KR102627336B1 - Method and apparatus for estimaing blood pressure using newural architecture search and deep supervision - Google Patents

Abstract

A blood pressure estimation device using neural structure search and deep supervision techniques according to an embodiment of the present invention includes a data preprocessor for increasing the depth of the input signal; a neural structure search unit for constructing the blood pressure estimation model by extracting first cells and second cells; and a deep supervisor for learning the blood pressure estimation model, wherein the first cell is a normal cell and the second cell is a reduction cell.

Description

Blood pressure estimation method and device using neural structure search and deep supervision technique {METHOD AND APPARATUS FOR ESTIMAING BLOOD PRESSURE USING NEWURAL ARCHITECTURE SEARCH AND DEEP SUPERVISION}

The present invention relates to a method and device for estimating blood pressure using neural structure search and deep supervision techniques.

Blood pressure is a very important indicator in diagnosing heart-related diseases, and in general, regular blood pressure measurement is required to prevent heart disease. High blood pressure is a common and dangerous disease that increases the risk of myocardial infarction and heart attack. At this time, the risk of myocardial infarction and heart attack leading to death is very high. Accordingly, technologies that can easily measure blood pressure have recently been developed.

Blood pressure measurement methods can be broadly divided into invasive methods and non-invasive methods.

First, the invasive method is a method of measuring accurate blood pressure values using an arterial catheter. However, because the invasive method may cause side effects such as infection, it is not used in general situations.

Among non-invasive blood pressure measurement methods, a representative method is a mercury sphygmomanometer-based method that uses a cuff wrapped around the upper arm, and this method is currently considered the standard for non-invasive blood pressure measurement methods. However, mercury sphygmomanometer-based measurement equipment is heavy and large in size and has low mobility, making it difficult to use it regularly in daily life.

Accordingly, blood pressure estimation methods using photoplethysmogram (PPG) and electrocardiogram (ECG) sensors have recently been actively studied.

This blood pressure estimation method can be easily implemented in a smart device or wearable device (eg, smart watch, etc.), and has the advantage of being able to measure blood pressure periodically without being affected by time or place. However, currently developed blood pressure estimation methods based on photoplethysmography and electrocardiogram have the following problems.

1) The blood pressure estimation method based on photoplethysmography and electrocardiogram estimates blood pressure using a calibration method based on the reference blood pressure value obtained from a cuff-based blood pressure measurement device. This reference blood pressure value must be updated periodically, which can cause great inconvenience to the user. A calibration-free method has been proposed to solve this problem, but the calibration-free method has the problem of low accuracy.

2) Entering the deep learning era, a deep learning-based blood pressure estimation method that automatically analyzes the relationship between blood pressure and input signals is being developed (S. Baek et al., "End-to-end blood pressure prediction via (see “fully convolutional networks,” IEEE Access, 2019). Designing deep and complex end-to-end neural networks that provide good estimation performance requires significant effort from trained experts. In other words, in the past, in order to find the optimal neural network structure, the process of experienced engineers constructing and testing models had to be repeated, thus requiring a lot of effort from engineers. However, because engineers cannot experimentally implement all models, there is a problem in that it cannot be guaranteed that the neural network model designed by humans is the optimal model.

Korean Patent Publication No. 10-2019-005724 ‘Blood pressure estimation model generation system and method and blood pressure estimation system and method’ (published on May 13, 2019) Korean Patent No. 10-2108961 ‘Blood pressure estimation device using image-based artificial intelligence deep learning’ (published on May 4, 2020) Korean Patent No. 10-2163217 ‘Electrocardiogram arrhythmia classification method and chapter using deep convolutional neural network’ (published on December 24, 2019)

The purpose of the present invention is to provide a blood pressure estimation method and device using neural structure search and deep supervision techniques that can automatically find the optimal model using neural structure search.

Another object of the present invention is to provide a neural structure search that can be learned using a deep supervision technique using the morphological information of the photoplethysmographic signal as the answer and a blood pressure estimation method and device using a deep supervision technique.

Another object of the present invention is to provide a blood pressure estimation method and device using neural structure search and deep supervision techniques that can improve blood pressure estimation performance in calibration-free.

A blood pressure estimation device using neural structure search and deep supervision techniques according to an embodiment of the present invention includes a data preprocessor for increasing the depth of the input signal; a neural structure search unit for constructing the blood pressure estimation model by extracting first cells and second cells; and a deep supervisor for learning the blood pressure estimation model, wherein the first cell is a normal cell and the second cell is a reduction cell.

dX(t)/dt

d²X(t)/dt²이고, X(t)는 상기 입력 신호이고,

는 순차결합(concatenation) 연산을 의미하는 것을 특징으로 한다. In addition, the data pre-processing unit may include a differentiation processing unit for differentiating the input signal to generate a first differential signal and a second differential signal; and a sequential combining unit for sequentially combining the input signal, the first differential signal, and the second differential signal, based on Equation 1, where Equation 1 is:

dX(t)/dt

d ² X(t)/dt ² , where X(t) is the input signal,

is characterized in that it means a sequential concatenation operation.

