Data-driven model for predicting progression risk of future diabetes related diseases in early stage of diabetes and construction method thereof

Abstract

The invention discloses a data-driven early-diabetic future diabetes-related disease progress risk prediction model and a construction method thereof, and the method comprises the steps: collecting clinical index data, including age, gender, BMI, WHR, HOMA-IR, HDL-C, TG, SBP, DBP, SCR and ALT, of early-diabetic patients in a training and verification queue; using an unsupervised soft clustering method combining dimension reduction based on UMAP, graph clustering and a Gaussian mixture model to identify the phenotypic heterogeneity of the prediabetes mellitus; obtaining the probability of the individual phenotype characteristics, evaluating the association between the probability and the development risk of the future diabetes related diseases in the early stage of diabetes, and constructing a development risk prediction model of the future diabetes related diseases in the early stage of urine diseases; and performing model optimization and robustness verification in the verification queue. According to the method, the heterogeneity of the prediabetes mellitus can be effectively identified, and accurate risk stratification and personalized prevention are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical data analysis and disease prediction, and particularly relates to a data-driven pre-diabetes future diabetes-related disease progression risk prediction model and a construction method thereof. BACKGROUND

[0002] Pre-diabetes is a highly prevalent metabolic abnormality state and an important risk stage of global diabetes and related complications.

[0003] Pre-diabetes has significant heterogeneity and complexity, which brings great challenges to identifying individuals who are prone to progress to T2D and its complications. Although early intervention can delay or prevent the occurrence of diabetes, the current diagnosis standard based on blood glucose is insufficient to capture its pathophysiological heterogeneity or predict disease progression and complications, and some individuals even maintain a stable metabolic state for a long time without progressing to diabetes. This clinical heterogeneity further suggests that pre-diabetes is not a single disease state, but contains multiple biologically different subtypes, and the underlying mechanisms and long-term risks have not been fully elucidated.

[0004] Traditional data-driven clustering methods are limited in application due to challenges in model selection, analysis framework and parameter estimation. These limitations are particularly evident in biological systems, as variables often exhibit nonlinear dependence relationships. Recent studies have proposed several new methods for classifying pre-diabetes populations, for example, Wagner et al. used K-means clustering method to identify six different clusters based on clinical and genetic variables, revealing the pathophysiological heterogeneity of pre-diabetes. On this basis, further studies have divided Chinese pre-diabetes individuals into six subgroups with different disease risk characteristics, emphasizing the significant heterogeneity within the pre-diabetes state. These findings suggest that multi-dimensional phenotype stratification can enhance disease risk assessment.

[0005] However, traditional clustering methods have limited ability to capture intra-cluster differences, nonlinear cluster boundaries, and high-dimensional complexity in biological data. Therefore, it is crucial to develop and validate more robust, reproducible and clinically meaningful subtype partitioning strategies, especially in large prospective cohorts, which can help better elucidate individual risk characteristics and promote personalized treatment and prevention of pre-diabetes. SUMMARY

[0006] The object of the present application is to provide a data-driven pre-diabetes future diabetes-related disease progression risk prediction model to achieve the following technical objectives: 1) to identify the individual heterogeneity of pre-diabetes patients based on commonly used clinical variables; 2) how the assignment probability of specific phenotypes predicts the disease progression of pre-diabetes individuals to diabetes and complications. The present application realizes more accurate risk stratification and personalized prevention by accurately evaluating the risk of pre-diabetes future diabetes-related disease progression.

[0007] Another object of the present application is to provide a construction method of the data-driven pre-diabetes future diabetes-related disease progression risk prediction model, which uses a soft clustering method combining UMAP dimension reduction, graph clustering and GMM, and divides the pre-diabetes population into six different subtypes based on nine continuous metabolic clinically available phenotypes, determines the probability of individual assignment to each subtype, and evaluates the risk of pre-diabetes future diabetes-related disease progression through it.

