System and method for predicting insulin resistance or pancreatic beta-cell function and computer readable medium thereof

Abstract

Systems and methods for predicting insulin resistance and / or β-cell function utilize machine learning models to predict decreased insulin resistance and / or β-cell function in individuals requiring such prediction based on a feature set extracted from a database. Therefore, clinicians or individuals can be alerted before an individual is diagnosed with diabetes to take necessary steps and adjust related medical or lifestyle modifications. Computer-readable media of the corresponding systems and methods are also provided.

Description

Technical Field

[0001] This invention relates to medical monitoring applications, particularly to predictive techniques for individual insulin resistance and / or β-cell function. Prior Technology

[0002] Insulin resistance (IR) is a condition characterized by reduced insulin action in cells of skeletal muscle, liver, and adipose tissue. IR increases when a given concentration of insulin is insufficiently responded to glucose. Many potential factors can contribute to IR, including obesity, stress, medications (such as steroids), pregnancy, insulin antibodies, and genetic defects in the insulin signaling pathway. IR is also associated with clinical conditions such as diabetes mellitus (DM), coronary artery disease, metabolic syndrome, polycystic ovary syndrome, and non-alcoholic fatty liver disease.

[0003] Insulin resistance (IR) is known as an early diagnostic indicator of diabetes mellitus (DM). IR is one of the primary manifestations of metabolic syndrome. Reduced insulin action (IR) leads to overworking β-cells, causing them to secrete more insulin as a compensatory mechanism to maintain blood glucose levels. As metabolic syndrome worsens, the number of β-cells decreases, and blood glucose concentration begins to rise. Eventually, the elevated blood glucose reaches the diagnostic definition of diabetes. Notably, β-cell dysfunction (reduced insulin secretion) can begin as early as 12 years before a DM diagnosis and persist throughout the disease process. Numerous clinical studies have shown that earlier intervention in diabetes leads to better prognosis; early control of diabetes is highly beneficial in reducing the residual effects of glycemic abnormalities (i.e., metabolic memory). Therefore, identifying IR (i.e., early detection of DM) or β-cell function is crucial for the clinical manifestations of DM.

[0004] In clinical practice, homeostasis testing includes the Homeostatism Model Assessment of Insulin Resistance (HOMA-IR) and the Homeostatism Model Assessment of β-cell function (HOMA-β). HOMA assesses the degree of β-cell deficiency and insulin sensitivity based on an equation composed of fasting insulin concentration and fasting blood glucose concentration. However, fasting insulin values ​​are not regularly monitored; while calculating HOMA-IR and HOMA-β requires fasting insulin values. Therefore, although IR and β-cell deficiency have important clinical significance, they cannot be effectively used as indicators in clinical practice.

[0005] In recent years, machine learning has been used to address the aforementioned situations. However, most studies are based on predictive models for diabetes, rather than predictive models for indicators of early diagnosis of DM in non-diabetic patients (i.e., IR and decreased β-cell function).

[0006] In view of the above, there is an urgent need in the field for a solution that uses artificial intelligence to predict IR and / or β-cell function in non-diabetic populations. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a system for predicting an individual's insulin resistance and / or β-cell function, comprising: a database for providing a dataset; a feature extraction module for collecting and processing features from the dataset to form an individual's feature set, the feature set including the individual's age, sex, race, and body mass index; and a model building and optimization module for building a machine learning model based on the feature set and predicting the individual's insulin resistance and / or β-cell function using the machine learning model.

[0008] The present invention also provides a method for predicting an individual's insulin resistance and / or β-cell function, comprising: a database providing a dataset; a feature extraction module collecting and processing features from the dataset to form a feature set of the individual, the feature set including the individual's age, sex, race, and body mass index; and a model building and optimization module building a machine learning model based on the feature set and predicting the individual's insulin resistance and / or β-cell function by means of the machine learning model.

[0009] The present invention also provides a computer-readable medium having computer-executable instructions stored thereon, which, when executed, can implement the method described in the present invention. Simple Explanation of the Diagram

[0010] Figure 1 is a schematic diagram of the system architecture and method for predicting insulin resistance and / or β-cell function according to an embodiment of the present invention. Figure 2A is a graph showing the receiver operating characteristic (ROC) curves of insulin resistance (IR) in patients with chronic kidney disease (CKD) using training groups obtained through extreme gradient boosting (XGboost), random forest, logistic regression, and deep neural network (DNN) embodiments of the present invention. Figure 2B is a graph of the ROC curve of the prediction of IR of CKD patients using internal validation sets through XGboost, Random Forest, Logistic Regression and DNN in an embodiment of the present invention. Figure 2C is a graph of the ROC curves of the test group used to predict the IR of CKD patients through XGboost, Random Forest, Logistic Regression and DNN in an embodiment of the present invention. Figure 3A is a bar chart showing the detailed feature importance of the XGboost prediction model according to an embodiment of the present invention. Figure 3B is a bar chart showing the detailed feature importance of the random forest prediction model according to an embodiment of the present invention. Figure 4A is a heatmap illustrating the positive and negative impacts of predicting insulin resistance using the Shapley Additive exPlanations (SHAP) of the XGboost prediction model, according to an embodiment of the present invention. Figure 4B is a heatmap illustrating the positive and negative impacts of using the SHAP value of a random forest prediction model to predict insulin resistance in an embodiment of the present invention. Figures 5A and 5B are Kaplan-Meier survival curves for predicting clinically significant (cardiovascular disease (CV) mortality and all-cause mortality) CV mortality with and without IR, based on the XGboost prediction model from the Taiwan Human Biobank according to an embodiment of the present invention. Figures 5C and 5D are Kaplan-Mayer survival curves for all-cause mortality with and without IR, based on the XGboost prediction model from the Taiwan Human Biobank, according to embodiments of the present invention. Figure 6A is a bar chart illustrating the interpretation of personalized IR risks and feature sets in an embodiment of the present invention. Figure 6B is a bar chart illustrating the interpretation of personalized β-cell defect risk and feature set in an embodiment of the present invention. Implementation

[0011] The following embodiments are provided to illustrate the present invention. Those skilled in the art will readily understand the advantages and effects of the present invention after reading its contents, and it can also be implemented or applied in other different embodiments. Therefore, any element or method within the scope of this disclosure can be combined with any other element or method disclosed in any embodiment of the present invention.

[0012] The relationships, structures, steps, and elements shown in the accompanying illustrations are solely for illustrative purposes and to aid those skilled in the art in understanding the invention, and are not intended to limit the scope of the invention. Therefore, any changes, modifications, or adjustments to these features without affecting the purpose and effects of the invention fall within the scope of the technical content disclosed herein.

[0013] In this document, when the described object "comprises", "includes", or "has" technical features, it may also include other modules, elements, components, structures, parts, devices, systems, steps, connections, elements, features, etc., unless otherwise stated, and other features shall not be excluded.

[0014] The terms "first," "second," "third," "fourth," etc., used herein are only for describing or distinguishing limitations such as modules, sets, databases, elements, components, structures, parts, devices, systems, steps, connections, elements, or features, and are not intended to limit the scope of the invention or the spatial order of elements. Furthermore, unless otherwise specified, the singular forms "a" and "the" used herein also include the plural forms, and the terms "or" and "and / or" are used interchangeably.

[0015] As used in this article, the terms “individual,” “participant,” “person,” and “patient” are used interchangeably.

[0016] As used in this article, the terms "fasting blood glucose" and "fasting plasma glucose" are used interchangeably.

[0017] As used in this article, the terms "decreased β-cell function", "decreased β-cell function" and "β-cell defect" are used interchangeably.

[0018] As used in this article, the terms "variable", "element" and "characteristic" are used interchangeably.

[0019] As used herein, the terms “comprising,” “including,” “having,” or any other variation thereof are intended to encompass non-exclusive inclusion. For example, a module, process, or method that comprises a list of elements is not necessarily limited to those elements, but may include other elements not expressly listed or inherent to such a module, process, or method.

