Cholesterol risk estimation device, cholesterol risk estimation method, and program

Abstract

To solve a problem that the measurement of LDL cholesterol or HDL cholesterol requires collecting blood from a subject and performing a biochemical analysis of the blood, which requires sticking a needle or the like into the skin of the subject in an invasive manner and involves a psychological or physical burden on the subject.SOLUTION: According to the present invention, a cholesterol estimation model is generated by machine learning on the basis of attribute information, non-invasive biological information, and blood test data of a plurality of subjects which have been acquired in advance. It is thus possible to estimate cholesterol in a non-invasive manner from attribute information and / or non-invasive biological information of a predetermined user.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a cholesterol risk estimation device, a cholesterol risk estimation method, and a program.

Background Art

[0002] Lipoproteins contained in blood can be classified into four types: chylomicrons, VLDL (Very Low Density Lipoprotein), LDL (Low Density Lipoprotein) cholesterol, and HDL (High Density Lipoprotein) cholesterol. HDL cholesterol is also called good cholesterol, and it is generally known that the risk of diseases such as arteriosclerosis increases when it is insufficient in the blood. According to the judgment criteria in the blood test established by the Japanese Society of Clinical Pathology, if the HDL cholesterol is 40 mg / dL or more, there is no abnormality, and if it is less than 40 mg / dL, it is determined that there is a risk. Also, LDL cholesterol is also called bad cholesterol, and it is generally known that the risk of diseases such as arteriosclerosis increases when it is present in large amounts in the blood. According to the judgment criteria in the blood test established by the Japanese Society of Clinical Pathology, if the LDL cholesterol is less than 120 mg / dL, there is no abnormality, and if it is 120 mg / dL or more, it is determined that there is a risk.

[0003] Conventionally, for the measurement of LDL cholesterol and HDL cholesterol, it has been necessary to collect the blood of a subject and perform biochemical analysis. However, this method has a problem that it is necessary to invasively pierce the skin of the subject with a needle or the like, which imposes a psychological or physical burden on the subject. As a method for non-invasively measuring lipids as the whole lipoprotein, the method described in Patent Document 1 is known.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

[0005] [Non-Patent Document 1] Psychology Research and Behavior Management 2011:4 81-86, Summary of the clinical investigations ESTeck Complex March, 20, 2010 Abstract [Non-Patent Document 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 Document 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 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] However, the method described in Patent Document 1 does not allow for the estimation of concentrations for each type of lipoprotein, such as LDL and HDL, making it difficult to determine whether or not individual LDL cholesterol and HDL cholesterol pose a health risk.

[0007] The present invention has been made in view of the above circumstances, and the object of the present invention is to provide a cholesterol risk estimation device, a cholesterol risk estimation method, and a program that can estimate the health risk of cholesterol, i.e., LDL cholesterol or HDL cholesterol, from non-invasive biological information with extremely high accuracy. [Means for solving the problem]

[0008] The cholesterol risk estimation device according to the present invention is characterized by comprising: an information acquisition unit that acquires attribute information and non-invasive biological information of a predetermined user; an estimation model storage unit that stores a cholesterol risk estimation model; and an estimation processing unit that uses the cholesterol risk estimation model to calculate an estimated cholesterol risk value for a predetermined user based on the attribute information and / or non-invasive biological information of the predetermined user.

[0009] The cholesterol risk estimation device according to the present invention is further characterized by comprising a training data storage unit that stores a training dataset, and a learning processing unit that generates a cholesterol risk estimation model by machine learning based on the training dataset.

[0010] Attribute information includes either age and / or sex, or a combination thereof, and non-invasive biometric information includes BMI, blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, or a combination thereof.

[0011] The training dataset is characterized by including subject attribute information, non-invasive biometric information, and cholesterol measurements taken from blood.

[0012] Non-invasive biological information is characterized by further including oxygen saturation (SpO2).

[0013] The estimation accuracy of the cholesterol risk estimate is characterized by an accuracy that allows for classification of the presence or absence of risk using an ROC_AUC of 0.7 or higher.

[0014] The learning processing unit is characterized by adding labels to the training dataset that indicate the presence or absence of cholesterol risk based on cholesterol measurement values ​​taken from blood, and if the difference between the number of labels indicating cholesterol risk and the number of labels indicating no cholesterol risk is greater than or equal to a predetermined value, it increases the sample data in the training dataset to reduce the difference.

[0015] The learning processing unit generates a first cholesterol risk estimation model and a second cholesterol risk estimation model using machine learning based on different types of training datasets, and the estimation processing unit calculates a cholesterol risk estimate for a given user using the first cholesterol risk estimation model and the second cholesterol risk estimation model.

[0016] The system further comprises a biometric information estimation unit that estimates at least one of the biometric information components included in the biometric information, such as BMI, blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation, and an information acquisition unit that acquires the biometric information estimated by the biometric information estimation unit as the biometric information of a predetermined user.

[0017] The system is characterized by comprising a cholesterol risk estimation device and a bio-information measurement device that further measures non-invasive biological information, thereby constituting a non-invasive cholesterol risk estimation system.