[0008] To achieve the above object, the present application adopts the following technical solutions:

[0009] In a first aspect of the present application, a construction method of a data-driven pre-diabetes future diabetes-related disease progression risk prediction model is provided, comprising the following steps:

[0010] S1: Collecting clinical index data of pre-diabetes patients in the training cohort and the validation cohort, including the age, gender and nine clinical phenotype variables of the patients, namely body mass index (BMI), waist-hip ratio (WHR), insulin resistance index (HOMA-IR), high-density lipoprotein cholesterol (HDL-C), triglyceride (TG), systolic blood pressure (SBP), diastolic blood pressure (DBP), serum creatinine (SCR) and alanine aminotransferase (ALT);

[0011] S2: Preprocessing the clinical index data of the training cohort and the validation cohort in step S1, including removing missing values and outliers, performing natural logarithmic transformation on skewed distribution variables, then establishing a linear regression model for each phenotype variable and correcting age and gender to obtain a standardized residual matrix;

[0012] S3: using UMAP algorithm to reduce the dimensionality of the standardized residual matrix obtained in step S2, constructing a two-dimensional projection of the heterogeneity of pre-diabetic patients, applying graph clustering algorithm for initial hard clustering, calculating the normalized eigen centrality score of each individual in its belonging cluster, and taking it as the weight to calculate the center and covariance matrix of each cluster, and constructing a Gaussian mixture distribution;

[0013] S4: verifying the independent running of the above steps S2-S3 in the validation queue; at the same time, optimizing the phenotype characteristics identified in step S3 in the training queue, retaining the phenotype characteristics that are stable and reproducible in both the training queue and at least one validation queue, verifying the robustness and generalization of the model;

[0014] S5: based on the robust phenotype characteristics, evaluating the risk of future diabetes-related diseases of pre-diabetic individuals, and constructing a data-driven risk prediction model for the progression of future diabetes-related diseases in pre-diabetic individuals;

[0015] Preferably, the pre-diabetes is defined as FPG of 5.6 mmol / L to 6.9 mmol / L, or PBG of 7.8 mmol / L to 11.0 mmol / L, or HbA1c of 5.7% to 6.4% in participants without diabetes.

[0016] Preferably, the future diabetes-related diseases include T2DM, CKD and CVD.

[0017] Preferably, in step S1, the training queue is 71,880 pre-diabetic participants from the Chinese Cardiovascular and Metabolic Diseases and Malignant Tumor Study (4C); the validation queue includes: 1,827 individuals with pre-diabetes from a community-based prospective cohort established in June-August 2009 in the Songnan area of Shanghai, China; and 5,293 pre-diabetic participants from a community-based prospective cohort established in March-June 2011 in Karamay, Xinjiang Uygur Autonomous Region, China; and clinical variable data of pre-diabetic population is collected.

[0018] Preferably, in step S2, the outliers are values outside the range of mean ± 5 standard deviations (> 5 SD); and the skew distribution variables include HOMA-IR, TG, ALT and SCR, and are subjected to natural logarithmic transformation.

[0019] Preferably, in step S3, the parameter settings of UMAP algorithm include: based on the total number of individuals (N total) Calculate the number of nearest neighbors, formula nn = max(10, 10 + 15 x (log 10 (N total )−4)) and set binary_edge_weights = TRUE and dens_scale = 1; the graph clustering algorithm comprises: finding stable initial seeds by Leading eigen vector algorithm, and clustering the two-dimensional projected neighborhood network using Leiden algorithm, each individual is only assigned to one cluster. Then calculate the normalized eigenvector centrality score of each individual in its belonging cluster to measure its importance in the cluster, these scores are used as weights to calculate the center point and covariance matrix of each cluster, which together constitute a Gaussian mixture distribution.

[0020] As preferred, in step S4, since the central region of the projection space is less stable in different iterations, a "non-significant feature" with mean zero and covariance matrix as the identity matrix is introduced as a reference distribution in GMM to enhance the robustness of clustering. The stable reproduction refers to: the individual assigned to a certain phenotypic feature probability > 75% in the training cohort, whose phenotypic feature probability > 75% in any validation cohort, and those who do not meet the above conditions are included in the "non-significant feature" distribution. The center and covariance matrix of the above clustering are fixed, and the weight of the feature typing is recalibrated accordingly, representing its proportion in the overall population, each individual has a probability score for each type, and the sum of the probabilities of all types is equal to 1.