[0020] As used herein, the terms "Homorexed Insulin Resistance Model Assessment (HOMA-IR)" and "Homorexed Beta Cell Function Homeostasis Model Assessment (HOMA-β)" are used to assess insulin resistance (IR) and beta cell function. HOMA-IR and HOMA-β are the most appropriate methods for assessing IR and decline in beta cell function. Specifically, variables used in calculating HOMA-IR include age, sex, race, year of assessment, BMI, and smoking status. Furthermore, beta cell function is strongly associated with the development of type 2 diabetes (DM). This association is statistically independent of impaired glucose tolerance, obesity, and body fat distribution. High values ​​of HOMA-IR and HOMA-β are independently associated with a high risk of prediabetes. HOMA-IR measures insulin resistance using the following formula: fasting insulin level (μU / ml) × fasting blood glucose level (mg / dL) / 405; HOMA-β assesses beta cell function using the following formula: (20 × fasting insulin level (μU / ml)) / (fasting blood glucose level (mg / dL) - 63). For adults, a HOMA-IR value of 2.5 is the cutoff value for representing IR; and a HOMA-β value of 66% is the cutoff value for representing decreased β-cell function. In some embodiments, HOMA-IR and HOMA-β values ​​may be calculated for individuals in a third database (i.e., a combined database of the first and second databases); and patients or participants may be categorized into non-IR, IR, non-β-cell deficiency, and β-cell deficiency groups based on their respective HOMA-IR values ​​(≤ or > 2.5) and HOMA-β values ​​(< 66%).

[0021] In at least one embodiment of the present invention, the database includes a first database and a second database. The data in the first database and the second database are derived from individuals of a first group and individuals of the second group, respectively. In some embodiments, the ethnicity of the first group differs from that of the second group. For example, the first group may be non-Asian, and the second group may be Asian, but the present invention is not limited thereto.

[0022] Research Methods

[0023] As shown in Figure 1, the system and method for predicting IR and / or β-cell function are illustrated, with the operational steps described by arrows and the following text. Specifically, the abbreviation "ML" in Figure 1 stands for Machine Learning.

[0024] This disclosure includes the research plan and participants.

[0025] In some embodiments, the data used include a first dataset from a first database (e.g., the National Health and Nutrition Examination Survey (NHANES) database), a second dataset from a second database (e.g., the MAJOR, MJ Health Resource Center database), and a fourth dataset from a fourth database (e.g., the Taiwan Biobank Database (TWB)).

[0026] In some embodiments, the first database is the NHANES database, which originates from a U.S. health program. This health program's health surveys are regularly initiated by the Centers for Disease Control and Prevention (CDC) and the National Center for Health Statistics (NCHS). The first dataset is freely available to the public for research purposes. This disclosure was approved by the NCHS Research Ethics Review Board, and all participants or their representatives provided written informed consent. Examinations included anthropometrics, health and nutrition questionnaires, and laboratory tests. Participants completed questionnaires through in-home interviews. This disclosure analyzes participants from the NHANES program between January 1, 1999, and December 31, 2012. Participants under 18 years of age, with incomplete laboratory data, or suffering from diabetes mellitus (DM) were excluded from the analysis.

[0027] In some embodiments, the second database is the MJ Health Resource Center database, which is the most detailed and accurate medical database in Taiwan (Republic of China). The Taiwan MJ Cohort resource is an ongoing and active prospective study of women and men who have participated in large-scale health check-up programs offered by the Taiwan MJ Health Management Organization (www.mjclinic.com.tw). The Taiwan MJ Group is the largest health management group in Asia. Detailed information on the Taiwan MJ Cohort study and data collection methods is available elsewhere. Since 1994, the MJ Cohort has included approximately 600,000 Taiwanese individuals. Participants complete standardized procedures, including health history questionnaires, multiphasic blood tests, and laboratory tests such as pulmonary function, electrocardiogram, urinalysis, and stool analysis. These procedures comply with ISO 9001 quality management standards. After the initial check-up, all participants are encouraged to return annually for the MJ Health Resource Center to update all data.

[0028] In some embodiments, step S1 indicates that the database merging module merges the first database and the second database into a third database, that is, the third database is a combination of the first database and the second database.

[0029] In some embodiments, the fourth database is the Taiwan Biobank Database (TWB), a project initiated by the Ministry of Health and Welfare of Taiwan to collect national genetic and laboratory data of the Taiwanese people. Taiwan possesses one of the world's most comprehensive health databases, covering up to 99% of the population. The TWB is a rich database for biomedical research. The TWB database currently contains 174,077 participants, 84,276,467 questionnaires, 72,076,086 weight / height data points, and 79,314,928 laboratory data points. Biomarkers and genetic data can be generated from urine and blood samples of all participants. This disclosure analyzes participants from the TWB population between December 10, 2008, and November 30, 2018.

[0030] In some embodiments, a fourth database, such as the TWB database, is used for external validation to determine whether IR or β-cell function estimated by a third database can predict mortality in non-diabetic individuals by implementing diagnostic algorithms.

[0031] In some embodiments, participants were collected from a first database, a second database, and a fourth database (e.g., the NHANES database (1999–2012), the MJ database (2008–2017), and the TWB database (October 10, 2008 to November 30, 2018). Baseline variables included age (years), sex, body mass index (BMI; weight (kg) divided by the square of height (m), HOMA-IR (except for the TWB database), total cholesterol (TC; mg / dL), high-density lipoprotein (HDL; mg / dL), triglyceride levels (mg / dL), fasting plasma glucose (FPG; 20 mg / dL), and glycated hemoglobin (HbA1c). Hemoglobin, HbA1c; unit is %). DM is defined according to the American Diabetes Association (ADA) guidelines for DM.

[0032] In some embodiments, a third feature set from a third database is used to enable machine learning to make optimal predictions about IR or β-cell function. The third feature set consists of factors commonly found in the art related to IR or β-cell function.

[0033] The machine learning model building process disclosed herein

[0034] In some embodiments, step S2 describes randomly selecting 55% of the patient population in the third database as the training group for model building, 15% as the internal validation group for hyperparameter optimization, and 30% as the independent test group.

[0035] In some embodiments, 78.6% (55 / 70) of the 70% of the patient population in the third database (i.e., 78.6% of the training group (55%) and the internal validation group (15%)) are used as predictive models to train the machine learning model to avoid overfitting. This rate (78.6%) is lower than that of other predictive models (80%). The rate used for internal validation (21.4%) (15 / 70) is higher than that of other predictive models (20%), indicating that the algorithm of this invention utilizes more patients to validate model performance.

[0036] In some embodiments, the Synthetic Minority Oversampling Technique (SMOTE) preprocessing algorithm is used to achieve sample balance between the target and non-target populations in the training group. In some embodiments, step S3 describes that the third feature set used to predict IR (HOMA-IR>2.5) is an easily accessible set of variables, which includes, but is not limited to: age, sex, race, weight, body mass index (BMI), fasting blood glucose (fg / FBG / FPG), glycated hemoglobin (HbA1c), triglyceride (TG), total cholesterol (TC / chol), and / or high-density lipoprotein (HDL)-cholesterol (HDL-cholesterol / hdlc). In some embodiments, step S3 describes that the third set of features used to predict β-cell function (HOMA-β value <66%) is an easily obtainable set of variables, which includes, but is not limited to: age, sex, race, body mass index (BMI), fasting blood glucose (fg / FBG / FPG), triglyceride (TG), total cholesterol (TC / chol), high-density lipoprotein (HDL)-cholesterol (HDL-cholesterol / hdlc), glutamate transaminase (GOT), glutamate transaminase (GPT), waist circumference (WC), total bilirubin (TB), albumin (alb), systolic blood pressure (sbp), diastolic blood pressure (dbp), and estimated glomerular filtration rate. Rate (eGFR) and / or serum creatinine (CRE) levels.

[0037] In some embodiments, a Deep Neural Network (DNN) can be used to estimate the first selected method. Other conventional machine learning algorithms, such as Random Forest (RF), extreme gradient boosting (XGboost), and logistic regression, can also be used. Their accuracy will be compared with the logistic regression model. After the model training is complete, another set is used for testing. For the training set, the Receiver Operating Characteristic (ROC) curves of different algorithms will be compared. The ROC curve and the area under the curve (AUC) are used to compare the classification performance of different classifiers. A target AUC value (>0.80) indicates that the model of the present invention is sufficient to predict IR and / or β cell function. The DNN structure is set as follows: 9 input layers → 18 intermediate hidden layers → 27 intermediate hidden layers → 36 intermediate hidden layers → one-dimensional output layer.

[0038] In some embodiments, a binary outcome of insulin resistance or β-cell function is set in the output layer. To avoid overfitting during deep learning model training, dropout layers are added between hidden layers. The dropout rate is set to 0.2. For the activation function, scaled exponential linear units are used in the intermediate layers, and hard sigmoid units are used in the output layer. During hyperparameter tuning for XGBoost and RF, grid search is used to identify the optimal combination of potential parameter values.

