KR20190126539A - System and method for estimating oscillometric blood pressure based on deep learning - Google Patents

Abstract

A method for measuring blood pressure using a two-stage adaptive ensemble deep neural network and a system thereof are disclosed. The method for measuring blood pressure using an ensemble deep neural network comprises the steps of: extracting feature data from a blood pressure related oscillometric signal measured by a blood pressure measuring device worn by a test subject; setting an input parameter of a deep neural network (DNN) model based on the extracted feature data; setting stethoscope blood pressure data measured when the blood pressure of the test subject is measured by the blood pressure measuring device as target data; and learning a non-linear relationship between a blood pressure vibration signal measured by the blood pressure measuring device and the stethoscope blood pressure data measured by a stethoscope based on the set input parameter, the deep neural network model, and the target data.

Description

Blood pressure measurement method and system using two-stage adaptive ensemble deepening neural network {SYSTEM AND METHOD FOR ESTIMATING OSCILLOMETRIC BLOOD PRESSURE BASED ON DEEP LEARNING}

Embodiments of the present invention relate to a technique for measuring blood pressure indicative of a human health state using a learning model.

Blood pressure is used as one of the most important indicators of a person's health, and there is an auscultatory method in which a trained doctor or nurse measures the blood pressure using a stethoscope. In the case of a blood pressure measurement method using a stethoscope, the oscillometric-based blood pressure measurement technology is recently used for easy use at home, work, and hospital due to the inconvenience of measuring by a professional.

Oscillometric-based blood pressure measurement technology uses a pressure cuff to measure the blood pressure vibration signal at the top of the wrist or forearm to measure blood pressure using a predetermined systolic and diastolic ratio. For example, oscillometric-based blood pressure measurement technology measures blood pressure based on the ratio of systolic blood pressure (SBP) and diastolic blood pressure (DBP) to maximum amplitude algorithm (MAA). . As such, the pre-defined contraction and diastolic ratios become less accurate as they measure continuously changing blood pressure, such as blood pressure, at a pre-defined empirical ratio based on the proportion of blood pressure testers prior to algorithm development. Is present. In other words, the oscillometric-based blood pressure measurement is relatively inaccurate compared to the blood pressure measurement using a stethoscope.

Accordingly, there is a need for a technique for measuring blood pressure based on an oscillometric signal and improving measurement accuracy without using a ratio defined in the assessment.

Korean Patent Registration No. 10-1778533 relates to a method for estimating oscillometric blood pressure based on machine learning, which obtains an envelope of an oscillometric waveform for a measurement based on a sample blood pressure measurement, and uses an nonparametric bootstrap to oscillate the measurement. Obtain a pseudomaximum amplitude based on the envelope's maximum amplitude of the metric waveform, use a nonparametric bootstrap to obtain a pseudo envelope based on the envelope of the oscillometric waveform for the measurement, associate the pseudo maximum amplitude with the pseudo envelope, Techniques for estimating systolic and diastolic blood pressure ratios for individual subjects using machine learning are described.

[1] S. Lee, and J.-H. Chang, Oscillometric Blood Pressure Estimation Based on Deep Learning, IEEE Transactions on Industrial Informatics, vol. 13, no. 2, pp. 461-472, Apr. 2017. [2] S. Lee, and J.-H. Chang, Deep Belief Networks Ensemble for Blood Pressure Estimation, IEEE ACCESS, vol. 5, pp. 9962-9972, Dec. 2017.

One embodiment of the present invention is to estimate the blood pressure of the subject based on an oscillometric signal using a deep neural network (DNN) without pre-defining the systolic and diastolic ratios.

In addition, by applying the stethoscope blood pressure data measured using an oscillometric signal and a stethoscope to the DNN to increase the accuracy of blood pressure measurement relative to when measuring the blood pressure based on the oscillometric signal.

A blood pressure measurement method using an ensemble deep neural network, comprising extracting feature data from a blood pressure related oscillometric signal measured by a blood pressure measuring device worn by a test subject, and using a deep neural network based on the extracted feature data. ) Setting input parameters of the model, setting the stethoscope blood pressure data measured when the blood pressure of the test subject is measured by the blood pressure measuring device as the target data, and setting the input parameters, the deep neural network model, and the target data. And learning a non-linear relationship between the blood pressure vibration signal measured by the blood pressure measuring device and the stethoscope blood pressure data measured based on a stethoscope.

