KR20200010991A - Apparatus and method for determining blood glucose concentration - Google Patents

Abstract

The present invention relates to a non-invasive method for determining blood glucose concentration. The method for determining blood glucose concentration according to an embodiment of the present invention can comprise the steps of: processing a near infrared spectrum and a pure spectrum to obtain a preprocessed near infrared spectrum; extracting a key feature set from the preprocessed near infrared spectrum; obtaining a plurality of homogenized feature sets of validation data; and determining the blood glucose concentration of a subject by using homogenized feature sets of training data and validation data for a plurality of subjects.

Description

Apparatus and method for determining blood glucose levels {APPARATUS AND METHOD FOR DETERMINING BLOOD GLUCOSE CONCENTRATION}

The present invention relates to a blood glucose monitoring technique, and more particularly, to a technique for performing general-purpose calibration to determine blood glucose concentration non-invasively.

Continuous Glucose monitoring (CGM) is used to check blood glucose levels in subjects at regular intervals. Continuous blood glucose monitoring can be performed invasive or non-invasive. In invasive methods, the skin of the human body is pierced to obtain a blood sample for examination.

However, non-invasive methods do not require the collection of blood samples to obtain blood glucose levels. Common methods used for non-invasive glucose monitoring include mid-infrared (Mid IR), near infrared (NIR) and Raman spectroscopy. Recently, near-infrared methods are commonly used for continuous glucose monitoring. Here, the infrared (IR) wavelength passes through the skin, and the absorption of the infrared wavelength by the subcutaneous portion of the skin helps to determine the glucose level.

The absorption of wavelength A by the sample is defined by the BEER Lambert law.

…(1)

… (One)

는 흡수계수이고,

는 샘플에서의 성분의 농도이며,

는 침투 깊이(penetration depth)를 나타낸다.here

Is the absorption coefficient,

Is the concentration of the component in the sample,

Denotes the penetration depth.

If the sample consists of different components, the total absorption is given by

…(2)

… (2)

The near-infrared absorption spectrum of skin consists of the absorption of several components such as water, fat, proteins (collagen and keratin), amino acids, elastin and glucose.

Non-invasive monitoring of blood glucose levels is very difficult because the concentration of glucose in the blood is several times lower than that of other components. Glucose information may be buried beneath the noise and drift components of the near infrared spectrum. The approximate sequence of different component concentrations is shown in the table below.

Constituent Water Fat Protein Elastin / Acid Glucose Order of concentration (~) 10 ⁰ 10 ^-1 10 ^-3 10 ^-3 10 ^-4

Existing blood glucose prediction mechanisms deal with measuring blood glucose using only specific IR spectroscopy, including a water absorption peak, and use transmissive electromagnetic (EM) radiation from the skin to the measurement region, such as blood vessels. Blood glucose is calculated by analyzing the collected light / EM and comparing it with a stored reference calibration curve. However, existing glucose determination mechanisms assume that background interference is common to all ranges of the near infrared region. These methods use different reference calibration curves depending on the subject, and thus universality is not guaranteed.

Other existing glucose determination mechanisms measure components in the blood via near infrared region (NIR) spectroscopy. It uses a spectrum substraction generator to generate a plurality of spectral subtractions from the spectra measured by the spectrum analyzer at different time points. Multiple regression models based on Partial Least Squares Regression (PLS) or Principal Component Regression (PCR) are used. However, existing glucose determination mechanisms do not distinguish between different blood components such as glucose / fat / lipids.

Other existing glucose determination mechanisms use near infrared region (NIR) spectroscopy and Monte Carlo simulations to remove noise from single test data. This method yields good results for same-day validation but poor performance for other date validations.

Disclosed is an apparatus and method for performing a calibration to determine blood glucose levels non-invasively.

According to an aspect of the present invention, there is provided a method for determining a blood glucose level by processing a near-infrared spectrum including at least one of training data of a plurality of subjects, calibration data of a subject, and verification data of a subject by a preprocessing unit. Obtaining a preprocessed near infrared spectrum, extracting a dominant feature set from the preprocessed near infrared spectrum by a feature set extractor, and extracting a plurality of homogenized feature sets of the verification data by a feature homogenizer. The method may include obtaining the blood glucose level by using the training data for the plurality of subjects and the homogenized feature set of the verification data by the obtaining and ensemble learners.

Acquiring the pre-processed near infrared spectrum may include: filtering the near infrared spectrum to remove unwanted components from the frequency domain by a frequency domain filtering unit; Acquiring the step, by the EMSC (Extended Multiplicative Scatter Correction unit), the step of removing components other than glucose and water in the near infrared spectrum by using the orthogonal pure spectrum and the time of the near infrared spectrum by the drifted remover Removing drift present in the domain.

The filtering of the near infrared spectrum may include filtering out noise present in the near infrared spectrum using a Savitzky-Golay (SG) filter and differentiating the near infrared spectrum with respect to the wavelength to remove linear drift in the wavelength domain of the near infrared spectrum. It may include the step.

