KR102472848B1 - METHOD AND APPARATUS FOR MEASURING NON-INVASIVE HbA1c USING LEARNING ALGORITHM - Google Patents

Abstract

The present invention relates to a non-invasive method and apparatus for measuring glycated hemoglobin using a learning algorithm, the method comprising: collecting learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients; constructing a prediction model for predicting the glycated hemoglobin level by using a learning algorithm based on the learning data; Collecting PPG signal data of a subject to be measured; and predicting the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model.

Description

Method and apparatus for non-invasive glycated hemoglobin measurement using learning algorithm {METHOD AND APPARATUS FOR MEASURING NON-INVASIVE HbA1c USING LEARNING ALGORITHM}

The present invention relates to a non-invasive glycated hemoglobin measurement method and apparatus, and more particularly, to a non-invasive glycated hemoglobin (HbA1c) concentration capable of accurately and easily non-invasively measuring the concentration of glycated hemoglobin (HbA1c) using a learning algorithm. It relates to a method and apparatus for measuring hemoglobin.

Diabetes is a metabolic disease characterized by hyperglycemia caused by secretion or dysfunction of insulin necessary for blood sugar control in the body. Chronic hyperglycemia due to diabetes causes damage and malfunction of each organ of the body. In particular, it causes microvascular complications in the retina, kidneys, and nerves, and macrovascular complications such as arteriosclerosis, cardiovascular, and cerebrovascular diseases, resulting in mortality. increases

However, diabetes can reduce the rate of exacerbation or complications of diabetes due to glycemic control, weight loss, and medication. Therefore, diabetic patients need to frequently measure their own blood sugar levels to manage blood sugar levels, and periodically receive glycated hemoglobin (HbA1C) tests, which are a treatment index as important as blood sugar levels in diabetic patients.

The glycated hemoglobin (HbA1c) test is a test to see how much hemoglobin in red blood cells, which plays a role in transporting oxygen in the blood, is glycated. reflects Normal people naturally have glucose, so hemoglobin in our blood is glycated to some extent. Depending on the test method, there is a difference between normal values, but usually up to 5.6% is normal.

In the case of a diabetic patient, since the concentration of glucose in the blood increases, the level of glycated hemoglobin, that is, glycated hemoglobin, also increases. Therefore, the direction of future treatment is determined by looking at this result, which reveals the level of blood sugar management so far.

On the other hand, the conventional method of measuring glycated hemoglobin (HbA1c) is to obtain a capillary blood sample by collecting blood from a vein in the arm of the subject or piercing the fingertip with a small, sharp needle, and using the obtained blood to measure the concentration of glycated hemoglobin (HbA1c). was measured. This invasive glycated hemoglobin measurement method increases the burden of blood collection on measurement subjects, and has problems of providing inaccurate values when red blood cell lifespan is short, pregnancy, or renal disease.

Korean Patent Registration No. 10-0871074 (2008.11.24)

The technical problem to be achieved by the present invention is to provide a non-invasive glycated hemoglobin measuring method and apparatus capable of accurately and easily non-invasively measuring the concentration of glycated hemoglobin (HbA1c) using a learning algorithm. have.

Among the embodiments, a non-invasive glycated hemoglobin measurement method using a learning algorithm includes collecting learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients; constructing a prediction model for predicting the glycated hemoglobin level by using a learning algorithm based on the learning data; Collecting PPG signal data of a subject to be measured; and predicting the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model.

The collecting of the training data may include filtering the PPG signal using at least one of predefined quality metrics; normalizing the filtered PPG signal; and normalizing the normalized PPG signal.

Building the predictive model may include using one of 1D models including a convolutional neural network (CNN), a recursive neural network (RNN), and a CNN-RNN as the learning algorithm.

Building the predictive model may include applying any one of tanh, ReLu, Leaky ReLu, Logistic/Sigmoid, ELU, Linear, and Softmax as an activation function to the learning algorithm. can

Building the predictive model may include applying any one of Adagrad, Adadelta, Adam, and stochastic gradient descent (SGD) as an optimizer to the learning algorithm.

Building the prediction model may include independently building candidate prediction models based on learning algorithms; verifying the candidate predictive models using a plurality of quality metrics including specificity, sensitivity and area under curve (AUC); and finally determining a candidate prediction model having the most optimal performance based on a result of the verification.

Building the predictive model may include constructing a first predictive model for predicting a blood sugar level based on the learning data; and constructing a second predictive model for predicting an average blood glucose level during a unit period based on the learning data and output data of the first predictive model.

