Blood glucose estimation device, blood glucose estimation method and program

Abstract

Conventionally, methods for measuring blood glucose levels include using blood samples or interstitial fluid to predict blood glucose values. However, both methods involve invasive procedures such as inserting needles into the skin, which can cause psychological or physiological burdens on the subject. According to the present invention, based on pre-obtained attribute information, non-invasive biological information, and blood test data from several subjects, a blood glucose estimation model is generated through machine learning. This model can non-invasively estimate blood glucose values ​​based on predetermined user attribute information and / or non-invasive biological information.

Description

[Technical Field]

[0001] This invention relates to a blood glucose value estimation device, a blood glucose value estimation method and program. [Previous Technology]

[0002] Conventionally, methods for measuring blood glucose levels include taking blood samples from a subject or predicting blood glucose levels from interstitial fluid. However, regardless of the method used, invasive procedures such as inserting needles into the skin of the subject are required, which can cause psychological or physiological burdens to the subject. A non-invasive method for estimating blood glucose levels is, for example, the method described in Patent Document 1. Patent Document 1 describes measuring the ratio of naturally stable carbon isotopes of isoprene contained in a breath sample and using it as a dynamic indicator of blood glucose levels. [Prior Art Documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2017-9597 [Non-Patent Document]

[0004] [Non-Patent Literature 1] Psychology Research and Behavior Management 2011:4 81-86, Summary of the clinical investigations ESTeck Complex March, 20, 2010 Abstract [Non-Patent Literature 2] RN Chua, YW Hau, CM Tiew and WL Hau, "Investigation of Attention Deficit / Hyperactivity Disorder Assessment Using Electro Interstitial Scan Based on Chronoamperometry Technique," in IEEE Access, vol. 7, pp. 144679-144690, 2019, doi: 10.1109 / ACCESS.2019.2938095. [Non-Patent Literature 3] Maarek A. Electro interstitial scan system: assessment of 10 years of research and development. Med Devices (Auckl). 2012; 5:23-30. doi:10.2147 / MDER.S29319 [Summary of the Invention]

[0005] (The problem that the invention aims to solve)

[0006] Here, the non-invasive blood glucose estimation by means of biomarker molecules in exhaled breath is accompanied by the determination of stable carbon isotope ratio, so it is difficult to easily measure blood glucose in health diagnosis and other purposes.

[0007] This invention was made in view of the above-described circumstances. The object of this invention is to provide a blood glucose estimation device, method, and program capable of estimating blood glucose levels with excellent accuracy based on information from non-invasive organisms. (Technical means to solve the problem)

[0008] The blood glucose estimation device of the present invention is characterized in that it comprises: an information acquisition unit, which acquires predetermined user attribute information and non-invasive biological information; an estimation model memory unit, which stores a blood glucose estimation model; and an estimation processing unit, which uses the blood glucose estimation model to calculate the predetermined user's blood glucose estimation value based on the predetermined user attribute information and / or non-invasive biological information.

[0009] The blood glucose estimation device is characterized in that it further comprises: a training data memory unit, which stores a training data set; and a learning processing unit, which generates the blood glucose estimation model based on the training data set by means of machine learning.

[0010] The blood glucose estimation device is characterized in that: the attribute information includes either age and gender, or a combination thereof, and the non-invasive biological information includes body mass index (BMI), blood pressure, pulse data, electrocardiogram data, biological impedance, or a combination thereof.

[0011] The blood glucose estimation device is characterized by: a non-invasive biological information system that further includes blood oxygen saturation (SpO2).

[0012] The blood glucose estimation device is characterized in that: the training dataset includes the subject's attribute information, non-invasive organism information and blood glucose values ​​measured from the blood, and the estimation accuracy of the blood glucose value meets the measurement accuracy specified in the international standard ISO 15197.

[0013] The blood glucose estimation device is characterized in that: the training dataset includes non-invasive biological information of the subject and blood glucose values ​​measured from the blood, and the estimation processing unit calculates the blood glucose risk instead of the blood glucose estimation value.

[0014] The blood glucose estimation device is characterized in that: the learning processing unit adds a label to the training dataset indicating whether or not there is a risk of blood glucose value based on the blood glucose value measured from the blood, and when the difference between the number of ...

[0015] The blood glucose estimation device is characterized in that: the learning processing unit generates a first blood glucose estimation model and a second blood glucose estimation model respectively by machine learning based on different types of training data sets, and the estimation processing unit uses the first blood glucose estimation model and the second blood glucose estimation model to calculate the estimated blood glucose value of the predetermined user.

[0016] The blood glucose estimation device is characterized in that it further comprises: a biological information estimation unit, which estimates at least one of the biological information including BMI, blood pressure, pulse wave data, electrocardiogram data, biological impedance and blood oxygen saturation, and the information acquisition unit acquires the biological information estimated by the biological information estimation unit as the biological information of the predetermined user.

[0017] The non-invasive blood glucose estimation system of the present invention is characterized in that, in addition to having a blood glucose estimation device, it further has a biological information measuring device for measuring non-invasive biological information, thereby constituting a non-invasive blood glucose estimation system.

[0018] The blood glucose estimation method of the present invention is characterized by comprising the following steps: memorizing a training dataset containing the subject's attribute information, non-invasive biological information, and blood glucose values ​​measured from blood; generating a blood glucose estimation model by machine learning based on the training dataset; and calculating the estimated blood glucose value of a given user based on the blood glucose estimation model and / or the user's attribute information and / or non-invasive biological information.

[0019] The program in this invention is characterized in that it causes a computer to perform the following steps: memorizing a training dataset containing the subject's attribute information, non-invasive biological information, and blood glucose levels measured from blood; generating a blood glucose level estimation model based on the training dataset through machine learning; and using the blood glucose level estimation model, calculating the estimated blood glucose level for a given user based on the user's attribute information and / or non-invasive biological information. (Effectiveness compared to prior art)

[0020] According to the present invention, blood glucose levels can be estimated with excellent accuracy using non-invasive biological information through machine learning.

