cycle. In addition to the above-described specific values, the length of the cardiac cycle may be changed within a reasonable range in order to accurately estimate the user's blood pressure.

6 is a diagram schematically showing a process in which a blood pressure related parameter calculated from a biological signal of a 90 cardiac cycle cardiac cycle is applied to an RNN.

In FIG. 6, the current time point is considered to be time t, and the estimated blood pressure 610 at time t may be output as a final result of the RNN. Referring to FIG. 6, in order to calculate the blood pressure-related parameters, as the 90 cardiac cycles are limited to the temporal length of the bio-signal, the blood pressure-related parameters for each time point from time t-90 to time t-1 are input to the RNN, In the hidden layer of the RNN, it is possible to estimate the blood pressure at the point t by repeating learning and testing, taking into account that the data at each point in time affect each other.

7 is a diagram for explaining blood pressure related parameters calculated from biological signals through ECG and PPG.

7 shows a waveform of a bio-signal according to the ECG at the top, and a waveform of the bio-signal according to the PPG at the bottom. When the feature information extraction unit 410 extracts a point (feature information) having a meaningful value from each biosignal, the parameter calculation unit 430 defines PTT or LASI in the feature information based on the previously stored parameter calculation criteria. And calculate. The length of the PTT and the length of the LASI may vary according to the detailed standard stored in the parameter calculation unit 430. Hereinafter, as a description of an optional embodiment of the parameter calculation unit 430 of FIG. 4, with reference to FIG. 7, what a1, a2, a3, AI, PPGarea, t1, and t2 mean will be described. .

First, t1 and t2 are parameters for determining how fast the pulse wave of blood pressure rises, and according to experimental findings, it is determined that they are parameters related to blood pressure, and thus are blood pressure related parameters adopted in the present invention. Referring to FIG. 7, it can be seen that t1 is the time taken to reach the position corresponding to 40% of the highest point from the lowest point of the pulse wave of the same period, and t2 is the time required to reach the highest point from the lowest point of the pulse wave of the same period. It can be seen that it is the time taken by subtracting t1 from the time it takes.

In addition, a1 and a2 are factors related to pulse pressure, which is an element of blood pressure, and according to experimental findings, the values of a1 and a2 tend to have a high correlation with the magnitude of the pulse pressure. Referring to FIG. 7, a1 may be defined as the difference between the lowest point of the pulse wave and the pulse wave when time elapses by t1, and a2 may be defined as the difference between the highest point and the lowest point of the pulse wave.

a3 is a blood pressure-related parameter that can be additionally calculated by a1 and a2, and in the present invention, the parameter calculating unit 430 is not a blood pressure-related parameter that is processed and calculated as an input value of an RNN, but it calculates a blood pressure-related parameter called AI. This is an auxiliary parameter required for this. Referring to FIG. 7, it can be seen that a3 denotes a pulse wave value when a time corresponding to LASI elapses from t2.

AI (augmentation index) is a blood pressure-related parameter that is directly related to LASI, and refers to the ratio of the magnitudes of forward and reflected pulse waves. More specifically, in AI, because the size of the reflected pulse wave varies depending on how much amplitude damping of the pulse wave occurs due to the elasticity of the blood vessel, AI can also be seen as a parameter related to blood vessel elasticity. Is the value obtained by dividing a3 by a2.

Finally, PPGarea (photoplethysmogram area) refers to the integral value (width) of the graph of the pulse wave integrated for one period when one cycle elapses as the pulse wave reaches the highest point from the lowest point and then returns to the lowest point again. Like a1 and a2, as a parameter related to the pulse pressure, it may be calculated by the parameter calculation unit 430 of the present invention.