According to an aspect, the extracting of the feature data may include generating an envelope signal connecting each peak of the signal to the oscillometric signal, and the oscillometric signal and the envelope. And extracting feature data from the signal.

According to another aspect, the feature data includes at least one of mean arterial pressure (MAP), maximum amplitude (MA), and area of the envelope signal (AE). and,

The target data may include reference systolic blood pressure (RSBP) and reference diastolic blood pressure (RDBP).

According to another aspect of the present invention, the setting of the input parameter may include: replicating the extracted feature data into a plurality of pieces based on a bootstrap technique, a bootstrap sample and the feature data generated by the duplication; Generating virtual feature data comprising a, and setting the virtual feature data as the input parameter.

According to another aspect, in the learning step, the first learning may be performed on the input data, based on a Restricted Boltzmann Machine (RBM) probability model.

According to another aspect, in the learning step, the weight for the second learning may be set based on the result of performing the first learning.

According to another aspect, the bootstrap sample may be updated based on the virtual feature data on which secondary learning is performed based on the set weight.

A blood pressure measuring system using an ensemble deepening neural network, characterized by extracting feature data from blood pressure-related oscillometric signals measured by a blood pressure measuring apparatus worn by a test subject, and deepening based on the extracted feature data. A setting unit for setting an input parameter of a neural network (DNN) model and setting the stethoscope blood pressure data measured when the blood pressure of the test subject is measured by the blood pressure measuring device as the target data, and the set input parameter, the deep neural network model, The apparatus may include a learning controller configured to learn a non-linear relationship between the blood pressure vibration signal measured by the blood pressure measuring apparatus and the stethoscope blood pressure data measured based on a stethoscope based on target data.

According to one aspect, the feature data extracting unit, for the oscillometric signal, generates an envelope signal (envelope) connecting each peak of the signal, and the feature data from the oscillometric signal and the envelope signal Can be extracted.

According to another aspect, the feature data includes at least one of mean arterial pressure (MAP), maximum amplitude (MA), and area of the envelope signal (AE). The target data may include RSBP (reference systolic blood pressure) and RDBP (reference diastolic blood pressure).

According to another aspect, the setting unit, the plurality of extracted feature data based on a bootstrap technique, a virtual feature data including a bootstrap sample and the feature data generated by the replication; And generate the virtual feature data as the input parameter.

According to another aspect, the learning controller may perform first-order learning on the input data based on a Restricted Boltzmann Machine (RBM) probability model.

According to another aspect, the learning controller may set a weight for secondary learning based on a result of performing the primary learning.

According to another aspect, the bootstrap sample may be updated based on the virtual feature data on which secondary learning is performed based on the set weight.

According to embodiments of the present invention, an embodiment of the present invention estimates blood pressure of a subject based on an oscillometric signal using a deep neural network (DNN), without prior definition of systolic and diastolic ratios. Blood pressure can be measured.

In addition, by applying the stethoscope blood pressure data measured using the oscillometric signal and the stethoscope to the DNN, it is possible to increase the accuracy than when measuring the blood pressure based on the oscillometric signal. In other words, the measurement of the blood pressure based on the oscillometric signal can be improved compared to that measured with a stethoscope.