In addition, the neural structure search unit may include a cell extractor for extracting a first cell and a second cell, each of which includes a plurality of nodes and a mixing operator connecting the plurality of nodes to each other; an operator weight setting unit configured to set a weight for each of a plurality of candidate operators for the mixing operator; As learning progresses, a weight update unit for updating the operator weights; When learning ends, an operator selection unit for selecting a candidate operator with the maximum operator weight; An evaluation unit for evaluating performance by applying the structure of the blood sugar estimation model to an evaluation set during the learning process; and a model constructor configured to construct the blood pressure estimation model by stacking a plurality of cells composed of the first cell and the second cell.

Additionally, the plurality of candidate operators include max pooling 1x7, avg. pooling 1x7, skip-connection, standard conv 1x11, dilation conv. It is characterized by one of 1x7 [dilation=1, 2, 3, 4], and none operation.

In addition, the evaluation unit performs bi-level optimization with operator weights as upper-level variables and training parameters as lower-level variables, based on Equation 2. The evaluation is performed using, and Equation 2 above is: , where w is a training parameter and α refers to an operator weight for candidate operator selection.

In addition, L in Equation 2 is defined according to Equation 3, and Equation 3 is: It is characterized by being.

In addition, the model constructor configures the blood pressure estimation model by stacking 8 cells consisting of the first cell and the second cell, and the first cell is the 1st, 2nd, 4th, 5th, 7th, and It is located in the 8th position, and the second cell is located in the 3rd and 6th positions.

In addition, the deep supervision unit includes a probability distribution array generator for generating a probability distribution array having a similar form to the input signal; a calculation unit for calculating skewness and kurtosis based on the probability distribution array; and a morphological information extraction unit for extracting morphological information of the input signal based on the skewness and kurtosis.

In addition, the morphological information extraction unit extracts a comprehensive loss through Equation 4, where Equation 4 is Ltotal=Lmain +r*Laux, Ltotal is the comprehensive loss, Lmain is the main loss, and Laux is the auxiliary loss. , r is the auxiliary loss weight, and the blood pressure estimation model is characterized in that it is learned in a direction that minimizes the overall loss.

In addition, the neural architecture search unit is characterized in that it constructs the structure of the blood pressure estimation model using a gradient descent-based neural architecture search method.

A blood pressure estimation method using a neural structure search and deep supervision technique according to an embodiment of the present invention includes the steps of a data preprocessor increasing the depth of an input signal; A neural structure search unit constructing a blood pressure estimation model by extracting first cells and second cells; and a deep supervisor training the blood pressure estimation model, wherein the first cell is a normal cell and the second cell is a reduction cell.

dX(t)/dt

d²X(t)/dt²이고, X(t)는 상기 입력 신호이고,

는 순차결합(concatenation) 연산을 의미하는 것을 특징으로 한다. In addition, the step of increasing the depth of the input signal includes: generating, by a differentiation processing unit, a first differential signal and a second differential signal by differentiating the input signal; And a sequential combining unit sequentially combining the input signal, the first differential signal, and the second differential signal based on Equation 1, where Equation 1 is:

dX(t)/dt

d ² X(t)/dt ² , where X(t) is the input signal,

is characterized in that it means a sequential concatenation operation.

In addition, constructing a blood pressure estimation model may include extracting, by a cell extraction unit, a first cell and a second cell, each of which includes a plurality of nodes and a mixing operator connecting the plurality of nodes to each other; setting, by an operator weight setting unit, a weight for each of a plurality of candidate operators for the mixing operator; A weight update unit updating the operator weight as learning progresses; selecting, by an operator selection unit, a candidate operator having the maximum operator weight when learning ends; An evaluation unit evaluating performance by applying the structure of the blood sugar estimation model to an evaluation set during a learning process; and a step of a model constructor constructing the blood pressure estimation model by stacking a plurality of cells composed of the first cell and the second cell.

In addition, the step of learning the blood pressure estimation model includes: generating, by a probability distribution array generator, a probability distribution array having a shape similar to that of the input signal; A calculation unit calculating skewness and kurtosis based on the probability distribution array; and a morphological information extraction unit extracting morphological information of the input signal based on the skewness and kurtosis.

A blood pressure estimation method using a neural structure search and deep supervision technique according to an embodiment of the present invention includes the steps of a data preprocessing unit preprocessing an input signal into a form suitable for a blood pressure estimation model; A neural structure search unit constructing the blood pressure estimation model by extracting a first cell and a second cell; and a deep supervisor training the blood pressure estimation model, wherein the neural architecture search unit uses a gradient descent-based neural architecture search method to determine the structure of the blood pressure estimation model. It is characterized by configuring.

The blood pressure estimation method and device using neural structure search and deep supervision techniques of the present invention have the effect of automatically finding the optimal model using neural structure search.

In addition, the blood pressure estimation method and device using the neural structure search and deep supervision technique of the present invention have the effect of being able to learn using the deep supervision technique using morphological information of the photoplethysmographic signal as the correct answer.

Additionally, the blood pressure estimation method and device using the neural structure search and deep supervision techniques of the present invention have the effect of improving blood pressure estimation performance in calibration-free mode.

In addition, the blood pressure estimation method and device using the neural structure search and deep supervision techniques of the present invention can obtain a model with more improved estimation performance in a short time than the model designed by existing engineers, and can save development costs due to engineers. There is a possible effect.