[0021] As preferred, in step S4, six stable feature patterns in each cohort are determined, including five patterns with significant phenotypic features: high blood pressure (HBP); high blood lipids, creatinine and liver enzymes (HLP_SCR_ALT); obesity and high insulin resistance (HOIR); relative health (RH); low HDL-C; and one non-significant feature (NSF) pattern.

[0022] As preferred, in step S5, compared with NSF, the risk of T2D significantly increases when the probability of HBP or HLP_SCR_ALT increases by 20%.

[0023] As preferred, in step S5, compared with NSF, the risk of CVD increases by 20% for each increase in the probability of HBP.

[0024] As preferred, in step S5, compared with NSF, higher probability of HBP and HLP_SCR_ALT is associated with increased risk of CKD, and RH has a lower risk of CKD.

[0025] In a second aspect, the application provides a data-driven pre-diabetes future diabetes-related disease progression risk prediction system, comprising:

[0026] A data collection module is configured to collect clinical index data of pre-diabetes patients, including age, gender, and nine clinical phenotype variables: BMI, WHR, HOMA-IR, HDL-C, TG, SBP, DBP, SCR, and ALT.

[0027] A data preprocessing module is configured to remove missing values and outliers, perform logarithmic conversion on skew distribution variables, and perform age and gender correction.

[0028] A data analysis module is configured to perform UMAP dimension reduction, graph clustering, and Gaussian mixture model calculation in a training cohort.

[0029] A data verification module is configured to optimize and verify the robustness of the phenotype characteristics in a validation cohort, determine the final model, and output the probability of an individual belonging to each phenotype characteristic.

[0030] A prediction model construction module is configured to calculate the allocation probability of pre-diabetes population belonging to different phenotypes, use the allocation probability as an independent variable to establish Cox and Logistic regression models, evaluate the risk of pre-diabetes future diabetes-related diseases, and construct a pre-diabetes future disease progression risk prediction model.

[0031] Preferably, the future diabetes-related diseases include T2DM, CKD, and CVD.

[0032] Compared with the prior art, the application has the following advantages:

[0033] 1. The application first applies a series of soft clustering methods combining UMAP dimension reduction, graph clustering, and GMM to a large-scale Chinese pre-diabetes population. Sufficient sample size can construct a robust phenotype structure, effectively capture high-dimensional nonlinear data structure in exploring disease progression and understanding disease heterogeneity, and identify stable phenotype characteristics with more biological significance and clinical interpretability. Using probability allocation instead of hard classification can better reflect the continuity and heterogeneity of the disease, reduce false classification, and perform dimension reduction on multi-dimensional continuous clinical variables of pre-diabetes population to describe disease heterogeneity, which helps to more personalized guide pre-diabetes care.

[0034] 2. The present application characterizes the phenotypic heterogeneity of individuals with prediabetes from a large nationwide population-based Chinese prospective cohort, rather than classifying individuals into specific subgroups, the probability of each participant in each feature subgroup reflects the continuity of metabolic heterogeneity, studies the clinical metabolic differences of these features and their association with the risk of T2D and related complications, and further confirms the robustness and generalizability of these features in two independent validation cohorts, providing new insights into the phenotypic structure of prediabetes and establishing a conceptual and methodological framework for early risk stratification and personalized prevention strategies.

[0035] 3. In the data-driven prediabetes future diabetes-related disease progression risk prediction model of the present application, the HBP and HLP_SCR_ALT phenotype features have increased risks of T2D and CKD, while the HBP feature has a higher risk of CVD, and prevention strategies based on this model will likely benefit high-risk populations of T2D, CVD, and CKD. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The flowchart of the construction method and the schematic diagram of the model structure of the data-driven prediabetes future diabetes-related disease progression risk prediction model in the examples; three independent prospective cohorts are included: the training cohort includes 71880 individuals with prediabetes from different regions of China; validation cohort 1 (XJ_2011) contains 5293 prediabetes patients from Karamay City, Xinjiang Uygur Autonomous Region, China; validation cohort 2 (SN_2009) contains 1827 prediabetes patients from Songnan District, Shanghai City, China; nine clinical variables are extracted as potential phenotype biomarkers for prediabetes classification, and the overall analysis steps are shown, including detailed description of the integrated algorithm used to convert biomarker features into probability feature maps.