[0018] The cholesterol risk estimation method of the present invention is characterized by comprising the steps of: storing a training dataset including subject attribute information, non-invasive biological information, and cholesterol measurement values ​​measured from blood; generating a cholesterol risk estimation model by machine learning based on the training dataset; and calculating a cholesterol risk estimate for a predetermined user based on the attribute information and / or non-invasive biological information of a predetermined user using the cholesterol risk estimation model.

[0019] The program according to the present invention causes a computer to execute steps of storing a training data set including attribute information of a subject, non-invasive biological information, and a measured value of cholesterol measured from blood; generating a cholesterol risk estimation model by machine learning based on the training data set; and calculating an estimated cholesterol risk value of a predetermined user based on the attribute information and / or non-invasive biological information of the predetermined user using the cholesterol risk estimation model.

Effect of the Invention

[0020] According to the present invention, there is provided a cholesterol risk estimation device, a cholesterol risk estimation method, and a program capable of extremely accurately estimating the risk to health of cholesterol, that is, LDL cholesterol or HDL cholesterol, by machine learning using non-invasive biological information.

Brief Description of the Drawings

[0021] [Figure 1] It is a block diagram showing a schematic configuration of a cholesterol risk estimation system. [Figure 2] It is a diagram for explaining ESG (electroscangram). [Figure 3] It is a hardware configuration diagram of a cholesterol risk estimation device. [Figure 4] It is a flowchart showing an execution procedure for generating an LDL cholesterol risk estimation model by machine learning. [Figure 5] It is a flowchart showing an execution procedure for LDL cholesterol risk estimation processing. [Figure 6] It is a flowchart showing an execution procedure for generating an HDL cholesterol risk estimation model by machine learning. [Figure 7] It is a flowchart showing an execution procedure for HDL cholesterol risk estimation processing. [Figure 8]This is the ROC_AUC curve of the estimation results in Example 1. [Figure 9] This is the ROC_AUC curve of the estimation results in Example 2. [Figure 10] This is the ROC_AUC curve of the estimation results in Example 3. [Modes for carrying out the invention]

[0022] The embodiments will be described below with reference to the drawings. Note that the embodiments are illustrative and the present invention is not limited to the configurations described below.

[0023] <Device Functions> The cholesterol risk estimation system 1 and cholesterol risk estimation device 30 according to this embodiment will be described with reference to Figures 1 to 7. Figure 1 is a block diagram showing the schematic configuration of the cholesterol risk estimation system 1 according to this embodiment. The cholesterol risk estimation system 1 comprises a terminal device 10, a biological information measuring device 20, a cholesterol risk estimation device 30, and a display device 39. Here, "user" refers to a person who uses the cholesterol risk estimation system to obtain a non-invasive estimate of their cholesterol risk. "Subject" refers to a person who, after following the prescribed procedures and obtaining consent, provides attribute information such as age and gender, non-invasive biological information, and cholesterol measurements taken from blood as training data to be used in the cholesterol risk estimation system.

[0024] The terminal device 10 can be any information terminal that allows user attribute information (name, ID, age, gender, etc.) to be input and that can output the input information to the cholesterol risk estimation device 30 via a wired or wireless communication network. Examples include tablet devices, smartphones, wearable devices, and other mobile devices, or a PC (Personal Computer). Height and weight may be measured by the biometric information measurement device 20 described later.

[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 by methods that do not require the insertion of instruments into the skin or into body orifices. Non-invasive biometric information can be measured using commercially available height measuring devices, weight measuring devices, blood pressure monitors, pulse oximeters, pulse wave meters, electrocardiograms, impedance measuring devices, galvanic skin measuring devices, etc. Alternatively, the ESTEC BC-3 (Ryobi Systems), which can simultaneously measure pulse wave data, electrocardiogram data, bioimpedance, and oxygen saturation (SpO2), can also be used. These devices can measure non-invasive biometric data without causing psychological or physical burden to the user.

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

[0027] BMI is calculated from height h [m] and weight w [kg] using the following formula. BMI = w / h 2 [kg / m 2 ]

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

[0029] Pulse wave data is obtained by irradiating a protruding part of the body, such as a finger, with red light (~660nm) from a Red LED and near-infrared light (~905nm) from an IR LED, and then measuring the transmitted light using a phototransistor. Pulse wave data includes one or a combination of pulse rate, elasticity index, peripheral vascular resistance, acceleration pulse wave, b / a, e / a, -d / a, Takazawa acceleration pulse wave aging index, ejection fraction, LVET, DEI (heavy pulse elasticity index). Here, the elasticity index is a value obtained by dividing height by the time from the peak of systole to the peak of diastolic pulse wave detection in the fingertip plethysmography. Peripheral vascular resistance is calculated by mean arterial pressure / cardiac output × 80. DEI (Duration Elasticity Index) is an index that shows the elasticity of diastolic blood vessels and can be measured with a PWV measuring device. 0.3 to 0.7 is normal, below 0.3 may indicate hypertension or arteriosclerosis, and above 0.7 may suggest acute anxiety neurosis. The acceleration pulse wave is the second derivative of the photoplethysmogram (SDPTG) of the photoplethysmogram (PTG). The acceleration pulse wave consists of an initial positive wave (wave a), an initial negative wave (wave b), a mid-systolic rise wave (wave c), a late-systolic fall wave (wave d), and an initial diastolic positive wave (wave e). The b / a, e / a, and -d / a ratios are calculated from the ratio of each wave height. Since b / a increases and c / a, d / a, and e / a decrease with age, the Takazawa acceleration pulse wave aging index (bcde) / a can be used to evaluate vascular aging. Ejection fraction is the proportion of blood sent from the ventricle in each heartbeat and is proportional to the acceleration pulse wave aging index. LVET is left ventricular ejection time, which is the time it takes for blood in the left ventricle to be ejected into the aorta after the aortic valve opens.