[0039] In some embodiments, the Gini Index is used to calculate feature importance, i.e., the importance of each variable in the third feature set. During model interpretation, the Shapley Additive exPlanations (SHAP) are used to explain the working principles of different machine learning models. Finally, since IR and β-cell function can be used to predict mortality in non-diabetic individuals, the predicted mortality rates are also tested in the IR prediction algorithm and β-cell function prediction algorithm disclosed in this paper.

[0040] In some embodiments, step S4 describes that after the above model achieves good diagnostic performance, the best model can be applied to a fourth database (e.g., the TWB database) for external validation to determine its possible clinical interpretation (e.g., cardiovascular disease (CV) mortality and all-cause mortality).

[0041] Explanation of IR and β-cell function: CV mortality and all-cause mortality

[0042] The clinical interpretation of IR and β-cell function was analyzed using the model that demonstrated the highest predictive ability after the above training. Mortality information was derived by linking data with active follow-up investigations and death certificates. All-cause mortality was coded as ICD-9 0001-E999 and ICD-10 A00-Y98 according to the International Classification of Diseases (ICD) 9th or 10th editions. The CV mortality codes correspond to CV diseases (ICD-9 390-456 and ICD-10 I00-I99), including coronary heart disease (ICD-9 140-414 and ICD-10 I20-I25), stroke (ICD-9 430-438 and ICD-10 I60-I69), ischemic stroke (ICD-9 434 and ICD-10 I63) and hemorrhagic stroke (ICD-9 4330-432 and ICD-10 I60-I62).

[0043] In some embodiments, a fourth database is used to provide a fourth data set and for external validation and evaluation of clinical interpretations. In some embodiments, clinical interpretations include cardiovascular mortality and all-cause mortality.

[0044] Statistical analysis of this disclosure

[0045] NHANES is a multivariate and complex survey. To represent the sample weighted data, weighted data needs to be calculated according to the analysis guidelines (US NHANES: Analysis Guidelines, 2011–2014 and 2015–2016, available online). However, raw unweighted data from the NHANES and MJ databases were used to build models for machine learning and deep learning. This invention does not use weighted data for two reasons. First, weighted data is typically used to estimate national incidence / prevalence. This disclosure does not require estimating national prevalence; it only requires training the aforementioned models using the relationship between inter-individual IR / β-cell function and the third feature set. Second, the MJ database does not provide corresponding weights, and data from the NHANES and MJ databases need to be combined. Therefore, unweighted data from NHANES were used. For unweighted data from the NHANES, MJ, and TWB databases, continuous variables are expressed as mean ± standard deviation (SD), and categorical data are expressed as percentages. For categorical variables, the chi-square test is used; for continuous variables, the independent t-test is used to assess differences in clinical variables between IR / β cell functional status. Univariate and multivariate logistic regressions using the restricted cubic spline approach identify nonlinear relationships between selected features and IR / β cell function, which can be used to compare association patterns between machine learning and traditional statistical methods.

[0046] In some embodiments, the feature extraction module collects and processes data sets by deriving at least one of the mean, standard deviation of the mean, quantity, percentage, coefficient of variation, and slope and R-squared of the linear regression, thereby serving as variables of the feature set and used to build a machine learning model.

[0047] In some embodiments, a fourth dataset from the TWB database is used as input to the algorithm described above, and the output is the predicted probability of IR / β-cell function for each participant. A predicted probability greater than 0.5 is defined as the cutoff point for determining whether a participant belongs to the IR or β-cell deficiency group. The predicted CV mortality or all-cause mortality can be compared among the IR, non-IR, β-cell deficiency, and non-β-cell deficiency groups using a Kaplan-Meier survival curve with log-rank and proportional hazard ration model tests, and time-to-event analysis tests. All p-values ​​obtained are two-sided tests and are considered significant when p < 0.05. Deep learning algorithms and other machine learning algorithms (including XGBoost, RF, and DNN) are executed in Keras (version 2.4.0), TensorFlow (version 1.10.0), and Python (version 3.6.5). Univariate and multivariate analyses of CV mortality and all-cause mortality were also performed. This disclosure also compares the predictive power of the algorithm with the Framingham score using the Concordance Index (C index) and misclassification statistics. All statistical analyses were performed using SAS for Windows (version 9.4; SAS, Cary, North Carolina). This disclosure was approved by the Ethics Committee of Taichung Veterans General Hospital, IRB No.: CE20023A. All procedures were performed in accordance with relevant guidelines and regulations.

[0048] result

[0049] Participants selected from the NHANES database, MJ database, and TWB database

[0050] Initially, this disclosure included 71,916 participants from the NHANES database (1999–2012) and 14,359 participants from the MJ database (2008–2017). After exclusion, the final analysis included 25,737 participants (14,211 from the NHANES database and 12,526 from the MJ database). Excluded participants were: those under 18 years of age (26,373 from the NHANES database and 72 from the MJ database); those with incomplete laboratory data (28,939 from the NHANES database and 657 from the MJ database); and those with a history of diabetes mellitus (DM) (2,393 from the NHANES database and 1,104 from the MJ database). Of all participants in the third database (a combined database of the NHANES and MJ databases), 14,705 were randomly selected as the training group, and 4,018 were selected as the internal validation group. In the training group, the models were trained using XGBoost, RFs, logistic regression, and DNN, and their AUCs were compared. The test group for model evaluation comprised 8,014 participants. Participants collected from the TWB database between December 10, 2008, and November 30, 2018, were used for external validation and clinical interpretation evaluation.

[0051] Baseline characteristics of participants in the Caucasian database (NHANES database) and the Asian database (Taiwan's MJ database).

[0052] The first database (NHANES database) and the second database (MJ database) included two ethnic groups: Caucasians and Asians. These two groups had similar age (44.84 years vs. 45.06 years) and sex ratios (48.29% vs. 50.44%). Furthermore, they had similar fasting blood glucose levels (91.54 mg / dL vs. 100.07 mg / dL), glycated hemoglobin levels (5.32% vs. 5.2%), and total cholesterol levels (198.45 mg / dL vs. 199.13 mg / dL). However, patients in the NHANES database had significantly higher insulin resistance than those in the MJ database (2.84 vs. 1.92). In addition, individuals from the NHANES database had higher insulin levels (12.1 mIU / L vs. 7.67 mIU / L), BMI (27.86 kg / m² vs. 23.78 kg / m²), and triglyceride levels (126.21 mg / dL vs. 116.59 mg / dL).

[0053] Based on HOMA-IR (≤2.5 or >2.5) and / or HOMA-β (<66%), the baseline characteristics of patients with or without DM were observed.

[0054] There were no statistically significant differences among the three groups (training, internal validation, and testing). Table 1 details the relevant information of all participants in the first, second, and third databases based on the presence or absence of insulin resistance. Participants in the NHANES and MJ databases had similar baseline characteristics, but participants from the NHANES database had higher mean HOMA-IR values ​​(2.84 vs. 1.92), faster insulin values ​​(12.1 mIU / L vs. 7.67 mIU / L), BMI (27.86 kg / m² vs. 23.78 kg / m²), and triglyceride values ​​(126.21 mg / dL vs. 116.59 mg / dL), and lower FPG values ​​(91.54 mg / dL vs. 100.07 mg / dL) and HDL-cholesterol values ​​(53.88 mg / dL vs. 57.77 mg / dL). In the merged NHANES and MJ databases (i.e., the third database), the IR group (IR>2.5) was significantly (p<0.0001) older (46.06±16.58 vs 45.28±15.34), had a higher proportion of males (54.3% vs 46.5%), higher HOMA-IR (4.53±2.94 vs 1.47±0.53), higher insulin concentrations (17.84±10.88 vs 6.17±2.17), and higher BMI (30.05±6). The values ​​were 0.05 vs. 24±4.03, higher HbA1c (5.43±0.41 vs. 5.21±0.41), higher FPG (99.08±10.62 vs. 93.92±9.78), lower HDL-cholesterol (49.19±12.97 vs. 59.2±15.43), higher TC (201.35±39.97 vs. 197.16±38.17), and higher triglyceride (155.51±112.06 vs. 104.25±78.04).

[0055] Furthermore, Table 2 details the relevant information of all participants in the first, second, and third databases based on β-cell function.