Implementation Method

[0022] Hereinafter, embodiments will be described with reference to drawings. Furthermore, these embodiments are illustrative only, and the present invention is not limited to the following configuration.

[0023] <Function of the Device> Referring to Figures 1 to 8, the blood glucose estimation system 1 and blood glucose estimation device 30 of this embodiment will be described. Figure 1 is a block diagram showing the general structure of the blood glucose estimation system 1 of this embodiment. The blood glucose estimation system 1 includes a terminal device 10, a biological information measuring device 20, a blood glucose estimation device 30, and a display device 39.

[0024] The terminal device 10 can be any information terminal that can input user attribute information (name, ID, age, gender, etc.) and height, weight, etc., and can output the input information to the blood glucose estimation device 30 via a wired or wireless communication network. Examples include mobile terminals such as tablet terminals, smartphones, and wearable terminals, or personal computers (PCs). Furthermore, height, weight, etc., can be measured by the biometric information measuring device 20 described below.

[0025] The biometric information measuring device 20 measures the user's non-invasive biometric information. Here, non-invasive biometric information refers to biometric information obtained using methods that do not require the insertion of instruments into the skin or openings in the body. Non-invasive biometric information can be measured using commercially available height and weight scales, blood pressure monitors, pulse oximeters, pulse wave meters, electrocardiographs, impedance meters, and skin conductance meters. Alternatively, the Estec BC-3 (dual-purpose system), which can simultaneously measure pulse wave data, electrocardiogram data, biometric impedance, and blood oxygen saturation (SpO2), can also be used. These devices can measure non-invasive biometric information without causing psychological or physiological burden to the user.

[0026] In an embodiment of the present invention, non-invasive biological information includes any one or a combination of BMI (body mass index), blood pressure, pulse wave data, electrocardiogram data and biological impedance, and may also include blood oxygen saturation (SpO2).

[0027] BMI is calculated based on height h[m] and weight w[kg] using the following formula: BMI=w / h2[kg / m2].

[0028] Blood pressure includes any one of systolic blood pressure, diastolic blood pressure, pulse pressure, and mean arterial pressure, or a combination thereof. Pulse pressure is calculated by the following formula: Pulse pressure = Systolic blood pressure - Diastolic blood pressure. Mean arterial pressure is calculated by the following formula: Mean arterial pressure = Diastolic blood pressure + Pulse pressure × 1 / 3.

[0029] Pulse data is measured using a pulse oximeter or pulse wave meter by irradiating red light (~660 nm) onto a prominent part of the body, such as the fingers, with a red light-emitting diode (Red LED) or near-infrared light (~905 nm) irradiated with an infrared light-emitting diode (IR LED), and the transmitted light is received by a photoelectric crystal. Pulse data includes pulse rate, elasticity index, peripheral vascular resistance, acceleration pulse wave, b / a, e / a, -d / a, Gauzer acceleration pulse wave aging index, ejection rate, left ventricular ejection time (LVET), and dicrotic-pulse elasticity index (DEI), or a combination thereof. Here, the elasticity index is the value obtained by dividing the time from the systolic peak of the fingertip volume pulse wave to the detected diastolic peak by height. Peripheral vascular resistance is calculated by multiplying mean arterial pressure / cardiac output by 80. The DEI (diplastic elasticity index) is an indicator of diastolic vascular elasticity and can be measured using a pulse wave velocity (PWV) measuring device. A value of 0.3 to 0.7 is considered normal; a value below 0.3 may indicate hypertension or arteriosclerosis, while a value above 0.7 suggests the possibility of acute anxiety neurosis. The acceleration pulse wave is the second derivative of the photoplethysmogram (SDPTG). The acceleration pulse wave consists of an initial positive wave (a wave), an initial negative wave (b wave), a mid-systolic rising wave (c wave), a late-systolic falling wave (d wave), and an initial diastolic positive wave (e wave). The b / a, e / a, and -d / a ratios are calculated based on the proportions of each wave height. With aging, the ratio of b / a increases while the ratios of c / a, d / a, and e / a decrease. Therefore, the aging of blood vessels can be assessed using the Gauzer acceleration pulse wave aging index (bcde) / a. Ejection rate is the proportion of blood pumped from the ventricles with each heartbeat and is directly proportional to the acceleration pulse wave aging index. LVET is the left ventricular ejection time, which is the time it takes for blood from the left ventricle to be ejected into the aorta after the aortic valve opens.

[0030] Electrocardiogram (ECG) data can be measured using electrocardiography (ECG) with electrodes or photoplethysmography (PPG). ECG data includes respiratory rate, heart rate, RR interval, standard deviation of RR interval, MxDMn ratio, power spectrum in the low-frequency band, power spectrum in the high-frequency band, heart rate variability index (Low Frequency / High Frequency), total power, or a combination thereof. Here, the RR interval refers to the interval from one QRS wave to the next on the ECG. The MxDMn ratio refers to the ratio of the longest to the shortest RR interval over a given time, and is an index of irregular heartbeats. Total power refers to the calculated total power of the power spectrum at frequencies of 0–0.4 Hz (Very Low Frequency, LF, HF) in a 2-minute minute measurement. This value reflects the overall activity of the autonomic nervous system, primarily dominated by sympathetic nerve activity. By calculating the power spectral density based on the electrocardiogram, it is possible to calculate the power spectral density ratio of high frequency (0.1875~0.50 Hz: HF), the power spectral density ratio of low frequency (0.05~0.1875 Hz: LF), the LF / HF ratio, and the power spectral density ratio of ultra-low frequency band (0~0.05 Hz: VLF).