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

[0031] The impedance (conductance) of a living organism can be measured, for example, by passing a weak electric current between two of six electrodes located on both legs, both hands, and the left and right foreheads. By passing current through two of the six electrodes, it is possible to measure (1) anode / cathode conductance (μS), (2) cathode / anode conductance (μS), (3) the difference between the conductance measured in (1) and the conductance measured in (2) (delta SCRA-SCRC), and (4) electrical conductivity (μS / m). Muscle mass, body fat mass, total water content, phase angle, and resistance can also be measured simultaneously. Furthermore, the dielectric constant (μSi) can be measured when current is passed between the right hand and the left forehead, and between the right and left foreheads. It is preferable to measure the impedance (conductance) of the living organism by utilizing 22 patterns of conductivity from the six electrodes.

[0032] Bioelectrical impedance includes body fat mass (kg), body fat mass (%), lean body mass, lean body mass percentage, muscle mass, total water content (kg), total water content (%), intracellular water content (%), cardiac output, 1 left forehead-2 left Hand / SCR A, 1 forehead left side-2 left Hand / Delta SCR C-SCR A, 5 Left hand - 6 Left foot / SCR A, 5 Left hand - 6 Left foot / Delta SCR C-SCR A, 13 Left foot - 14 Right foot / SCR C, 13 Left foot - 14 Right foot / SCR A, 15 Right hand - 16 Left side of forehead / Delta SCR C-SCR A, 15 Right hand - 16 Left side of forehead SCR C, 19 Right foot - 20 Left hand / Delta SCR C-SCR A, ESG 2+4+15+17 (μS / m), ESG 6+13+19 (%), ESG 6+8+19+21 (%), ESG 6+8+19+21 (μS / m), ESG This includes one or a combination of the following: 9+10 (μS / m), ESG9+10 (%), left foot conductance, R (Ω), phase angle, permittivity of the forehead path, electrical conductivity to the forehead path (9), permittivity of the one-handed to one-handed path, electrical conductivity from hand to hand (11,12), stroke volume (cardiac output ÷ heart rate).

[0033] Here, SCR is an abbreviation for skin conductance response, and ESG is an abbreviation for electroscangram. The "+" in ESG2+4+15+17 indicates which electrode attached to the body was used for measurement. For example, ESG2+4+15+17 means the average conductance measured on the left hand when current was passed from the left hand to the left forehead, on the right hand when current was passed from the right hand to the right forehead, on the right hand when current was passed from the right hand to the left forehead, and on the left hand when current was passed from the left hand to the right forehead, as shown in Figure 2. These conductances are described in detail in Non-Patent Document 1. The measurement method for ESG (electroscangram) is described in detail in Non-Patent Documents 2 and 3. "1 Left side of forehead-2 left Hand / SCR A is a cathode on the left side of the forehead, and the second side is a cathode. left "5 Left Hand - 6 Left Foot / Delta SCR C - SCR A" is the conductance (or conductivity) of the path measured when electricity flows with the hand facing the anode, and "5 Left Hand - 6 Left Foot / Delta SCR C - SCR A" is the difference in conductance measured when current is passed between "5 Left Hand" and "6 Left Foot" in anode-cathode and cathode-anode directions.

[0034] BMI and blood pressure can be measured using a height and weight scale and a blood pressure monitor. Non-invasive biological information may also include oxygen transport capacity calculated from SpO2 and cardiac output.

[0035] The measured non-invasive biological information is output to the cholesterol risk estimation device 30 via a wired or wireless communication network. The biological information measuring device 20 may be a built-in measuring device or a portable measuring device such as a wearable terminal.

[0036] The cholesterol risk estimation device 30 comprises a first acquisition unit 31, a second acquisition unit 32, a user data storage unit 33, a training data storage unit 34, a learning processing unit 35, an estimation model storage unit 36, an estimation processing unit 37, and an estimation data storage unit 38. The first acquisition unit 31 acquires user attribute information from the terminal device 10. The second acquisition unit 32 acquires non-invasive biological information of the user from the biological information measurement device 20.

[0037] The user data storage unit 33 stores user attribute information and non-invasive biological information acquired from the first acquisition unit 31 and the second acquisition unit 32.

[0038] The training data storage unit 34 stores multiple training datasets for machine learning, each consisting of attribute information of multiple subjects acquired in advance, non-invasive biological information, and sample test information of cholesterol (LDL cholesterol or HDL cholesterol) obtained from blood tests. The sample test information may also include test information obtained from blood, urine, stool, etc.