In FIG. 7, PTT1, PTT2, and PTT3 used to calculate t1 and t2 define the pulse wave transmission time in various ways, and according to the embodiment, it will be replaced by the pulse arrival time instead of the pulse wave delivery time. May be. The reason that the pulse wave arrival time can replace the pulse wave arrival time is that the pulse wave transmission time (PTT) can be calculated by dividing the blood vessel length by the pulse wave transmission rate, and the pulse wave arrival time (PAT) is the pulse wave transmission time (PTT) and It is defined as the sum of the pre-ejection time, and since the pre-ejection time is a parameter that is not related to the vascular elasticity, the pulse wave transmission time and the pulse wave arrival time have the same tendency that is related to blood pressure and are factors related to blood pressure Because it is self-evident.

Hereinafter, FIG. 4 will be described again.

The class classification unit 450 normalizes the blood pressures at each time point in the past period to calculate a normalized blood pressure, and calculates the normalized blood pressure at least two or more according to the normalized blood pressure value. Classify by class.

According to the present invention, the class classification unit 450 classifies normalized blood pressures at a past point into at least two or more classes, so that the correlation between the section blood pressure estimated according to the present invention and the actual blood pressure is clearly revealed. In addition, the difference between the estimated section blood pressure and the actual blood pressure at the future point can be minimized.

In the estimating apparatus 400 according to the present invention, the class classification unit 450 may be omitted according to embodiments. For convenience of explanation, a description of the classes classified by the class classification unit 450 is the blood pressure estimation control unit 470 ) Will be further explained in the description.

The blood pressure estimation control unit 470 inputs the blood pressure related parameters of the past period calculated by the parameter calculation unit 430 in a time series to the circulatory neural network (RNN) so that the estimated blood pressure of the user in the future period including the present time is output. Control. More specifically, the blood pressure estimation control unit 470 comprehensively controls a machine learning process that repeats training and tests so that appropriate interval blood pressure can be estimated by using a blood pressure related parameter as an input value.

8 is a diagram for explaining a function performed by the class classification unit.

Referring to the waveform 810 of the user's actual blood pressure in FIG. 8, the user's actual blood pressure dynamically changes with the passage of time. On the other hand, referring to FIG. 8, when estimating the user's blood pressure according to a conventional blood pressure estimation technique, it can be seen that the blood pressure 830 estimated by the regression method has a constant constant value regardless of the passage of time. In general, blood pressure maintains a large number of averages and changes around the average. This is a result of a tendency to misunderstand that the cost function is optimized when the neural network estimates the average when training a neural network.

According to the method of estimating the user's blood pressure according to the conventional blood pressure estimation technique, after inputting the data of the past period into the RNN, the blood pressure estimation control unit 470 calculates the estimated value and Repeat until the MSE (Mean Square Error) with the actual value is minimized. It will be apparent to those skilled in the art that, unless regression is used in the conventional blood pressure estimation technique, various cost functions other than the above-described MSE can be used.

Equation 1 represents an equation of the MSE used in a conventional machine learning process.

은 추정 혈압값을 의미한다.In Equation 1, y denotes an actual blood pressure value,

Means an estimated blood pressure value.

When the blood pressure estimation control unit 470 repeats the process of minimizing the MSE according to the mathematical characteristic of Equation 1 to estimate the section blood pressure, the estimated blood pressure is generally expressed as a constant constant value regardless of the passage of time. In other words, since the actual blood pressure is a time-varying function, but the estimated blood pressure is expressed as a time-invariant function, there is a limit that the correlation between the estimated blood pressure and the actual blood pressure at each time point cannot be revealed in a graph at all.

FIG. 9 is a diagram showing additional waveforms of blood pressure estimated by the class classification method in FIG. 8.

Referring to the waveform 850 of the blood pressure estimated by the class classification method in FIG. 9, it can be seen that the blood pressure of the user is very similar to the actual blood pressure of the user and changes dynamically as time passes. As shown in FIG. 9, a method for the blood pressure estimation control unit 470 to estimate section blood pressure using a class classification method is as follows.

First, the class classification unit 450 normalizes the blood pressure values for each time point in the past period to calculate a normalized blood pressure, and calculates the normalized blood pressure at least according to the normalized blood pressure value. Classify into two or more classes.