1 illustrates waveforms of an oscillometric signal and an envelope signal according to an embodiment of the present invention.
2 is a flowchart illustrating a blood pressure measurement method using a two-stage ensemble deep neural network according to an embodiment of the present invention.
3 is a block diagram illustrating an internal configuration of a blood pressure measuring system according to an exemplary embodiment of the present invention.
4 is a graph illustrating a cumulative function distribution (CDF) of virtual feature data according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating a structure of a two-stage ensemble deep neural network (DNN) according to one embodiment of the present invention.
FIG. 6 is a diagram illustrating performance of target data and a two-stage adaptive ensemble deep neural network model according to an embodiment of the present invention.
FIG. 7 is a graph showing a standard error error (SDE) according to an increase in a virtual blood pressure sample according to one embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention relates to a technique for measuring blood pressure using a two-stage adaptive ensemble deepening neural network (DNN), wherein an expert such as a nurse or a doctor uses a stethoscope at the same time as an oscillometric signal, which is a vibration signal measuring blood pressure in a blood pressure measuring device. The present invention relates to a technique for estimating blood pressure by applying auscultation blood pressure data, which is a measurement of blood pressure, to a deep neural network. In particular, the present invention relates to a technique for learning a complex nonlinear relationship between a blood pressure vibration signal (ie, an oscillometric signal) measured by a blood pressure monitor and auscultation blood pressure data using a deep neural network model based on bootstrap technology. In other words, the insufficient amount of input feature data solves the weaknesses based on the deep neural network technology, and there are still many problems of random initialization allocation, such as training data sets, weights, and deflections, which may cause unstable estimation in the pre-training stage. The present invention relates to a technique to be solved through an adaptive ensemble deep neural network using bootstrap technology.

In the present embodiments, the 'blood pressure measuring device' is a device for measuring the blood pressure of the test subject by being worn on the wrist, the forearm of the test subject, for example, manufactured in the form of a wearable watch or a wearable smart pressure sensor. The blood pressure and heart rate signal may be extracted based on the blood pressure of the subject.

The full name of the term used in the two-stage ensemble deep neural network-based learning algorithm presented in the present embodiments may be as shown in Table 1 below.

1 illustrates waveforms of an oscillometric signal and an envelope signal according to an embodiment of the present invention.

Referring to 110 of FIG. 1, as a blood pressure of a test subject is measured, a cuff pressure received from a blood pressure monitor may be indicated.

In FIG. 1, an envelope signal 121 may be generated by connecting respective peaks of an oscillometric wave signal. Here, the oscillometric signal may be obtained by filtering and detrending the cuff pressure 110.

FIG. 2 is a flowchart illustrating a blood pressure measuring method using a two-stage ensemble deep neural network in an embodiment of the present invention, and FIG. 3 is a block diagram illustrating an internal configuration of a blood pressure measuring system according to an embodiment of the present invention. It is also.

The blood pressure measuring system 300 of FIG. 3 may include a feature data extracting unit 310, a setting unit 320, and a learning control unit 330. The blood pressure measuring system 300 may be a blood pressure measuring device worn by a test subject. The blood pressure vibration signal measured from (not shown) can be received through a wired or wireless network. For example, the blood pressure vibration signal may be received through short range wireless communication such as Bluetooth or Zigbee. Each of steps 210 to 240 of FIG. 2 is performed by the feature data extracting unit 310, the setting unit 320, and the learning control unit 330, which are components of the blood pressure measuring system 300 of FIG. 3. Can be performed.

2 and 3 illustrate a case in which a wearable watch of a wearable watch type is used as the blood pressure measuring device, but this is an example, which corresponds to an embodiment, and is mounted on the forearm of the test subject in addition to the wearable watch type. A device including a wearable smart pressure sensor may be used to measure.

In operation 210, the feature data extractor 310 may extract feature data from a blood pressure-related oscillometric signal measured by a wearable device worn by a test subject.

For example, the feature data extractor 310 may receive a blood pressure vibration signal of the test subject, which is being measured for a predetermined time (eg, 1 minute, 5 minutes, etc.), from the wearable device. Here, the blood pressure vibration signal may include a cuff pressure. Then, the feature data extractor 310 may generate an oscillometric wave signal by performing signal processing such as filtering and trending on the blood pressure vibration signal.

In operation 211, the feature data extractor 310 may generate an envelope signal by connecting each peak of the signal to the oscillometric signal.

In operation 212, the feature data extractor 310 may extract feature data for blood pressure estimation from an oscillometric signal and an envelope signal.