The predicting of the glycated hemoglobin level may include predicting an instantaneous blood glucose level of the measurement subject by applying the PPG signal data to the first prediction model; and predicting the average blood glucose level of the measurement subject by applying the PPG signal data and the instantaneous blood glucose level to the second prediction model.

Among embodiments, a non-invasive glycated hemoglobin measurement apparatus using a learning algorithm includes a learning data collection unit that collects learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients; a predictive model builder constructing a predictive model for predicting the HbA1c level by using a learning algorithm based on the learning data; Measurement data collection unit for collecting PPG signal data of the subject to be measured; and a glycated hemoglobin predictor for predicting the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model.

The learning data includes at least one feature extracted from the PPG signal, a first physiological state related to whether the patient is in an fasting state based on a time point at which the PPG signal is measured, and a second physiological state related to whether the corresponding patient is in a medication state state can be included.

The predictive model builder may use any one of 1D models including a convolution neural network (CNN), a recursive neural network (RNN), and a CNN-RNN as the learning algorithm.

The predictive model builder may determine a predictive model having the most optimal performance according to a verification result based on a quality metric among a plurality of predictive models each independently built based on the training data.

The disclosed technology may have the following effects. However, it does not mean that a specific embodiment must include all of the following effects or only the following effects, so it should not be understood that the scope of rights of the disclosed technology is limited thereby.

The method and apparatus for noninvasively measuring glycated hemoglobin using a learning algorithm according to an embodiment of the present invention can measure the concentration of glycated hemoglobin (HbA1c) accurately and easily and noninvasively using a learning algorithm.

1 is a diagram for explaining a non-invasive glycated hemoglobin measurement system according to the present invention.
FIG. 2 is a diagram explaining the system configuration of the glycated hemoglobin measuring device of FIG. 1 .
FIG. 3 is a view explaining the functional configuration of the glycated hemoglobin measuring device of FIG. 1 .
4 is a flowchart illustrating a method for non-invasively measuring glycated hemoglobin according to the present invention.
5 is a diagram illustrating a 1D CNN model as a learning algorithm.
6 is a diagram illustrating a comparison between A1c and blood glucose.

Since the description of the present invention is only an embodiment for structural or functional description, the scope of the present invention should not be construed as being limited by the embodiments described in the text. That is, since the embodiment can be changed in various ways and can have various forms, it should be understood that the scope of the present invention includes equivalents capable of realizing the technical idea. In addition, since the object or effect presented in the present invention does not mean that a specific embodiment should include all of them or only such effects, the scope of the present invention should not be construed as being limited thereto.

Meanwhile, the meaning of terms described in this application should be understood as follows.

Terms such as "first" and "second" are used to distinguish one component from another, and the scope of rights should not be limited by these terms. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element.

It should be understood that when an element is referred to as being “connected” to another element, it may be directly connected to the other element, but other elements may exist in the middle. On the other hand, when an element is referred to as being "directly connected" to another element, it should be understood that no intervening elements exist. Meanwhile, other expressions describing the relationship between components, such as “between” and “immediately between” or “adjacent to” and “directly adjacent to” should be interpreted similarly.

Expressions in the singular number should be understood to include plural expressions unless the context clearly dictates otherwise, and terms such as “comprise” or “have” refer to an embodied feature, number, step, operation, component, part, or these. It should be understood that it is intended to indicate that a combination exists, and does not preclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

In each step, the identification code (eg, a, b, c, etc.) is used for convenience of explanation, and the identification code does not describe the order of each step, and each step clearly follows a specific order in context. Unless otherwise specified, it may occur in a different order than specified. That is, each step may occur in the same order as specified, may be performed substantially simultaneously, or may be performed in the reverse order.

The present invention can be implemented as computer readable code on a computer readable recording medium, and the computer readable recording medium includes all types of recording devices storing data that can be read by a computer system. . Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices. In addition, the computer-readable recording medium may be distributed to computer systems connected through a network, so that computer-readable codes may be stored and executed in a distributed manner.

All terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs, unless defined otherwise. Terms defined in commonly used dictionaries should be interpreted as consistent with meanings in the context of the related art, and cannot be interpreted as having ideal or excessively formal meanings unless explicitly defined in the present application.

The first requirement of machine learning can correspond to large and varied data sets. To achieve this goal, it is necessary to collect very large data sets (e.g. around 5k might be best). In other words, as more data is applied to the learning process, the built learning model can predict more generalized results.

The second requirement of machine learning may correspond to sufficient hardware specifications to perform complex and prolific operations. Deep learning works with huge data sets, so the RAM that holds the data for training can be very important, with 32/64 GB of RAM being used in most ML research.