[0031] The impedance (conductance) of a living organism can be measured by passing a weak current between two of six electrodes, such as those on both feet, both hands, and both foreheads. By passing a current between two of the six electrodes, (i) the conductance of the anode / cathode (μS), (ii) the conductance of the cathode / anode (μS), (iii) the difference between the conductance measured by (i) and the conductance measured by (ii) (∆SCR A-SCR C), and (iv) the conductivity (μS / m) can be measured simultaneously. Furthermore, muscle mass, body fat mass, total water content, phase angle, and resistance value can also be measured. Additionally, the dielectric constant (μSi) when current flows between the right and left hands, and between the right and left foreheads can also be measured. Preferably, the impedance (conductance) of the living organism is measured using the electrical conductivity of 22 patterns from the six electrodes.

[0032] Bio-body impedance includes body fat mass (kg), body fat mass (%), lean body mass, lean body percentage, muscle mass, total water content (kg), total water content (%), intracellular water content (%), cardiac output, 1 frontal left-2 right hand / SCR A, 1 frontal left-2 right hand / ∆SCR C-SCR A, 5 left hand-6 left foot / ∆SCR C-SCR A, 13 left foot-14 right foot / SCR A, 15 right hand-16 frontal left / ∆SCR C-SCR A, 15 right hand-16 frontal left / ∆SCR C-SCR A, 19 right foot-20 left hand / ∆SCR C-SCR A. ESG2+4+15+17 (μS / m), ESG6+13+19 (%), ESG6+8+19+21 (%), ESG6+8+19+21 (μS / m), ESG9+10 (μS / m), ESG9+10 (%), conductance of the left foot, R (Ω), phase angle, dielectric constant of the frontal path, conductivity of the frontal path (9), dielectric constant of the single-hand-to-single-hand path, conductivity from hand to hand (11, 12), any one of the following: stroke volume (cardiac output ÷ heart rate), or a combination thereof.

[0033] Here, SCR is short for skin conductance response, and ESG is short for electron scanning spectroscopy. The "+" in ESG2+4+15+17 indicates which electrode on the body was used to measure the conductance. For example, ESG2+4+15+17 refers to the average conductance measured in Figure 2 using the left hand when an electric current is applied from the left hand to the left forehead, the right hand when an electric current is applied from the right hand to the right forehead, the right hand when an electric current is applied from the right hand to the left forehead, and the left hand when an electric current is applied from the left hand to the right forehead. These conductances are described in detail in Non-Patent Document 1. Furthermore, the method for measuring ESG (electron scanning spectroscopy) is described in detail in Non-Patent Documents 2 and 3. "1 front left - 2 right hand / SCR A" refers to the conductivity (or conductivity) of the path measured when current flows through "1 front left" as the cathode and "2 right hand" as the anode. "5 left hand - 6 left foot / ∆SCR C - SCR A" refers to the difference in conductivity measured when current flows through "5 left hand" and "6 left foot" in an anode-cathode and cathode-anode configuration.

[0034] BMI and blood pressure can be measured using a height and weight scale and a blood pressure monitor. Furthermore, information on non-invasive organisms can also include oxygen transport calculated based on SpO2 and cardiac output.

[0035] The non-invasive biological information measured is output to the blood glucose estimation device 30 via a wired or wireless communication network. The biological information measuring device 20 can be an installed measuring device or a portable measuring device such as a wearable terminal.

[0036] The blood glucose estimation device 30 includes a first acquisition unit 31, a second acquisition unit 32, a user data memory unit 33, a training data memory unit 34, a learning processing unit 35, an estimation model memory unit 36, an estimation processing unit 37, and an estimation data memory unit 38. The first acquisition unit 31 acquires the user's attribute information via the terminal device 10. Furthermore, the second acquisition unit 32 acquires the user's non-invasive biological information via the biometric information measuring device 20.

[0037] The user data memory unit 33 stores the user's attribute information and non-invasive biological information obtained by the first acquisition unit 31 and the second acquisition unit 32.

[0038] The training data memory unit 34 stores several training data sets consisting of attribute information of several subjects obtained in advance, information on non-invasive organisms, and physical examination information such as blood glucose levels obtained through blood tests, and uses this as a training data set for machine learning. Furthermore, the physical examination information may also include examination information obtained from blood, urine, stool, etc.

[0039] The learning processing unit 35 acquires the training data set stored in the training data memory unit 34 and uses the training data set to create a blood glucose value estimation model. Specifically, the acquired training data set is normalized, and the relationship between attribute information and non-invasive biological information and blood glucose values ​​is learned through machine learning methods such as neural networks (NN), XGBoost, and linear regression. The learning processing unit 35 generates an estimation model that estimates blood glucose values ​​based on non-invasive biological information. The estimation model memory unit 36 ​​stores the blood glucose value estimation model generated by the learning processing unit 35. Non-invasive biological information includes any one or a combination of BMI, blood pressure, pulse wave data, electrocardiogram data, and biological impedance. Furthermore, blood oxygen saturation (SpO2) is included as needed.

[0040] The estimation processing unit 37 uses the estimation model generated by the learning processing unit 35 to estimate the user's blood glucose level based on the user's attribute information and / or non-invasive biological information. Then, the estimated blood glucose level is stored in the estimation data storage unit 38.

[0041] The display device 39 can display the estimated blood glucose value along with the user's attribute information and non-invasive biological information. Furthermore, this data can also be displayed on the user's terminal device 10.