For example, the class classification unit 450 normalizes and projects the blood pressure values for each time point in the past period on the normal distribution curve as shown in FIG. 5, and then, the blood pressure values of different sections in class 1, class 2, and class 3, respectively. Can be classified as belonging. Here, the number of classes is at least two or more.

Class number Location of x-values on the normal distribution curve Class 1 Normalized blood pressure <-1.3 Class 2 -1.3 <normalized blood pressure <-0.85 Class 3 -0.85 <normalized blood pressure <-0.52

Table 2 is a table showing an example in which classes are classified.

In Table 2, the number of classes is limited to three, but when the present invention is actually implemented, the number of classes may be more than three.

The blood pressure estimating control unit 470 repeats until the cross entropy calculated based on the softmax of the class reaches the minimum value, and controls the blood pressure at the point where the minimum value becomes the minimum value to be output as the estimated blood pressure of the user. .

Equation 2 is an equation representing softmax. In Equation 2, x is a feature of the target value for obtaining a softmax value, i is the order of features (i-th), and j is the total number of features.

In order for the blood pressure estimation control unit 470 to repeat the learning and then stop and output the result value, it is necessary to know how accurate the estimated value is. In addition, even if the value estimated by the blood pressure estimation control unit 470 is accurate due to the nature of machine learning, an estimated value for whether the probability of finally outputting the estimated value is 50% or 70% is required. 470) is an expected value to output a specific value through Equation 2 according to the definition of softmax, and controls to output a value between 0 and 1.

Equation 3 is an equation representing cross entropy. In Equation 3, H denotes cross entropy, L denotes an actual classification value, S denotes a result value of Softmax, and i denotes a class number.

The blood pressure estimating control unit 470 sets the cross entropy as a classification loss function and repeats a process of minimizing the cross entropy, thereby controlling an estimated section blood pressure similar to the actual blood pressure to be output.

The calculation method for softmax and cross entropy is a widely known method, and will be omitted in this specification.

As described above, according to the present invention, by repeating the process of training and testing the RNN in a format in which blood pressure is divided into step levels and classified, the interval blood pressure that changes over time is estimated, unlike the results according to the conventional regression estimation. And the estimated interval blood pressure shows a pattern very similar to the actual blood pressure. In addition, the estimated section blood pressure changes with time, and thus, it is distinguished from the conventional technology in that a correlation between the estimated section blood pressure and the actual blood pressure can be numerically confirmed.

10A is an example of a diagram schematically showing the accuracy of an estimation method according to the present invention.

FIG. 10A is a diagram showing a result of estimating blood pressure using bio-signals (electrocardiogram, optical volume pulse wave) of critically ill patients in MIMIC of a physionet, one of the bio-signal databases. More specifically, FIG. 10A shows the result of estimating systolic blood pressure, and the difference between the estimated systolic blood pressure and the actual blood pressure is a widely known standard, the Association for The Advancement of Medical Instrumentation (AAMI) standard and BHS. (British Hypertension Society).

10A shows the class and estimation of the actual blood pressure (True BP) when the normalized blood pressure is calculated according to the present invention and the calculated normalized blood pressure is classified into at least two or more classes according to the value of the normalized blood pressure. Blood pressure (Predicted BP) is displayed simultaneously. Each numerical value divided into a total of 100 boxes in FIG. 10A indicates a ratio of estimated classes.

For example, in the box corresponding to the actual blood pressure 1 and the estimated blood pressure 1 of FIG. 10A, 0.49 is included, and the value of 0.49 is when there are 100 samples with an actual blood pressure of class 1, by applying the present invention. This means that 49 were estimated as Class 1. Therefore, 1 can always be calculated by summing all the ratios on the horizontal.