Extracting a set of key features from the pre-processed near infrared spectra, calculating the correlation between the glucose concentration of the training data and each of the pre-processed near infrared spectra in the training data and based on the maximum value of the correlation calculated from the training data. And selecting the primary feature set comprising at least one pre-processed near infrared spectrum corresponding to each of the training data, calibration data and verification data.

Acquiring a plurality of homogenized feature sets of the validation data includes obtaining a single homogenized feature set of the validation data, homogenized with respect to the primary feature set of the training data and the calibration data of the single subject, and And repeating the steps to obtain a plurality of homogenized feature sets.

Acquiring a single homogenized feature set of the validation data comprises obtaining a linear approximation relationship for each feature in the primary feature set with respect to the corresponding glucose concentrations of the training data and the calibration data, maintaining the training data as a reference. Calculating calibration factors for each feature in the primary feature set for calibration data and mapping each feature in the primary feature set of the validation data in a manner that is subtracted from the corresponding adjustment factor to form a single homogenized version of the validation data. Obtaining a feature set.

Computing the adjustment factor for each feature in the key feature set includes calculating an average glucose value for the training data, using a linear approximation relationship at each mean in the key feature set of training data and calibration data at the mean glucose value. Calculating near infrared feature values for the feature, and subtracting the near infrared feature value of the calibration data from the near infrared feature value of the training data to obtain an adjustment factor for each feature in the primary feature set.

Determining the blood glucose level of the subject includes training an ensemble of each regression model trained separately from the main feature set in a separate subject's training data, and inputting each of the regression models to the input of the validation data. Calculating a plurality of regression outputs from an ensemble of regression models taken as a corresponding homogenized feature set, using a plurality of weights of a meta-classifier for each of the regression models using training data of a plurality of subjects. Estimating, and determining the blood glucose concentration of the subject by weighted averaging the plurality of regression outputs based on the weights of the meta-classifier.

Estimating a plurality of weights of the meta-classifier for each of the regression models comprises: generating a plurality of weights from the ensemble of the regression models taking input for each of the regression models as a set of training data of a plurality of subjects. Calculating the model outputs, calculating the correlation between the plurality of model outputs and the glucose concentration of the training data of the plurality of subjects, calculating the plurality of model correlations and scaling the plurality of model correlations, thereby regression model Obtaining a plurality of weights of the meta-classifier corresponding to each of the two.

According to one aspect, the pre-processing unit to obtain a pre-processed near infrared spectrum by processing a near-infrared spectrum including at least one of training data of the plurality of subjects, calibration data of the subject and verification data of the subject, and pure spectrum; A feature set extractor for extracting a dominant feature set from a pre-processed near infrared spectrum, a feature homogenizer for obtaining a plurality of homogenized feature sets of verification data, and training data of a plurality of objects and homogenization of the verification data It may include an ensemble learning unit for determining the blood glucose concentration of the subject using the set feature.

Pre-processing unit is a frequency domain filtering unit for filtering out the near-infrared spectrum to remove unwanted components in the frequency domain, a data whitening unit for obtaining an orthogonal pure spectrum after the transformation is applied to the pure spectrum, present in the near-infrared spectrum using an orthogonal pure spectrum It may include an EMSC portion for removing components other than glucose and water and a drift removal portion for removing drift existing in the time domain of the near infrared spectrum.

The frequency domain filtering unit may filter out noise present in the near infrared spectrum using a Savitzky-Golay (SG) filter, and may remove linear drift present in the wavelength domain of the near infrared spectrum by differentiating the near infrared spectrum with respect to the wavelength.

The feature set extracting unit calculates a correlation between the glucose concentration of the training data and each of the pre-processed near infrared spectra in the training data, and based on the maximum value of the correlation calculated from the training data, the training data, the calibration data and The main feature set may be selected to include at least one pre-processed near infrared ray corresponding to each of the verification data.

The feature homogenizer repeats the process of obtaining a single homogenized feature set of the verification data, which is homogenized with respect to the training data of the single subject and the main feature set of the calibration data, to obtain a plurality of homogenized feature sets for the verification data. can do.

The feature homogenizer obtains a linear approximation relationship for each feature in the main feature set with respect to the corresponding glucose concentrations of training data and calibration data, and maintains the training data as a reference in the key feature set. By calculating the adjustment factor for each feature and mapping each feature in the main feature set of the validation data in a subtracted manner from the corresponding adjustment factor, a single homogenized feature set for the validation data can be obtained.

The feature homogenizer calculates an average glucose value for training data, calculates near infrared feature values for each feature in the main feature set of the training data and the calibration data at the average glucose value using a linear approximation relationship, and trains The adjustment factor for each feature in the main feature set can be obtained by subtracting the near infrared feature value of the calibration data from the near infrared feature values of the data.