[0039] The learning processing unit 35 retrieves the training dataset stored in the training data storage unit 34 and uses the training dataset to create a cholesterol risk estimation model. Specifically, when estimating LDL cholesterol risk, the acquired training dataset is used to learn the relationship between attribute information and non-invasive biological information and LDL cholesterol risk through machine learning using gradient boosting such as XGBoost. When estimating HDL cholesterol risk, the acquired training dataset is standardized or normalized, and the relationship between attribute information and non-invasive biological information and HDL cholesterol risk is learned through machine learning using logistic regression. The estimation model storage unit 36 ​​stores the cholesterol risk estimation model generated by the learning processing unit 35. Non-invasive biometric data includes BMI, blood pressure, pulse wave data, electrocardiogram data, and bioelectrical impedance, or a combination thereof. Additionally, oxygen saturation (SpO2) may be 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 cholesterol risk based on predetermined user attribute information and / or non-invasive biological information. The estimated cholesterol risk is then stored in the estimation data storage unit 38.

[0041] The display device 39 can display estimated cholesterol risk values ​​along with user attribute information and non-invasive biological information. This data may also be displayed on the user's terminal device 10.

[0042] <Device Hardware Configuration> Figure 3 is a hardware configuration diagram of the cholesterol risk estimation device 30. As shown in Figure 3, the cholesterol risk estimation device 30 is composed of a computer 300 having one or more processors 301, a memory 302, a storage 303, an input / output port 304, and a communication port 305. The processor 301 performs processing related to cholesterol estimation according to this embodiment by executing a program. The memory 302 temporarily stores the program and the calculation results of the program. The storage 303 stores the program that executes the processing by the cholesterol risk estimation device 30. The storage 303 can be anything that is readable by the computer, and various recording media such as recording media (magnetic disks, optical disks, etc.), random access memory, flash-type memory, and read-only memory can be used. The input / output port 304 inputs information from the terminal device 10 and the biological information measuring device 20, and outputs cholesterol estimates to the display device 39. The communication port 305 transmits and receives data with other information terminals such as computers (not shown). Wireless communication and wired communication can be used as the communication method. Furthermore, the cholesterol risk estimation device 30 can be implemented using a commercially available desktop PC or notebook PC, and the time required to calculate the estimated cholesterol risk using the estimation model is only a few seconds. In addition, when the cholesterol risk estimation device 30's processor 301 is in operation, the first acquisition unit 31, the second acquisition unit 32, the learning processing unit 35, the estimation processing unit 37, etc., are all functioning.

[0043] By generating a cholesterol estimation model using machine learning based on non-invasive biometric data including BMI (Body Mass Index), blood pressure, pulse wave data, electrocardiogram data, and bioelectrical impedance, it becomes possible to estimate cholesterol levels without performing blood tests. Furthermore, as shown below, it is possible to determine whether cholesterol levels have normalized or not even when the number of data points included in non-invasive biological data is limited.

[0044] <Generation of an LDL cholesterol risk estimation model using machine learning> Figure 4 is a flowchart showing the procedure for generating and executing an LDL cholesterol risk estimation model using machine learning. In step ST101, the learning processing unit 35 preprocesses the input data (for example, the training dataset described above). Specifically, the learning processing unit 35 converts LDL cholesterol obtained from blood tests into 0 (no risk) for LDL cholesterol less than 120 mg / dl and 1 (risk) for LDL cholesterol of 120 mg / dl or more. Furthermore, if there is a discrepancy and imbalance between the number of people classified as 0 and the number of people classified as 1, the learning processing unit 35 may apply SMOTE (Chawla, NV. et al. 2002) to the training data to artificially generate training samples. That is, in the input data (training dataset), labels indicating the presence or absence of LDL cholesterol risk (for example, 0 or 1 as described above) are added based on LDL cholesterol obtained from blood tests, and if the difference between the number of people with LDL cholesterol risk (label: 1) and the number of people with LDL cholesterol risk (label: 0) is greater than or equal to a predetermined value, the sample data in the training dataset may be increased (generated) to reduce this difference. In step ST102, the learning processing unit 35 performs machine learning using gradient boosting decision trees. For machine learning using gradient boosting regression trees, software libraries such as XGBoost, CatBoost, and LightBGM can be used. Also, the risk value obtained by comparing the LDL cholesterol value obtained by blood test with a predetermined threshold (0: no risk, 1: at risk) was compared with the LDL cholesterol risk value estimated by machine learning, and each parameter of XGBoost (max_depth, subsample, colsample_bytree, learning_rate) was adjusted so that the f1 score was maximized, and an estimation model was generated. max_depth is the depth of the decision tree, subsample is the ratio of samples randomly extracted in each tree, colsample_bytree is the ratio of columns randomly extracted in each tree, and learning_rate indicates the learning rate. However, max_depth was adjusted within the range of 1 to 10, subsample was adjusted within the range of 0.1 to 1.0, colsample_bytree was adjusted within the range of 0.3 to 1.0, and learning_rate was adjusted within the range of 0.1 to 0.7. The learning processing unit 35 stores the LDL cholesterol risk estimation model generated by the above learning processing in the estimation model storage unit 36. Note that the above machine learning algorithms are just examples and are not limited thereto.