[0042] <Hardware Configuration of the Device> Figure 3 is a hardware configuration diagram of the blood glucose estimation device 30. As shown in Figure 3, the blood glucose estimation device 30 is composed of a computer 300 having one or more processors 301, memory 302, storage 303, input / output ports 304, and communication ports 305. The processor 301 performs processing related to blood glucose estimation in this embodiment by executing a program. The memory 302 temporarily stores the program and the results of the program's calculations. The storage 303 stores the program that executes the processing performed by the blood glucose estimation device 30. The storage 303 can be any storage device as long as it can be read by a computer, such as recording media (magnetic disk, optical disk, etc.), random access memory, flash memory, read-only memory, and other various recording media. The input / output port 304 is used to input information from the terminal device 10 and the biometric information measuring device 20, or to output the blood glucose estimation value to the display device 39. Communication port 305 is used for sending and receiving data between other information terminals, such as computers, not shown. Communication can be conducted via wireless or wired communication. Furthermore, the blood glucose estimation device 30 can be installed on a commercially available desktop or laptop PC, and the time required to calculate the estimated blood glucose value using the estimation model is several seconds. Moreover, when the processor 301 of the blood glucose estimation device 30 is in operation, the aforementioned first acquisition unit 31, second acquisition unit 32, learning processing unit 35, and estimation processing unit 37 perform their functions.

[0043] <Generating a Blood Glucose Prediction Model via Machine Learning> Figure 4 is a flowchart showing the execution steps of generating a blood glucose prediction model via machine learning. In step ST101, the learning processing unit 35 preprocesses the input data (e.g., the training dataset mentioned above). Specifically, the learning processing unit 35 converts the gender attribute information of each subject into a one-hot vector. Furthermore, the learning processing unit 35 normalizes attribute information other than gender, non-invasive biological information of each subject, and physical examination information of blood glucose values ​​obtained from blood tests using Yeo-Johnson transformation or Boolean-box transformation. In step ST102, the learning processing unit 35 performs machine learning in parallel using linear regression, neural networks (NN), and gradient boosting regression trees. Alternatively, any one of linear regression, neural networks (NN), and gradient boosting regression trees can be used, or a combination of both can be used. Furthermore, grid search can also be used for hyperparameter tuning of gradient boosting. For machine learning using gradient boosting regression trees, software libraries such as XGBoost, CatBoost, and LightBGM can be used. In step ST103, the learning processing unit 35 performs unsampled learning (ridge regression) on the learning results obtained through linear regression, neural networks (NN), and gradient boosting regression trees. Moreover, if any of linear regression, neural networks (NN), or gradient boosting regression trees is selected in step ST102, unsampled learning (ridge regression) in step ST103 is not required. The learning processing unit 35 stores the blood glucose value estimation model generated by the above learning processing in the estimation model memory unit 36.

[0044] Machine learning performed by linear regression can, for example, utilize linear regression provided by Scikit-learn, an open-source machine learning library for Python. Furthermore, principal component analysis can be used to compress the dimensionality as needed. Moreover, the above-mentioned machine learning algorithm is merely one example and is not limited to it.

[0045] Figure 5 shows the construction of the neural network (NN) used as the presumption model. Rectangles represent layers that perform data transformation, and rounded rectangles represent input and output data. D is the number of inspection items. The presumption model in Figure 5 transforms data through four layers (Compile A1~Compile A3, Compile B) in the order of D-dimensional, 96-dimensional, 96-dimensional, 96-dimensional, and 1-dimensional.

[0046] Furthermore, as shown in Figure 5, Compile A1 to A3 each contain a fully connected layer "Linear" for fully connected processing, a "Keranel regulizar" for regularization, and a ReLU layer "ReLU" for ReLU processing. Compile B contains a fully connected layer "Linear" for fully connected processing and an "Adam" layer for optimization. The input units of the fully connected layers in Compile A1 correspond to the input layers, and the output units in Compile B correspond to the output layers. The units located between these correspond to intermediate layers (hidden layers). The intermediate layers include Dropout layers that control a portion of the input values ​​to 0 to prevent overlearning.

[0047] When XGBoost is used in machine learning via gradient boosting regression trees, the residual between the blood glucose value obtained from the blood test and the estimated blood glucose value is calculated, and the parameters of XGBoost (max_depth, subsample, colsample_bytree, learning_rate) are adjusted to minimize the mean squared error, thereby generating the estimation model. max_depth is the depth of the decision tree, subsample is the proportion of randomly selected samples from each tree, colsample_bytree is the proportion of randomly selected rows from each tree, and learning_rate represents the learning rate. Specifically, max_depth is adjusted within the range of 1 to 10, subsample is adjusted within the range of 0.1 to 1.0, colsample_bytree is adjusted within the range of 0.3 to 1.0, and learning_rate is adjusted within the range of 0.1 to 0.7.

[0048] <Using a Blood Glucose Estimation Model to Estimate Blood Glucose Values> Figure 6 is a flowchart showing the execution steps of the blood glucose estimation process. In step ST201, the first acquisition unit 31 of the blood glucose estimation device 30 acquires the user's attribute information via the terminal device 10. In step ST202, the second acquisition unit 32 of the blood glucose estimation device 30 acquires the user's non-invasive biological information. Then, the user's attribute information and non-invasive biological information are stored in the user data memory unit 33. Next, in step ST203, the blood glucose estimation model stored in the estimation model memory unit 36 ​​is used to calculate the estimated blood glucose value by the estimation processing unit 37. In step ST204, the calculated estimated blood glucose value is stored in the estimation data memory unit 38, and in step ST205, the estimated blood glucose value is output to an external terminal such as a display device 39 for display.