The third requirement of machine learning may correspond to a framework that can be selected after hardware setup. Currently, Pytorch's deep learning framework developed by Facebook Inc. is popular. When hardware, software, and framework selection are completed, a model most suitable for data among various models can be determined and used for the measurement of glycated hemoglobin.

1 is a diagram for explaining a non-invasive glycated hemoglobin measurement system according to the present invention.

Referring to FIG. 1 , the non-invasive glycated hemoglobin measurement system 100 may include a user terminal 110, a glycated hemoglobin measurement device 130, and a database 150.

The user terminal 110 may correspond to a computing device capable of providing learning data or checking a glycated hemoglobin measurement result, and may be implemented as a smart phone, a laptop computer, or a computer, but is not limited thereto, and may be implemented in various ways such as a tablet PC. It can also be implemented as a device. The user terminal 110 may be connected to the glycated hemoglobin measuring device 130 through a network, and a plurality of user terminals 110 may be simultaneously connected to the glycated hemoglobin measuring device 130 .

In one embodiment, the user terminal 110 may measure the PPG signal from the user. For example, the user terminal 110 may acquire an image of a specific body part (fingertip) of the user through a camera module, and generate a PPG signal by performing signal processing on the image. In addition, a color component and an intensity component may be selectively extracted and used from an image in a signal processing process, and a flash module may be used together when capturing an image through a camera module, if necessary. In this case, the color component may include red, blue, green, and the like.

In one embodiment, the user terminal 110 may provide a query related to PPG signal measurement from the user, and as a response, information on a first physiological state related to fasting state and a second physiological state related to taking medication can receive That is, the user may additionally input information about whether the user's current state is before or after a meal and information about whether the current state is before or after taking a medicine.

In one embodiment, the user terminal 110 may collect PPG signals by interworking with dedicated terminals for PPG signal measurement. For example, a dedicated terminal for measuring a PPG signal may include a sensor module such as an SEN or analog front end (AFE). That is, the user terminal 110 may collect the PPG signal measured by the dedicated terminal and provide it to the glycated hemoglobin measuring device 130 .

The glycated hemoglobin measuring device 130 is implemented as a server corresponding to a computer or program capable of non-invasively measuring the concentration of glycated hemoglobin by constructing a learning model for predicting the concentration of glycated hemoglobin using a learning algorithm and utilizing the learning algorithm. It can be. The glycated hemoglobin measurement device 130 may be connected to the user terminal 110 through a network to exchange information. In one embodiment, the glycated hemoglobin measuring device 130 may store data necessary for measuring the concentration of glycated hemoglobin in association with the database 150 .

The database 150 may correspond to a storage device that stores various pieces of information necessary for the operation of the glycated hemoglobin measuring device 130 . The database 150 may store learning data and learning models for machine learning, respectively, and may store various learning algorithms used in the learning process and information about their parameters, but is not limited thereto, and the glycated hemoglobin measuring device In the process of performing non-invasive glycated hemoglobin measurement (130), information collected or processed in various forms may be stored.

FIG. 2 is a diagram explaining the system configuration of the glycated hemoglobin measuring device of FIG. 1 .

Referring to FIG. 2 , the glycosylated hemoglobin measuring apparatus 130 may include a processor 210, a memory 230, a user input/output unit 250, and a network input/output unit 270.

The processor 210 may execute a procedure for processing each step in the process of the glycated hemoglobin measurement device 130 operating, and manage the memory 230 read or written throughout the process, and the memory ( Synchronization time between the volatile memory and the non-volatile memory in 230) can be scheduled. The processor 210 may control overall operations of the glycated hemoglobin measuring device 130, and is electrically connected to the memory 230, the user input/output unit 250, and the network input/output unit 270 to control data flow therebetween. can do. The processor 210 may be implemented as a central processing unit (CPU) of the glycated hemoglobin measuring device 130 .

The memory 230 is implemented as a non-volatile memory such as a solid state drive (SSD) or a hard disk drive (HDD) and may include an auxiliary memory device used to store all data necessary for the glycated hemoglobin measuring device 130. , may include a main memory implemented as a volatile memory such as RAM (Random Access Memory).

The user input/output unit 250 may include an environment for receiving user input and an environment for outputting specific information to the user. For example, the user input/output unit 250 may include an input device including an adapter such as a touch pad, a touch screen, an on-screen keyboard, or a pointing device, and an output device including an adapter such as a monitor or touch screen. In one embodiment, the user input/output unit 250 may correspond to a computing device connected through a remote connection, and in such a case, the HbA1c measurement device 130 may be implemented as an independent server.