[0045] <Estimation of LDL Cholesterol Risk Using the LDL Cholesterol Risk Estimation Model> As shown in Figure 5, in step ST201, the first acquisition unit 31 of the cholesterol risk estimation device 30 acquires user attribute information from the terminal device 10. In step ST202, the second acquisition unit 32 of the cholesterol risk estimation device 30 acquires the user's non-invasive biological information. The user's attribute information and non-invasive biological information are then stored in the user data storage unit 33. Then, in step ST203, the estimation processing unit 37 uses the LDL cholesterol risk estimation model stored in the estimation model storage unit 36 ​​to calculate the probability of belonging to class 0 (no risk) or class 1 (risk present), i.e., the LDL cholesterol risk estimate. In step ST204, the calculated LDL cholesterol risk estimate is stored in the estimation data storage unit 38, and in step ST205, the LDL cholesterol risk estimate is output to an external terminal such as the display device 39 for display.

[0046] <Generation of HDL cholesterol risk estimation models using machine learning> Figure 6 is a flowchart showing the procedure for generating and executing an HDL cholesterol risk estimation model using machine learning. In step ST301, the learning processing unit 35 preprocesses the input data. Specifically, the learning processing unit 35 converts HDL cholesterol obtained from blood tests into 0 (no risk) for HDL cholesterol of 40 mg / dl or more and 1 (risk) for HDL cholesterol less than 40 mg / dl. Furthermore, if there is a discrepancy and imbalance between the number of people classified as 0 and the number of people classified as 1, the learning processing unit 35 may apply SMOTE (Chawla, NV. et al. 2002) to the training data to artificially generate training samples. That is, in the input data (training dataset), labels indicating the presence or absence of HDL cholesterol risk (e.g., 0 or 1 as described above) are added based on HDL cholesterol obtained from blood tests, and if the difference between the number of people with HDL cholesterol risk (label: 1) and the number of people with HDL cholesterol risk (label: 0) is greater than a predetermined value, the sample data in the training dataset may be increased (generated) to reduce this difference. Further data standardization and normalization are performed. In step ST302, the learning processing unit 35 performs machine learning using logistic regression. For machine learning using logistic regression, for example, Logistic Regression provided in Scikit-learn, an open-source machine learning library for Python, can be used. Alternatively, the dimensionality may be reduced by principal component analysis if necessary. Further, the risk value (0: no risk, 1: risk) obtained by comparing the HDL cholesterol value obtained by blood test with a predetermined threshold value was compared with the HDL cholesterol risk value estimated by machine learning, and each parameter (C, regularization method, max_iter, solber) of Logistic Regression was adjusted so that the f1 score was maximized to generate an HDL cholesterol risk estimation model. Here, C is a trade-off parameter that determines the strength of regularization, and the larger the value, the weaker the regularization strength. The regularization method means L1 regularization or L2 regularization, and this is selected. max_iter is the maximum number of times to repeat learning. With solber, a convergence method (e.g., L-BFGS method, Newton CG method, liblinear, sag, and saga) that minimizes the cross-entropy error is selected. In Examples 2 and 3 below, the liblinear method was selected. The learning processing unit 35 stores the HDL cholesterol risk estimation model generated by the above learning processing in the estimation model storage unit 36. Note that the above machine learning algorithms are just examples and are not limited to these.

[0047] <Estimation of HDL Cholesterol Risk Using the HDL Cholesterol Risk Estimation Model> As shown in Figure 7, in step ST401, the first acquisition unit 31 of the cholesterol risk estimation device 30 acquires user attribute information from the terminal device 10. In step ST402, the second acquisition unit 32 of the cholesterol risk estimation device 30 acquires the user's non-invasive biological information. The user's attribute information and non-invasive biological information are then stored in the user data storage unit 33. Then, in step ST403, the estimation processing unit 37 uses the HDL cholesterol risk estimation model stored in the estimation model storage unit 36 ​​to calculate the probability of belonging to class 0 (no risk) or class 1 (risk present), i.e., the HDL cholesterol risk estimate. In step ST404, the calculated HDL cholesterol risk estimate is stored in the estimation data storage unit 38, and in step ST405, the HDL cholesterol risk estimate is output to an external terminal such as the display device 39 for display.

[0048] <Example (LDL cholesterol risk estimation)> Examples of LDL cholesterol risk estimation are described below. However, the embodiments of LDL cholesterol risk estimation in the present invention are not limited to the following examples. Attribute information includes ID, name, age, gender, or a combination thereof, while non-invasive biometric information includes BMI, blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2), or a combination thereof. Height and weight, which are the basis for calculating BMI, were measured using a height meter and a weight scale, respectively, and blood pressure was measured using a blood pressure monitor. Pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) were measured using an ESTEC BC-3 (Ryobi Systems). Alternatively, a commercially available pulse wave meter, electrocardiograph, impedance measuring device, and pulse oximeter may be used in combination instead of the ESTEC BC-3. Furthermore, the above-mentioned non-invasive biometric information may be acquired using a designated wearable terminal. Bioelectrical impedance (conductance) was measured by passing a weak electric current between two of six electrodes located on both feet, both hands, and both sides of the forehead. The voltage and current were set to 1.28V and 200 μA, and conductance was measured for 32 milliseconds per second. With current flowing through two of the six electrodes, the following measurements were taken: (1) anode / cathode conductance (μS), (2) cathode / anode conductance (μS), (3) the difference between the conductance measured in (1) and the conductance measured in (2) (delta SCRA-SCRC), and (4) electrical conductivity (μS / m). In addition, muscle mass, body fat mass, total water content, phase angle, and resistance values ​​were measured, and the dielectric constant (μSi) was also measured when current was passed between the right hand and the left forehead, and between the right and left forehead.