[0056] Table 1: Baseline data of all participants in the first database (NHANES database), the second database (MJ database), and the third database (a combination of the NHANES and MJ databases) based on the presence or absence of insulin resistance. total HOMA-IR ≤ 2.5 HOMA-IR >2.5 p-value [NHANES] [database] N 14211 8082 6129 Age (years) 44.84 (44.29-45.4) 44.23 (43.58-44.88) 45.73 (45.12-46.34) <0.0001 Male, n (%) 6802(48.29) 3730 (44.98) 3072(53.12) <0.0001 Female, n (%) 7409(51.71) 4352 (55.02) 3057 (46.88) <0.0001 HOMA-IR 2.84 (2.77-2.91) 1.48 (1.46-1.5) 4.83 (4.7-4.96) <0.0001 Insulin concentration (mIU / L) 12.1ma ins 6.51ma in 19.46a insu <0.0001 BMI (kg / m²) 27.86 (27.72-28) 25.39 (25.26-25.52) 31.45 (31.22-31.69) <0.0001 Glycated hemoglobin level (%) 5.32 (5.31-5.34) 5.25 (May 23-27) 5.43 (5.41-5.44) <0.0001 Fasting blood glucose level (mg / dl) 91.54 (91.22-91.86) 88.27 (87.93-88.6) 96.3 (95.92-96.68) <0.0001 HDL cholesterol level (mg / dL) 53.88 (53.47-54.3) 58.03 (57.52-58.54) 47.84 (47.39-48.29) <0.0001 Cholesterol level (mg / dL) 198.45 (197.36-199.53) 196.53 (195.25-197.82) 201.23 (199.71-202.76) <0.0001 Triglyceride level (mg / dL) 126.21 (123.6-128.83) 104.22 (101.12-107.32) 158.25 (153.55-162.96) <0.0001 [MJ] [database] N 12526 9824 2702 Age (years) 45.06±11.49 45.14±11.55 44.73±11.26 <0.0001 Male, n (%) 6318 (50.44) 4595 (46.77) 1723 (63.77) <0.0001 Female, n (%) 6208 (49.56) 5229(53.23) 979 (36.23) <0.0001 HOMA-IR 1.92±1.26 1.44±0.51 3.68±1.58 <0.0001 Insulin concentration (mIU / L) 7.67ma in 5.88ma in 14.14a ins <0.0001 BMI (kg / m²) 23.78±3.76 22.81±3.09 27.3±3.89 <0.0001 Glycated hemoglobin level (%) 5.2±0.44 5.16±0.43 5.38±0.44 <0.0001 Fasting blood glucose level (mg / dl) 100.07±8.3 98.57±7.72 105.49±8.08 <0.0001 HDL cholesterol level (mg / dL) 57.77±14.52 59.93±14.66 49.9±10.8 <0.0001 Cholesterol level (mg / dL) 199.13±35.28 197.77±34.9 204.09±36.18 <0.0001 Triglyceride level (mg / dL) 116.59±88.9 104.01±73.04 162.33±120.74 <0.0001 [Merge NHANES database and MJ database] N 26737 17906 8831 Age (years) 45.54±15.76 45.28±15.34 46.06±16.58 <0.0001 Male, n (%) 13120 (49.07) 8325 (46.49) 4795(54.3) <0.0001 Female, n (%) 13617 (50.93) 9581(53.51) 4036(45.7) <0.0001 HOMA-IR 2.48±2.26 1.47±0.53 4.53±2.94 <0.0001 Insulin concentration (mIU / L) 10.02±8.51 6.172±8.5 17.84±8.518 <0.0001 BMI (kg / m²) 26±5.57 24±4.03 30.05±6.05 <0.0001 Glycated hemoglobin level (%) 5.28±0.42 5.21±0.41 5.43±0.41 <0.0001 Fasting blood glucose level (mg / dl) 95.62±10.35 93.92±9.78 99.08±10.62 <0.0001 HDL cholesterol level (mg / dL) 55.89±15.4 59.2±15.43 49.19±12.97 <0.0001 Cholesterol level (mg / dL) 198.54±38.82 197.16±38.17 201.35±39.97 <0.0001 Triglyceride level (mg / dL) 121.18±93.85 104.25±78.04 155.51±112.06 <0.0001

[0057] Table 2: Baseline data of all participants based on β-cell function in the first database (NHANES database), the second database (MJ database), and the third database (a combination of the NHANES and MJ databases). [NHANES] [database] total HOMA-β>66 HOMA-β≤66 p-value N 11792 9065 2727 HOMA-β 139.39±267.86 167.05±299.96 47.45±13.06 <.0001 Age (years) 44.6±18.99 43.07±18.65 49.68±19.23 <.0001 Waist circumference (cm) 95.58±15.25 97.89±15.47 87.9±11.52 <.0001 BMI (kg / m²) 27.83±6.09 28.87±6.22 24.39±4.04 <.0001 Systolic blood pressure (mmHg) 121.65±18.5 121.33±17.99 122.72±20.05 0.0012 Diastolic blood pressure (mmHg) 68.9±13.01 69.15±12.9 68.08±13.33 0.0002 Fasting blood glucose level (mg / dl) 96.6±10.25 96.09±10.33 98.3±9.77 <.0001 HDL cholesterol level (mg / dL) 54.39±15.81 52.52±14.92 60.62±17.05 <.0001 Cholesterol level (mg / dL) 195.73±42.13 195.5±42.05 196.5±42.42 0.28 Triglyceride level (mg / dL) 121.24±94.11 129.02±99.96 95.35±64.89 <.0001 eGFR (mL / min / 1.73m²) 101.84±24.47 103.89±24.53 95.04±22.98 <.0001 Serum creatinine level (mg / dL) 0.84±0.36 0.82±0.33 0.9±0.45 <.0001 Albumin level (g / dL) 4.26±0.37 4.24±0.38 4.3±0.33 <.0001 Total bilirubin level (mg / dL) 0.75±0.33 0.72±0.3 0.84±0.4 <.0001 AST (aspartate transaminase) value 25.68±24.92 25.56±21.28 26.1±34.34 0.3236 ALT (alanine transaminase) value 25.38±28.7 26.15±24.35 22.83±39.8 <.0001 Male, n(%) 5728 (48.58) 4157 (45.86) 1571 (57.61) <.0001 Female, n(%) 6064 (51.42) 4908 (54.14) 1156 (42.39) [MJ] [database] total HOMA-β>66 HOMA-β≤66 p-value N 13005 6427 6578 HOMA-β 75.07±44.96 103.11±49 47.68±11.5 <.0001 Age (years) 45.09±11.52 43.12±11.14 47.01±11.56 <.0001 Waist circumference (cm) 78.63±10.43 82.1±10.77 75.25±8.86 <.0001 BMI (kg / m²) 23.8±3.79 25.23±4.04 22.41±2.91 <.0001 Systolic blood pressure (mmHg) 115.82±16.93 117.75±16.88 113.93±16.77 <.0001 Diastolic blood pressure (mmHg) 73.98±11.16 75.24±11.4 72.75±10.77 <.0001 Fasting blood glucose level (mg / dl) 100.17±8.37 99.53±8.33 100.79±8.36 <.0001 HDL cholesterol level (mg / dL) 57.74±14.5 54.1±13 61.3±14.99 <.0001 Cholesterol level (mg / dL) 199.1±35.14 199.75±35.06 198.48±35.21 0.039 Triglyceride level (mg / dL) 116.85±88.57 135.42±93.8 98.71±79.05 <.0001 eGFR (mL / min / 1.73m²) 81.12±12.54 81.72±12.87 80.54±12.19 <.0001 Serum creatinine level (mg / dL) 0.96±0.22 0.97±0.25 0.95±0.19 <.0001 Albumin level (g / dL) 4.41±0.2 4.42±0.2 4.39±0.2 <.0001 Total bilirubin level (mg / dL) 0.99±0.38 0.93±0.36 1.06±0.39 <.0001 AST (aspartate transaminase) value 23.99±11.05 25.23±13 22.78±8.57 <.0001 ALT (alanine transaminase) value 28.54±22.57 33.95±28 23.24±13.57 <.0001 Male, n(%) 6585 (50.63) 3393 (52.79) 3192 (48.53) <.0001 Female, n(%) 6420 (49.37) 3034 (47.21) 3386 (51.47) [Merge NHANES database and MJ database] total HOMA-β>66 HOMA-β≤66 p-value N 24797 15492 9305 HOMA-β 105.66±190.29 140.52±233.74 47.61±11.98 <.0001 Age (years) 44.85±15.53 43.09±15.97 47.79±14.29 <.0001 Waist circumference (cm) 86.69±15.47 91.34±15.77 78.96±11.3 <.0001 BMI (kg / m²) 25.72±5.4 27.36±5.71 22.99±3.4 <.0001 Systolic blood pressure (mmHg) 118.59±17.93 119.84±17.63 116.51±18.23 <.0001 Diastolic blood pressure (mmHg) 71.56±12.34 71.67±12.66 71.38±11.77 0.0644 Fasting blood glucose level (mg / dl) 98.47±9.48 97.52±9.7 100.06±8.87 <.0001 HDL cholesterol level (mg / dL) 56.15±15.23 53.18±14.17 61.1±15.63 <.0001 Cholesterol level (mg / dL) 197.5±38.66 197.26±39.36 197.9±37.48 0.2066 Triglyceride level (mg / dL) 118.94±91.27 131.67±97.5 97.73±75.19 <.0001 eGFR (mL / min / 1.73m²) 90.98±21.78 94.69±23.24 84.79±17.42 <.0001 Serum creatinine level (mg / dL) 0.9±0.3 0.88±0.3 0.93±0.29 <.0001 Albumin level (g / dL) 4.33±0.3 4.32±0.33 4.36±0.25 <.0001 Total bilirubin level (mg / dL) 0.87±0.38 0.8±0.35 0.99±0.4 <.0001 AST (aspartate transaminase) value 24.8±18.97 25.42±18.31 23.75±19.99 <.0001 ALT (alanine transaminase) value 27.03±25.72 29.38±26.21 23.12±24.38 <.0001 Male, n(%) 12313 (49.66) 7550 (48.73) 4763 (51.19) 0.0002 Female, n(%) 12484 (50.34) 7942 (51.27) 4542 (48.81)