1 is a diagram showing a partial configuration of a blood pressure estimation device according to an embodiment of the present invention.
Figure 2 is a diagram for explaining a data preprocessing unit according to an embodiment of the present invention.
Figure 3 is a diagram for explaining a neural structure search unit according to an embodiment of the present invention.
Figure 4 is a diagram for explaining the deep supervision unit according to an embodiment of the present invention.
Figure 5 is a flowchart for explaining a method for estimating blood pressure according to an embodiment of the present invention.
Figure 6 is a flowchart showing in detail the blood pressure estimation method according to an embodiment of the present invention.
Figure 7 is a diagram showing a blood pressure estimation method according to an embodiment of the present invention.
Figure 8 is a diagram showing a blood pressure estimation method according to an embodiment of the present invention.
Figure 9 is a diagram for explaining in detail the cell structure according to an embodiment of the present invention.
Figure 10 is a diagram for explaining the structure of a first cell according to an embodiment of the present invention.
Figure 11 is a diagram for explaining the structure of a second cell according to an embodiment of the present invention.
Figure 12 is a diagram for explaining the function of the second cell according to an embodiment of the present invention.
Figure 13 is a diagram showing a probability distribution array according to an embodiment of the present invention.
Figure 14 is a diagram showing the structure of a blood pressure estimation model according to an embodiment of the present invention.
Figure 15 is a diagram showing the performance of a blood pressure estimation device according to an embodiment of the present invention.
Figure 16 is a diagram showing the results of comparing the estimation performance of a blood pressure estimation device according to an embodiment of the present invention with that of a conventional blood pressure estimation device.

The present invention will be described in more detail.

Hereinafter, with reference to the attached drawings, embodiments of the present invention and other matters necessary for those skilled in the art to easily understand the contents of the present invention will be described in detail. However, since the present invention can be implemented in various different forms within the scope set forth in the claims, the embodiments described below are merely illustrative, regardless of whether they are expressed or not.

Like reference numerals refer to like elements. Additionally, in the drawings, the thickness, proportions, and dimensions of components are exaggerated for effective explanation of technical content. “And/or” may include any one or more combinations that the associated configurations may define.

Terms such as first, second, etc. may be used to describe various components, but the components should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another. For example, a first component may be named a second component, and similarly, the second component may also be named a first component without departing from the scope of the present invention. Singular expressions may include plural expressions, unless the context clearly indicates otherwise.

Additionally, terms such as “below,” “on the lower side,” “above,” and “on the upper side” are used to describe the relationship between the components shown in the drawings. The above terms are relative concepts and are explained based on the direction indicated in the drawings.

Terms such as “include” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but do not include one or more other features, numbers, or steps. , it should be understood that it does not exclude in advance the possibility of the existence or addition of operations, components, parts, or combinations thereof.

In other words, the present invention is not limited to the embodiments disclosed below, but can be implemented in various different forms. In the description below, when a part is connected to another part, it is directly connected. In addition, it may also include cases where they are electrically connected with another element in between. In addition, it should be noted that the same components in the drawings are indicated with the same reference numbers and symbols as much as possible, even if they are shown in different drawings.

In this specification, neural architecture search is a field of automated machine learning and is a method of deriving an optimal neural network structure through learning. Neural structure search can easily find the optimal neural network structure while minimizing the engineer's work and verification process.

Neural architecture search methods include reinforcement learning-based neural architecture search, evolutionary algorithm-based neural architecture search, and gradient descent-based neural architecture search. It can be classified into three methods: architecture search.

NASNet (B. Zoph et al., "Learning transferable architectures for scalable image recognition," IEEE conference on computer vision and pattern recognition, 2018.) is a representative model for reinforcement learning-based neural structure search. AmoebaNet (E. Real et al., "Regularized evolution for image classifier architecture search," AAAI, 2019) is a representative model for neural architecture search based on evolutionary algorithms. DARTS; differentiable architecture search (H. Liu et al., "DARTS: Differentiable architecture search," ICLR, 2019) is a model representing neural architecture search based on gradient descent.

The model found using the neural structure search method optimizes the model structure only within a limited search space. Therefore, although performance varies depending on the defined search space, it has the advantage of being able to extract a model that at least converges to the local optima within the corresponding search space.

In this specification, deep supervision refers to a method developed to solve the gradient vanishing problem that occurs when the model for estimation has a very deep structure. The deep supervision technique connects to specific points in the model to calculate an auxiliary loss, and adds this calculated value to the main loss to arrive at the final loss. The best known model using deep supervision techniques is GoogLeNet (C. Szegedy et al., "Going deeper with convolutions," IEEE conference on computer vision and pattern recognition, 2015).

GoogLeNet shows multiple auxiliary heads connected to specific points in the model. GoogLeNet uses a very deep neural network structure and calculates an auxiliary loss for each specific point to solve the gradient vanishing problem and enable learning to the bottom.

The reason why the deep supervision technique is used in the present invention is that a very deep neural network structure can be created by combining normal cells and reduction cells extracted through neural structure search.

In general, the purpose of calculating in the auxiliary loss and the purpose of calculating in the main loss are the same.

However, the present invention proposes a method to make learning more effective by modifying the existing deep supervision technique.

Figure 1 is a diagram showing a blood pressure estimation device 10 according to an embodiment of the present invention.

Referring to FIG. 1 , the blood pressure estimation device 10 may include a data preprocessor 100, a neural structure search unit 200, a deep supervisor 300, and a blood pressure estimation model 400.