For example, the feature data extractor 310 may calculate an estimated mean arterial pressure (MAP), a maximum amplitude (MA), an area under the envelope signal (AE), or the like. It can extract as feature data. In addition to these three types of feature data, the feature data extracting unit 310 is described in the non-patent literature [1] S. Lee, and J.-H. Chang, Oscillometric Blood Pressure Estimation Based on Deep Learning, IEEE Transactions on Industrial Informatics, vol . 13, no. 2, pp. 461-472, Apr. 2017. and [2] S. Lee, and J.-H. Chang, Deep Belief Networks Ensemble for Blood Pressure Estimation, IEEE ACCESS, vol. 5, pp. 9962-9972, Dec. Based on the feature data extraction method presented in 2017. 11 kinds of feature data can be extracted from the oscillometric signal and the envelope signal.

In operation 220, the setting unit 320 may set an input parameter of the deep neural network (DNN) model based on the extracted feature data. In this case, a large amount of data is required for deep neural network learning, and the data obtained through deep blood pressure measurement may not be sufficient for deep neural network learning. Accordingly, the setting unit 320 may additionally generate data enough to perform deep neural network learning by using a bootstrap technique.

In operation 221, the setting unit 320 may generate a bootstrap sample by replicating the extracted feature data into a plurality of pieces based on the bootstrap technique.

In operation 222, the setting unit 320 may generate virtual feature data including the generated bootstrap sample and feature data.

For example, the blood pressure data can obtain auscultation blood pressure data used as the 425 blood pressure measurement signal and label 5 times from actual 85 test subjects on an international protocol basis, and the non-patent literature [1] S.Lee, and J .-H. Chang, Oscillometric Blood Pressure Estimation Based on Deep Learning, IEEE Transactions on Industrial Informatics, vol. 13, no. 2, pp. 461-472, Apr. 2017. and [2] S. Lee, and J.-H. Chang, Deep Belief Networks Ensemble for Blood Pressure Estimation, IEEE ACCESS, vol. 5, pp. 9962-9972, Dec. Based on the virtual feature data generation method presented in 2017 , parameter-based bootstrap using statistical feature values (ie mean and standard deviation) of selected feature data as a large amount of data is needed to train the deep neural network model The technique can be used to generate 8500 virtual feature data, 1000 times for 85 people. For example, as shown in FIG. 4, the cumulative distribution (CDF) of the virtual feature data (ie, the virtual feature vector) has a pattern similar to the Gaussian distribution.

In operation 223, the setting unit 320 may set the virtual feature data as an input parameter of the deep neural network model.

In operation 230, the setting unit 320 may set the stethoscope blood pressure data as the target data.

For example, when measuring the blood pressure of the subject in the blood pressure measuring device, at the same time or immediately measured by the blood pressure measuring device, the blood pressure may be measured for a predetermined time period using a stethoscope by a specialist such as a nurse or a doctor. The measured blood pressure information may be input through a user terminal (eg, a smartphone, a PC, etc.) possessed by an expert, and the stethoscope blood pressure data may be generated as an average of the input data. For example, when measured five times by one minute, stethoscope blood pressure data may be generated as an average of blood pressure measurement data using five stethoscopes. The setting unit 320 may set the stethoscope blood pressure data as the target data of the deep neural network (DNN) model. For example, RSBP and RDBP shown in Table 2 above may be set as target data.

In operation 240, the learning controller 330 may learn a nonlinear relationship between the blood pressure vibration signal (ie, the oscillometric signal) measured in the blood pressure measurement and the stethoscope blood pressure data based on the set input parameter, the deep neural network model, and the target data. .

Table 2 below shows the results of validating the original feature data through statistical analysis of the virtual feature data (that is, the virtual feature vector), and may indicate the degree of convergence between the actual feature data and the virtual feature data. Here, the actual feature data may represent data measured by the wearable measuring apparatus worn by the test subject, and the virtual feature data may represent data generated through duplication based on the actual feature data.

For example, Table 2 shows the results of validating using the Kolmogorov-Smirnov test technique, and may indicate a prediction error between the actual feature data and the virtual feature data. The degree of convergence between the actual feature data and the virtual feature data shown in Table 2 may be expressed by Equation 1 below.