[0049] Using the ESTEC BC-3, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) were measured for each subject for 2 minutes. For the measurements, a device with electrocardiogram, pulse wave meter, and pulse oximeter functions was attached to the left index finger of each subject, and two electrodes were attached to their forehead. The subjects were seated in a chair with both hands and feet placed on electrode plates.

[0050] <Learning Model 1> In learning model 1, machine learning was performed using a gradient boosting decision tree, as shown in Figure 4. In this process, the following data was selected and used as attribute information and non-invasive biological data. (A) Attribute information ·age (B) Non-invasive biometric data Blood pressure... Diastolic blood pressure • Pulse wave data... Elasticity index, • Electrocardiogram data: heart rate, standard deviation of RR interval • Bioelectrical impedance: Body fat mass (kg), muscle mass, total water content (%), cardiac output, 5 left hand - 6 left foot / SCRA, ESG 2+4+15+17 (μS / m), ESG 9+10 (μS / m), ESG 9+10 (%), R (Ω) Here, ESG9+10 represents the average impedance measured at the locations shown in Figure 2. [μS / m] is the unit of the measured average value, and [%] is the value obtained by scaling the measured average value to within the range that can normally be measured. Furthermore, non-invasive biometric data also includes cardiac output, which is included in bioelectrical impedance, and oxygen transport capacity, which is estimated from oxygen saturation (SpO2).

[0051] <Example 1> In Example 1, an LDL cholesterol risk estimation model was generated using machine learning with the above-mentioned learning model 1, based on (1) attribute information of 712 subjects, (2) non-invasive biological information measured by a height and weight scale, blood pressure monitor, and ESTEC BC-3, and (3) LDL cholesterol obtained from blood tests performed on the same day as the non-invasive biological information measurement. The estimation accuracy of the LDL cholesterol risk estimation model was then evaluated using the ROC_AUC curve. As a result, the ROC_AUC was 0.71, exceeding 0.7, which indicates good classification. Figure 8 shows the ROC_AUC curve for the estimation results of Example 1.

[0052] <Example (HDL cholesterol risk estimation)> Examples of HDL cholesterol risk estimation are described below. However, the embodiments of HDL cholesterol risk estimation in the present invention are not limited to the following examples. Attribute information includes ID, name, age, gender, or a combination thereof, while non-invasive biometric information includes BMI, blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2), or a combination thereof. Height and weight, which are the basis for calculating BMI, were measured using a height meter and a weight scale, respectively, and blood pressure was measured using a blood pressure monitor. Pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) were measured using an ESTEC BC-3 (Ryobi Systems). Alternatively, a commercially available pulse wave meter, electrocardiograph, impedance measuring device, and pulse oximeter may be used in combination instead of the ESTEC BC-3. Furthermore, the above-mentioned non-invasive biometric information may be acquired using a designated wearable terminal. Bioelectrical impedance (conductance) was measured by passing a weak electric current between two of six electrodes located on both feet, both hands, and both sides of the forehead. The voltage and current were set to 1.28V and 200 μA, and conductance was measured for 32 milliseconds per second. With current flowing through two of the six electrodes, the following measurements were taken: (1) anode / cathode conductance (μS), (2) cathode / anode conductance (μS), (3) the difference between the conductance measured in (1) and the conductance measured in (2) (delta SCRA-SCRC), and (4) electrical conductivity (μS / m). In addition, muscle mass, body fat mass, total water content, phase angle, and resistance values ​​were measured, and the dielectric constant (μSi) was also measured when current was passed between the right hand and the left forehead, and between the right and left forehead.

[0053] Using the ESTEC BC-3, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) were measured for each subject for 2 minutes. For the measurements, a device with electrocardiogram, pulse wave meter, and pulse oximeter functions was attached to the left index finger of each subject, and two electrodes were attached to their forehead. The subjects were seated in a chair with both hands and feet placed on electrode plates.

[0054] <Learning Model 2> In learning model 2, machine learning was performed using logistic regression, as shown in Figure 6. In this process, the following data was selected and used as attribute information and non-invasive biological data. (A) Attribute information ·sex (B) Non-invasive biometric data Blood pressure... pulse pressure • Pulse wave data: peripheral vascular resistance, pulse rate, e / a, acceleration pulse wave aging index • ECG data: LF / HF • Bioelectrical impedance: Lean body mass (%), 15 right hand - 16 left forehead / SCRC, ESG 9 + 10 (μS / m), dielectric constant of the forehead path • Pulse oximeter... SpO2 Furthermore, non-invasive biodata includes cardiac output, which is included in bioelectrical impedance, as well as oxygen transport capacity estimated from oxygen saturation (SpO2), and pulse pressure / pulse rate calculated from pulse rate and pulse pressure.