The data preprocessor 100 may preprocess the input signal into a form suitable for the blood pressure estimation model.

For example, the data preprocessor 100 can increase the depth of the input signal. In this specification, increasing the depth of a signal means increasing the depth, which is one of the characteristics of the signal.

Specifically, the data preprocessor 100 may generate a first-order differential signal by differentiating the input signal and generate a second-order differential signal by differentiating the first-order differential signal. Additionally, the data preprocessor 100 can increase the depth of the input signal by combining the original signal, first differential signal, and second differential signal.

Depending on the embodiment, the input signal may be a PPG signal. That is, the data preprocessor 100 can increase the depth of the input signal, which is a PPG signal. That is, the data preprocessor 100 can increase the depth of the input signal by sequentially combining the PPG, the first differential signal of the PPG, and the second differential signal of the PPG. In this specification, the PGG signal refers to a photoplethysmogram (PPG) signal.

The data preprocessor 100 may cut the PPG signal into 512 sample units and generate a first differential signal and a second differential signal based on this. Additionally, the data preprocessor 100 may sequentially combine the original signal, first differential signal, and second differential signal. Through this, the blood pressure estimation method and device according to an embodiment of the present invention can increase the amount of information that the model can use and improve estimation power by increasing the channel of the input signal.

The target is to cut the ABP (arterial blood pressure) signal obtained through an arterial catheter into the same part as the PPG signal in units of 512 samples, and determine the average of the upper peaks as SBP and the average of the lower peaks as DBP.

In the architecture search stage, only the target is used, and since the model for SBP and the model for DBP are searched respectively, the shape of the target during learning is m x 1.

In the architecture evaluation stage, both the target and the aux_target are used, and the shape of the aux_target is m x 2. In this case, the constant 2 means skewness and kurtosis.

The neural structure search unit 200 may construct the blood pressure estimation model 400 by extracting the first cell and the second cell. At this time, the first cell may be a normal cell, and the second cell may be a reduction cell. At this time, the second cell may serve to reduce the size of the feature map by half.

For example, the neural architecture search unit 200 may configure the structure of the optimal blood pressure estimation model 400 using a gradient descent-based neural architecture search method. Depending on the embodiment, the neural architecture search unit 200 may use DARTS (Differentiable architecture search), one of the gradient descent-based neural architecture search methods.

The deep supervisor 300 can learn a blood pressure estimation model.

The blood pressure estimation model 400 can estimate the patient's blood pressure based on the input signal. For example, the structure of the blood pressure estimation model 400 constructed by the neural structure search unit 200 may be a structure in which first cells are stacked deeply and then second cells are inserted at 1/3 and 2/3 points.

The blood pressure estimation model 400 may use the patient's PPG signal as an input signal. At this time, the PPG signal may have a sampling rate of 125Hz.

The output of the blood pressure estimation model 400 is as follows.

The architecture search stage can return the structure of the architecture as output. Specifically, it returns a numeric array in the form of an array, and the numbers in the array can mean the structure of the architecture for each position.

In the architecture evaluation stage, an architecture can be created using the architecture structure received in the search stage, and the estimated performance can then be returned by learning the architecture.

Figure 2 is a diagram showing the data preprocessing unit 100 according to an embodiment of the present invention.

Referring to FIG. 2, the data pre-processing unit 100 may include a differential processing unit 110 and a sequential combining unit 120.

The differentiation processing unit 110 may generate a first differential signal and a second differential signal by differentiating the input signal.

The sequential combiner 120 may sequentially combine the input signal, the first differential signal, and the second differential signal based on Equation 1.

dX(t)/dt

d²X(t)/dt², 여기서, X(t)는 입력 신호이고,

는 순차결합(concatenation) 연산을 의미한다. [Equation 1] X(t) = X(t)

dX(t)/dt

d ² X(t)/dt ² , where X(t) is the input signal,

means sequential concatenation operation.

Therefore, the input signal that ultimately enters the input of the blood pressure estimation model is

At this time, the multiplied constant 3 means three-dimensional data that appears because the first differential signal and the second differential signal are sequentially combined. The constant 512 refers to the length of the one-dimensional signal. The input shape of the model is the same in the architecture search stage and architecture evaluation stage.

Figure 3 is a diagram showing a neural structure search unit 200 according to an embodiment of the present invention.

Referring to FIG. 3, the neural structure search unit 200 includes a cell extraction unit 210, an operator weight setting unit 220, a weight updating unit 230, an operator selection unit 240, an evaluation unit 250, and It may include a model component 260.

The cell extractor 210 may extract the first cell and the second cell. Each of the first cell and the second cell may include a plurality of nodes and a mixing operator connecting the plurality of nodes to each other. For example, at least one of the first cell and the second cell may be composed of four nodes and a mixing operator. At this time, the mixing operator may be selected from a plurality of candidate operators. Mixed operators can define operations that connect nodes.

Because a mixing operator is selected from within a plurality of candidate operators, a search space can be defined according to the type of candidate operator.

The operator weight setting unit 220 may set a weight for each of a plurality of candidate operators for a mixed operator. Depending on the embodiment, 9 candidate operators may be used in the present invention. If too many candidate operators are used, the weights corresponding to each candidate operator may be too small, so learning may not be performed properly, and the search space may become large, causing an issue where learning time takes an excessively long time.