As another example, the box value of actual blood pressure 0 and estimated blood pressure 9 in FIG. 10A is 0.021, and if there are a total of 1000 actual blood pressure samples classified as class 0, 21 of them are estimated as class 9 when applied to the present invention. Means that you did. According to the characteristics of the present invention described above, when the number of samples is adjusted to an appropriate number, and the blood pressure-related biosignals collected for an appropriate time become the first input data, a MIMIC diagram having a higher estimation degree than that of FIG. 10A can be calculated. have.

10B and 10C are diagrams showing MIMIC diagrams corresponding to systolic and diastolic phases as a result of applying the present invention.

10B is an example of the result of estimating the diastolic blood pressure, and FIG. 10C shows an example of the result of estimating the systolic blood pressure as shown in FIG. 10A, see FIGS. 10B and 10C In this case, it can be seen that a high estimation rate is shown for class 0 and class 9, and when the values of adjacent classes are summed up in the remaining classes, a high estimation rate similar to that of class 0 and class 9 is shown. When the values of adjacent classes are summed up, the fact that all classes show similar estimation rates can be interpreted as meaning that the actual blood pressure value is well estimated within an allowable error range. As a result, according to FIGS. 10A to 10C, the present invention In estimating blood pressure, it can be seen that the user's blood pressure can be accurately estimated with a high probability without being limited to a specific class.

11 is another example of a diagram schematically showing the accuracy of an estimation method according to the present invention.

FIG. 11 schematically shows that there is little difference between the actual blood pressure and the estimated interval blood pressure indicating a large variance due to a large amount of blood pressure change over time. In FIG. 11, the horizontal axis (x-axis) denotes the cardiac cycle, and the vertical axis (y-axis) denotes the blood pressure value.

12 is another example of a diagram schematically showing the accuracy of an estimation method according to the present invention.

FIG. 12 schematically shows that there is almost no difference between the actual blood pressure and the estimated interval blood pressure, which is highly undulating over time, as in FIG. 11.

11 and 12, it can be seen that the method for estimating section blood pressure according to the present invention is well applied regardless of the width or rate of change in actual blood pressure.

13 is a flowchart illustrating an example of a method for estimating section blood pressure according to the present invention.

Since FIG. 13 may be implemented by the section blood pressure estimation apparatus 400 according to FIG. 4, it will be described with reference to FIG. 4, and descriptions overlapped with those described in FIG. 4 will be omitted below.

The feature information extracting unit 410 receives the user's bio-signals measured in at least one or more methods in the past period, analyzes the received bio-signals, and extracts feature information for each time point at which the bio-signals are measured (S1310).

The parameter calculation unit 430 calculates a blood pressure related parameter for the past period based on the feature information extracted in step S1310 (S1330).

The class classification unit 450 normalizes the blood pressures at each time point in the past period to calculate a normalized blood pressure, and calculates the normalized blood pressure at least two or more according to the normalized blood pressure value. Classified into a class (S1350).

The blood pressure estimation control unit 470 inputs the blood pressure related parameters of the past period calculated by the parameter calculation unit 430 in a time series to the circulatory neural network (RNN) so that the estimated blood pressure of the user in the future period including the present time is output. Control. More specifically, the blood pressure estimation control unit 470 collectively controls the machine learning process of repeating training and testing so that appropriate section blood pressure can be estimated by using a blood pressure related parameter as an input value (S1370). ).

As described above, according to the present invention, in order to estimate blood pressure, by preprocessing data input to RNN through a series of processes, it is possible to accurately estimate the blood pressure at a future point from various biological signals of the user. It is possible. More specifically, in the present invention, the feature information extracted from the bio-signal is normalized, and the blood pressure-related parameters are calculated from the normalized feature information according to a preset criterion, and a classification process based on a class rather than a regression is used. By repeatedly performing RNN learning, it is possible to accurately estimate blood pressure compared to any conventionally known biosignal-based blood pressure estimation technique.

The embodiment according to the present invention described above may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded in a computer-readable medium. In this case, the medium is a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical recording medium such as a CD-ROM and a DVD, a magnetic-optical medium such as a floptical disk, and a ROM. A hardware device specially configured to store and execute program instructions, such as, RAM, flash memory, and the like.