[Equation 1]

는 실제 특징데이터(즉, 특징벡터) X를 기반으로 부트스트랩 기법을 적용한 조건부 확률을 나타내고, 가상특징데이터의 평균값과 실제 특징데이터의 차이는 실제 특징데이터의 평균값과 실제 특징데이터 값의 차이와 동일하며, 좌측

과 우측

의 차이가 거의 0에 수렴하는 것을 알 수 있다. 위의 수학식 1에서

는 진실(true)을 나타내는 파라미터에 해당할 수 있다.In Equation 1 above,

Represents the conditional probability applying the bootstrap technique based on the actual feature data (i.e., feature vector) X, and the difference between the mean value of the virtual feature data and the feature data is equal to the difference between the mean feature value and the feature data Left

And right

It can be seen that the difference of converges to almost zero. In Equation 1 above

May correspond to a parameter representing true.

가 계산될 수 있다.Based on Equation 1 above, the prediction error as shown in Equation 2 below

Can be calculated.

[Equation 2]

Referring to Equation 2 and Table 2 above, the difference between the mean values MAP and AE of the actual feature data and the mean values of the virtual feature data may be calculated as a prediction error. In Table 2, in the case of RSBP and RDBP, which are target data, an error between the actual mean value and the virtual mean value may converge to almost zero.

When using the feature data shown in Table 2 above, the error of the standard deviation of the virtual feature data can be calculated based on Equation 3 below.

[Equation 3]

Referring to Table 2 above, the error (MAP: 0.016) of the standard deviation of the virtual feature data calculated based on Equation 3 above is relatively more compared with the error (MAP: 0.036) of the standard deviation of the actual feature data. It can be seen that it has a small value.

FIG. 5 is a diagram illustrating a structure of a two-stage ensemble deep neural network (DNN) according to one embodiment of the present invention.

In 510 of FIG. 5, an arrow may indicate a direction of information flow through a Restricted Boltzmann Machine (RBM).

)과 출력층(Y) 상이에 복수개의 유닛, 즉, 은닉층(H)이 계층적으로 존재할 수 있다. 최상층의 층간은 무향 에지(undirected edge)로 연결되고, 나머지 층간은 아래방향(directed edge)로 연결될 수 있다. 그리고, 정보는 위에서 아래 방향으로 전달되며, 최하위층인 입력층

의 상태구조가 결정될 수 있다.Referring to 520 of FIG. 5, the input layer (

) And a plurality of units, that is, the hidden layer (H) may exist hierarchically between the output layer (Y) and the output layer (Y). The interlayers of the uppermost layer may be connected by undirected edges, and the remaining interlayers may be connected by directed edges. The information is transmitted from top to bottom, and the input layer is the lowest layer.

The state structure of can be determined.

로 표현되고, 심화신경망의 모든 유닛(즉, 은닉층)의 결합 확률분포는 아래의 수학식 4와 같이 표현될 수 있다.In FIG. 5, the two-stage ensemble deep neural network model may perform classification and estimation by considering a state of the highest layer determined when the virtual feature data is set in the input layer as a representation of the data. Units (ie, hidden layers) of each layer in the upward direction from the lowest input layer

The combined probability distribution of all units (ie, hidden layers) of the deep neural network may be expressed by Equation 4 below.

[Equation 4]

부터

까지의 층(layer)은 이웃하는 윗층과 아래 방향 유향 에지(undirected edge)로 연결되고, 연결된 층 사이의 조건부 분포는 위의 수학식 4에서

에 해당할 수 있다. 그리고, 최상위층은 RBM(Restricted Boltzmann Machine)의 분포함수

로 표현될 수 있다.Lowest input layer

from

The layers up to are connected by neighboring upper layers and undirected edges, and the conditional distribution between the connected layers is expressed in Equation 4 above.

It may correspond to. And, the top layer is the distribution function of RBM (Restricted Boltzmann Machine)

It can be expressed as.

Table 3 below shows an algorithm of a two-step adaptive deep neural network regression model for effectively finding a nonlinear relationship between input data and target data.