At this time, each of the plurality of candidate operators is max pooling 1x7, avg. pooling 1x7, skip-connection, standard conv 1x11, dilation conv. It can be any one of 1x7 [dilation=1, 2, 3, 4], and none operation. Weights may be assigned to a plurality of candidate operators. As the blood pressure estimation model 400 is learned, the weights are updated, and the candidate operator with the highest weight may be finally selected. Therefore, to design a mixed operator, a list of candidate operators must first be defined.

The weight update unit 230 may update operator weights as learning progresses.

When learning ends, the operator selection unit 240 may select a candidate operator with the maximum operator weight.

The evaluation unit 250 may evaluate performance by applying the structure of the blood sugar estimation model to an evaluation set during the learning process. Additionally, the evaluation unit 250 of the neural structure search unit 200 may learn to improve the structure of the blood pressure estimation model based on the confirmation result.

For example, the evaluation unit 250 performs bi-level optimization (bi-level optimization) with operator weights as upper-level variables and training parameters as lower-level variables, based on Equation 2. Evaluation can be performed using optimization.

[Equation 2] , where w is a training parameter and α may mean an operator weight for candidate operator selection.

L in Equation 2 can be defined according to Equation 3.

[Equation 3]

The model constructor 260 may construct a blood pressure estimation model by stacking a plurality of cells composed of first cells and second cells.

For example, the model constructor 260 may construct a blood pressure estimation model by stacking 8 cells consisting of a first cell and a second cell. The first cell may be located at the 1st, 2nd, 4th, 5th, 7th, and 8th positions, and the second cell may be located at the 3rd and 6th positions. The second cell may serve to reduce the size of the feature map.

Figure 4 is a diagram showing the deep supervision unit 300 according to an embodiment of the present invention. The deep supervisor 300 of the present invention can ultimately learn the blood pressure estimation model 400. The deep supervision unit 300 uses one auxiliary head and can be connected to a second reduction cell. In general, the target of the auxiliary loss is the same as the target of the main loss. However, the blood pressure estimation method and device using neural structure search and deep supervision techniques of the present invention use morphological information of the input signal as the goal of auxiliary loss for more effective learning. For example, the deep supervisor 300 may extract morphological information of the input signal using skewness and kurtosis.

Referring to FIG. 4 , the deep supervisor 300 may include a probability distribution array generator 310, a calculation unit 320, and a morphological information extraction unit 330.

The probability distribution array generator 310 may generate a probability distribution array having a similar shape to the input signal.

The calculation unit 320 may calculate skewness and kurtosis based on the probability distribution array. For example, skewness and kurtosis may be information indicating the form of a probability distribution. Skewness and kurtosis cannot be obtained directly from the input signal.

The morphological information extraction unit 330 may extract morphological information of the input signal based on skewness and kurtosis.

For example, the morphological information extraction unit 330 can extract morphological information of the input signal by calculating the comprehensive loss through Equation 4.

[Equation 4] - Ltotal=Lmain +r*Laux

In Equation 4 above, Ltotal may refer to the overall loss, Lmain may refer to the main loss, Laux may refer to the auxiliary loss, and r may refer to the auxiliary loss weight.

The main loss may refer to the difference between the target blood pressure and the estimated blood pressure, and the auxiliary loss may refer to the distance difference between the estimated vector and the two-dimensional vector consisting of kurtosis and skewness, which are auxiliary targets.

At this time, the auxiliary loss weight may be a value between 0.1 and 0.6. Preferably, the auxiliary loss weight may be 0.2.

The blood pressure estimation model can be learned in a way that minimizes overall loss.

If you set the morphological information of the input signal as the goal of the auxiliary loss, the deep part of the model can be well learned. Therefore, the blood pressure estimation method and device using the neural structure search and deep supervision technique of the present invention not only prevents the gradient vanishing problem, but also learns to estimate the shape of the signal at the bottom of the model, so that the entire model Blood pressure estimation performance can be improved. This is possible because the shape of the PPG signal is related to blood pressure.

The morphological information of PPG, which is used as an aux_target of Modified Deep Supervision, is not used in the architecture search stage, but can be used in the architecture evaluation stage. . Morphological information of PPG can be obtained by cutting the input signal cut into 512 sample units again into each pulse (the most basic form of the PPG signal), converting it into a pulse-like histogram, and approximating the skewness and kurtosis values.

Figure 5 is a flowchart showing a blood pressure estimation method according to an embodiment of the present invention.

Below, with reference to FIGS. 1 to 5, a blood pressure estimation method is described.

First, the data preprocessor 100 may preprocess the input signal into a form suitable for the blood pressure estimation model 400. For example, the data preprocessor 100 can increase the depth of the input signal (S10).

The neural structure search unit 200 may construct the blood pressure estimation model 400 by extracting the first cell and the second cell (S20). The first cell is a normal cell, and the second cell is a reduction cell.

For example, the neural architecture search unit 200 may construct the structure of the blood pressure estimation model using a gradient descent-based neural architecture search method.

The deep supervisor 300 may train the blood pressure estimation model 400 (S30).

The blood pressure estimation model 400 can estimate the blood pressure value based on the input signal (S40).

Figure 6 is a flowchart showing in detail the blood pressure estimation method according to an embodiment of the present invention.