[0037] Figure 2 UMAP two-dimensional projection map of biomarker differences in different cohorts in the examples.

[0038] Figure 3 Comparison chart of different cohort phenotype features in the examples; the cluster centers are obtained by applying the same clustering method to the 3 cohorts respectively, the points represent the average residual value of each clinical variable in each clustering feature, and the lines represent the distance from zero.

[0039] Figure 4 Validation chart of the phenotype features identified in the 4C study in the validation cohort in the examples; the nodes are the cluster groups found by running the same clustering algorithm in each cohort; the numbers and Figure 3the same number in ; connecting clusters from the 4C_2011 model with clusters from another cohort, provided that individuals with high probability of being assigned to a 4C_2011 cluster (>75%) also have high median probability of being assigned to a cluster of another validation cohort (>75%), only 4C_2011 clusters with edges to at least one cluster of a validation cohort are considered reproducible.

[0040] Figure 5 Figure 6. Phenotype signatures identified in the 4C study and stable across the replication cohorts; A: UMAP two-dimensional projection plot, colors represent feature assignments; B: cluster weights; C: forest plot of the average phenotype biomarker residual for each feature; points and error bars represent the estimated average residual for each biomarker and its 95% confidence interval; dashed lines represent residuals of 0; NSF: no significant features; HBP: hypertension; HLP_SCR_ALT: high lipids, serum creatinine, and alanine aminotransferase; HOIR: high obesity and insulin resistance; RH: relatively healthy; LHDL: low high-density lipoprotein cholesterol.

[0041] Figure 6 Figure 7. Association between phenotype signature probabilities and future cardiovascular metabolic disease risk in prediabetic patients; estimated disease risk for individuals with 10% or 20% increase in the probability of belonging to each different phenotype signature, while subtracting the NSF combined probability accordingly, presented as random effect values from the multi-cohort meta-analysis. DETAILED DESCRIPTION

[0042] In order to more fully understand the technical solutions, objectives and advantages of the present application, the following further describes the technical effects produced by the present application in detail and completely in combination with the drawings and specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all. It should be noted that other embodiments obtained by those of ordinary skill in the art without departing from the concept of the present application all belong to the protection scope of the present application.

[0043] The 4C study in the following embodiments is a prospective cohort study for Chinese adults aged 40 and above, and the research subjects are older prediabetic patients. Prediabetes is an intermediate stage of diabetes development and has multiple levels of heterogeneity, emphasizing accurate risk assessment to achieve early intervention, effective management, and targeted prevention of diabetes and complications.

[0044] An innovative unsupervised learning framework was adopted in the following examples, integrating UMAP dimensionality reduction, graph clustering, and Gaussian mixture model, based on three large-scale Chinese prospective cohort studies, including a training cohort (4C study) and two independent validation cohorts (XJ_2011 and SN_2009), to evaluate the heterogeneity of 79,000 prediabetic patients according to 9 clinical phenotype variables of BMI, WHR, HOMA-IR, HDL-C, TG, SBP, DBP, SCR, and ALT, explore different risks of progressing to T2D, CKD, and CVD, and construct risk prediction models. The T2D and CKD risks of the hypertensive (HBP) and hyperlipidemia-creatinine-ALT (HLP_SCR_ALT) phenotype characteristics were increased, while the CVD risk of the HBP characteristic was higher. The prevention strategies based on this model would bring benefits to the high-risk population of T2D, CVD, and CKD.

[0045] Example 1

[0046] 1. Materials and Methods

[0047] Cohort and variable definition, three community-based prospective cohort datasets were used for analysis:

[0048] Training cohort: 71,880 prediabetic participants in the Chinese Cohort Study of Cardiovascular and Metabolic Diseases and Malignancies (4C) at baseline in 2011. The 4C study is a nationwide, multi-center, prospective, population-based study in Chinese adults, aiming to examine the relationship between blood glucose parameters and clinical outcomes, including diabetes, cardiovascular disease, chronic kidney disease, and mortality.