In Table 3, the input data is virtual feature data generated for the oscillometric-based blood pressure estimation, and the target data may correspond to systolic and diastolic blood pressure data, which are stethoscope blood pressure data. Here, the virtual feature data set as input data may be generated using a bootstrap technique. For example, pre-training of deep neural networks may result in unstable blood pressure estimation due to lack of training data, and random pre-assignment of variables such as weights and biases in pre-training may lead to instability of learning, fit, and standard deviation of estimation errors. Can be. Accordingly, by training the deep neural network model in two stages using the bootstrap technique, the standard deviation of the error and the mean absolute error can be relatively lowered when the blood pressure is estimated using the existing single deep neural network regression model. Referring to Table 3 above, in step 1 to step 2, the feature data (i.e., the actual feature data) is extracted from the oscillometric signal, and the feature data is input parameters (i.e., the deep neural network model). Input data). The stethoscope blood pressure data can be set as the target data.

)를 계산할 수 있다.Subsequently, in step 3, the mean and standard deviation of the specific data and the target data for generating the virtual feature data (

) Can be calculated.

In step 4, a predetermined initial weight value may be assigned, and virtual feature data may be calculated, that is, generated by applying the bootstrap technique in steps 7-9. For example, bootstrap samples may be generated by copying the actual feature data, and an average value of the generated bootstrap samples may be calculated. The virtual feature data may be generated to include the generated bootstrap sample and the actual feature data. Then, in step 10, pre-training (ie, primary learning) may be performed based on the virtual feature data and the RBM.

In step 12, the result of pre-training, that is, an average value of each output value (for example, the ensemble 50) in which the primary learning is performed may be calculated and set as the output of weights and biases. .

As such, after the pre-training, that is, the primary learning is performed through the steps 1 to 12, the secondary learning may be performed.

In steps 16-18, initial virtual feature data generated in the pre-training step may be allocated for secondary learning. In steps 19 to 25, the weights for the second learning may be updated based on the initial virtual feature data generated through the first learning. Then, by applying a two-stage technique based on the error, the distribution of the actual training data set can be adaptively adjusted before each new bootstrap sample is obtained. Here, in each estimator using the deep neural network, as illustrated in step 19, the bootstrap training sample may be updated according to the weighted training sample. That is, the bootstrap sample can be obtained from a distribution that is repeatedly updated by a relative error function for subsequent estimators. In addition, the instance and the weight of each estimator may be adjusted according to the complexity. That is, the probability that an instance with a large previous error will appear in the next bootstrap sample may increase. Accordingly, the two-stage adaptive ensemble deep neural network model can reduce the variance of the mean error and increase the reliability of the result for the deep neural network regression estimator. The error function of the deep neural network may be expressed as Equation 5 below.

[Equation 5]

과 목표혈압값(즉, 청진 혈압데이터)

간의 차이를 기반으로 최소평균오차가 계산될 수 있다.Referring to Equation 5 above, the blood pressure value estimated based on the deep neural network model

And target blood pressure (ie stethoscope blood pressure data)

The minimum mean error can be calculated based on the difference between them.

[Equation 6]

Equation 6 above may be used to calculate the weight to be updated. The backpropagation technique and weight update for the error may be performed to reduce the estimation error based on the ensemble weights and biases of the first stage pre-training, that is, the weights and biases calculated through the first learning. At this time, the weight update and the bias variable for the error may be updated based on Equation 6 above.

In Equation 6, ε may represent a learning rate, η may represent a momentum variable, and L may represent the number of hidden layers.

FIG. 6 is a diagram illustrating performance of target data and a two-stage adaptive ensemble deep neural network model according to an embodiment of the present invention.

That is, FIG. 6 may show the data shown in Table 5 below and the standard error result based on the two-step adaptive ensemble deep neural network model (DBN-DNN) compared to the target data (RSBP and RDBP), which are target blood pressure values.