The ensemble learning unit trains an ensemble of each of the regression models trained separately from the main feature set in the separate subject training data, and takes the input to each of the regression models as the corresponding homogenized feature set of the validation data. Compute a plurality of regression outputs from the ensemble of regression models, estimate the plurality of weights of the meta-classifier for each of the regression models using the training data of the plurality of subjects, and the weight of the meta-classifier The blood glucose concentration of the subject can be determined by weighted average of the plurality of regression outputs based on the results.

The ensemble learning unit calculates a plurality of model outputs from an ensemble of regression models taking an input for each of the regression models as a set of training data of the plurality of subjects, and calculates the glucose concentration of the plurality of model outputs and the training data of the plurality of subjects. Computing a plurality of model correlations by calculating a correlation therebetween, and scaling a plurality of model correlations, can obtain a plurality of weights of the meta-classifier corresponding to each of the regression models.

The foregoing generally describes various aspects of the invention, and the following detailed description will be provided for better understanding. In this regard, it should be understood that the embodiments herein are not limited to the methods of use or applications described and illustrated herein. Any other advantages and objects of the embodiments become apparent or apparent from the detailed description or examples contained herein are within the scope of the embodiments disclosed herein.

Calibration can be performed to determine blood glucose levels non-invasively.

Embodiments herein are shown in the accompanying drawings, wherein like reference numerals designate corresponding parts in the various figures. Embodiments of the present specification will be better understood from the following description with reference to the drawings.
1 is a block diagram illustrating various units of an electronic device for calibrating and predicting a blood sugar level, according to embodiments.
2A is a block diagram illustrating the flow of training and calibration data for calibrating and predicting blood glucose levels, in accordance with embodiments.
2B is a block diagram illustrating the flow of verification data for predicting blood glucose concentration, in accordance with embodiments.
3A and 3B are plots illustrating feature homogenization of verification data that maintains training data as a reference, in accordance with embodiments.
4 is a plot showing a linear approximation to training data and calibration data for a 109th feature (part of key feature set), according to embodiments.
FIG. 5 is a plot showing an adjustment factor for the 109th feature of calibration data, in accordance with embodiments, wherein the adjustment factor is calculated for calibration data maintained relative to the training data.
FIG. 6 is a plot showing homogenized 109 th feature of the validation data and calibration data, in accordance with embodiments, wherein the 109 th feature is homogenized over the training data.
FIG. 7 is a block diagram illustrating an ensemble learner for determining a blood glucose level of a subject, according to example embodiments.

Embodiments, various features, and advantageous details thereof, are described more fully with reference to the non-limiting embodiments shown in the accompanying drawings and described in the following detailed description. Descriptions of known components and processing techniques have been omitted so as not to unnecessarily obscure the embodiments herein. The description herein is merely intended to facilitate understanding of how the exemplary embodiments of the present disclosure may be practiced and to enable those skilled in the art to practice the exemplary embodiments of the present disclosure. Accordingly, the present disclosure should not be construed as limiting the scope of the exemplary embodiments herein.

Embodiments relate to a method and apparatus for performing a calibration to determine blood glucose levels non-invasively. The method for determining blood glucose concentrations includes processing the near infrared (NIR) spectrum and the pure spectrum to obtain a pretreated near infrared spectrum. In this case, the near infrared spectrum may include at least one of training data of a plurality of objects, calibration data of a subject, and verification data of the subject. The method of determining blood glucose concentration also includes extracting a dominant feature set from the pre-processed near infrared spectrum. In addition, the method of determining blood glucose concentration may include obtaining a plurality of homogenized feature sets for the validation data. The method may also include determining a blood glucose level of the subject using the homogenized feature sets of the training data and the validation data of the plurality of subjects.

1 through 7, like reference numerals denote corresponding features throughout the drawings.

As used herein, 'subject' is not limited, but may be any living thing or part of life that contains blood in the body, such as humans, animals, birds, aquatic organisms, and the like.

1 is a block diagram illustrating various configurations of an electronic device for calibrating and predicting a blood sugar level, according to embodiments. In one embodiment, the electronic device 100 includes at least one mobile phone, smartphone, tablet, phablet, personal digital assistant, wearable computing device, Internet of Things (IOT) device, glucose monitoring device. , At least one of a glucometer or other electronic devices, but is not limited thereto.

Embodiments may provide a method and an electronic device for performing a calibration to non-invasively determine a blood glucose level of a subject.

The electronic device 100 may include a preprocessor 110, a feature set extractor 120, a calibration unit 130, and a memory 140. In addition, the calibration unit 130 may include a feature homogenizer 132 and an ensemble learner 134.

The preprocessor 110 may process the near infrared spectrum. In this case, the near infrared spectrum may include at least one of training data for a plurality of objects, calibration data of a subject, and verification data of the subject. In one embodiment, the near infrared spectra and pure spectra are pre-processed near infrared spectra by removing noise components, time variant drift components, frequency variant drift components, and other components present in the near infrared spectrum. Can be processed to obtain.