[0049] Validation cohort 1: 5,293 prediabetic participants in the Xinjiang Uygur Autonomous Region Study (XJ_2011).

[0050] Validation cohort 2: 1,827 prediabetic participants in the Shanghai Songnan Study (SN_2009).

[0051] All participants signed informed consent forms, and the study was approved by the Ethics Committee of Ruijin Hospital, School of Medicine, Shanghai Jiao Tong University.

[0052] 2. Clinical index information collection

[0053] Collected clinical indices included: age, sex, BMI, WHR, HOMA-IR, HDL-C, TG, SBP, DBP, SCR, ALT. In the 4C and XJ_2011 studies, HbA1c was measured by high-performance liquid chromatography using a VARIANT II hemoglobin testing system (Bio-Rad Laboratories). Serum insulin, creatinine, ALT, total cholesterol (TC), HDL-C, and TG were measured by an automated analyzer (ARCHITECT ci16200, Abbott Laboratories, Chicago, IL, USA). In the SN_2009 study, plasma glucose, serum creatinine, ALT, and lipids were measured by an automated analyzer (ADVIA VR 1650 Chemistry System, Bayer Corporation, Leverkusen, Germany). Serum insulin was measured by electrochemiluminescence (Roche Diagnostics, Basel, Switzerland). Homeostatic model assessment of insulin resistance (HOMA-IR) was calculated as: fasting insulin (μΙU / mL) x fasting glucose (mg / dL) / 405.

[0054] 3. Prediabetes was defined as FPG 5.6 to 6.9 mmol / L, or PBG 7.8 to 11.0 mmol / L, or HbA1c 5.7 to 6.4% in participants without diabetes.

[0055] 4. Outcome definitions

[0056] In the 4C study, T2D was defined as FPG > 7.0 mmol / L, or 2h-PG > 11.1 mmol / L, or HbA1c > 6.5%, or physician-diagnosed diabetes. In the XJ_2011 and SN_2009 studies, new-onset diabetes was defined as physician-diagnosed diabetes based on the National Health Insurance system.

[0057] In the 4C study, the primary composite outcome of adverse renal events included the occurrence of renal failure requiring dialysis or renal replacement therapy, death due to renal causes, eGFR < 60 mL / min / 1.73 m 2 at follow-up, or a decrease in eGFR > 25% from baseline. In the XJ_2011 and SN_2009 studies, the occurrence of renal events was defined as death due to renal causes, eGFR < 60 mL / min / 1.73 m 2 at follow-up, or a decrease in eGFR > 25% from baseline.

[0058] Major cardiovascular events included: first non-fatal myocardial infarction, non-fatal stroke, hospitalization or treatment for heart failure, and cardiovascular death. Information on vital status and clinical outcomes was collected from the National Disease Surveillance Point System and local death and disease registries of the National Health Insurance System.

[0059] 5. Statistical analysis

[0060] The method flowchart and model construction process are shown in FIG. 1. Figure 1 Individuals with missing values in any of the clustering variables or with extreme outliers (> 5 SD) were removed. HOMA-IR, TG, ALT, and SCR were natural log-transformed to correct skewness. For each phenotype variable, a linear regression model (dependent variable: phenotype value, independent variables: age and sex) was established to obtain the residuals. This step aimed to eliminate the effects of age and sex on the phenotype variable and enhance the biological interpretability of subsequent clustering analysis. The residuals were standardized (mean = 0, SD = 1) to form a standardized residual matrix.

[0061] UMAP dimension reduction was performed using the R package uwot. Parameter settings: the number of nearest neighbors was calculated based on the total number of individuals in each cohort (N total ), with the formula nn = max(10, 10 + 15 × (log 10 (N total ) - 4)), which ensured that the number of nearest neighbors was adaptively adjusted with the size of the cohort. binary_edge_weights = TRUE and dens_scale = 1 were set. On the UMAP projection, the Leading eigen vector algorithm was used to find stable initial seeds, and the Leiden algorithm was used to perform graph clustering on the two-dimensional projected neighborhood network to obtain the initial hard clustering results. Subsequently, the normalized eigen centrality scores of each individual in its own cluster were calculated to measure its importance in the cluster. These scores were used as weights to calculate the center point and covariance matrix of each cluster, which together constitute a Gaussian mixture distribution.