In Figure 6, based on the oscillometric blood pressure measurement data from BioSine Technology Co., Ltd. located in Toronto, Canada, used in the experiment for performance evaluation. Here, blood pressure data was approved by the Ontario Research Council of Ontario, Canada, and all participants submitted a blood pressure measurement agreement based on the criteria of the Research Ethics Committee. Blood pressure measurement samples required for the experiment were extracted from 85 cardiovascular disease test subjects aged 12 to 80 years, and 85 test subjects may include 48 males and 37 females. Oscillometric blood pressure measurements are wrist-worn blood pressure for all subjects in accordance with the American National standard manual, electronic or automated sphygmonanometers / association for the advancement of medical instrumentation (ANSI / AAMI) standard recommendations. It is assumed that five blood pressure measurements using a measuring device were obtained (five measurements, 85 subjects = 425 blood pressure measurement samples). In addition, it is assumed that two nurses simultaneously measure a systolic and diastolic blood pressure using a stethoscope blood pressure monitor (ie, a stethoscope) and use the calculated average value as a target blood pressure value. As a result, systolic blood pressure was 78 ~ 147 mmHg, diastolic blood pressure was 42 ~ 99mmHg. At this time, the blood pressure measurement process is measured five times for one experimenter, after one measurement, it can be proceeded by taking a 1 minute rest. Table 4 below shows the variable configuration of the two-stage adaptive deep neural network model.

Referring to Table 4 above, the variables used in the two-stage adaptive deepening neural network include 11 input feature vectors, double blood pressure value vectors (diastolic and systolic) 2, 100 samples of each virtual feature vector, and the actual number of samples. 5, hidden layer 3, and the like. At this time, the performance can be verified based on the mean error (ME), the standard deviation of the mean error (SDE), and the mean absolute error (MAE) shown in Table 5 below to evaluate the performance of the two-stage adaptive deep neural network. .

The AAMI protocol recognizes a blood pressure measurement device based on a standard error (SDE) of 8 mmHg within 5 mmHg of the mean error (ME). Referring to the performance evaluation of the blood pressure measurement methods in Table 5, the mean error (ME) was found to be within 5 mmHg in all methods, and the standard error (SDE) maximum amplitude algorithm (MAA) 9.03mmHg and 7.40 The diastolic and systolic standard errors of mmHg can be identified. And, feed-forward neural network (FFNN) 7.61mmHg and 6.79mmHg, support vector regression (SVR) 7.11mmHg and 6.14mmHg, Gaussian mixture regression (GMR) 6.12 mmHg and 5.29mmHg, single deep belief network-DBN-DNN 6.20mmHg and 5.26mmHg, and two-stage adaptive ensemble deep neural network model (DBN-DNN ensemble) 5.74mmHg and 4.68mmHg You can check the error performance. According to Table 5, the two-stage adaptive ensemble deep neural network model (DBN-DNN ensemble) shows a standard error reduction of about 11% and about 18% compared to the single deep neural network regression model. In addition, referring to the root mean square error (RMS) in Table 5, it can be seen that even in error measurement, the two-stage adaptive ensemble deep neural network model (DBN-DNN ensemble) shows a lower error than other methods.

FIG. 7 is a graph showing a standard error error (SDE) according to an increase in a virtual blood pressure sample according to one embodiment of the present invention.

Grades can be assigned for each mean absolute error (MAE) calculated based on the blood pressure measurement sample 425 data of all subjects using the British hypertension society (BHS) assessment method shown in Table 6 below. For example, the measuring method may be calculated in three groups for each method, MAE 5mmHg or less, 10mmHg or less, 15mmHg or less, and the grades A to D grades can be given. At this time, a grade A may be given if the measured blood pressure sample shows 60%, 10mmHg or less, 85%, or 15% or less and 95% or less of the MAE 5mmHg based on the DBN-DNN model. Although DBN-DNN and other blood pressure measurement methods were all grade A based on the BHS standard, the performance measurement results of DBN-DNN were 71.06%, 90.82%, 95.53% systolic (SBP) as 81.18% 96.24%, 99.29%. In addition, it can be seen that it has relatively better performance than other methods.

Method according to the embodiment is implemented in the form of program instructions that can be executed by various computer means may be recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions recorded on the media may be those specially designed and constructed for the purposes of the embodiments, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer readable recording media include magnetic media such as hard disks, floppy disks and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine code, such as produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Although the embodiments have been described by the limited embodiments and the drawings as described above, various modifications and variations are possible to those skilled in the art from the above description. For example, the described techniques may be performed in a different order than the described method, and / or components of the described systems, structures, devices, circuits, etc. may be combined or combined in a different form than the described method, or other components. Or even if replaced or substituted by equivalents, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are within the scope of the claims that follow.