Hereinafter, with reference to FIGS. 1 to 6, the step (S10) of increasing the depth of the input signal of FIG. 5 will be described in detail.

First, the differentiation processing unit 110 may generate a first differential signal and a second differential signal by differentiating the input signal (S110).

The sequential combining unit 120 may sequentially combine the input signal, the first differential signal, and the second differential signal (S120).

Figure 7 is a flowchart showing in detail the blood pressure estimation method according to an embodiment of the present invention.

Below, with reference to FIGS. 1 to 7 , the step S20 of constructing the blood pressure estimation model 400 of FIG. 5 will be described in detail.

First, the cell extractor 210 may extract a first cell and a second cell, each of which includes a plurality of nodes and a mixing operator connecting the plurality of nodes to each other (S210).

The operator weight setting unit 220 may set a weight for each of a plurality of candidate operators for the mixed operator (S220).

The weight update unit 230 may update the operator weight as learning progresses (S230).

When learning ends, the operator selection unit 240 may select a candidate operator with the maximum operator weight (S240).

The evaluation unit 250 may evaluate performance by applying the structure of the blood sugar estimation model to the evaluation set during the learning process (S250).

The model constructor 260 may construct a blood pressure estimation model by stacking a plurality of cells consisting of a first cell and a second cell (S260).

Figure 8 is a flowchart showing in detail the blood pressure estimation method according to an embodiment of the present invention.

Below, with reference to FIGS. 1 to 8 , the step S30 of learning the blood pressure estimation model 400 of FIG. 5 will be described in detail.

First, the probability distribution array generator 310 may generate a probability distribution array having a similar form to the input signal (S310).

The calculation unit 320 may calculate skewness and kurtosis based on the probability distribution array (S320).

The morphological information extraction unit 330 may extract morphological information of the input signal based on skewness and kurtosis (S330).

Figure 9 is a diagram showing a cell structure according to an embodiment of the present invention.

Referring to FIG. 9, either the first cell or the second cell may be composed of four nodes and a mixing operator.

A mixed operator consists of several candidate operators, and each candidate operator is combined with a weight, which is the probability that the candidate operation will be selected, and the weight can be updated as learning progresses. When learning is completed, the blood pressure estimation model can be confirmed by the mixing operator leaving only candidate operators with high weight and removing other candidate operators.

Figure 10 is a diagram showing the structure of a first cell according to an embodiment of the present invention.

Referring to FIG. 10, the structure of the first cell, that is, a normal cell, extracted through a neural architecture search method is shown.

In the cell structure described in FIG. 9, as learning progresses, candidate operations with high weights are converted and extracted. ReLU (rectified linear unit) was used as the activation function, and batch-normalization was applied after the operation, except for pooling operation and skip-connection.

Figure 11 is a diagram showing the structure of a second cell according to an embodiment of the present invention.

Referring to FIG. 11, the structure of a second cell, that is, a reduction cell, extracted through a neural architecture search method is shown.

In the cell structure described in FIG. 9, as learning progresses, candidate operations with high weights are converted and extracted. ReLU (rectified linear unit) was used as the activation function, and batch-normalization was applied after the operation, except for pooling operation and skip-connection.

Figure 12 is a diagram showing the function of the second cell according to an embodiment of the present invention.

Referring to FIG. 12, the normal cell, which is the first cell, maintains the size of the feature map, but the reduction cell, which is the second cell, can reduce the size of the feature map by half. The second cell may exist at 1/3 or 2/3 positions in the final model structure.

Figure 13 is a diagram showing a probability distribution array according to an embodiment of the present invention.

Referring to FIG. 13, a histogram, which is a probability distribution array similar to a signal created to obtain skewness and kurtosis, which are morphological information of an input signal, is shown. The graph shown on the left represents the input signal, and the histogram shown on the right represents the probability distribution array.

Using the probability distribution array corresponding to the histogram, the morphological information of the left signal can be extracted.

Figure 14 is a diagram showing the structure of a blood pressure estimation model according to an embodiment of the present invention.

Referring to Figure 14, the structure of the blood pressure estimation model is shown. A blood pressure estimation model can be constructed by combining normal cells and reduction cells extracted through neural architecture search. As shown, the left part of the second reduction cell represents a deep supervision technique that estimates the morphological information of the signal.

Figure 15 is a diagram showing the performance of a blood pressure estimation device according to an embodiment of the present invention.

Referring to Figure 15, when using a general deep supervision technique based on SBP, there was a small difference in estimation performance of about 0.1, but when a deep supervision technique using morphological information of the signal was applied, a large difference of about 0.9 was seen. It can be seen that there is a difference in estimated performance.

Figure 16 is a diagram showing the results of comparing the estimation performance of a blood pressure estimation device according to an embodiment of the present invention with that of a conventional blood pressure estimation device. It can be seen that the invented algorithm uses only a single PPG signal and achieves more improved performance than the estimation model using a composite signal of electrocardiogram (ECG) and PPG.

Through the above-described method, the blood pressure estimation method and device using neural structure search and deep supervision techniques of the present invention have the effect of automatically finding the optimal model using neural structure search.

In addition, the blood pressure estimation method and device using the neural structure search and deep supervision technique of the present invention have the effect of being able to learn using the deep supervision technique using morphological information of the photoplethysmographic signal as the correct answer.