[0062] A "non-significant feature" with mean zero and covariance matrix identity was introduced in GMM as a reference distribution to enhance the robustness of clustering, as the central region of the low-dimensional projection space was less stable across iterations. The same parameter configuration was used in both validation cohorts and the same clustering analysis pipeline was run, and the results were compared with the test cohort. Individuals assigned to a certain phenotypic feature with probability >75% in the training cohort were retained into the final model if they also had a probability >75% in any of the validation cohorts. Those who did not meet the above condition were included in the "non-significant feature" distribution. Subsequently, the weight of each cluster was adjusted and the validated stable clustering was applied to subsequent analysis. Each cluster was named according to the average residual error of all its biomarkers.

[0063] Regression models were used to further assess the risk of individuals with different phenotypic features to progress to T2DM, CKD and CVD, and to construct risk prediction models for future disease progression in individuals with prediabetes. For each outcome, two models were constructed: (1) a base model containing a series of known risk factors; (2) a multivariate model including the log-transformed odds ratio of the probability of phenotypic assignment in addition to the base model. Since the probability of phenotypic assignment itself has a constraint that the sum is 1, the probability of each individual having a significant feature was divided by the probability of NSF, and the natural logarithm of these quotients was taken as the predictor variable. For CVD, the base model included the predictors in the PREVENT-base model, as well as the phenotypic indicators used in the clustering analysis but not included in the PREVENT-base (including: age, sex, HOMA-IR, BMI, WHR, TG, ALT, SCR, TC, HDL-C, SBP, eGFR, current smoking, use of antihypertensive and lipid-lowering drugs). For T2D and CKD, the base model included all variables involved in the clustering analysis, with the addition of current smoking status, use of antihypertensive and lipid-lowering drugs, and other covariates that might affect the selected phenotype. In the multivariate model, the change in disease risk was assessed when the probability of different phenotypic clusters increased by 10% or 20% (and the probability of NSF cluster decreased accordingly). Likelihood ratio tests were used to assess the difference in C-statistics, and further quantify the proportion of additional variance explained by the inclusion of the probability of phenotypic assignment, expressed as a percentage of the total explained variance.

[0064] 6. Results

[0065] The baseline characteristics of the training and validation cohorts are shown in Table 1.

[0066] Table 1: Baseline characteristics of study participants

[0067]

[0068] Note: Data are presented as mean ± standard deviation, number (percentage), or median (interquartile range).

[0069] 6.1. Visualization and clustering of phenotypic features

[0070] The same analysis was performed in each of the three independent cohorts (Table 2) Figure 2 and Figure 3 Based on the reproducible phenotyping results in each cohort (Table 2) Figure 4 six phenotypic features were ultimately identified: Non-significant feature (NSF), Relative healthy (RH), Hypertension (HBP), High lipids-creatinine-ALT (HLP_SCR ALT), Low HDL-C (LHDL), and High obesity insulin resistance (HOIR) (Table 3) Figure 5 Individuals with significant phenotypic features were significantly different from NSF individuals (Table 4) Figure 5 C.

[0071] 6.2. Association of different phenotypic features with clinical outcomes

[0072] The change in disease risk when the probability of belonging to different phenotypic classes was increased by 10% or 20% (and the probability of belonging to the NSF class was decreased accordingly) was evaluated in a hypothetical 55-year-old male with prediabetes (Table 5) Figure 6 When the probability of belonging to the HBP or HLP_SCR ALT classes was increased by 20%, the risk of T2D was significantly higher than in the NSF class, with corresponding odds ratios (OR, 95% confidence interval) of 1.09 (1.07-1.11) and 1.10 (1.06-1.14), respectively Figure 6 Similarly, a 20% increase in the probability of belonging to the HBP class was associated with an increased risk of CVD, with a hazard ratio (HR) of 1.06 (95% CI: 1.00-1.12). In addition, higher probabilities of belonging to the HBP and HLP_SCR ALT classes were also associated with an increased risk of CKD, with odds ratios of 1.12 (1.09-1.15) and 1.10 (1.04-1.16), respectively, compared to the NSF class. Conversely, the relative healthy (RH) class was associated with a decreased risk of T2D and CKD, with odds ratios of 0.87 (95% CI: 0.84-0.89) and 0.81 (95% CI: 0.77-0.85), respectively.