The network input/output unit 270 includes an environment for connecting to an external device or system through a network, and includes, for example, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), and a VAN ( An adapter for communication such as Value Added Network) may be included.

FIG. 3 is a view explaining the functional configuration of the glycated hemoglobin measuring device of FIG. 1 .

Referring to FIG. 3 , the glycated hemoglobin measuring device 130 includes a learning data collection unit 310, a predictive model construction unit 330, a measurement data collection unit 350, a glycated hemoglobin prediction unit 370, and a control unit 390. can include

The learning data collection unit 310 may collect learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients. Here, the PPG signal may include transmissive and reflective data. That is, the learning data collection unit 310 may collect PPG signals of actual patients or normal people and information on the level (or concentration value) of HbA1c at the time of measurement through various sources, and based on this, an artificial intelligence model is created. You can create learning data for building. In this case, the level of glycated hemoglobin may be obtained through a hospital or a commercial medical device. Information collected from patients and normal people may include various types of biometric information, and the learning data collection unit 310 may perform a preprocessing operation to extract and process valid data necessary for generating learning data from the corresponding data.

In one embodiment, the learning data includes at least one feature extracted from the PPG signal, a first physiological state related to whether the patient is fasting or not, and whether the patient is taking medication based on a time point when the PPG signal is measured. It may include a second physiological state related to. For example, the features extracted from the PPG signal include wavelength, frequency, peak, peak interval, energy level, heart rate (HR), and heart rate variability (HRV). , pulse wave velocity (PWV, Pulse Wave Velocity) and AC/DC. In addition, the first physiological state may correspond to information about whether the patient is fasting at the time of measuring the PPG signal, and the second physiological state may correspond to information about whether the patient is in a medication state at the time of measuring the PPG signal. have.

That is, the glycated hemoglobin measuring device 130 may build a predictive model by simply learning only the patient's PPG signal, but may build a predictive model by including additional data as learning data to increase predictive accuracy. That is, the blood sugar level or glycated hemoglobin level of the measurement subject may vary depending on the fasting state or the drug-taking state, and the prediction model additionally learns various physiological information in addition to the PPG signal to obtain a better understanding of the glycated hemoglobin level according to the physical condition of the measurement subject. It can generate highly accurate prediction results.

In one embodiment, the learning data collection unit 310 may filter the PPG signal using at least one of predefined quality metrics, normalize the filtered PPG signal, and standardize the normalized PPG signal. have. The learning data collection unit 310 may independently perform a filtering operation, a normalization operation, and a standardization operation through data preprocessing operations, and may perform a composite operation as a result of the combination in a specific order, if necessary.

More specifically, various filters such as a Kalman filter and an adaptive filter may be used for filtering the PPG signal, and a predefined signal quality metric (SQM) may be used. For example, the quality metric may include various statistical or spectral characteristics of the raw PPG signal or the filtered PPG signal. In addition, the learning data collection unit 310 may perform a normalization operation on the PPG signal through Equation 1 below.

[Equation 1]

Here, x _new1 corresponds to a normalization score, x corresponds to a specific PPG signal, x _min corresponds to a minimum value of PPG signals, and x _max corresponds to a maximum value of PPG signals.

In addition, the learning data collection unit 310 may perform a standardization (or z-scoring) operation on the PPG signal through Equation 2 below.

[Equation 2]

Here, x _new2 is a standardized score, x is a specific PPG signal, mean is the average of the PPG signals, and σ is the standard deviation of the PPG signals.

In an embodiment, the learning data collection unit 310 may extract at least one feature from the PPG signal and sequentially perform filtering, normalization, and standardization operations on the at least one feature. That is, the glycated hemoglobin measurement apparatus 130 may build a prediction model for predicting the glycated hemoglobin level based on at least one feature signal extracted by the learning data collection unit 310 .

The predictive model builder 330 may build a predictive model for predicting the HbA1c level (or concentration value) by using a learning algorithm based on the learning data. That is, the predictive model builder 330 may measure the concentration of glycated hemoglobin of the subject in a non-invasive manner and with high accuracy by using an artificial intelligence technique including deep learning to predict the concentration of glycated hemoglobin. Meanwhile, various learning algorithms may be used for the predictive model, but only algorithms having good predictive performance among a plurality of learning algorithms may be selectively applied to the HbA1c measurement, if necessary.

In one embodiment, the predictive model building unit 330 may use any one of 1D models including a convolution neural network (CNN), a recursive neural network (RNN), and a CNN-RNN as a learning algorithm. That is, the predictive model builder 330 may perform an operation for selecting a predictive model, and may adopt a 1D model from among various learning models. Also, the predictive model builder 330 may adopt any one of various 1D models.