Additionally, the blood pressure estimation method and device using the neural structure search and deep supervision techniques of the present invention have the effect of improving blood pressure estimation performance in calibration-free mode.

The functional operations described herein and embodiments of the subject matter described above may be implemented in digital electronic circuits, computer software, firmware or hardware, including the structures disclosed herein and their structural equivalents, or in a combination of one or more of these. possible.

Embodiments of the subject matter described herein may relate to one or more computer program products, that is, computer program instructions encoded on a tangible program medium for execution by or to control the operation of a data processing device. It can be implemented as a module. The tangible program medium may be a radio signal or a computer-readable medium. A radio signal is an artificially generated signal, such as a machine-generated electrical, optical or electromagnetic signal, that is generated to encode information for transmission to a suitable receiver device for execution by a computer. The computer-readable medium may be a machine-readable storage device, a machine-readable storage substrate, a memory device, a combination of materials that affect a machine-readable radio wave signal, or a combination of one or more of these.

A computer program (also known as a program, software, software application, script, or code) may be written in any form of a programming language, including a compiled or interpreted language, or an a priori or procedural language, as a stand-alone program or module; It can be deployed in any form, including components, subroutines, or other units suitable for use in a computer environment.

Computer programs do not necessarily correspond to files on a file device. A program may be stored within a single file that serves the requested program, or within multiple interacting files (e.g., more than one file storing a module, subprogram, or portion of code), or within a file holding other programs or data. Some (e.g., one or more scripts stored within a markup language document) may be stored.

The computer program may be deployed to run on a single computer or multiple computers located at one site or distributed across multiple sites and interconnected by a communications network.

Additionally, the logic flow and structural block diagram described in this patent document describe corresponding actions and/or specific methods supported by corresponding functions and steps supported by the disclosed structural means, It can also be used to build software structures and algorithms and their equivalents.

The processes and logic flows described herein can be performed by one or more programmable processors, each of which executes one or more computer programs to perform functions by operating on input data and generating output.

Processors suitable for executing computer programs include, for example, any one or more processors of both general-purpose and special-purpose microprocessors and any type of digital computer. Typically, the processor will receive instructions and data from read-only memory, random access memory, or both.

The core elements of a computer are one or more memory devices for storing instructions and data and a processor for executing instructions. Additionally, a computer generally includes one or more mass storage devices for storing data, such as magnetic, magneto-optical or optical disks, operable to receive data from, transfer data to, or perform both such operations. It may be combined with or include these. However, a computer does not need to have such a device.

The present description sets forth the best mode of the invention and provides examples to illustrate the invention and to enable any person skilled in the art to make or use the invention. The specification prepared in this way does not limit the present invention to the specific terms presented.

Although the present invention has been described above with reference to preferred embodiments, those skilled in the art or have ordinary knowledge in the relevant technical field should not deviate from the spirit and technical scope of the present invention as set forth in the claims to be described later. It will be understood that the present invention can be modified and changed in various ways within the scope of the present invention.

Therefore, the technical scope of the present invention should not be limited to what is described in the detailed description of the specification, but should be defined by the scope of the patent claims.

10: Blood pressure estimation device 100: Data preprocessing unit
110: Differential processing unit 120: Sequential coupling unit
200: Neural structure search unit 210: Cell extraction unit
220: operator weight setting unit 230: weight updating unit
240: operator selection unit 250: evaluation unit
260: model configuration unit 300: deep supervision unit
310: Probability distribution array generation unit 320: Calculation unit
330: Morphological information extraction unit 400: Blood pressure estimation model

Claims (15)

A data preprocessor to increase the depth of the input signal containing information about blood pressure;
a neural structure search unit for constructing a blood pressure estimation model by extracting the first cell and the second cell; and
A deep supervisor that extracts morphological information of the input signal and uses it as an auxiliary loss target to learn the blood pressure estimation model,
The first cell is a normal cell, the second cell is a reduction cell,
The blood pressure estimation model is constructed by combining the normal cell and the reduction cell, and is characterized in that it receives the input signal as an input and estimates the blood pressure value using a deep supervision technique.
Blood pressure estimation device using neural structure search and deep supervision techniques. According to paragraph 1,
The data preprocessing unit,
a differentiation processing unit for differentiating the input signal to generate a first differential signal and a second differential signal; and
Based on Equation 1, it includes a sequential combining unit for sequentially combining the input signal, the first differential signal, and the second differential signal,
Equation 1 above is:
X(t) = X(t)