[0058] Assessment of baseline data and clinical interpretation from the TWB database for external validation

[0059] In some embodiments, a fourth dataset from a fourth database (e.g., the TWB database) is fed into the XGBoost algorithm to be divided into predicted non-IR groups (n=86,283) and IR groups (n=8,774). Participants predicted to be in the IR group were significantly (p<0.0001) younger (47.51±10.81 years vs. 49.42±10.86 years), had a higher proportion of males (48.80% vs. 33.39%), and exhibited higher BMI (30.27±3.544 vs. 30.27±3.54), higher HbA1c levels (5.83±0.32% vs. 5.57±0.33%), higher FPG levels (99.35±9.44 mg / dL vs. 90.98±6.99 mg / dL), lower HDL-cholesterol levels (43.39±8.59 mg / dL vs. 56.32±13.25 mg / dL), and higher TC levels (200.22±35.75). The levels were 195.44 ± 34.98 mg / dL and higher triglyceride levels (201.95 ± 150.76 mg / dL vs. 100.54 ± 67.6 mg / dL).

[0060] Predictions from machine learning models

[0061] Table 3 illustrates the results of the internal validation and test groups using XGboost, RF, logistic regression, and DNN algorithms. Figure 2A shows the ROC curves for the training group. Figure 2B shows that the AUC of all ROCs in the internal validation group is >0.8 (the highest being XGboost, 0.87). According to Table 3, in the internal validation group, XGboost also has the highest accuracy (0.80), sensitivity (0.63), specificity (0.89), positive predictive value (PPV; 0.74), and negative predictive value (NPV; 0.8). Figure 2C shows that the AUC of all ROCs in the test group is also >0.80 (the highest being XGboost, 0.88). Furthermore, according to Table 3, XGboost also has the highest accuracy (0.81), sensitivity (0.64), and NPV (0.83) in the test group, while RF has the highest specificity (0.90) and PPV (0.75).

[0062] Table 3: Area Under Curve (AUC), Accuracy, Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV) for Different Models of Receiver Operating Characteristic (ROC) Curves XGboost Random Forest (RF) Logical Regression DNN Internal Validation Group ROC's AUC 0.87 0.86 0.82 0.86 Accuracy 0.80 0.80 0.78 0.80 Sensitivity 0.63 0.60 0.56 0.61 Specificity 0.89 0.89 0.89 0.89 PPV 0.74 0.74 0.71 0.74 NPV 0.83 0.82 0.80 0.82 test group ROC's AUC 0.88 0.87 0.83 0.87 Accuracy 0.81 0.81 0.78 0.81 Sensitivity 0.64 0.62 0.56 0.64 Specificity 0.89 0.90 0.88 0.89 PPV 0.74 0.75 0.70 0.75 NPV 0.83 0.83 0.80 0.83 Comparison of AUC between models XGboost for RF XGboost for DNN XGboost's logical regression Internal Validation Group 0.0007 0.0264 <0.0001 test group <0.0001 0.0022 <0.0001

[0063] Tables 4 and 5 illustrate the AUC of the ROC curves for predicting IR and β cell function using different feature sets through machine learning models, thus demonstrating the predictive power.

[0064] In some embodiments, the machine learning model for predicting IR and β-cell function can be implemented using classification algorithms and additional algorithms, but the invention is not limited thereto. Classification algorithms in some embodiments may be, but are not limited to: XGBoost, RF, logistic regression, and / or DNN. Additional algorithms in some embodiments may be, but are not limited to: linear regression, multinomial regression, support vector regression, decision tree regression, random forest regression, ridge regression, Lasso regression, elastic network regression, logistic regression, decision tree regression, gradient boosting, and / or K-nearest neighbor regression.

[0065] In some embodiments, the true values ​​of an individual's HOMA-IR and / or HOMA-β can be determined by a classification algorithm, the predicted probabilities of the individual's IR / β cell function, and additional algorithms. The classification algorithm can be, but is not limited to: XGBoost, RF, logistic regression, and / or DNN. Additional algorithms can be, but are not limited to: linear regression, multinomial regression, support vector regression, decision tree regression, random forest regression, ridge regression, Lasso regression, elastic network regression, logistic regression, decision tree regression, gradient boosting, and / or K-nearest neighbor regression.

[0066] In some embodiments, the model building and optimization module builds a machine learning model based on a feature set, thereby predicting insulin resistance and β-cell function in non-diabetic patients by classifying them into insulin resistance group, non-insulin resistance group, β-cell deficiency group and non-β-cell deficiency group, and generating corresponding predicted values ​​and classification performance values.

[0067] In some embodiments, the classification performance value is the area under the receiver operating characteristic curve of the machine learning model.

[0068] Table 4: Area Under Curve (AUC) of Receiver Operating Characteristic (ROC) curves predicted by machine learning models for IR. "N / A" indicates that this data point was not displayed. Status number Feature set Race AUC of the ROC curve 1-1 Age, gender, race and BMI Asian groups 0.8293 1-2 Age, gender, race and BMI Caucasian groups 0.793 2-1 Age, gender, race and BMI Asian groups 0.865 (if FPG value is selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level 0.8364 (if HbA1c value is selected) 2-2 Age, gender, race and BMI Caucasian groups 0.839 (if FPG value is selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level 0.8041 (if HbA1c value is selected) 3-1 Age, sex, race, BMI, and triglyceride levels Asian groups 0.8463 3-2 Age, sex, race, BMI, and triglyceride levels Caucasian groups 0.8175 4-1 Age, sex, race, BMI, and triglyceride levels Asian groups 0.8471 (if TC value is selected) Total cholesterol (TC) or HDL cholesterol level 0.8518 (if HDL-cholesterol value is selected) 4-2 Age, sex, race, BMI, and triglyceride levels Caucasian groups 0.8202 (if TC value is selected) Total cholesterol (TC) or HDL cholesterol level 0.8231 (if HDL-cholesterol value is selected) 5-1 Age, gender, race and BMI Asian groups N / A Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). N / A N / A 5-2 Age, gender, race and BMI Caucasian groups N / A Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). N / A N / A 6-1 Age, gender, race and BMI Asian groups 0.8666 (if FPG and TC values ​​are selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level. 0.8736 (if FPG and HDL-cholesterol values ​​are selected) Total cholesterol (TC) or HDL cholesterol level 0.8369 (if HbA1c and TC values ​​are selected) 0.8473 (if HbA1c and HDL-cholesterol values ​​are selected) 6-2 Age, gender, race and BMI Caucasian groups 0.8409 (if FPG and TC values ​​are selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level 0.8509 (if FPG and HDL-cholesterol values ​​are selected) Total cholesterol (TC) or HDL cholesterol level 0.8041 (if HbA1c and TC values ​​are selected) 0.8213 (if HbA1c and HDL-cholesterol values ​​are selected) 7-1 Age, sex, race, BMI, and fasting plasma glucose (FPG) Asian groups N / A (if TC, GOT, and GPT values ​​are selected) N / A (if HDL-cholesterol, GOT, and GPT levels are selected) Total cholesterol (TC) or HDL cholesterol level N / A (if TC and TB values ​​are selected) N / A (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). N / A (if TC and alb values ​​are selected) N / A (if HDL-cholesterol and alb levels are selected) 7-2 Age, sex, race, BMI, and fasting plasma glucose (FPG) Caucasian groups N / A (if TC, GOT, and GPT values ​​are selected) N / A (if HDL-cholesterol, GOT, and GPT levels are selected) Total cholesterol (TC) or HDL cholesterol level N / A (if TC and TB values ​​are selected) N / A (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). N / A (if TC and alb values ​​are selected) N / A (if HDL-cholesterol and alb levels are selected) 8-1 Age, sex, race, BMI, triglyceride level, and fasting plasma glucose (FPG) Asian groups N / A (if TC, GOT, and GPT values ​​are selected) N / A (if HDL-cholesterol, GOT, and GPT levels are selected) Total cholesterol level (TC HDL) N / A (if TC and TB values ​​are selected) N / A (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). N / A (if TC and alb values ​​are selected) N / A (if HDL-cholesterol and alb levels are selected) 8-2 Age, sex, race, BMI, triglycerides, and fasting plasma glucose (FPG) levels. Caucasian groups N / A (if TC, GOT, and GPT values ​​are selected) N / A (if HDL-cholesterol, GOT, and GPT levels are selected) Total cholesterol (TC) and HDL cholesterol levels N / A (if TC and TB values ​​are selected) N / A (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). N / A (if TC and alb values ​​are selected) N / A (if HDL-cholesterol and alb levels are selected)