More specifically, CNNs are designed for learning in the image domain (i.e., where the pixels are spatially distributed), but can provide excellent efficiency for time domain data as well. For example, a 1D (Dimensional) CNN may be used to predict a blood glucose level (BGL), and a filtered PPG signal may be used as input data. Also, CNN may be used to classify diabetic patients and non-diabetic patients based on PPG signals.

RNNs can be used to extract features in the time domain. That is, RNN can extract temporal motion from an input sequence according to time. Meanwhile, a long short-term memory (LSTM) derived from an RNN may be used to predict a blood sugar level (BGL) at a future time point based on previous time series data. More specifically, LSTM is a type of RNN, but unlike RNN, which stores only the most recent information in that it can choose whether to remember or forget information, it can selectively store and learn long-term and short-term information, Accordingly, time series prediction performance may have a good advantage. In addition, a gated recurrent unit (GRU) may be used instead of an RNN.

CNN-RNN may correspond to a fusion structure of CNN and RNN, and based on continuous blood glucose concentration data and other physiological data (eg, age, blood pressure, eating pattern, etc.), the blood glucose level (BGL) at a future time point ) can be used to predict In addition, CNN-RNN can be used to predict diabetes based on RR-interval (eg, heart rate variability (HRV) derived from electrocardiogram (ECG)). Meanwhile, CNN-RNN may be replaced with CNN-LSTM.

In one embodiment, the predictive model building unit 330 may apply any one of tanh, ReLu, Leaky ReLu, Logistic/Sigmoid, ELU, Linear, and Softmax as an activation function to the learning algorithm. can

More specifically, the hyperbolic tangent (tanh) may provide an output in the range [-1,1], and in this case, the center of the output may correspond to 0. The slope of the hyperbolic tangent may be greater than the sigmoid. The hyperbolic tangent can be mainly used with LSTM/RNN architectures, and can be expressed as in Equation 3 below.

[Equation 3]

The rectified linear unit (ReLu) outputs 0 when the input value is less than 0, and outputs the input value as it is when it is greater than 0, and can be expressed as in Equation 4 below.

[Equation 4]

Leaky ReLu is a transform function of ReLu, which can be effective when negative numbers are included in the input, and can be expressed as in Equation 5 below.

[Equation 5]

Here, α is a controlling parameter.

Logistic/Sigmoid can provide an output in the [0,1] range, can be used as a regressor, and can be expressed as in Equation 6 below.

[Equation 6]

The exponential linear unit (ELU) outputs the input value in the positive part without refining to prevent the loss of gradient problem, and can be expressed as in Equation 7 below.

[Equation 7]

Here, α is a controlling parameter.

A linear function can be used when the output is proportional to the input, and can be expressed as f(x) = x.

Softmax can calculate the probability of each target class for all possible target classes, and in a later step, the calculated probability can be used to determine the target class for a given input, expressed as Equation 8 below: It can be.

[Equation 8]

In one embodiment, the predictive model builder 330 may apply any one of Adagrad, Adadelta, Adam, and stochastic gradient descent (SGD) as an optimizer to the learning algorithm. Here, the optimization algorithm may play a role of controlling the overall performance of the actual predictive model. Therefore, selecting an accurate and efficient optimization algorithm can be very difficult.

More specifically, Adagrad may correspond to an abbreviation of adaptive gradient, and the parameter (w) may indicate a slower learning rate with a higher gradient and a faster learning rate with a lower gradient. learning rate can be expressed. Gradient (g), learning rate (α), and weight (W) can be connected by iteration of t as shown in Equation 9 below.

[Equation 9]

은 10^-4 ~ 10^-8정도의 작은 값으로서 0으로 나누는 것을 방지하기 위한 작은 값이다.From here,

is a small value of about 10 ^-4 to 10 ^-8 and is a small value to prevent division by zero.

Adadelta may correspond to an abbreviation of Adaptive Delta, and may correspond to a variation of the stochastic gradient descent algorithm. Adadelta can use an adaptive learning rate in the learning process to reduce overfitting. Adadelta can be more powerful than Adagrad in terms of learning rate based on window movement of gradient update.

Adam may correspond to an abbreviation of adaptive moment estimation. Adam can correspond to a combination of RMSProp and momentum, and can compute an adaptive learning rate for each parameter.

Stochastic Gradient Descent (SGD) can solve the problem of cycling through a local minima for a large data set through randomly sampled samples from the large data set. Stochastic gradient descent can reach a global minimum, but the process can be slow and noisy.