[0073] 6.3. Evaluation of the performance of the risk prediction model

[0074] To evaluate the incremental value of phenotyping in disease prediction, the probability of assignment of individuals to a profile was incorporated into the base model. In the 4C_2011 study, the addition of profile information significantly improved model performance, as indicated by the increase in likelihood ratio test results and C-statistics (Table 2). The proportion of additional variance explained by phenotyping in the prediction of T2D, CVD, and CKD events ranged from 0.97% to 16.82%. Although in the XJ_2011 and SN_2009 studies, profile information also explained some additional variance, these improvements did not reach statistical significance according to the likelihood ratio test results.

[0075] Table 2: Comparison of models without and with inclusion of probabilities of assignment to a phenotypic profile

[0076]

[0077] 7. Discussion and conclusion

[0078] In this study, we used an unsupervised clustering approach to elucidate the significant heterogeneity of individuals with prediabetes in a nationwide, population-based prospective Chinese cohort. The clustering structure proved to be robust and was validated in two independent cohorts. The innovative application of a combination of UMAP, graph clustering, and GMM approaches identified 6 distinct metabolic branches based on 9 clinical phenotypic variables that effectively represented the differences in metabolic profiles and highlighted different risks of T2DM, CKD, and CVD events. Notably, the RH profile was consistently associated with lower disease risk, while the HLP_SCR_ALT and HBP profiles indicated varying degrees of higher cardiovascular metabolic burden. Moreover, the inclusion of profile assignment probabilities significantly improved disease prediction models, indicating that phenotyping has additional clinical utility beyond conventional biomarkers.

[0079] The above only is the preferred embodiment of the present application, and is not used to limit the present application, for those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for constructing a data-driven model to predict the risk of future diabetes-related disease progression in prediabetes, characterized in that, Includes the following steps: S1: Collect clinical data of prediabetic patients in the training and validation cohorts, including patients’ age, sex and nine clinical phenotypic variables, namely body mass index (BMI), waist-to-hip ratio (WHR), insulin resistance index (HOMA-IR), high-density lipoprotein cholesterol (HDL-C), triglycerides (TG), systolic blood pressure (SBP), diastolic blood pressure (DBP), serum creatinine (SCR) and alanine aminotransferase (ALT). S2: Preprocess the clinical indicator data of the training and validation queues mentioned in step S1, including removing missing and outlier values, performing natural logarithmic transformation on skewed distribution variables, and then establishing a linear regression model for each phenotypic variable and correcting for age and gender to obtain a standardized residual matrix. S3: Use the UMAP algorithm to reduce the dimensionality of the standardized residual matrix obtained in step S2, construct a two-dimensional projection of the heterogeneity of prediabetic patients, apply the graph clustering algorithm to perform initial hard clustering, calculate the normalized eigenvector centrality score of each individual in its own cluster, and use it as weight to calculate the centroid and covariance matrix of each cluster, forming a Gaussian mixture distribution. S4: The validation queue independently runs the above steps S2-S3, while optimizing the phenotypic features identified in step S3 in the training queue, retaining the phenotypic features that are stably reproduced in both the training queue and at least one validation queue, and verifying the robustness and generalization of the model. S5: Assess the risk of future diabetes-related diseases in individuals with prediabetes based on robust phenotypic characteristics, and construct a data-driven model for predicting the risk of future diabetes-related disease progression in individuals with prediabetes. The term prediabetes is defined as: in participants without diabetes, FPG of 5.6 mmol / L to 6.9 mmol / L, PBG of 7.8 mmol / L to 11.0 mmol / L, or HbA1c of 5.7% to 6.4%.

2. The method for constructing a data-driven model for predicting the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, The future diabetes-related diseases mentioned include T2DM, CKD, and CVD.