dX(t)/dt

^{d 2}
X(t) is the input signal,

is characterized in that it means a sequential concatenation operation,
Blood pressure estimation device using neural structure search and deep supervision techniques. According to paragraph 2,
The neural structure search unit,
a cell extractor for extracting a first cell and a second cell, each of which includes a plurality of nodes and a mixing operator connecting the plurality of nodes to each other;
an operator weight setting unit configured to set a weight for each of a plurality of candidate operators for the mixing operator;
As learning progresses, a weight update unit for updating the operator weights;
When learning ends, an operator selection unit for selecting a candidate operator with the maximum operator weight;
An evaluation unit for evaluating performance by applying the structure of the blood pressure estimation model to an evaluation set during the learning process; and
Characterized in that it includes a model constructor for constructing the blood pressure estimation model by stacking a plurality of cells composed of the first cell and the second cell.
Blood pressure estimation device using neural structure search and deep supervision techniques. According to paragraph 3,
The plurality of candidate operators are max pooling 1x7, avg. pooling 1x7, skip-connection, standard conv 1x11, dilation conv. Characterized by any one of 1x7 [dilation=1, 2, 3, 4], and none operation,
Blood pressure estimation device using neural structure search and deep supervision techniques. According to clause 4,
The evaluation department,
Based on Equation 2, evaluation is performed using bi-level optimization with operator weights as upper-level variables and training parameters as lower-level variables. ,
Equation 2 above is:
ego,
Here, w is a training parameter, and α refers to the operator weight for candidate operator selection,
Blood pressure estimation device using neural structure search and deep supervision techniques. According to clause 5,
L in Equation 2 is defined according to Equation 3,
Equation 3 above is:
Characterized in that,
Blood pressure estimation device using neural structure search and deep supervision techniques. According to paragraph 3,
The model constructor constructs the blood pressure estimation model by stacking 8 cells consisting of the first cell and the second cell,
The first cell is located at the 1st, 2nd, 4th, 5th, 7th and 8th positions,
Characterized in that the second cell is located in the 3rd and 6th positions,
Blood pressure estimation device using neural structure search and deep supervision techniques. According to paragraph 3,
The in-depth supervision department said,
a probability distribution array generator for generating a probability distribution array having a shape corresponding to the input signal;
a calculation unit for calculating skewness and kurtosis based on the probability distribution array; and
Characterized in that it comprises a morphological information extraction unit for extracting morphological information of the input signal based on the skewness and the kurtosis,
Blood pressure estimation device using neural structure search and deep supervision techniques. According to clause 8,
The morphological information extraction unit extracts the comprehensive loss through Equation 4,
Equation 4 above is:
Ltotal=Lmain +r*Laux,
Ltotal is the overall loss, Lmain is the main loss, Laux is the secondary loss, r is the secondary loss weight,
The blood pressure estimation model is characterized in that it is learned in a way that minimizes the overall loss.
Blood pressure estimation device using neural structure search and deep supervision techniques. According to paragraph 1,
The neural architecture search unit is characterized in that it configures the structure of the blood pressure estimation model using a gradient descent-based neural architecture search method.
Blood pressure estimation device using neural structure search and deep supervision techniques. A data preprocessing unit increasing the depth of an input signal including information about blood pressure;
A neural structure search unit constructing a blood pressure estimation model by extracting first cells and second cells; and
A deep supervisor extracting morphological information of the input signal and using it as an auxiliary loss target to learn the blood pressure estimation model,
The first cell is a normal cell, the second cell is a reduction cell,
The blood pressure estimation model is constructed by combining the normal cell and the reduction cell, and is characterized in that it receives the input signal as an input and estimates the blood pressure value using a deep supervision technique.
Blood pressure estimation method using neural structure search and deep supervision techniques. According to clause 11,
The step of increasing the depth of the input signal is,
A differentiation processing unit, differentiating the input signal to generate a first differential signal and a second differential signal; and
A sequential combining unit sequentially combines the input signal, the first differential signal, and the second differential signal based on Equation 1,
Equation 1 above is:
X(t) = X(t)

dX(t)/dt

^{d 2}
X(t) is the input signal,

is characterized in that it means a sequential concatenation operation,
Blood pressure estimation method using neural structure search and deep supervision techniques. According to clause 12,
The steps to construct a blood pressure estimation model are:
Extracting, by a cell extraction unit, a first cell and a second cell, each of which includes a plurality of nodes and a mixing operator connecting the plurality of nodes to each other;
setting, by an operator weight setting unit, a weight for each of a plurality of candidate operators for the mixing operator;
A weight update unit updating the operator weight as learning progresses;
selecting, by an operator selection unit, a candidate operator having the maximum operator weight when learning ends;
An evaluation unit evaluating performance by applying the structure of the blood pressure estimation model to an evaluation set during a learning process; and
Characterized in that the model construction unit includes the step of constructing the blood pressure estimation model by stacking a plurality of cells consisting of the first cell and the second cell.
Blood pressure estimation method using neural structure search and deep supervision techniques. According to clause 12,
The step of learning the blood pressure estimation model is,
generating, by a probability distribution array generator, a probability distribution array having a shape corresponding to the input signal;
A calculation unit calculating skewness and kurtosis based on the probability distribution array; and
Characterized in that the morphological information extraction unit includes a step of extracting morphological information of the input signal based on the skewness and the kurtosis.
Blood pressure estimation method using neural structure search and deep supervision techniques. A data preprocessing unit preprocessing an input signal containing information about blood pressure into a form to be used in a blood pressure estimation model;
Constructing the blood pressure estimation model by extracting, by a neural structure search unit, a first cell that is a normal cell and a second cell that is a reduction cell; and
A deep supervisor extracting morphological information of the input signal and using it as an auxiliary loss target to train the blood pressure estimation model,
The neural architecture search unit configures the structure of the blood pressure estimation model using a gradient descent-based neural architecture search method,
The blood pressure estimation model is constructed by combining the normal cell and the reduction cell, and is characterized in that it receives the input signal as an input and estimates the blood pressure value using a deep supervision technique.
Blood pressure estimation method using neural structure search and deep supervision techniques.