[0069] Table 5: Area Under Curve (AUC) of Receiver Operating Characteristic (ROC) curves predicting β-cell function using machine learning models. "N / A" indicates that this data point was not displayed. Status number Feature set Race AUC of the ROC curve 1-1 Age, gender, race and BMI Asian groups 0.7578 1-2 Age, gender, race and BMI Caucasian groups 0.7812 2-1 Age, gender, race and BMI Asian groups 0.7735 (if FPG value is selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level N / A (if HbA1c value is selected) 2-2 Age, gender, race and BMI Caucasian groups 0.7928 (if FPG value is selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level N / A (if HbA1c value is selected) 3-1 Age, sex, race, BMI, and triglyceride levels Asian groups 0.7812 3-2 Age, sex, race, BMI, and triglyceride levels Caucasian groups 0.8034 4-1 Age, sex, race, BMI, and triglyceride levels Asian groups 0.7823 (if TC value is selected) Total cholesterol (TC) or HDL cholesterol level 0.7833 (if HDL-cholesterol value is selected) 4-2 Age, sex, race, BMI, and triglyceride levels Caucasian groups 0.8056 (if TC value is selected) Total cholesterol (TC) or HDL cholesterol level 0.8083 (if HDL-cholesterol value is selected) 5-1 Age, gender, race and BMI Asian groups 0.7814 (if GOT and GPT values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). 0.7587 (if the alb value is selected) 0.763 (if TB value is selected) 5-2 Age, gender, race and BMI Caucasian groups 0.7897 (if GOT and GPT values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). 0.7823 (if the alb value is selected) 0.7921 (if TB value is selected) 6-1 Age, gender, race and BMI Asian groups 0.7736 (if FPG and TC values ​​are selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level 0.785 (if FPG and HDL-cholesterol values ​​are selected) Total cholesterol (TC) or HDL cholesterol level N / A (if HbA1c and TC values ​​are selected) N / A (if HbA1c and HDL-cholesterol values ​​are selected) 6-2 Age, gender, race and BMI Caucasian groups 0.7948 (if FPG and TC values ​​are selected) Fasting plasma glucose (FPG) or glycated hemoglobin (HbA1c) level 0.8129 (if FPG and HDL-cholesterol values ​​are selected) Total cholesterol (TC) or HDL cholesterol level N / A (if HbA1c and TC values ​​are selected) N / A (if HbA1c and HDL-cholesterol values ​​are selected) 7-1 Age, sex, race, BMI, and fasting plasma glucose (FPG) Asian groups 0.8024 (if TC, GOT, and GPT values ​​are selected) 0.8089 (if HDL-cholesterol, GOT, and GPT values ​​are selected) Total cholesterol (TC) or HDL cholesterol level 0.7839 (if TC and TB values ​​are selected) 0.7902 (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). 0.7787 (if TC and alb values ​​are selected) 0.7903 (if HDL-cholesterol and ALB values ​​are selected) 7-2 Age, sex, race, BMI, and fasting plasma glucose (FPG) Caucasian groups 0.8069 (if TC, GOT, and GPT values ​​are selected) 0.8189 (if HDL-cholesterol, GOT, and GPT values ​​are selected) Total cholesterol (TC) or HDL cholesterol level 0.8034 (if TC and TB values ​​are selected) 0.8186 (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). 0.795 (if TC and alb values ​​are selected) 0.8164 (if HDL-cholesterol and ALB values ​​are selected) 8-1 Age, sex, race, BMI, triglyceride level, and fasting plasma glucose (FPG) Asian groups 0.8205 (if TC, GOT, and GPT values ​​are selected) 0.8198 (if HDL-cholesterol, GOT, and GPT values ​​are selected) Total cholesterol level (TC or HDL-cholesterol level) 0.8076 (if TC and TB values ​​are selected) 0.8075 (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). 0.8037 (if TC and alb values ​​are selected) 0.8044 (if HDL-cholesterol and ALB values ​​are selected) 8-2 Age, sex, race, BMI, triglyceride level, and fasting plasma glucose (FPG) Caucasian groups 0.8228 (if TC, GOT, and GPT values ​​are selected) 0.8279 (if HDL-cholesterol, GOT, and GPT values ​​are selected) Total cholesterol (TC) or HDL cholesterol level 0.8225 (if TC and TB values ​​are selected) 0.8285 (if HDL-cholesterol and TB values ​​are selected) Liver function tests (GOT and GPT), total bilirubin (TB), and albumin (alb). 0.8176 (if TC and alb values ​​are selected) 0.8254 (if HDL-cholesterol and ALB values ​​are selected)

[0070] Importance of parameter correlation in XGBoost and Random Forest (RF) algorithms

[0071] Figures 3A and 3B illustrate the correlation importance of each feature in the feature set studied using XGboost and RF prediction models, respectively. Among these features, the importance of all features is similar for both the XGboost and RF algorithms. BMI has the greatest impact on IR (0.43 for XGboost and 0.47 for RF), and its impact is significantly higher than other features. Besides BMI, the importance of variables affecting IR decreases in the following order: FPG value (0.12 for XGboost and 0.13 for RF), HDL value (0.11 for XGboost and 0.10 for RF), and triglyceride value (0.097 for XGboost and 0.13 for RF).

[0072] Referring to Figures 4A and 4B, for both the XGboost and RF algorithms, the feature's contribution to the algorithm (i.e., the SHAP value) exhibits a positive or negative correlation with IR (Figure 4A is for XGboost, and Figure 4B is for RF). In both algorithms, all features are clearly distinguished as positively or negatively correlated with IR, and similar patterns are observed. Some features, such as BMI, FPG value, triglyceride level, gender, and HbA1c value, show a positive impact on IR. Other features, such as race, age, HDL-C cholesterol level, and total cholesterol level, have a negative impact on IR.

[0073] Specifically, SHAP (SHapley Additive exPlanations) is a method for attributing the contribution of each feature to the predictions of a machine learning model. Calculating the SHAP value involves averaging the contribution of each feature across all features. Therefore, in some embodiments, the SHAP value of each feature can be used to estimate the importance of each feature in the machine learning model. As shown in Figures 4A and 4B, positive SHAP values ​​are associated with increased IR predictions, while negative SHAP values ​​are associated with decreased IR predictions. Features with higher values ​​are associated with red, and features with lower values ​​are associated with blue. Figure 4A has 9 bands, and Figure 4B has 9 bands; the band for each feature represents the distribution of all individuals in this invention. Taking the first band in Figure 4A as an example, the higher the BMI, the higher the SHAP value, indicating that the higher the BMI, the higher the probability of the machine learning model predicting IR. Further, taking the sixth band in Figure 4A as an example, the lower the HDL-cholesterol value, the higher the SHAP value, indicating that the lower the HDL-cholesterol value, the higher the probability of the machine learning model predicting IR. In addition, the characteristic value for Asians is 1, while the characteristic value for non-Asians is 0.

[0074] The following features showed an increasing trend (SHAP values ​​increased with increasing features): BMI, FPG, triglyceride, and HbA1c. The following features showed a decreasing trend (SHAP values ​​decreased with increasing features): HDL, age, and total cholesterol. All patterns of SHAP values ​​derived from machine learning resembled correlation patterns.