3. The method for constructing a data-driven model for predicting the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, In step S1, the training cohort consists of 71,880 prediabetic participants from the China Cardiovascular Metabolic Disease and Malignant Tumor Study; the validation cohort includes: a community-based prospective cohort established in Songnan District, Shanghai, China from June to August 2009, comprising 1,827 individuals with prediabetes; and a community-based prospective cohort established in Karamay District, Xinjiang Uygur Autonomous Region, China from March to June 2011, comprising 5,293 prediabetic participants; clinical variable data of the prediabetic population are collected.

4. The method for constructing a data-driven prediction model for the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, In step S2, the outlier values ​​are those that exceed the mean by ±5 standard deviations; the skewed distribution variables include HOMA-IR, TG, ALT, and SCR, and are transformed by their natural logarithm.

5. The method for constructing a data-driven prediction model for the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, In step S3, the parameter settings of the UMAP algorithm include: based on the total number of people N in each queue. total The formula for calculating the number of nearest neighbors is nn = max(10, 10 + 15 × (log n)). 10 (N total )−4)), and set binary_edge_weights=TRUE and dens_scale=1; The graph clustering algorithm includes: finding a stable initial seed using the Leading eigenvector algorithm; using the Leiden algorithm to cluster the neighborhood network of the two-dimensional projection, with each individual being assigned to only one cluster; then calculating the normalized eigenvector centrality score of each individual in its respective cluster, which is used as a weight to calculate the centroid and covariance matrix of each cluster, which together form a Gaussian mixture distribution.

6. The method for constructing a data-driven prediction model for the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, In step S4, a "non-significant feature" with zero mean and an identity matrix is ​​introduced into the GMM as a reference distribution. The stable reproducibility means that individuals assigned to a certain phenotypic feature with a probability greater than 75% in the training queue also have a phenotypic feature probability greater than 75% in any validation queue. Individuals that do not meet the above conditions are included in the "non-significant feature" distribution. The center and covariance matrix of the above clusters are fixed, and the weights of the feature subtypes are recalibrated to represent their proportion in the overall population. Each individual has a probability score for each subtype, and the sum of the probabilities of all subtypes is equal to 1.

7. The method for constructing a data-driven prediction model for the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, In step S4, stable characteristic patterns are identified in each cohort, including five patterns with significant phenotypic features: hypertension (HBP); hyperlipidemia, creatinine, and liver enzymes (HLP, SCR, ALT); obesity and high insulin resistance (HOIR); relative health (RH); low high-density lipoprotein cholesterol (HDL-C); and one pattern with non-significant NSF.

8. The method for constructing a data-driven prediction model for the risk of future diabetes-related disease progression in prediabetes according to claim 1, characterized in that, In step S5, when the probability of HBP or HLP_SCR_ALT increases by 20% compared to NSF, the risk of having T2D is significantly increased, while the risk of having T2D is low for RH. Compared to NSF, for every 20% increase in HBP probability, the risk of CVD increases; Compared to NSF, the high incidence of HBP and HLP_SCR_ALT is associated with an increased risk of CKD, while RH has a low risk of CKD.

9. A data-driven system for predicting the risk of future diabetes-related disease progression in prediabetes, characterized in that, include: The data acquisition module is used to collect clinical indicator data of patients with prediabetes, including age, gender, and nine clinical phenotypic variables: BMI, WHR, HOMA-IR, HDL-C, TG, SBP, DBP, SCR, and ALT. The data preprocessing module is used to remove missing and outlier values, perform logarithmic transformation on skewed distribution variables, and correct for age and gender. The data analysis module is used to perform UMAP dimensionality reduction, graph clustering, and Gaussian mixture model calculations in the training queue; The data validation module is used to optimize and validate the robustness of phenotypic features in the validation queue, determine the final model, and output the probability that an individual belongs to each phenotypic feature. The predictive model building module is used to calculate the probability of different phenotypes in people with prediabetes. This probability is used as an independent variable to build Cox and Logistic regression models to assess the risk of future diabetes-related diseases in people with prediabetes and to build a predictive model for the future progression of prediabetes.

10. The data-driven prediabetes-related disease progression risk prediction system according to claim 9, characterized in that, The future diabetes-related diseases mentioned include T2DM, CKD, and CVD.

Images