The feature set extractor 120 may extract a dominant feature set from the pre-processed NIR spectrum. In this case, the main feature set may include at least one pre-processed near infrared spectrum corresponding to each of the training data, the calibration data, and the verification data.

The preprocessor 110 does not remove any wavelength or any time sample present in the near infrared spectrum. In one example, the preprocessing unit 110 filters the time sample, and thus the pre-processed NIR spectrum includes 129 time samples, in which 129 wavelengths may exist in the NIR spectrum. Since each wavelength can be considered a feature, the near infrared spectrum will have 129 features.

In addition, the feature homogenizer 132 may obtain a plurality of homogenized feature sets for the verification data. At this time, each feature set may be homogenized with respect to the main feature set in the training data of the distinct object and the main feature set in the calibration data of the subject. In one embodiment, a method of obtaining a plurality of homogenized feature sets for validation data comprises taking a single homogenized feature set of homogenized validation data for a single feature set of training data and calibration data for a single object. It may include the step of obtaining. In addition, the method of obtaining a plurality of homogenized feature sets may repeat the above process to obtain a plurality of homogenized feature sets with respect to the verification data.

In addition, the ensemble learner 134 may determine a blood sugar level of a subject by training an ensemble of regression models trained separately with main feature sets of separate target training data. The method of determining the present blood glucose concentration may also include calculating a plurality of regression outputs from an ensemble of regression models that take a corresponding homogenized feature set of the validation data as input to each of the regression models. In addition, the method of determining the present blood glucose concentration may include estimating a plurality of weights of a meta-classifier for each regression model using training data of a plurality of subjects. In addition, the method of determining the present blood glucose concentration may include determining the blood sugar concentration of the subject by weighted averaging the plurality of regression outputs based on weights of the meta classifier.

The memory 140 may store training data, pure spectrum, calibration data, verification data, and the determined glucose concentration of the subject. Memory 140 may include one or more computer readable storage media. Memory 140 may include non-volatile storage elements. Examples of such nonvolatile storage elements may include the form of magnetic hard disks, optical disks, floppy disks, flash memory, or electrically programmable memory (EPROM) or electrically erasable and programmable memory (EEPROM). In addition, the memory 140 may be considered as a non-transitory storage medium in some examples. The term “non-transitory” may indicate that the storage medium is not implemented with a carrier wave or a propagated signal. However, the term "non-transitory" should not be interpreted to mean that the memory 140 is immovable. In some examples, memory 140 may be configured to store a greater amount of information than memory. In certain instances, non-transitory storage media may store data that may change over time (eg, in random access memory (RAM) or cache).

Although FIG. 1 illustrates an exemplary configuration of an electronic device 100, it is to be understood that other embodiments are not limited to the electronic device 100. In other embodiments, the electronic device 100 may include fewer or more units. In addition, the label and name of each structure are used for the purpose of description only, and do not limit the scope of the present invention. One or more units may be combined together to perform the same or substantially similar function in the electronic device 100.

2A is a block diagram illustrating the flow of training and calibration data for calibrating and predicting blood glucose levels in accordance with embodiments. Embodiments provide an electronic device 100 that performs universal calibration to non-invasively determine blood glucose levels. The electronic device 100 includes a preprocessor 110, a feature set extractor 120, and a calibration unit 130.

The preprocessing unit 110 may include a frequency domain filtering unit 112, a data whitening unit 114, an extended multiplicative scatter correction (EMSC) unit 116, and a drift removing unit 118. In addition, the calibration unit 130 may include a feature homogenizer 132 and an ensemble learner 134.

The preprocessor 110 may receive a near infrared (NIR) spectrum from a plurality of objects. The near infrared spectrum may include at least one of training data of a plurality of objects, calibration data of a subject, and verification data of the subject.

The frequency domain filtering unit 112 may remove unwanted components such as noise and frequency change drift components in the frequency domain. The frequency domain filtering unit 112 may filter noise present in the near infrared spectrum by using a Savitzky-Golay (SG) filter. The derivative to the wavelength of the near infrared spectrum can eliminate the linear drift present in the wavelength domain of the near infrared spectrum.

, 정규화된 순수 스펙트럼을

라고 가정하면 고유 벡터 E 및 D는 다음을 사용하여 계산될 수 있다.The data whitening unit 114 may obtain an orthogonal pure spectrum after applying the transform to the pure spectrum. The data whitening unit 114 may include an average and standard deviation normalization of the pure spectrum, calculate an eigenvector of the normalized pure spectrum, and apply a whitening transform to obtain an orthogonal pure spectrum. Pure spectrum

Normalized pure spectrum

Eigenvectors E and D can be computed using

를 얻기 위한 화이트닝 변환은 다음과 같이 주어질 수 있다.Orthogonal Pure Spectrum

The whitening transform to obtain is given by

과 동일한 화이트닝된 근적외선 스펙트럼 데이터의 공분산 행렬

을 만들 것이다.Also, data whitening

Covariance matrix of whitened near infrared spectral data equal to

Will make

가 직교이며 서로의 투영(projection)이 0임을 보장할 수 있다.This is the whitened near infrared spectrum

We can guarantee that is orthogonal and that the projections of each other are zero.