[0075] The XGboost algorithm used to predict cardiovascular mortality and all-cause mortality was externally validated by a fourth database (i.e., the Taiwan Human Biobank).

[0076] Since the XGboost model has the best predictive value for IR, this disclosure uses the XGboost algorithm to perform external validation on the fourth dataset of the TWB database. At this time, participants from the TWB database will be predicted and divided into the HOMA-IR(+)(IR) group and the HOMA-IR(-)(non-IR) group.

[0077] Referring to Figures 5A, 5B, 5C, and 5D, the Kaplan-Mayer survival curves (CV mortality rates in Figures 5A and 5B; and all-cause mortality rates in Figures 5C and 5D) show significant differences between the predicted IR and non-IR groups. Therefore, the XGboost model has a clear clinical interpretation for predicting IR, particularly for predicting more severe CV and all-cause mortality rates. Potential relevant factors from the multivariate analysis are shown in Table 6.

[0078] Table 6: Univariate and multivariate analyses of factors associated with CV mortality and all-cause mortality from the TWB database. Predicting CV mortality rate using the Taiwan Human Biobank Univariate analysis Multivariate analysis Risk ratio (95% confidence interval) p-value Risk ratio (95% confidence interval) p-value age 1.072 (1.057-1.088) <0.0001 1.069 (1.054-1.085) <0.0001 Male-to-female ratio 3.657 (2.771-4.825) <0.0001 3.075 (2.273-4.161) <0.0001 Body Mass Index (kg / m²) 1.098 (1.065-1.132) <0.0001 1.077 (1.035-1.121) 0.0002 Glycated hemoglobin level (%) 3.097 (2.083-4.604) 0.996-1.001) 0.1666 <0.0001 1.573 (1.009-2.453) 0.0455 Fasting blood glucose level (mg / dl) 1.043 (1.027-1.059) <0.0001 0.995 (0.977 - 1.014) 0.6078 HDL cholesterol level (mg / dL) 0.969 (0.959-0.98) <0.0001 0.99 (0.976 - 1.004) 0.1673 Cholesterol level (mg / dL) 0.999 (0.995-1.003) 0.5459 0.999 (0.994 - 1.003) 0.5037 Triglyceride level (mg / dL) 1.001 (1-1.002) 0.0701 0.998 (0.996 - 1.001) 0.1666 Predicting all-cause mortality using the Taiwan Human Biodatabase age 1.072 (1.065-1.079) <0.0001 1.074 (1.067-1.081) <0.0001 Male-to-female ratio 2.196 (1.95-2.474) <0.0001 1.914 (1.679-2.182) <0.0001 Body Mass Index (kg / m²) 1.047 (1.031-1.063) <0.0001 1.029 (1.01-1.049) 0.0032 Glycated hemoglobin level (%) 1.708 (1.43-2.041) <0.0001 0.937 (0.77-1.14) 0.5147 Fasting blood glucose level (mg / dl) 1.029 (1.022-1.037) <0.0001 0.996 (0.988 - 1.005) 0.3826 HDL cholesterol level (mg / dL) 0.981 (0.976-0.986) <0.0001 0.996 (0.99 - 1.002) 0.196 Cholesterol level (mg / dL) 0.998 (0.996-1) 0.0142 0.996 (0.994-0.998) <0.0001 Triglyceride level (mg / dL) 1.001 (1-1.001) <0.0001 1(1-1.001) 0.2482

[0079] In Table 7, for univariate and multivariate analyses, the hazard ratios (HRs) for all-cause mortality and cardiovascular mortality were significantly higher when HOMA-IR>2.5 and HOMA-IR≤2.5.

[0080] Table 7: Time-based event analysis using the proportional HR model Risk ratio (95% confidence interval) p-value Univariate analysis All-cause mortality rate HOMA-IR >2.5 vs. HOMA-IR ≤2.5 1.443 (1.168-1.783) 0.0007 CV mortality rate HOMA-IR >2.5 vs. HOMA-IR ≤2.5 2.394 (1.620-3.538) <0.0001 Multivariate analysis All-cause mortality rate age 1.069 (1.061-1.077) <0.0001 female ratio 0.454 (0.395-0.522) <0.0001 HOMA-IR >2.5 vs. HOMA-IR ≤2.5 1.474 (1.192-1.822) 0.0003 CV mortality rate age 1.077 (1.059-1.095) <0.0001 female ratio 0.302 (0.219-0.415) <0.0001 HOMA-IR >2.5 vs. HOMA-IR ≤2.5 2.4 (1.620-3.556) <0.0001

[0081] Compared to the Fleminghan score for CV mortality (as shown in Table 8), XGboost had a lower predictive power (in terms of consistency 0.547 vs. 0.643 and 0.557). However, in Table 9, for condition 1 (cutoff value: 10% risk of 10-year CV estimated using Fleminghan score), in the HOMAIR > 2.5 (XGboost) + & ≤ 10% (Fleminghan score) group, there was still an increased risk of mortality even if the Fleminghan score was ≤ 10% (all-cause mortality: 1.556 (95% confidence interval, CI: 1.186–20.41), p = 0.0014) (CV mortality: 3.374 (95% CI: 2.051–5.552, p < 0.0001)). Similarly, in condition 2 (cutoff value: using Flemingham score to estimate the 20% risk of 10-year CV), in the HOMAIR>2.5(XGboost)+&≤20%(Flemingham score) group, there was still an increased risk of mortality even if the Flemingham score was ≤20% (all-cause mortality: 1.433 (95% CI: 1.141–1.8, p = 0.002)) (CV mortality: 2.655 (95% CI: 1.752–4.023, p < 0.0001)).

[0082] Table 8: Concordance Index (C-index) for comparing the prediction algorithm (XGBoost) and the Fleminghan score in predicting CV mortality rates. Predictive power of the two methods Predictive power of CV mortality (C index) XGboost's predicted IR+ Consistency = 0.547 (Standard error (se) = 0.024) Moderate risk Flemingham score (>10% / 10 years) Consistency = 0.643 (se = 0.021) Moderate risk Flemingham score (>20% / 10 years) Consistency = 0.557 (se = 0.015)

[0083] Table 9: Misclassification Statistics of Flemingham Score for XGBoost and CV Mortality HR (95% CI) p value Scenario 1. Cutoff value: Using the Flemingham score to estimate the 10-year CV 10% risk. All-cause mortality rate age 1.065 (1.057-1.074) <0.0001 female ratio 0.491 (0.422-0.571) <0.0001 HOMAIR ≤ 2.5 (XGboost) and ≤ 10% (Flemingham score) 1 (Reference value) HOMAIR > 2.5 (XGboost) and ≤ 10% (Flemingham score) 1.556 (1.186 - 20.41) 0.0014 HOMAIR ≤ 2.5 (XGboost) and > 10% (Flemingham score) 1.3 (1.074-1.573) 0.0071 HOMAIR >2.5 (XGboost) and >10% (Flemingham score) 1.609 (1.609-1.152) 0.0052 CV mortality rate age 1.604 (1.045-1.084) <0.0001 female ratio 0.383 (0.27-0.545) <0.0001 HOMAIR ≤ 2.5 (XGboost) and ≤ 10% (Flemingham score) 1 (Reference value) HOMAIR > 2.5 (XGboost) and ≤ 10% (Flemingham score) 3.374 (2.051-5.552) <0.0001 HOMAIR ≤ 2.5 (XGboost) and > 10% (Flemingham score) 2.223 (1.48-3.339) 0.0001 HOMAIR >2.5 (XGboost) and >10% (Flemingham score) 2.70 (1.431-5.286) 0.0024 Scenario 2. Cutoff value: Use the Flemingham score to estimate the 20% risk of the 10-year CV. All-cause mortality rate age 1.067 (1.059-1.075) <0.0001 female ratio 0.471 (0.408-0.544) <0.0001 HOMAIR ≤ 2.5 (XGboost) and ≤ 20% (Flemingham score) 1 (Reference value) HOMAIR > 2.5 (XGboost) and ≤ 20% (Flemingham score) 1.433 (1.141-1.8) 0.002 HOMAIR ≤ 2.5 (XGboost) and > 20% (Flemingham score) 1.408 (1.025-1.933) 0.0346 HOMAIR >2.5 (XGboost) and >20% (Flemingham score) 2.102 (1.229-3.595) 0.0067 CV mortality rate age 1.071 (1.053-1.09) <0.0001 female ratio 0.331 (0.237-0.462) <0.0001 HOMAIR ≤ 2.5 (XGboost) and ≤ 20% (Flemingham score) 1 (Reference value) HOMAIR > 2.5 (XGboost) and ≤ 20% (Flemingham score) 2.655 (1.752-4.023) <0.0001 HOMAIR ≤ 2.5 (XGboost) and > 20% (Flemingham score) 2.45 (1.418-4.232) 0.0013 HOMAIR >2.5 (XGboost) and >20% (Flemingham score) 2.095 (0.657-6.682) 0.2115