In one embodiment, the predictive model builder 330 uses binary cross entropy, categorical entropy, and Kullback-Leibler as a loss function for the learning algorithm. ), Mean Squared Error (MSE), and Mean Absolute Error (Mean Absolute Error) can be applied. The loss function can vary depending on the type of output, and in general loss can be divided into two types: stochastic loss (when the output is between 0 and 1) and regression loss (when a continuous value is estimated).

More specifically, binary cross entropy can be used when only two classes exist. If y is a binary indicator and p is a predicted probability, the loss can be expressed as in Equation 10 below.

[Equation 10]

-(ylog(p) + (1 - y)log(1 - p))

In the case of categorical entropy with multiple classes (M = 2), the cross entropy loss function can be defined as in Equation 11 below.

[Equation 11]

Here, o is the probability observation of class c.

The Kullback-Leibler may be mainly used in a generative model or an autoencoder, and may be expressed as in Equation 12 below.

[Equation 12]

Here, y _true is the actual probability distribution and y _pred is the predicted probability distribution.

Mean Squared Error (MSE) can be used for regression purposes and can be expressed as loss = (y _true - y _pred ) ² .

Mean Absolute Error (Mean Absolute Error) can be used for regression analysis and can be expressed as loss = |y _true - y _pred |.

In one embodiment, the predictive model builder 330 independently builds candidate predictive models based on learning algorithms, and a plurality of qualities including specificity, sensitivity, and area under curve (AUC). Candidate prediction models may be verified using metrics, and a candidate prediction model having the most optimal performance may be finally determined based on a result of the verification. That is, the prediction model builder 330 may specify at least one quality metric, calculate performance scores of candidate prediction models according to the quality metric, and finally determine a candidate prediction model having the highest performance score. Meanwhile, the quality metric may include a true positive rate (TPR) and a false positive rate (FPR).

Specificity and sensitivity may correspond to statistical indicators for prediction. That is, the specificity may correspond to the probability that the prediction result is negative when the prediction result is negative, and the sensitivity may correspond to the probability that the prediction result is positive when the prediction result is positive. The area under curve (AUC) can provide an aggregate measure of performance across all possible classification thresholds.

In one embodiment, the prediction model builder 330 builds a first prediction model for predicting the blood sugar level based on the learning data, and based on the learning data and the output data of the first prediction model, the average for a unit period. A second prediction model for predicting the blood sugar level may be constructed. That is, the predictive model builder 330 may configure a learning model for predicting the HbA1c level step by step. For example, a predictive model built in advance by the predictive model builder 330 may include a plurality of sub-models and may include first and second predictive models. The first prediction model may predict the instantaneous blood sugar level based on the PPG signal, and the second prediction model may predict the average blood sugar level for a unit period based on the PPG signal and the instantaneous blood sugar level predicted by the first prediction model. have.

The measurement data collection unit 350 may collect PPG signal data of the measurement target. The measurement data collection unit 350 may collect PPG signal data of the measurement target in various ways, and may collect both transmission type and reflection type data. For example, PPG signal data may be measured through the user terminal 110 of the measurement target, and the corresponding data may be transmitted from the user terminal to the measurement data collection unit 350 . Meanwhile, the measurement data collection unit 350 may perform filtering, normalization, and standardization operations on PPG signal data in order to improve the accuracy of measurement results.

The glycated hemoglobin predictor 370 may predict the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model. That is, the glycated hemoglobin predictor 370 may convert the PPG signal data of the measurement subject according to the input data format of the prediction model, and predict the glycated hemoglobin level (or concentration value) of the measurement subject based on the output of the prediction model. can

In one embodiment, the glycated hemoglobin predictor 370 predicts the instantaneous blood sugar level of the measurement subject by applying the PPG signal data to the first prediction model, and measures the measurement by applying the PPG signal data and the instantaneous blood sugar level to the second prediction model. A subject's average blood sugar level can be predicted. A predictive model for predicting the level of glycated hemoglobin may be composed of a plurality of sub-models, and first and second predictive models may be built in advance by the predictive model builder 330 . In this case, the glycated hemoglobin predictor 370 may predict the glycated hemoglobin level by sequentially applying the first and second prediction models to the PPG signal data obtained from the measurement subject.

The control unit 390 controls the overall operation of the glycated hemoglobin measurement device 130, and includes the learning data collection unit 310, the prediction model construction unit 330, the measurement data collection unit 350, and the glycated hemoglobin prediction unit 370. It can manage control flow or data flow between

4 is a flowchart illustrating a method for non-invasively measuring glycated hemoglobin according to the present invention.