Once noise and drift are removed from the near infrared spectrum and the orthogonalized pure spectrum, the near infrared spectrum and the orthogonal pure spectrum can be further provided to the EMSC unit 116.

를 다른 혈액 성분들에 대한 다양한 순수 스펙트럼

를 포함하는 임의의 근적외선 스펙트럼이라 하면,

는 아래와 같이 임의의 파장에서 단순 선형 회귀분석(simple linear regression)을 통해 해결될 수 있다.The EMSC unit 116 may remove components other than glucose and water present in the near infrared spectrum using the orthogonal pure spectrum. The EMSC unit 116 may apply an extended multiplicative scatter correction (EMSC) method to the near infrared spectrum, and may regress the components in the near infrared spectrum using an orthogonal pure spectrum. For example,

Various pure spectra for different blood components

Any near infrared spectrum including

Can be solved through simple linear regression at any wavelength as follows.

는 혈액 성분의 강도(strength)이고

는 DC 성분이다.here

Is the strength of the blood components

Is a DC component.

을 포도당 스펙트럼이라 하면, 포도당 스펙트럼은 특정 스펙트럼에서 다른 성분들을 차감하여 획득할 수 있다.

When the glucose spectrum, the glucose spectrum can be obtained by subtracting other components from a specific spectrum.

Other components include, but are not limited to, for example, absorption water, absorption fat, absorption collagen, absorption keratin, absorption acid, and the like.

The near infrared spectrum may include only glucose components and drift components in the time domain. The drift removing unit 118 may remove drift existing in the time domain of the near infrared spectrum. Since the near infrared spectrum may now include only glucose components, the near infrared spectrum may be provided to the feature set extractor 120.

The feature set extractor 120 may extract a main feature set from the pre-processed near infrared spectrum. The main feature set may include at least one preprocessed near infrared spectrum corresponding to each of the training data, calibration data, and verification data. The output of the preprocessor 110 may be a pre-processed near infrared spectrum, and the pre-processed near infrared spectrum may include a near infrared spectrum at a predefined number of wavelengths (in this example, the number of wavelengths is 129). The preprocessor 110 may maintain 129 wavelengths each time with samples captured in the near infrared spectrum. Since each wavelength may be considered a feature, the near-infrared spectrum processed as described above may have the same number of features as the number of wavelengths.

For example, the wavelength range of the near infrared spectrum may be 1000 to 2400 nm. Assume that one of the features can be the spectrum captured at 1200 nm. The main feature set extracted by the feature set extractor 120 will select several features from the total number of wavelengths (129 in this example). These features are selected based on the calculated correlation between the glucose concentration of the training data and each feature of the training data. These features / wavelengths form the main feature set and are commonly maintained for training data, calibration data and verification data.

In addition, the feature set extractor 120 may divide the NIR data into two sets. In this case, one may be training data and calibration data and the other may be verification data. Training data and calibration data may be provided to the feature homogenizer 132.

The feature homogenizer 132 may calculate a correlation between the glucose concentration of the training data and each of the pre-processed near infrared spectra in the training data. The feature homogenizer 132 may repeat calculating correlations to obtain a plurality of homogenized feature sets with respect to the verification data.

The step of obtaining the single uniform feature set of the verification data by the feature homogenizer 132 may obtain a linear approximation relationship for each feature of the main feature set with respect to the corresponding glucose concentrations of the training data and the calibration data. . In addition, obtaining a single uniform feature set may calculate an adjustment factor for each feature of the primary feature set relative to the calibration data maintained based on the training data. In addition, obtaining a single uniform feature set may map each feature of the primary feature set of the validation data in a manner that is subtracted from the corresponding adjustment factor to obtain a single homogenized feature set of the validation data.

The ensemble learner 134 may divide the main feature set of the training data into different subsets based on the individual target training data to train the ensemble of the regression models. This partitioning can be performed based on individual subject training.

For example, the overall data may include the near infrared spectrum and the glucose concentrations of Subject 1, Subject 2, Subject 3, and Subject 4 corresponding thereto. Subject 1, Subject 2, and Subject 3 may provide training data, and Subject 4 may be a subject. Thus, the training data may be the near-infrared spectrum and corresponding glucose concentrations of the subjects 1, 2 and 3 corresponding thereto. The calibration data will consist of several initial time samples of the subject (in this example, subject 4) near infrared spectra (eg, ˜4-5 may be sufficient) and corresponding glucose concentrations. The validation data will consist of the remaining time samples of the subject (in this example, subject 4) near infrared spectrum and the corresponding glucose concentration.

