A method for constructing a subtype of CKM, a prediction method and a prediction system

Abstract

The application discloses a CKM subtype construction method, a prediction method and a prediction system, belongs to the technical field of medical care information science, and uses clinical detection data to calculate CHG reflecting sugar-fat metabolism load and SII reflecting systematic immune inflammation state. Linear mixed effect models are used to extract individual longitudinal levels and change trends of CHG and SII. Gaussian mixed models are used to identify different CKM longitudinal trajectory subtypes. According to a prediction model based on the trajectory subtypes, the progress outcome of CKM is further associated and predicted, thereby providing auxiliary reference for CKM progress risk stratification and clinical management. The application provides reference for clinical follow-up, risk monitoring and individualized management.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, specifically to a method for constructing subtypes of CKM, a prediction method, and a prediction system. Background Technology

[0002] Cardiovascular-renal-metabolic syndrome (CKM) is a health disorder resulting from the pathophysiological interactions between obesity, type 2 diabetes mellitus (T2DM), chronic kidney disease (CKD), and cardiovascular disease (CVD). CKM has been reported to have a profound impact on morbidity and mortality.

[0003] Publication CN117976223A provides a method for constructing subtypes of cardiovascular-metabolic comorbidities and their application in cardiovascular disease risk prediction. It selects 21 biomarkers and phenotypic indicators across eight categories: obesity, glucose metabolism, insulin metabolism, lipid metabolism, uric acid metabolism, liver function, kidney function, and blood pressure for modeling. Subsequently, an unsupervised K-means clustering algorithm is used, combined with the DB and KL indices, to determine the optimal number of cluster subtypes, thus constructing multiple cardiovascular-metabolic comorbidity subtypes. These subtypes are used for long-term cardiovascular disease risk prediction. However, its application and classification target are not cardiovascular-renal-metabolic syndrome.

[0004] Publication document CN119943361A provides a cardiovascular disease risk assessment system and method based on cardiorenal metabolic indicators. It uses seven indicators—age, sex, smoking status, systolic blood pressure, fasting blood glucose, non-high-density lipoprotein cholesterol, and estimated glomerular filtration rate—as predictive factors to output the risk of total cardiovascular disease, atherosclerotic cardiovascular disease, and heart failure, respectively. Its focus is on predicting the probability of cardiovascular events, rather than identifying clinical subtypes within CKM patients.

[0005] Current technologies lack methods and systems for clinically stratified CKM and risk assessment of progression based on routine biochemical and blood test data. Summary of the Invention

[0006] To address the aforementioned technical problems in the existing technology, this invention provides a method for constructing CKM subtypes, a prediction method, and a prediction system. Based on clinical test data, CKM subtypes are constructed, and CKM is predicted based on the constructed subtypes.

[0007] The first aspect of this invention discloses a method for constructing CKM subtypes, comprising the following steps: collecting clinical test data; preprocessing the clinical test data to obtain a first dataset; calculating the natural logarithmic values ​​of the cholesterol-high-density lipoprotein-glycemic index (CHG) and the systemic immune inflammation index (SII) based on the first dataset; constructing a first linear mixed-effects model and a second linear mixed-effects model based on CHG and log SII, respectively; extracting the longitudinal horizontal features of CHG and the longitudinal trend features of CHG from the intercept and slope of the first linear mixed-effects model; extracting the longitudinal horizontal features of log SII and the longitudinal trend features of log SII from the intercept and slope of the second linear mixed-effects model; constructing a feature vector based on the longitudinal horizontal features of CHG, the longitudinal trend features of CHG, the longitudinal horizontal features of log SII, and the longitudinal trend features of log SII; and performing unsupervised clustering on the feature vector based on a Gaussian mixture model to obtain a clustering model and the corresponding trajectory subtypes.

[0008] Preferred methods for obtaining subject trajectory subtypes include: Based on the clustering model and the clinical data of the subjects, the posterior probability of belonging to each trajectory subtype is calculated; By maximizing the posterior probability, the trajectory subtype to which the subject belongs can be obtained.

[0009] Preferably, the trajectory subtype to which the subject belongs is obtained by maximizing the posterior probability and the posterior probability threshold; that is, it is required that the posterior probability is maximized to be greater than the posterior probability threshold. The total number of trajectory subtypes is four.

[0010] The preferred method for calculating the cholesterol-high-density lipoprotein-glycemic index is as follows: CHG=ln[(TC×FBG) / (2×HDL_C)] (1); Where TC is total cholesterol, FBG is fasting blood glucose, and HDL_C is high-density lipoprotein cholesterol; the formula for calculating the systemic immune inflammation index SII is: SII=PLT×NEUT / LYMPH (2); Among them, PLT represents platelet count, NEUT represents neutrophil count, and LYMPH represents lymphocyte count.