[0049] <Example (Estimated Blood Glucose Value)> The following describes an embodiment of the invention. However, the invention is not limited to the following embodiments. Attribute information includes ID, name, age, and gender; non-invasive biological information includes BMI, blood pressure, pulse wave data, electrocardiogram data, bioimpedance, and blood oxygen saturation (SpO2). Height and weight, which serve as the basis for calculating BMI, are measured using a height meter and a weight scale, respectively, while blood pressure is measured using a blood pressure monitor. Furthermore, pulse wave data, electrocardiogram data, bioimpedance, and blood oxygen saturation (SpO2) are measured using the EstecBC-3 (dual-purpose system). Alternatively, commercially available pulse wave meters, electrocardiographs, impedance measuring devices, and pulse oximeters can be used in combination instead of the EstecBC-3. Furthermore, the aforementioned non-invasive biological information can be obtained using a pre-defined wearable terminal. Bioimpedance (conductance) was measured by passing a weak current between two of six electrodes located on the feet, hands, and left and right foreheads. The voltage and current were set to 1.28 V and 200 μA, respectively, and the conductivity was measured every 1 second for 32 milliseconds. Current was passed between two of the six electrodes, and the following measurements were taken: (i) the conductance at the anode / cathode (μS), (ii) the conductance at the cathode / anode (μS), (iii) the difference between the conductance measured in (i) and (ii) (∆SCR A-SCR C), and (iv) the conductivity (μS / m). Muscle mass, body fat mass, total water content, phase angle, and resistance were also measured. The dielectric constant (μSi) was also measured when current was applied between the right and left hands and between the right and left foreheads.

[0050] Using the Estec BC-3, pulse wave data, electrocardiogram data, bioimpedance and blood oxygen saturation (SpO2) were measured for each subject for 2 minutes. During the measurement, an instrument with the functions of electrocardiogram, pulse wave meter and pulse oximeter was installed on the index finger of the subject's left hand, and two electrodes were installed on the forehead. The subject then placed his / her hands and feet on the electrode plates while sitting in a chair.

[0051] <Learning Model 1> In Learning Model 1, as shown in Figure 4, machine learning performed by linear regression, neural networks (NN), and gradient boosting regression trees (XGBoost) is executed in parallel, and non-sampling learning (ridge regression) is then performed on the learning results. At this time, the data shown below are selected and used as attribute information and non-invasive organism data. (A) Attribute Information・Gender (B) Non-invasive Biological Data・BMI・Blood Pressure… Systolic Blood Pressure, Diastolic Blood Pressure, Pulse Pressure・Pulse Wave Data…Pulse, Ejection Rate, Elasticity Index, LVET (Left Ventricular Ejection Time), b / a, -d / a, Acceleration Pulse Wave Aging Index・ECG Data…Respiratory Rate, MxDMn Ratio, LF / HF, Total Power・Biological Impedance…Muscle Mass, ESG2+4+15+17 (μS / m), ESG6+8+19+21 [%], ESG9+10 (μS / m), ESG9+10 [%], R (Ω), Total Water Content (%), Cardiac Output・Oxygen Saturation (SpO2). Here, ESG9+10 refers to the average impedance measured at the location shown in Figure 2. [μS / m] is the unit of the measured average, and [%] is the value obtained by scaling down the measured average within the normally measurable range.

[0052] Furthermore, the non-invasive organism data further includes data on oxygen transport and pulse pressure / pulse rate estimated based on cardiac output and blood oxygen saturation (SpO2) contained in the organism's impedance.

[0053] <Learning Model 2> In Learning Model 2, machine learning is performed using a gradient boosting regression tree (XGBoost). At this time, the following data are selected and used as attribute information and non-invasive biological data. (A) Attribute Information・Age・Gender (B) Non-invasive Biological Data・BMI・Blood Pressure…Diastolic blood pressure, mean arterial pressure, pulse pressure・Pulse Wave Data…Pulse, elasticity index, b / a, ejection rate, acceleration pulse wave aging index・ECG Data…Heart rate, RR interval, low-frequency power spectrum, LF / HF・Bioimpedance…Body fat mass (%), muscle mass, ESG2+4+15+17 (μS / m), ESG9+10 (μS / m), ESG9+10 (%), R (Ω), cardiac output・Oxygen saturation (SpO2).

[0054] Furthermore, the non-invasive organism data further includes data on oxygen transport and pulse pressure / pulse rate estimated based on cardiac output and blood oxygen saturation (SpO2) contained in the organism's impedance.

[0055] <Example 1> In Example 1, a blood glucose estimation model was generated by machine learning through the above-mentioned learning model 1, using (1) attribute information of a total of 555 subjects, (2) non-invasive biological information measured by a height and weight scale, a blood pressure monitor and an Estec BC-3, and (3) training data obtained by blood glucose values ​​obtained from blood tests conducted on the same day as the non-invasive biological information measurement. Then, the estimation accuracy of the blood glucose estimation model was evaluated to see if it met the measurement accuracy specified in ISO 15197. Specifically, it was confirmed that more than 95% of the estimated blood glucose values ​​converged within ±15 mg / dL when the blood glucose value was below 100 mg / dL, and within ±15% when the blood glucose value was above 100 mg / dL. The results showed that the average error between the estimated and actual blood glucose values ​​was -0.7, the standard deviation was 7.43, and the standard error was 0.63. Based on the ISO 15197 standard, the accuracy was over 95% and 97%.

[0056] <Example 2> In Example 2, a blood glucose estimation model was generated by machine learning through the above-mentioned learning model 2, using (1) attribute information of a total of 711 subjects, (2) non-invasive biological information measured by a height and weight scale, a blood pressure monitor and an Estec BC-3, and (3) training data obtained by blood glucose values ​​obtained from blood tests conducted on the same day as the non-invasive biological information measurement. Then, the estimation accuracy of the blood glucose estimation model was evaluated to see if it met the measurement accuracy specified in ISO 15197. Specifically, it was confirmed whether more than 95% of the estimated blood glucose values ​​converged within ±15 mg / dL when the blood glucose value was below 100 mg / dL, and whether it converged within ±15% when the blood glucose value was above 100 mg / dL. The results showed that the average error between the estimated and actual blood glucose values ​​was 0.18, the standard deviation was 7.77, and the standard error was 0.58. Based on the ISO 15197 standard, the accuracy was over 95% and 95.4%.

[0057] As described above, based on non-invasive biological data including BMI (body mass index), blood pressure, pulse wave data, electrocardiogram data and biological impedance, a blood glucose estimation model is generated through machine learning, thereby estimating blood glucose levels without the need for blood tests.