Meanwhile, the computer program may be specially designed and configured for the present invention, or may be known and usable to those skilled in the computer software field. Examples of the computer program may include not only machine language codes produced by a compiler but also high-level language codes that can be executed by a computer using an interpreter or the like.

The specific implementations described in the present invention are examples, and do not limit the scope of the present invention in any way. For brevity of the specification, descriptions of conventional electronic configurations, control systems, software, and other functional aspects of the systems may be omitted. In addition, the connection or connection members of the lines between the components shown in the drawings exemplarily represent functional connections and/or physical or circuit connections, and in an actual device, various functional connections that can be replaced or additionally It may be referred to as a connection, or circuit connections. In addition, if there is no specific mention such as “essential” or “importantly”, it may not be an essential component for the application of the present invention.

In the specification of the present invention (especially in the claims), the use of the term “above” and a similar reference term may correspond to both the singular and the plural. In addition, when a range is described in the present invention, the invention to which an individual value falling within the range is applied (unless otherwise stated), and each individual value constituting the range is described in the detailed description of the invention. Same as Finally, if there is no explicit order or contrary to the steps constituting the method according to the present invention, the steps may be performed in a suitable order. The present invention is not necessarily limited according to the order of description of the steps. In the present invention, the use of all examples or illustrative terms (for example, etc.) is merely for describing the present invention in detail, and the scope of the present invention is limited by the above examples or illustrative terms, unless limited by the claims. It does not become. In addition, those skilled in the art can recognize that various modifications, combinations, and changes may be configured according to design conditions and factors within the scope of the appended claims or their equivalents.

Claims (16)