Referring to FIG. 4 , the glycated hemoglobin measurement apparatus 130 may collect learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients through the learning data collection unit 310 (step S410). The glycated hemoglobin measurement apparatus 130 may build a prediction model for predicting the glycated hemoglobin level by using a learning algorithm based on the learning data through the predictive model builder 330 (step S430). In addition, the glycated hemoglobin measurement device 130 may collect PPG signal data of the measurement subject through the measurement data collection unit 350 (step S450). The glycated hemoglobin measuring device 130 may predict the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model through the glycated hemoglobin predictor 370 (step S470).

In one embodiment, the glycated hemoglobin measuring device 130 may measure the glycated hemoglobin level by utilizing a state model together with an artificial intelligence model. Here, the state model may perform an operation of classifying the current state of the measurement target into any one of a plurality of predefined states by utilizing information (eg, physiological information) collected from the measurement target. have. That is, the glycated hemoglobin measuring device 130 may predict the current state through a state model based on the information collected from the measurement subject, and measure the glycated hemoglobin level through an artificial intelligence model based on the current state. Through this, the glycated hemoglobin measuring device 130 may reflect various conditions of the measurement subject to the glycated hemoglobin level measurement, and may provide a more accurate measurement value.

In one embodiment, the glycated hemoglobin measuring device 130 may build an independent predictive model for each state classified by the state model, and perform a first analysis step of performing state classification through the state model and the predictive model. The glycated hemoglobin level of the subject to be measured may be measured by sequentially performing the second analysis step of measuring glycated hemoglobin through the glycated hemoglobin level. To this end, the glycated hemoglobin measurement apparatus 130 may pre-classify learning data for each state classified by the state model, and build a prediction model for each state by learning only the classified learning data.

5 is a diagram illustrating a 1D CNN model as a learning algorithm.

Referring to FIG. 5 , the glycated hemoglobin measuring apparatus 130 may non-invasively measure the glycated hemoglobin level of the measurement subject using the 1D CNN model 510 . At this time, in the 1D CNN model 510, the input size may correspond to 1 * k (k is the number of samples in one frame). Also, ReLu and Adam can be used as activation functions and model optimization algorithms.

In FIG. 5, N may correspond to the number of filters and may increase to 2 ^m , such as 8, 16, 32, or 64. m (where m is a positive integer) may correspond to the size of the filter, typically the length of the window, and may correspond to 3*3 or 5*5 in the 2D model. In the case of a 1D model, the height of the filter may be regarded as the size of the filter, and may have a value between 1 and 3, but is not necessarily limited thereto.

s may correspond to the size of a stride, and may typically correspond to 1. p may correspond to the size of padding, and zero padding or other padding techniques may be used. c may correspond to the number of neurons in a dense layer, where c = 1 in the last dense layer of a regression problem, and c = no. of classes in the last dense layer of a classification problem. .

In addition, the batch size may correspond to 32, and the initial learning rate may correspond to 0.001 to 0.01. Meanwhile, the learning rate may vary according to the number of iterations. The loss function may correspond to mean absolute error (MAE) in the case of regression, and may correspond to categorical cross entropy in the case of classification. The composition of the data set may consist of 80% for learning, 10% for verification, and 10% for testing, and the corresponding ratio may be applied differently if necessary.

In addition, the glycated hemoglobin measuring apparatus 130 may apply a dropout layer having a probability of 0.2 to a place having a high parameter number in order to solve an overfitting problem. The glycated hemoglobin measurement device 130 may prevent overfitting by utilizing batch normalization (BN).

6 is a diagram illustrating a comparison between A1c and blood glucose.

Referring to FIG. 6 , the glycated hemoglobin measuring device 130 may measure the glycated hemoglobin level of the measurement subject using a pre-constructed prediction model. Meanwhile, the glycated hemoglobin measuring device 130 may determine whether or not the glycated hemoglobin is diabetic by using the predicted glycated hemoglobin level. According to the comparison table between A1c and blood sugar in FIG. 6, 4% to 5.6% corresponds to the normal range, 5.7% to 6.4% corresponds to the pre-diabetes stage, and 6.5% or more corresponds to diabetes. can do. That is, the glycated hemoglobin measurement apparatus 130 may measure the glycated hemoglobin level through a predictive model, and generate a classification result regarding whether or not the glycated hemoglobin is diabetic according to the glycated hemoglobin level.

Although the above has been described with reference to preferred embodiments of the present invention, those skilled in the art will variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the claims below. You will understand that it can be done.