[0055] <Learning Model 3> In learning model 3, machine learning was performed using logistic regression, as shown in Figure 6. For this purpose, the following non-invasive biological data was selected and used. In this process, the following data was selected and used as attribute information and non-invasive biological data. (A) Attribute information ·sex (B) Non-invasive biometric data · BMI • ECG data... RR interval, • Bioelectrical impedance: Cardiac output, 13 left foot - 14 right foot / SCRC, R (Ω)

[0056] Furthermore, non-invasive biometric data also includes cardiac output, which is included in bioelectrical impedance, and oxygen transport capacity, which is estimated from oxygen saturation (SpO2).

[0057] <Example 2> In Example 2, an HDL cholesterol risk estimation model was generated using machine learning with the above-mentioned learning model 2, based on (1) attribute information of 321 subjects, (2) non-invasive biological information measured by a height and weight scale, blood pressure monitor, and ESTEC BC-3, and (3) HDL cholesterol obtained from blood tests performed on the same day as the non-invasive biological information measurement. The estimation accuracy of the HDL cholesterol risk estimation model was then evaluated using the ROC_AUC curve. As a result, the ROC_AUC was 0.78 The result was above 0.7, indicating good classification. The ROC_AUC curve for the estimation results of Example 2 is shown in Figure 9.

[0058] <Example 3> In Example 3, an HDL cholesterol risk estimation model was generated using machine learning with the above-mentioned learning model 3, based on (1) attribute information of a total of 712 subjects, (2) non-invasive biological information measured by a height and weight scale, blood pressure monitor, and ESTEC BC-3, and (3) HDL cholesterol obtained from blood tests performed on the same day as the non-invasive biological information measurement. The estimation accuracy of the HDL cholesterol risk estimation model was then evaluated using the ROC_AUC curve. The result showed a ROC_AUC of 0.87, exceeding 0.8, which indicates extremely good classification. Figure 10 shows the ROC_AUC curve for the estimation results of Example 3.

[0059] (modified version) In the embodiments described above, an example of estimating HDL cholesterol risk using learning model 2 or learning model 3 was explained, but HDL cholesterol risk may be estimated using multiple learning models. This allows for more accurate estimation than using a single learning model to estimate HDL cholesterol risk.

[0060] Furthermore, in the embodiments and examples described above, there have been cases where the estimation accuracy improved when BMI was used as the dataset for the learning model. Therefore, a function unit for estimating BMI may be provided in the cholesterol risk estimation device. While BMI is generally not obtained through wearable devices and is instead calculated from user-entered height and weight, estimating BMI allows for the acquisition of cholesterol risk information by obtaining only biometric data, thus improving user convenience. The method for estimating BMI is not particularly limited, but for example, BMI is known to correlate with the tilt of the user's abdomen (at a predetermined position). Therefore, for example, a predetermined accelerometer may be placed on the user's abdomen (or a wristband-type wearable device equipped with an accelerometer may be placed on the abdomen), and the tilt of the abdomen may be determined based on the data output from the accelerometer, and the BMI may be estimated.

[0061] Furthermore, as described above, pulse wave data and oxygen saturation may also be estimated using a wristband-type wearable device equipped with a pulse wave sensor or a medium oxygen concentration sensor.

[0062] Furthermore, blood pressure may also be estimated using a wristband-type wearable device, similar to the method described above. This is because it is known that there is a correlation between the velocity of the pulse wave sent through the arteries by the heartbeat and blood pressure. Therefore, blood pressure may be estimated using a predetermined sensor that measures the velocity of the pulse wave sent through the arteries by the heartbeat.

[0063] Furthermore, as described above, electrocardiogram data may also be estimated using a wristband-type wearable device. For example, electrocardiogram data can be estimated based on data obtained from electrodes provided on the side of the wristband-type wearable device opposite to the display side and electrodes provided on the display side. Specifically, electrocardiogram data can be estimated from data obtained when the wrist of the hand wearing the wristband-type wearable device (e.g., left hand) touches the electrodes provided on the opposite side, and the fingertips of the hand opposite to the hand wearing the device (e.g., right hand) touch the electrodes provided on the display side.

[0064] Furthermore, similar to the above, bioimpedance may also be estimated using a wristband-type wearable device equipped with various electrodes. For example, bioimpedance can be estimated based on biometric information obtained from the chest and wrist using a wristband-type wearable device.

[0065] Furthermore, the biometric information to be estimated may be estimated using a classifier generated with various machine learning algorithms, with biometric information obtainable from a wearable device and the biometric information to be estimated as described above (BMI, blood pressure, pulse wave data, electrocardiogram data, and at least one of the biometric information such as bioelectrical impedance and oxygen saturation) as training data. In this case, the above-mentioned 2 The acquisition unit may acquire estimated biological information.

[0066] (others) Furthermore, for example, the series of processes described above can be executed by hardware or by software. In other words, the functional configuration described above is merely illustrative and not particularly limiting. That is, it is sufficient for the information processing system to have the functionality to execute the series of processes described above as a whole, and the type of functional block used to realize this functionality is not particularly limited to the example above. Also, the location of the functional blocks is also shown in the diagram. 1It is not limited to this and is optional. For example, the functional blocks of the server may be transferred to other terminals or devices. Conversely, the functional blocks of other terminals or devices may be transferred to the server. Also, a single functional block may consist of hardware alone, software alone, or a combination of both.