[0058] Furthermore, as shown below, even if the number of each data included in the data of non-invasive organisms is limited, it is still possible to determine the "blood glucose risk", that is, whether the blood glucose level is normalized.

[0059] <Generating a Blood Glucose Risk Prediction Model via Machine Learning> Figure 7 is a flowchart illustrating the execution steps of generating a blood glucose risk prediction model via machine learning. In step ST301, the learning processing unit 35 preprocesses the input data (e.g., the training dataset mentioned above). Specifically, the learning processing unit 35 converts blood glucose values ​​obtained from blood tests from values ​​less than 100 mg / ml to 0 (no risk) and values ​​greater than 100 mg / ml to 1 (risk). Furthermore, when the number of people classified as 0 differs from and is unbalanced with the number of people classified as 1, the learning processing unit 35 can also apply SMOTE (Chawla, NV. et al. 2002) to the learning data to manually generate training samples. That is, labels indicating the presence or absence of blood glucose risk (e.g., 0 or 1 as mentioned above) based on blood glucose levels measured from blood can be added to the input data (training dataset). When the difference between the number of individuals with blood glucose risk (label: 1) and the number of individuals without blood glucose risk (label: 0) exceeds a predetermined value, the number of samples in the training dataset is increased (generated) to reduce this difference. In step ST302, the learning processing unit 35 performs machine learning using logistic regression. For example, logistic regression provided by Scikit-learn, an open-source machine learning library for Python, can be used for machine learning using logistic regression. Furthermore, if necessary, principal component analysis can be used to compress the dimensionality. Additionally, by comparing the blood glucose risk obtained from blood tests with the blood glucose risk inferred through machine learning, the parameters of logistic regression (C, regularization method, max_iter, solber) are adjusted to maximize the F1 score, thereby generating a blood glucose risk estimation model. Here, C is a trade-off parameter determining the strength of regularization; a larger value results in weaker regularization. The regularization method refers to either L1 or L2 regularization, and a choice is made between them. `max_iter` represents the maximum number of repetitions. From the `solber` parameter, a convergence method that minimizes the cross-entropy error is selected (e.g., L-BFGS, Newton-CG, bilinear (liblinear), sag, and saga). Furthermore, the bilinear method was chosen in Examples 3 and 4 below. Moreover, the above machine learning algorithm is merely one example and is not limited to others.

[0060] <Using a Blood Glucose Risk Prediction Model to Predict Blood Glucose Risk> As shown in Figure 8, in step ST401, the first acquisition unit 31 of the blood glucose prediction device 30 acquires the user's attribute information from the terminal device 10. In step ST402, the second acquisition unit 32 of the blood glucose prediction device 30 acquires the user's non-invasive biological information. Then, the user's attribute information and non-invasive biological information are stored in the user data memory unit 33. Next, in step ST403, using the blood glucose risk prediction model stored in the prediction model memory unit 36, the prediction processing unit 37 calculates the blood glucose risk, i.e., the probability of belonging to level 0 (no risk) or level 1 (risky). In step ST404, the calculated blood glucose risk prediction value is stored in the prediction data memory unit 38. In step ST405, the blood glucose risk prediction value is output to an external terminal such as a display device 39 for display. Furthermore, the blood glucose estimation device 30 can be installed on a commercially available desktop or laptop PC, and the time required to calculate the estimated blood glucose value using the estimation model is a few seconds.

[0061] <Example (Blood Glucose Risk Prediction)> The following describes an example of blood glucose risk prediction. However, the form of blood glucose risk prediction in this invention is not limited to the following examples. Attribute information includes ID, name, age, and gender. Non-invasive biological information includes BMI, blood pressure, pulse wave data, electrocardiogram data, bioimpedance, and blood oxygen saturation (SpO2). Height and weight, which serve as the basis for calculating BMI, are measured using a height meter and a weight scale, respectively, while blood pressure is measured using a blood pressure monitor. Furthermore, pulse wave data, electrocardiogram data, bioimpedance, and blood oxygen saturation (SpO2) are measured using the EstecBC-3 (dual-purpose system). Alternatively, commercially available pulse wave meters, electrocardiographs, impedance measuring devices, and pulse oximeters can be used in combination instead of the EstecBC-3. Also, a pre-defined wearable terminal can be used to acquire the aforementioned non-invasive biological information. Bioimpedance (conductance) was measured by passing a weak current between two of six electrodes located on the feet, hands, and left and right foreheads. The voltage and current were set to 1.28 V and 200 μA, respectively, and the conductivity was measured every 1 second for 32 milliseconds. Current was passed between two of the six electrodes, and the following measurements were taken: (i) the conductance at the anode / cathode (μS), (ii) the conductance at the cathode / anode (μS), (iii) the difference between the conductance measured in (i) and (ii) (∆SCR A-SCR C), and (iv) the conductivity (μS / m). Muscle mass, body fat mass, total water content, phase angle, and resistance were also measured. The dielectric constant (μSi) was also measured when current was applied between the right and left hands and between the right and left foreheads.

[0062] Using the Estec BC-3, pulse wave data, electrocardiogram data, bioimpedance and blood oxygen saturation (SpO2) were measured for each subject for 2 minutes. During the measurement, an instrument with the functions of electrocardiogram, pulse wave meter and pulse oximeter was installed on the index finger of the subject's left hand, and two electrodes were installed on the forehead. The subject then placed his / her hands and feet on the electrode plates while sitting in a chair.

[0063] <Learning Model 3> In Learning Model 3, as shown in Figure 7, machine learning is performed using logistic regression. At this time, the data shown below are selected and used as non-invasive biological data. Furthermore, unlike Learning Models 1 and 2 above, subject attribute information is not required in Learning Model 3. (B) Non-invasive biological data・blood pressure…systolic blood pressure・pulse wave data…elasticity index・electrocardiogram data…respiratory rate, heart rate, low-frequency power spectrum・biological impedance…body fat mass (kg), muscle mass, total water content (%), 19 right foot - 20 left hand / ∆SCR C-SCR A, average values ​​of ESG6+13+19 (%) and ESG6+8+19+21 (%).