[0084] Referring to Figures 6A and 6B, the diagram illustrates personalized IR risk (Figure 6A), personalized β-cell defect risk (Figure 6B), and the results of the feature set description. Specifically, WC is an abbreviation for waist circumference; TG is an abbreviation for triglyceride level; GOT is an abbreviation for aspartate transaminase level; TB is an abbreviation for total bilirubin level; GPT is an abbreviation for alanine transaminase level; ALB is an abbreviation for albumin level; CHOOL is an abbreviation for total cholesterol level; DPB is an abbreviation for diastolic blood pressure; CRE is an abbreviation for creatinine level; HDLC is an abbreviation for high-density lipoprotein cholesterol level; FG is an abbreviation for fasting blood glucose level; SBP is an abbreviation for systolic blood pressure; MDRD is an abbreviation for the diet adjustment formula for kidney disease; and EGFR is an abbreviation for estimated glomerular filtration rate. Patients and healthcare professionals can adjust their medical or personal lifestyle based on the above results.

[0085] In some embodiments, the model building and optimization module builds a machine learning model based on a feature set of a third dataset to predict β-cell function in non-diabetic patients. The feature set includes age, sex, race, body mass index (BMI), fasting blood glucose (fg / FBG / FPG), triglyceride (TG), total cholesterol (TC / chol), high-density lipoprotein (HDL)-cholesterol (HDL-cholesterol / hdlc), aspartate transaminase (GOT), alanine transaminase (GPT), waist circumference (wc), total bilirubin (TB), albumin (alb), systolic blood pressure (sbp), diastolic blood pressure (dbp), estimated glomerular filtration rate (eGFR), and / or creatinine (cre).

[0086] In some embodiments, machine learning uses a third dataset from a third database (i.e., a combination of the NHANES and MJ databases) to find a good model with high diagnostic performance (AUC of 0.88) for predicting IR and β-cell function in non-diabetic populations, based on a feature set that is readily available from patients.

[0087] In some embodiments, the IR predicted by this cross-ethnic algorithm is associated with clinical interpretation validated using the TWB database, specifically: severe CV mortality and all-cause mortality. Furthermore, although the predictive power for CV mortality risk is lower (lower C-index) compared to the Fleminghan score, patients with HOMA-IR > 2.5 (below the lower CV risk prediction estimated by the Fleminghan score) still have a higher risk of CV mortality and higher all-cause mortality. The Fleminghan score cannot fully predict CV mortality and all-cause mortality, especially in populations without DM. This underscores the importance of this disclosure.

[0088] Numerous machine learning studies related to diabetes mellitus (DM) have been published. However, based on the results of the UKPDS, ADVANCE, VADT, and ACCORD studies, early glucose control should be implemented to achieve good vascular outcomes. According to the DM spectrum, identifying insulin resistance (IR) and β-cell function is crucial. However, calculating HOMA-IR and HOMA-β requires patient insulin levels, which are not regularly collected in clinical practice. Therefore, establishing predictive models for IR and β-cell function (without insulin level recordings) is essential.

[0089] The predictive machine learning model of this invention has two outstanding advantages: it has better predictive power (highest AUC) and is more general (it only uses easily available features and can be applied across ethnicities).

[0090] The significant impact of fasting insulin concentration on HOMA-IR underscores the importance of this invention. First, fasting insulin values ​​were not used in calculating HOMR-IR in this disclosure, despite its greatest influence on IR. Second, significant differences in HOMA-IR values ​​were observed among different ethnic groups (e.g., Asians and Caucasians) (NHANES group (Caucasians): 2.84 (2.77–2.91), MJ database (Asians): 1.92 ± 1.26). This difference in IR is primarily attributed to fasting insulin values ​​(12.1 ± 10.35 mIU / L in the NHANES database, 7.67 ± 4.76 in the MJ database) rather than FPG values ​​(91.54 mg / dL in the NHANES database, 100.07 ± 8.3 in the MJ database). This invention aims to train a machine learning model to predict IR for different ethnic groups without requiring fasting insulin values. The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made in accordance with the scope of the claims of the present invention shall be covered by the present invention.

[0091] S1~S4: Steps

Claims

1. A system for predicting an individual's insulin resistance or β-cell function, comprising: A database is provided to offer a dataset; a feature extraction module is used to collect and process features from the dataset to form a feature set for the individual, wherein the feature set includes the individual's age, sex, ethnicity, and body mass index, and includes the individual's total cholesterol and / or high-density lipoprotein cholesterol levels, and the individual is a non-diabetic individual; and a model building and optimization module is used to build a machine learning model based on the feature set, thereby predicting the individual's insulin resistance and / or β-cell function by classifying the non-diabetic individual into an insulin resistance group and a non-insulin resistance group or a β-cell deficiency group and a non-β-cell deficiency group.

2. The system as described in claim 1, wherein, This feature set also includes the individual's fasting blood glucose level and / or glycated hemoglobin level.

3. The system as described in claim 1, wherein, This feature set also includes the individual's triglyceride level.

4. The system as described in claim 1, wherein, The feature set also includes the individual's fasting blood glucose level and at least one of the groups consisting of the individual's aspartate transaminase level, alanine transaminase level, total bilirubin level, and albumin level.

5. The system as described in claim 1, wherein, This characteristic value also includes the individual's aspartate transaminase level, alanine transaminase level, total bilirubin level, and / or albumin level.

6. The system as described in claim 5, wherein, This feature set also includes the individual's fasting blood glucose level.

7. The system as described in claim 1, wherein, The machine learning model is trained by labeling features associated with a decline in insulin resistance and / or β-cell function in the individual.

8. The system as described in claim 1, wherein, The database includes: a first database containing data from the individual in a first group; and a second database containing data from the individual in a second group, wherein the race of the first group is different from the race of the second group.

9. A method for predicting an individual's insulin resistance or β-cell function, comprising: The database provides a dataset; the feature extraction module collects and processes features from the dataset to form a feature set for the individual, wherein the feature set includes the individual's age, sex, ethnicity, and body mass index, and includes the individual's total cholesterol and / or high-density lipoprotein cholesterol levels, and the individual is a non-diabetic individual; and the model building and optimization module builds a machine learning model based on the feature set, thereby predicting the individual's insulin resistance and / or β-cell function by classifying the non-diabetic individual into an insulin resistance group and a non-insulin resistance group or a β-cell deficiency group and a non-β-cell deficiency group.

10. The method as described in claim 9, wherein, This feature set also includes the individual's fasting blood glucose level and / or glycated hemoglobin level.

11. The method as described in claim 9, wherein, This feature set also includes the individual's triglyceride level.

12. The method as described in claim 9, wherein, The feature set also includes the individual's fasting blood glucose level and at least one of the groups consisting of the individual's aspartate transaminase level, alanine transaminase level, total bilirubin level, and albumin level.

13. The method as described in claim 9, wherein, This characteristic value also includes the individual's aspartate transaminase level, alanine transaminase level, total bilirubin level, and / or albumin level.

14. The method as described in claim 13, wherein, This feature set also includes the individual's fasting blood glucose level.

15. The method as described in claim 9, wherein, The model building and optimization module builds a machine learning model based on the feature set and uses the machine learning model to predict the individual's insulin resistance and / or β-cell function, including: the machine learning model generates corresponding predicted values ​​for each of the insulin resistance group, the non-insulin resistance group, the β-cell defect group, and the non-β-cell defect group, and generates the classification performance value of the machine learning model, wherein the classification performance value is the area under the receiver operating characteristic curve of the prediction performed by the machine learning model.

16. The method as described in claim 9, wherein, The machine learning model is trained by labeling features associated with a decline in insulin resistance and / or β-cell function in the individual.

17. The method as described in claim 9, wherein, The database includes: a first database containing data from the individual in a first group; and a second database containing data from the individual in a second group, wherein the race of the first group is different from the race of the second group.

18. A computer-readable medium having computer-executable instructions stored thereon, which, when executed, enable the implementation of the method as described in any one of claims 9 to 17.

Images