The separation of training data needed to train the ensemble of the regression model may be based on different data for each subject (in this example, subject 1, subject 2, subject 3) contributing to the training data, as shown in FIG. 2A. Can be. It can be seen that the homogenization of the verification data in the feature homogenizer 132 is performed several times based on the division of the training data (3 in this example). This may lead to the generation of a plurality of homogenized feature sets, each of which keeps separate, separate training data as a reference. In the examples herein, the data of subject 4 will be homogenized separately for the data of subject 1, subject 2 and subject 3. The plurality of homogenized feature sets of the validation data can be used as input to the corresponding trained regression model to calculate the plurality of regression outputs as shown in FIG. 2B. As shown in Figs. 2A and 2B, the paths to the feature set extractor are the same for the training data, the calibration data and the verification data.

와 시간 t에서 훈련 데이터

의 주요 특징 셋의 임의의 특징 및 포도당 농도

에 대하여, 검증 데이터

의 대응하는 특징은 도 3에 도시된 바와 같이 훈련 데이터에 대해 시프트될 수 있다. 도 3a는

에 대해 플롯된

를 나타내며, 이는 결과적으로 도 3b와 같이 훈련 데이터

에 대해 매핑된다. 도 3b에 도시된 매핑된 데이터는 더 나은 포도당 예측을 제공한다.3A and 3B are plots illustrating feature homogenization of validation data that maintains training data as a reference in accordance with embodiments. In one embodiment, the feature homogenizer 132 may obtain a plurality of homogenized feature sets of the verification data. Each of the plurality of homogenized feature sets is homogenized with respect to the main feature set of the individual subject's validation data and the subject's training feature set. For example, wavelength

And training data at time t

Random Features and Glucose Concentrations of Three Main Features

Verification data

The corresponding feature of may be shifted with respect to the training data as shown in FIG. 3. 3a

Plotted against

This results in training data as shown in Figure 3b.

Is mapped to. The mapped data shown in FIG. 3B provides better glucose prediction.

The following notations are used throughout the specification to indicate the following.

는 참조 벡터(훈련 데이터)에 대한 주요 특징 셋의 특징이다.

Is a feature of the main feature set for the reference vector (training data).

및

는 각각 파장 및 시간 인덱스를 나타내기 위해 사용된다.

And

Are used to represent wavelength and time index, respectively.

는 시간 t에서의 포도당 농도이다.

Is the glucose concentration at time t.

및

는 각각 시간

에서 피검체의 대응하는 특징에 대한 초기의 몇 개 값들 및 대응하는 포도당 농도이다.

And

Each time

In the initial few values for the corresponding feature of the subject and the corresponding glucose concentration.

는 피검체의 대응하는 특징에 대한 나머지 몇 개 값들이다.

를 제외함.

Are the remaining few values for the corresponding feature of the subject.

Except.

는 5분 간격으로 5 개의 포도당 측정치들로부터 각각 보간된 20 개의 값들을 갖는다.

Has 20 values each interpolated from five glucose measurements at 5-minute intervals.

는 다음과 같이 포도당 농도에 대하여 근사될 수 있다.4 is a plot showing a linear approximation relationship to the 109th feature with respect to the corresponding glucose concentration of training data and calibration data, in accordance with embodiments. Embodiments provide a method of obtaining a linear approximation of features with respect to corresponding glucose concentrations of training data and calibration data. Each pre-treated near-infrared feature in the main feature set,

Can be approximated for glucose concentration as follows.

는

에 대한 선형 근사이고,

및

는 λ번째 특징에 대한 기울기 및 절편이다.here,

Is

Is a linear approximation for,

And

Is the slope and intercept for the λ th feature.

는 두 선형 근사값들에 대해 거의 동일해야 한다.In one embodiment, the features of the key feature set for training data and calibration data can be approximated linearly to their corresponding glucose concentrations.

Should be nearly identical for the two linear approximations.

는 선형 근사 관계를 이용한 훈련 데이터의 평균 포도당 농도에서의 훈련 데이터 및 캘리브레이션 데이터의 주요 특징 셋의 각 특징에 대한 근적외선 특징 값들 사이의 거리이다. 예를 들어, 기준 평균 포도당

에 대하여,

는 다음과 같다.FIG. 5 is a plot showing adjustment factors for the 109th feature of the primary feature set of calibration data that maintains training data as a reference in accordance with embodiments. Embodiments calculate adjustment factors for the characteristics of the calibration data based on the training data. Adjustment factor

Is the distance between near infrared feature values for each feature of the key feature set of the training data and the calibration data at the mean glucose concentration of the training data using the linear approximation relationship. For example, the baseline average glucose

about,

Is as follows.

는

에서 모든 관련된

에 대하여 계산될 수 있다.

값들은 검증 데이터를 매핑하는데 사용될 수 있다. 도 5에서, 기준 평균 포도당

이고, 109번째 특징에 대한 캘리브레이션 데이터의 조정 인자는 다음과 같다.Adjustment factor

Is

All related in

Can be calculated for.

The values can be used to map validation data. In Figure 5, the reference mean glucose

And the adjustment factor of the calibration data for the 109th feature is as follows.