100: non-invasive glycated hemoglobin measurement system
110: user terminal 130: glycated hemoglobin measuring device
150: database
210: processor 230: memory
250: user input/output unit 270: network input/output unit
310: learning data collection unit 330: prediction model building unit
350: measurement data collection unit 370: glycated hemoglobin prediction unit
390: control unit
510: 1D CNN model

Claims (12)

In the non-invasive glycated hemoglobin measurement method using a learning algorithm performed in a non-invasive glycated hemoglobin measuring device using a learning algorithm including a learning data collection unit, a prediction model construction unit, a measurement data collection unit, and a glycated hemoglobin prediction unit,
Through the learning data collection unit, learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients is collected, and the learning data includes at least one feature extracted from the PPG signal based on each of transmission type and reflection type data ( feature) is included, while additional data on fasting and taking medications is not included;
constructing a prediction model for predicting the level of glycated hemoglobin using a learning algorithm based on the learning data through the prediction model building unit;
Collecting PPG signal data of a measurement target through the measurement data collection unit; and
Predicting the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model through the glycated hemoglobin prediction unit,
The step of building the prediction model includes using any one of 1D models including a convolution neural network (CNN), a recursive neural network (RNN), and a CNN-RNN as the learning algorithm,
Building the prediction model may include independently building candidate prediction models based on learning algorithms composed of the 1D models; verifying the candidate predictive models using a plurality of quality metrics including specificity, sensitivity and area under curve (AUC); And finally determining a candidate prediction model with the most optimal performance based on the verification result,
The step of estimating the level of glycated hemoglobin comprises the step of determining whether or not the measurement subject has diabetes using the predicted glycated hemoglobin level.
The method of claim 1, wherein collecting the learning data
filtering the PPG signal using at least one of predefined quality metrics;
normalizing the filtered PPG signal; and
A non-invasive glycated hemoglobin measurement method using a learning algorithm, comprising the step of normalizing the normalized PPG signal.
delete The method of claim 1, wherein the step of building the predictive model
Applying any one of tanh, ReLu, Leaky ReLu, Logistic/Sigmoid, ELU, Linear and Softmax as an activation function to the learning algorithm using a learning algorithm comprising the step of applying Non-invasive glycated hemoglobin measurement method.
The method of claim 4, wherein the step of building the predictive model
A non-invasive glycated hemoglobin measurement method using a learning algorithm comprising the step of applying any one of Adagrad, Adadelta, Adam, and stochastic gradient descent (SGD) as an optimizer to the learning algorithm.
delete The method of claim 1, wherein the step of building the predictive model
constructing a first predictive model for predicting a blood glucose level based on the learning data; and
Non-invasive glycated hemoglobin measurement using a learning algorithm comprising the step of constructing a second predictive model for predicting the average blood glucose level for a unit period based on the learning data and the output data of the first predictive model. Way.
The method of claim 7, wherein the step of predicting the level of glycated hemoglobin
predicting an instantaneous blood sugar level of the measurement subject by applying the PPG signal data to the first predictive model; and
The non-invasive glycated hemoglobin measurement method using a learning algorithm comprising the step of predicting the average blood glucose level of the measurement subject by applying the PPG signal data and the instantaneous blood glucose level to the second prediction model.
Learning data including PPG signals and glycated hemoglobin (HbA1c) levels of patients is collected, and the learning data includes at least one feature extracted from the PPG signal based on each of the transmission data and the reflection data. a learning data collection unit that does not include additional data on whether or not you are in a state and medication state;
a predictive model builder constructing a predictive model for predicting the HbA1c level by using a learning algorithm based on the learning data;
Measurement data collection unit for collecting PPG signal data of the subject to be measured; and
A glycated hemoglobin predictor for predicting the glycated hemoglobin level of the measurement subject by applying the PPG signal data to the prediction model,
The predictive model building unit uses one of 1D models including a convolution neural network (CNN), a recursive neural network (RNN), and a CNN-RNN as the learning algorithm,
independently constructing candidate prediction models based on learning algorithms composed of the 1D models by the prediction model building unit; verifying the candidate predictive models using a plurality of quality metrics including specificity, sensitivity and area under curve (AUC); and finally determining a candidate prediction model having the most optimal performance based on a result of the verification,
The non-invasive glycated hemoglobin measuring device using a learning algorithm, characterized in that the glycated hemoglobin prediction unit determines whether or not the measurement subject has diabetes using the predicted glycated hemoglobin level.
The method of claim 9, wherein the learning data
At least one feature extracted from the PPG signal, a first physiological state related to whether the patient is in an empty stomach state based on a time point when the PPG signal is measured, and a second physiological state related to whether the patient is in a medication state A non-invasive glycated hemoglobin measuring device using a learning algorithm, characterized in that.
delete delete

Images