[0064] <Learning Model 4> In Learning Model 4, as shown in Figure 7, machine learning is performed using logistic regression. At this time, the data shown below are selected and used as non-invasive biological data. Furthermore, unlike Learning Models 1 and 2 above, subject attribute information is not required in Learning Model 4. (B) Non-invasive biological data・BMI・Blood pressure…systolic blood pressure・Pulse wave data…elasticity index, peripheral vascular resistance・ECG data…respiratory rate, heart rate, low-frequency power spectrum, LF / HF・Bioimpedance…body fat mass (%), muscle mass, total water content (%), 3 frontal right - 4 left hand / SCR A, average value of ESG6+13+19 (%) and ESG6+8+19+21 (%), total conductance value of both feet.

[0065] Furthermore, the non-invasive organism data further includes the oxygen transport volume estimated based on the cardiac output and blood oxygen saturation (SpO2) contained in the organism's impedance.

[0066] <Example 3> In Example 3, a blood glucose risk prediction model was generated using a training dataset of (1) non-invasive biometric information measured by a height and weight scale, a blood pressure monitor, and an Estec BC-3, and (2) blood glucose values ​​obtained from blood tests conducted on the same day as the non-invasive biometric information measurement, from 321 subjects. The model was then machine-learned using the learning model 3 described above. The prediction accuracy of the blood glucose risk prediction model was then evaluated using a ROC_AUC curve. The ROC_AUC was 0.78, indicating good classification and exceeding 0.7. The ROC_AUC curve for the prediction results of Example 3 is shown in Figure 9.

[0067] <Example 4> In Example 4, a blood glucose risk prediction model was generated using a training dataset of (1) non-invasive biometric information measured by a height and weight scale, a blood pressure monitor, and an Estec BC-3, and (2) blood glucose values ​​obtained from blood tests conducted on the same day as the non-invasive biometric information measurements, from a total of 711 subjects. The prediction accuracy of the blood glucose risk prediction model was then evaluated using a ROC_AUC curve. The ROC_AUC was 0.79, indicating good classification and exceeding 0.7. The ROC_AUC curve for the prediction results of Example 4 is shown in Figure 10.

[0068] (Variation Example) In the above embodiments, although examples of using learning model 1 or learning model 2 to estimate blood glucose values ​​and using learning model 3 or learning model 4 to estimate blood glucose risk have been described, it is also possible to use several learning models to estimate blood glucose values ​​or blood glucose risk. In this way, estimations can be made with higher accuracy compared to using a single learning model to estimate blood glucose values ​​or blood glucose risk.

[0069] Furthermore, as seen in the above embodiments and examples, examples exist where the estimation accuracy is improved when BMI is used as the dataset for the learning model. Therefore, a function to estimate the BMI can also be provided in the blood glucose estimation device. BMI is generally not obtainable through wearable terminals, but is calculated based on the height and weight input by the user. By estimating the BMI, the aforementioned blood glucose value or blood glucose risk can be obtained simply by acquiring biological information, thus improving convenience for the user. The method of BMI estimation is not particularly limited. For example, it is known that BMI is correlated with the tilt of the user's abdomen (at a predetermined position). Therefore, for example, a predetermined accelerometer can be installed on the user's abdomen (or a wristband-type wearable terminal with an accelerometer can be placed against the abdomen), and the tilt of the abdomen can be calculated based on the data output from the accelerometer to estimate the BMI.

[0070] Similarly, as described above, a wristband-type wearable terminal equipped with a pulse wave sensor or a blood oxygen concentration sensor can also be used to estimate pulse wave data and blood oxygen saturation.

[0071] Similarly, as described above, blood pressure can also be estimated using a wristband-type wearable device. This is because it is known that the velocity of the pulse wave transmitted through the arteries by the heartbeat is correlated with blood pressure, so blood pressure can also be estimated using a predetermined sensor that measures the velocity of the pulse wave transmitted through the arteries by the heartbeat.

[0072] Similarly, as described above, wearable devices such as wristbands can also be used to estimate electrocardiogram (ECG) data. For example, ECG data can be estimated based on data obtained from electrodes disposed on the side of the wristband-type wearable device opposite to the display surface and electrodes disposed on the display surface side. Specifically, the wrist of the hand wearing the wristband-type wearable device (e.g., the left hand) can be connected to the electrodes disposed on the opposite side, and the fingertips of the hand opposite to the wearing hand (e.g., the right hand) can be connected to the electrodes disposed on the display surface side, and ECG data can be estimated based on the data obtained therefrom.

[0073] Similarly, bioimpedance can also be estimated using a wristband-type wearable terminal equipped with various electrodes. For example, a wristband-type wearable terminal can be used to estimate bioimpedance based on bio-information obtained from the chest and wrist.

[0074] Furthermore, biological information obtainable through a wearable terminal and biological information of the aforementioned presumed object (at least one of BMI, blood pressure, pulse wave data, electrocardiogram data, biological impedance, and blood oxygen saturation) can be used as teaching data, and a classifier generated using various machine learning algorithms can be used to presumably determine the biological information of the aforementioned presumed object. Furthermore, in this case, the first acquisition unit can also acquire the presumed biological information.

[0075] (Other) Furthermore, for example, the above series of processes can be executed by hardware or software. In other words, the above functional configuration is merely illustrative and not particularly limited. That is, as long as the information processing system has the function of executing the above series of processes as a whole, it is sufficient; the specific functional modules used to implement this function are not particularly limited to the above example. Furthermore, the location of the functional modules is not particularly limited to that shown in Figure 4 and can be arbitrarily set. For example, the functional modules of the server can be ported to other terminals or devices. Conversely, the functional modules of other terminals or devices can be ported to the server. Furthermore, a functional module can be composed of a single hardware component, a single software component, or a combination of the above.