는 각

에서

로부터 차감될 수 있다.FIG. 6 is a plot showing the 109 th feature homogenized for validation data and the preprocessed 109 th feature of training data, in accordance with embodiments. FIG. Embodiments herein map a feature of the validation data based on the calculated adjustment factor to homogenize that feature of the validation data. Adjustment factor

Is each

in

Can be deducted from.

는 회귀 모델들의 앙상블을 훈련시키기 위해 회귀 모델들의 앙상블에 전송될 수 있다. 여기서,

는 피검체의 혈당 농도를 결정하기 위해 사용될 수 있다.

May be sent to the ensemble of regression models to train the ensemble of regression models. here,

May be used to determine the blood glucose level of the subject.

7 generates an ensemble of regression models, calculates an ensemble of regression outputs using the ensemble of regression models, and measures the blood glucose concentration of the subject through a weighted average of the plurality of regression outputs based on weights of the meta-classifier. It is a block diagram which shows the ensemble learning part 134 to determine.

For example, referring to FIG. 7, when the verification data S4 and the training data S1, S2, and S3, the verification data S4 are homogenized separately for each of the training data S1, S2, and S3. The plurality of homogenized feature sets S4 (Hom) of the validation data S4 are each input to the corresponding trained regression model. At this time, the results Pred. 1, Pred. 2, and Pred. 3 predicted for the verification data S4 through each regression model are applied to each of the regression models using the plurality of training data S1, S2, and S3. A plurality of weights of the meta-classifiers may be estimated. In addition, blood glucose concentration (S4 glucose result) can be determined by combining a plurality of regression outputs based on weights of the meta-classifier, eg, by weighted average.

Embodiments may be implemented through at least one software program running on at least one hardware device and may perform network management functions to control components. The components shown in FIG. 1 may be at least one of a hardware device or a combination of hardware devices and software modules.

Meanwhile, the embodiments may be implemented by computer readable codes on a computer readable recording medium. Computer-readable recording media include all kinds of recording devices that store data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disks, optical data storage devices, and the like, which may also be implemented in the form of carrier waves (for example, transmission over the Internet). Include. The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion. And functional programs, codes and code segments for implementing the embodiments can be easily inferred by programmers in the art to which the present invention belongs.

Those skilled in the art will appreciate that the present disclosure may be embodied in other specific forms without changing the technical spirit or essential features disclosed. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

100: electronic device 110: preprocessor
112: frequency domain filtering unit 114: data whitening unit
116: EMSC part 118: drift removal unit
120: feature set extraction unit 130: calibration unit
132: feature homogenizer 134: ensemble learning unit
140: memory

Claims (18)