Claims (14)

In the blood pressure measurement method using an ensemble deep neural network,
Extracting feature data from the blood pressure-related oscillometric signal measured by the blood pressure measuring apparatus worn by the test subject;
Setting input parameters of a deep neural network (DNN) model based on the extracted feature data;
Setting the stethoscope blood pressure data measured when the blood pressure of the test subject is measured by the blood pressure measuring device as target data; And
Learning a non-linear relationship between the blood pressure vibration signal measured by the blood pressure measuring device and the stethoscope blood pressure data measured based on a stethoscope based on the set input parameter, a deep neural network model, and target data;
Blood pressure measuring method comprising a. The method of claim 1,
Extracting the feature data,
Generating an envelope signal connecting the respective peaks of the signal to the oscillometric signal; And
Extracting feature data from the oscillometric signal and the envelope signal
Blood pressure measuring method comprising a. The method of claim 1,
The feature data includes at least one of mean arterial pressure (MAP), maximum amplitude (MA), area of envelope signal (AE),
The target data includes reference systolic blood pressure (RSBP) and reference diastolic blood pressure (RDBP).
Blood pressure measuring method characterized in that. The method of claim 1,
The setting of the input parameter,
Replicating the extracted feature data into a plurality of based on a bootstrap technique;
Generating virtual feature data including a bootstrap sample generated by the duplication and the feature data; And
Setting the virtual feature data as the input parameter
Blood pressure measuring method comprising a. The method of claim 4, wherein
The learning step,
Performing first-order learning on the input parameter based on a Restricted Boltzmann Machine (RBM) probability model
Blood pressure measuring method comprising a. The method of claim 5,
The learning step,
Setting weights for the second learning based on the result of performing the first learning
Blood pressure measuring method further comprising. The method of claim 6,
The bootstrap sample is updated based on the virtual feature data on which secondary learning is performed based on the set weight.
Blood pressure measuring method characterized in that. In the blood pressure measurement system using the ensemble deep neural network,
A feature data extraction unit for extracting feature data from the blood pressure related oscillometric signal measured by the blood pressure measuring apparatus worn by the test subject;
A setting unit configured to set an input parameter of a deep neural network (DNN) model based on the extracted feature data, and set the stethoscope blood pressure data measured when the blood pressure of the test subject is measured by the blood pressure measuring device as target data; And
A learning control unit learning a non-linear relationship between the blood pressure vibration signal measured by the blood pressure measuring device and the stethoscope blood pressure data measured based on a stethoscope based on the set input parameter, the deep neural network model, and target data.
Blood pressure measuring system comprising a. The method of claim 8,
The feature data extraction unit,
Generating an envelope signal connecting the respective peaks of the signal to the oscillometric signal, and extracting feature data from the oscillometric signal and the envelope signal;
Blood pressure measuring system, characterized in that. The method of claim 8,
The feature data includes at least one of mean arterial pressure (MAP), maximum amplitude (MA), area of envelope signal (AE),
The target data includes reference systolic blood pressure (RSBP) and reference diastolic blood pressure (RDBP).
Blood pressure measuring system, characterized in that. The method of claim 8,
The setting unit,
Replicate the extracted feature data into a plurality of based on a bootstrap technique, generate virtual feature data including the bootstrap sample and the feature data generated through the duplication, and convert the virtual feature data into the Set by input parameters
Blood pressure measuring system, characterized in that. The method of claim 11,
The learning control unit,
Performing first-order learning on the input parameters based on a Restricted Boltzmann Machine (RBM) probability model
Blood pressure measuring system, characterized in that. The method of claim 12,
The learning control unit,
Setting weights for secondary learning based on a result of performing the primary learning
Blood pressure measuring system, characterized in that. The method of claim 13,
The bootstrap sample is updated based on the virtual feature data on which secondary learning is performed based on the set weight.
Blood pressure measuring system, characterized in that.

Images