[0067] When a series of processes are executed by software, the programs that make up that software are installed on a computer or other device from a network or storage medium. The computer may be a computer built into dedicated hardware. Alternatively, the computer may be a computer capable of performing various functions by installing various programs, such as a server, a general-purpose smartphone, or a personal computer.

[0068] Such recording media containing programs may consist not only of removable media (not shown) distributed separately from the main device to provide programs to users, but also of recording media provided to users in a state where they are pre-installed in the main device. Since programs can be distributed via a network, the recording media may be installed on or accessible from a computer connected to or capable of connecting to a network.

[0069] In this specification, the step of describing a program to be recorded on a recording medium includes not only processes that are performed chronologically in that order, but also processes that are not necessarily performed chronologically, but are executed in parallel or individually. Furthermore, in this specification, the term "system" refers to an overall system composed of multiple devices, means, etc. [Explanation of symbols]

[0070] 1. Cholesterol Risk Estimation System 10 Terminal devices 20. Biological Information Measurement Device 30 Cholesterol Risk Estimation Device 31 First acquisition part 32 Second acquisition part 33 User data storage unit 34 Training data storage unit 35 Learning Processing Unit 36 Estimated Model Memory Unit 37 Estimation Processing Unit 38 Estimated data storage unit 39 Display device 300 Computers 301 Processor 302 memory 303 Storage 304 input / output ports 305 Communication Port

Claims

1. An information acquisition unit that acquires attribute information and non-invasive biometric information of a specified user, An estimation model storage unit that stores cholesterol risk estimation models, An estimation processing unit calculates an estimated cholesterol risk value for a predetermined user based on the attribute information and non-invasive biological information of the predetermined user using the cholesterol risk estimation model, Equipped with, The attribute information includes at least gender, The aforementioned non-invasive biological information is (1) Blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) or (2) BMI, electrocardiogram data, and bioelectrical impedance including, A cholesterol risk estimation device characterized by the following features.

2. The aforementioned attribute information further includes age, The cholesterol risk estimation device according to feature 1.

3. The estimation accuracy of the cholesterol risk estimate is such that the presence or absence of risk can be classified by an ROC_AUC of 0.7 or higher. A cholesterol risk estimation device according to claim 1 or 2.

4. A training data storage unit that stores the training dataset, A learning processing unit that generates the cholesterol risk estimation model by machine learning based on the training dataset, A cholesterol risk estimation device according to any one of claims 1 to 3, further comprising the above.

5. The training dataset includes subject attribute information, non-invasive biometric information, and cholesterol measurements taken from blood. The cholesterol risk estimation device according to feature 4.

6. The cholesterol risk estimation device according to claim 5, characterized in that the non-invasive biological information further includes oxygen saturation (SpO2).

7. The aforementioned learning processing unit, In the aforementioned training dataset, labels indicating the presence or absence of cholesterol risk are added based on cholesterol measurement values ​​measured from blood. If, in the above label, the difference between the number of labels with cholesterol risk and the number of labels without cholesterol risk is greater than or equal to a predetermined value, the sample data in the training dataset is increased to reduce the difference. A cholesterol risk estimation device according to any one of claims 4 to 6.

8. The learning processing unit generates a first cholesterol risk estimation model and a second cholesterol risk estimation model, respectively, using machine learning based on different types of training datasets. The estimation processing unit calculates the estimated cholesterol risk value for a predetermined user using the first cholesterol risk estimation model and the second cholesterol risk estimation model. A cholesterol risk estimation device according to any one of claims 4 to 7.

9. The system further comprises a biometric information estimation unit that estimates at least one of the biometric information included in the biometric information, such as BMI, blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation. The information acquisition unit acquires the biometric information estimated by the biometric information estimation unit as the biometric information of the predetermined user. A cholesterol risk estimation device according to any one of claims 1 to 8.

10. The cholesterol risk estimation device according to any one of claims 1 to 9, A bio-information measuring device that measures non-invasive biological information, A non-invasive cholesterol risk estimation system characterized by comprising the following features.

11. The steps include storing a training dataset containing subject attribute information, non-invasive biometric information, and cholesterol measurements taken from blood, The steps include generating a cholesterol risk estimation model using machine learning based on the aforementioned training dataset, The method includes the step of using the cholesterol risk estimation model to calculate an estimated cholesterol risk for a predetermined user based on the user's attribute information and non-invasive biometric information. The attribute information includes at least gender, The aforementioned non-invasive biological information is (1) Blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) or (2) BMI, electrocardiogram data, and bioelectrical impedance, including, A method for estimating cholesterol risk characterized by the following features.

12. The steps include storing a training dataset containing subject attribute information, non-invasive biometric information, and cholesterol measurements taken from blood, The steps include generating a cholesterol risk estimation model using machine learning based on the aforementioned training dataset, Using the cholesterol risk estimation model, the steps include calculating an estimated cholesterol risk for a predetermined user based on the user's attribute information and non-invasive biological information, Have the computer run it, The attribute information includes at least gender, The aforementioned non-invasive biological information is (1) Blood pressure, pulse wave data, electrocardiogram data, bioelectrical impedance, and oxygen saturation (SpO2) or (2) BMI, electrocardiogram data, and bioelectrical impedance, including, A program characterized by the following features.

Images