Obtaining, by the preprocessing unit, a near infrared spectrum and a pure spectrum including at least one of training data of a plurality of objects, calibration data of a subject, and verification data of the subject, to obtain a preprocessed near infrared spectrum;
Extracting, by a feature set extractor, a dominant feature set from the preprocessed near infrared spectrum;
Obtaining, by a feature homogenizer, a plurality of homogenized feature sets of the verification data; And
Determining, by an ensemble learner, blood glucose levels of the subject using training data for the plurality of subjects and a homogenized feature set of the verification data. The method of claim 1,
Acquiring the pre-processed near infrared spectrum,
Filtering, by a frequency domain filtering unit, the near infrared spectrum to remove unwanted components in the frequency domain;
Acquiring an orthogonal pure spectrum after converting the pure spectrum by a data whitening unit;
Removing components other than glucose and water from the near infrared spectrum by using the orthogonal pure spectrum by an Extended Multiplicative Scatter Correction unit (EMMS); And
Removing, by the drift remover, drift present in the time domain of the near infrared spectrum. The method of claim 2,
The filtering of the near infrared spectrum may include:
Filtering noise present in the near infrared spectrum using a Savitzky-Golay (SG) filter; And
Removing the linear drift present in the wavelength domain of the near infrared spectrum by differentiating the near infrared spectrum with respect to the wavelength. The method of claim 1,
Extracting a set of key features from the pre-processed near infrared spectrum,
Calculating a correlation between the glucose concentration of the training data and each of the pre-processed near infrared spectra in the training data; And
Selecting a key feature set comprising at least one preprocessed near infrared spectrum corresponding to each of the training data, the calibration data, and the verification data, based on the maximum value of the correlation calculated from the training data; How to predict blood sugar levels. The method of claim 1,
Obtaining a plurality of homogenized feature sets of the verification data,
Obtaining a single homogenized feature set of the validation data, homogenized with respect to the training data of the single subject and the primary feature set of the calibration data; And
Repeating said step to obtain a plurality of homogenized feature sets for said validation data. The method of claim 5,
Obtaining a single homogenized feature set of the verification data
Obtaining a linear approximation relationship for each feature in the set of key features with respect to corresponding glucose concentrations in the training data and the calibration data;
Calculating an adjustment factor for each feature in the set of key features for the calibration data holding the training data as a reference; And
Mapping each feature in the primary feature set of the validation data in a manner that is subtracted from the corresponding adjustment factor to obtain a single homogenized feature set of the validation data. The method of claim 6,
Computing the adjustment factor for each feature in the set of key features,
Calculating an average glucose value for the training data;
Calculating near-infrared feature values for each feature in the key feature set of the training data and the calibration data at the average glucose value using the linear approximation relationship; And
Subtracting the near infrared feature value of the calibration data from the near infrared feature value of the training data to obtain an adjustment factor for each feature in the primary feature set. The method of claim 1,
Determining the blood sugar level of the subject,
Training an ensemble of regression models each separately trained with a set of key features in the training data of separate subjects; And
Calculating a plurality of regression outputs from an ensemble of regression models that takes an input to each of the regression models as a corresponding homogenized feature set of the validation data;
Estimating a plurality of weights of a meta-classifier for each of the regression models using training data of a plurality of subjects; And
Determining the blood glucose concentration of the subject by weighted averaging a plurality of regression outputs based on the weights of the meta-classifier. The method of claim 8,
Estimating a plurality of weights of the meta-classifier for each of the regression models,
Calculating a plurality of model outputs from an ensemble of regression models that takes a set of training data of a plurality of subjects as input to each of the regression models;
Calculating a plurality of model correlations by calculating a correlation between the plurality of model outputs and the glucose concentration of the training data of the plurality of subjects; And
Scaling a plurality of model correlations to obtain a plurality of weights of a meta-classifier corresponding to each of the regression models. A pre-processing unit processing a near-infrared spectrum including at least one of training data of a plurality of subjects, calibration data of a subject, and verification data of the subject, and a pure spectrum to obtain a pre-processed near-infrared spectrum;
A feature set extraction unit for extracting a dominant feature set from the pre-processed near infrared spectrum;
A feature homogenizer for obtaining a plurality of homogenized feature sets of the verification data; And
And an ensemble learner configured to determine a blood glucose level of a subject by using the training data of the plurality of subjects and the homogenized feature set of the verification data. The method of claim 10,
The pretreatment unit
A frequency domain filtering unit filtering the near infrared spectrum to remove unwanted components from the frequency domain;
A data whitening unit configured to obtain an orthogonal pure spectrum after applying a transform to the pure spectrum;
An EMSC unit for removing components other than glucose and water present in the near infrared spectrum by using the orthogonal pure spectrum; And
And a drift removing unit for removing drift existing in the time domain of the near infrared spectrum. The method of claim 11,
The frequency domain filtering unit
Savitzky-Golay (SG) filter is used to filter noise present in the near infrared spectrum,
And differentiating the near infrared spectrum with respect to wavelength to remove linear drift present in the wavelength domain of the near infrared spectrum. The method of claim 10,
The feature set extraction unit,
Calculate a correlation between the glucose concentration of the training data and each of the pre-processed near infrared spectra in the training data,
And based on the maximum value of the correlation calculated from the training data, selecting the primary feature set comprising at least one pre-processed near infrared ray corresponding to each of the training data, the calibration data and the verification data. The method of claim 10,
The feature homogenization unit
Repeating the process of obtaining a single homogenized feature set of the verification data, homogenized with respect to the training data of the single subject and the primary feature set of the calibration data, to obtain a plurality of homogenized feature sets for the verification data. , Electronic device. The method of claim 14,
The feature homogenization unit,
Obtaining a linear approximation relationship for each feature in the key feature set with respect to the corresponding glucose concentrations in the training data and the calibration data,
Calculating adjustment factors for each feature in the primary feature set for the calibration data holding the training data as a reference,
By mapping each feature in the main feature set of the verification data in a subtracted manner from the corresponding adjustment factor,
Obtain a single homogenized feature set for the verification data. The method of claim 15,
The feature homogenization unit,
Calculate an average glucose value for the training data,
Using the linear approximation relationship to calculate near infrared feature values for each feature in the key feature set of the training data and the calibration data at the mean glucose value,
And subtract the near infrared feature value of the calibration data from the near infrared feature values of the training data to obtain an adjustment factor for each feature in the primary feature set. The method of claim 10,
The ensemble learning unit,
Train an ensemble of regression models, each trained separately, with a set of key features in separate subject's training data,
Calculate a plurality of regression outputs from an ensemble of regression models that take a corresponding homogenized feature set of the validation data as input to each of the regression models,
Using a plurality of subjects' training data to estimate a plurality of weights of a meta-classifier for each of the regression models,
And determine a blood glucose concentration of the subject by weighted averaging a plurality of regression outputs based on weights of the meta-classifier. The method of claim 17,
The ensemble learning unit,
Calculate a plurality of model outputs from an ensemble of regression models taking a set of training data of a plurality of subjects as input to each of the regression models,
Calculating a plurality of model correlations by calculating a correlation between the plurality of model outputs and the glucose concentration of the training data of the plurality of subjects,
Obtaining a plurality of weights of the meta-classifier corresponding to each of the regression models by scaling the plurality of model correlations.

Images