A feature information extraction step of receiving a biosignal of a user measured in at least one or more ways in a past period, analyzing the received biosignal, and extracting feature information for each time point at which the biosignal was measured;
A parameter calculating step of calculating a blood pressure related parameter for the past period based on the extracted feature information; And
A blood pressure estimation control step of inputting the calculated blood pressure related parameters into a recurrent neural network (RNN) in time series to output the estimated blood pressure of the user in a future period including the present time point,
The biosignal is
It is the user's biosignal measured through the optical volumetric pulse wave test (PPG),
The blood pressure related parameter calculated from the characteristic information of the bio-signal according to the optical volume pulse wave test,
1) the user's heart rate,
2) the first time it takes to reach a position corresponding to 40% of the highest point from the lowest point of the pulse wave of the user's blood pressure,
3) a second time obtained by subtracting the first time from the time required to reach the highest point from the lowest point of the pulse wave of the user's blood pressure,
4) a first gap, which is the difference between the lowest point of the user's pulse wave and the pulse wave value at the point of time elapsed by a first time from the lowest point of the pulse wave,
5) the second gap, which is the gap between the highest point and the lowest point of the pulse wave,
6) LASI (large artery stiffness index)
7) AI (augmentation index) obtained by dividing the pulse wave value at the time point at which the LASI elapses in the second time by the second gap, and,
8) A method for estimating section blood pressure using a circulatory neural network, characterized in that it includes a PPGarea (photoplethysmogram area). The method of claim 1,
The feature information extraction step,
A method for estimating section blood pressure using a circulatory neural network, characterized in that the characteristic information is extracted by applying an ensemble average to the biological signal. The method of claim 1,
The feature information extraction step,
A method for estimating section blood pressure using a circulatory neural network, characterized in that noise is removed from the extracted feature information through a support vector machine (SVM) or a section in which noise is greater than or equal to a preset value is detected. The method of claim 1,
The biosignal is
In addition to the biosignal measured through the optical volumetric pulse wave test, it is measured by at least one of an electrocardiogram (ECG), an earthquake beat curve test, an impedance test, a cardiac ballistic test (BCG), a branched volume wave test, and an ultrasound test. Sectional blood pressure estimation method using a circulatory neural network, characterized in that it further comprises a bio-signal. delete The method of claim 1,
The biosignal is
In addition to the biosignal measured through the optical volumetric pulse wave test, it further includes a biosignal of the user measured through an electrocardiogram (ECG),
The parameter calculation step,
The blood pressure related to the difference between the time points of the pulse transit time (PTT) calculated from the characteristic information of the biological signal according to the electrocardiogram test and the pulse transit time calculated from the characteristic information of the biological signal according to the optical volumetric pulse wave test. Sectional blood pressure estimation method using a circulatory neural network, characterized in that calculated as a parameter. delete The method of claim 1,
The above method,
A class classification step of calculating normalized blood pressure by normalizing blood pressures at each time point in the past period, and classifying the normalized blood pressures into at least two or more classes according to the value of the normalized blood pressure. and,
The blood pressure estimation control step,
Sectional blood pressure estimation method using a circulatory neural network, characterized in that repeating until a cross entropy calculated based on the softmax of the class reaches a minimum value, and controlling the estimated blood pressure of the user to be output The method of claim 1,
The biosignal is
A method for estimating section blood pressure using a circulatory neural network, characterized in that it is a bio-signal for 90 cardiac cycles. The method of claim 1,
The biosignal is
A method for estimating section blood pressure using a circulatory neural network, characterized in that it is a biological signal for a cardiac cycle of 60 or more and 90 or less. The method of claim 1,
The biosignal is
A method for estimating section blood pressure using a circulatory neural network, characterized in that it is a biological signal for 150 cardiac cycles. The method of claim 1,
The biosignal is
Sectional blood pressure estimation method using a circulatory neural network, characterized in that it is a biological signal measured for a time period of 60 seconds or more and 90 seconds or less. A computer-readable recording medium storing a program for implementing the method according to any one of claims 1 to 4, 6 and 8 to 12. A feature information extracting end for receiving a user's bio-signals measured in at least one or more methods in the past period, analyzing the received bio-signals, and extracting feature information for each time point at which the bio-signals are measured;
A parameter calculator that calculates a blood pressure related parameter for the past period based on the extracted feature information; And
A blood pressure estimation controller configured to input the calculated blood pressure related parameters into a recurrent neural network (RNN) in a time series to output the estimated blood pressure of the user in a future period including a present time point,
The biosignal is
It is the user's biosignal measured through the optical volumetric pulse wave test (PPG),
The blood pressure related parameter calculated from the characteristic information of the bio-signal according to the optical volume pulse wave test,
1) the user's heart rate,
2) the first time it takes to reach a position corresponding to 40% of the highest point from the lowest point of the pulse wave of the user's blood pressure,
3) a second time obtained by subtracting the first time from the time required to reach the highest point from the lowest point of the pulse wave of the user's blood pressure,
4) a first gap, which is a difference between the lowest point of the user's pulse wave and the pulse wave value at the point of time elapsed by a first time from the lowest point of the pulse wave,
5) the second gap, which is the gap between the highest point and the lowest point of the pulse wave,
6) LASI (large artery stiffness index)
7) AI (augmentation index) obtained by dividing the pulse wave value at the time point at which the LASI elapses in the second time by the second gap, and,
8) A device for estimating section blood pressure using a circulatory neural network, characterized in that it includes a PPGarea (photoplethysmogram area). The method of claim 14,
The device,
Further comprising a class classification unit for calculating a normalized blood pressure by normalizing blood pressures for each time point in the past period, and classifying the normalized blood pressures into at least two or more classes according to the normalized blood pressure value, ,
The blood pressure estimation control unit,
Sectional blood pressure estimation device using a circulatory neural network, characterized in that it repeats until a cross entropy calculated based on a softmax of the class reaches a minimum value, and controls the estimated blood pressure of the user to be output. . The method of claim 14,
The biosignal is
A device for estimating section blood pressure using a circulatory neural network, characterized in that it is a bio-signal for 90 cardiac cycles.

Images