[0076] In cases where a series of processes are performed by software, the program constituting the software is installed from the network or recording media onto a computer, etc. The computer may be a computer assembled in dedicated hardware. Alternatively, the computer may be a computer that can perform various functions by installing various programs, such as a server, other general-purpose smartphone, or personal computer.

[0077] The recording medium containing such a program can be either a portable medium (not shown) that is separately configured from the device body for providing the program to users, or a recording medium that is provided to users in a state of pre-assembled in the device body. The program can be transmitted via a network, therefore the recording medium can also be a computer that is connected to or can be connected to a network.

[0078] Furthermore, in this specification, the steps of the program recorded in the recording medium naturally include processing performed in sequence and in a time series, as well as processing that is not necessarily performed in a time series but is performed in parallel or individually. Also, in this specification, the term "system" means an overall device composed of several devices or means. [Simplified Explanation of the Diagram]

[0021] Figure 1 is a block diagram showing the general structure of the blood glucose estimation system. Figure 2 is a diagram illustrating an Electron Scan Gram (ESG). Figure 3 is a hardware diagram of the blood glucose estimation device. Figure 4 is a flowchart showing the execution steps of generating a blood glucose estimation model through machine learning. Figure 5 is a hierarchical structure of a neural network. Figure 6 is a flowchart showing the execution steps of the blood glucose estimation processing. Figure 7 is a flowchart showing the execution steps of generating a blood glucose risk estimation model through machine learning. Figure 8 is a flowchart showing the execution steps of the blood glucose risk estimation processing. Figure 9 is the ROC-AUC curve of the estimation results in Example 3. Figure 10 is the ROC-AUC curve of the estimation results in Example 4.

Claims

1. A blood glucose estimation device, characterized in that it comprises: an information acquisition unit for acquiring predetermined user attribute information and non-invasive biological information; an estimation model memory unit for storing a blood glucose estimation model, which takes the predetermined user attribute information and / or non-invasive biological information as input and outputs a blood glucose estimation value; and an estimation processing unit for using the blood glucose estimation model to calculate the predetermined user's blood glucose estimation value based on the predetermined user attribute information and / or non-invasive biological information, wherein the blood glucose estimation model is generated by machine learning based on a training dataset, the training dataset including attribute information, non-invasive biological information, and blood glucose values ​​measured from blood of several subjects obtained in advance; the non-invasive biological information includes bioimpedance, blood pressure, and pulse wave data.

2. As in request item 1, the blood glucose level estimation device, wherein, The aforementioned attribute information includes either age or gender, or a combination thereof, and the aforementioned non-invasive organism information further includes either BMI or electrocardiogram data, or a combination thereof.

3. As in request item 2, the blood glucose level estimation device, wherein, The aforementioned non-invasive biological information system further includes blood oxygen saturation (SpO2).

4. The blood glucose estimation device according to any one of claims 1 to 3 further comprises: a training data memory unit that stores a training data set; and a learning processing unit that generates the blood glucose estimation model based on the training data set by means of machine learning.

5. As in request item 4, the blood glucose level estimation device, wherein, The above blood glucose estimation model outputs an estimated blood glucose value, which meets the measurement accuracy specified in the international standard ISO 15197.

6. As in request item 4, the blood glucose level estimation device, wherein, The aforementioned presumption processing unit calculates a risk presumption value for blood glucose levels, which indicates whether blood glucose levels have normalized.

7. As in request item 6, the blood glucose level estimation device, wherein, The aforementioned learning processing unit adds labels to the aforementioned training dataset indicating the presence or absence of blood glucose risk based on blood glucose levels measured from blood. When the difference between the number of individuals with the aforementioned blood glucose risk and the number without the aforementioned blood glucose risk exceeds a predetermined value, the unit manually generates and adds sample data to the aforementioned training dataset to reduce the aforementioned difference.

8. The blood glucose estimation device as described in request item 4, wherein, The learning processing unit generates a first blood glucose value estimation model and a second blood glucose value estimation model through machine learning based on different types of training datasets. The estimation processing unit uses the first blood glucose value estimation model and the second blood glucose value estimation model to calculate the estimated blood glucose value of the given user.

9. The blood glucose estimation device of claim 1 further includes a biometric information estimation unit, which estimates at least one of the biometric information, including BMI, blood pressure, pulse wave data, electrocardiogram data, biometric impedance and blood oxygen saturation, and the information acquisition unit acquires the biometric information estimated by the biometric information estimation unit as the biometric information of the predetermined user.

10. A non-invasive blood glucose estimation system, characterized in that: it comprises the blood glucose estimation device of any one of claims 1 to 9, and a biological information measuring device for measuring non-invasive biological information.

11. A method for estimating blood glucose levels, which is a computer-executed method for estimating blood glucose levels, comprising the following steps: memorizing a training dataset containing attribute information, non-invasive biological information, and blood glucose levels measured from blood samples of several pre-obtained subjects; generating a blood glucose level estimation model by machine learning based on the training dataset; and calculating the estimated blood glucose level of a predetermined user based on the blood glucose level estimation model and / or the non-invasive biological information of that user; wherein the non-invasive biological information includes bioimpedance, blood pressure, and pulse wave data.

12. A blood glucose estimation program that enables a computer to perform the following steps: memorizing a training dataset containing attribute information of several subjects obtained in advance, non-invasive biological information, and blood glucose values ​​measured from blood; generating a blood glucose estimation model by machine learning based on the training dataset; and calculating the estimated blood glucose value of a predetermined user based on the predetermined user's attribute information and / or non-invasive biological information using the blood glucose estimation model; wherein the non-invasive biological information includes bioimpedance, blood pressure, and pulse wave data.

Images