[0011] Preferably, after standardizing CHG and log SII, a first linear mixed-effects model and a second linear mixed-effects model are constructed, respectively.

[0012] A second aspect of the present invention provides a CKM prediction method, comprising the following steps: Based on the above trajectory subtypes and CKM progression outcomes, a prediction model is constructed. Predictive models are used to forecast the clinical test data of subjects to obtain the risk of CKM progression and / or survival probability.

[0013] Preferably, the predictive model includes KM survival curves and / or a first model based on Cox proportional hazards.

[0014] Preferably, the predictive model includes a second model based on Cox proportional hazards, and the covariates or independent variables of the second model also include age, gender, cardiorenal related indicators and comorbidities.

[0015] A third aspect of the present invention provides a prediction system for implementing the above-described prediction method, comprising an input module, a subtype identification module, and a prediction module; The input module is used to collect the clinical test data of the subjects to obtain the first dataset; The subtype identification module is used to calculate the posterior probability of a subject belonging to each trajectory subtype based on the first dataset and a clustering model based on a Gaussian mixture model; by maximizing the posterior probability, the trajectory subtype to which the subject belongs is obtained. The prediction module is used to predict the risk of CKM progression and / or survival probability based on the clinical test data of the subjects according to the prediction model.

[0016] Preferably, the prediction system further includes a first training module and a second training module; The first training module is used to perform unsupervised clustering of feature vectors based on Gaussian mixture models to obtain clustering models and corresponding trajectory subtypes; The second training module is used to train the training set with trajectory subtypes as variables to obtain a prediction model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: Using clinical test data, CHG (cholesterol-induced glucose and lipid metabolism) and SII (systemic immune inflammation) were calculated. A linear mixed-effects model was used to extract individual longitudinal levels and trends of CHG and SII. A Gaussian mixture model was employed to identify different longitudinal trajectory subtypes of chronic CKM (chronic inflammatory disease). Based on a trajectory subtype-based prediction model, CKM progression outcomes were further correlated and predicted, providing auxiliary references for CKM progression risk stratification and clinical management. This also provides a reference for clinical follow-up, risk monitoring, and individualized management. Attached Figure Description

[0018] Figure 1 This is a flowchart of the CKM subtype construction and prediction method of the present invention; Figure 2 It is the trajectory diagram of a linear mixed-effects model; Figure 3 It is based on the KM curve chart; Figure 4This is a time-dependent ROC analysis chart; Figure 5 This is a chart of net reclassification index analysis; Figure 6 It is a trajectory plot with a posterior probability exceeding 0.7; Figure 7 It is a trajectory plot with a posterior probability exceeding 0.8; Figure 8 It is a KM curve with a posterior probability exceeding 0.7; Figure 9 It is a KM curve with a posterior probability exceeding 0.8; Figure 10 It is a Cox regression forest plot with a posterior probability exceeding 0.7; Figure 11 It is a Cox regression forest plot with a posterior probability exceeding 0.8; Figure 12 This is a logic block diagram of the prediction system of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 provides a method for constructing CKM subtypes, such as Figure 1 As shown, it includes the following steps: Step S1: Collect clinical test data; preprocess the clinical test data to obtain the first dataset.

[0021] Step S2: Based on the first dataset, calculate the natural logarithm of the cholesterol-high-density lipoprotein-glycemic index CHG and the systemic immune inflammation index SII, log SII.

[0022] The formula for calculating CHG is as follows: CHG=ln[(TC×FBG) / (2×HDL_C)] (1); Among them, TC is total cholesterol, FBG is fasting blood glucose, and HDL_C is high-density lipoprotein cholesterol.

[0023] The formula for calculating SII is: SII=PLT×NEUT / LYMPH (2); Where PLT represents platelet count, NEUT represents neutrophil count, and LYMPH represents lymphocyte count. Since SII often shows a right-skewed distribution in clinical data, a natural logarithmic transformation is performed on SII to obtain log SII.

[0024] Step S3: Construct the first linear mixed-effects model LMM1 and the second linear mixed-effects model LMM2 based on the cholesterol-high-density lipoprotein-glycemic index CHG and the natural logarithm log SII, respectively.

[0025] Step S4: Extract the longitudinal horizontal features (intercept) and longitudinal trend features (slope) of CHG from the first linear mixed-effects model LMM1; extract the longitudinal horizontal features (intercept) and longitudinal trend features (slope) of log SII from the second linear mixed-effects model LMM2.

[0026] Among them, the longitudinal horizontal characteristic (intercept) can be understood as the long-term load difference of the subjects relative to the overall average level; the longitudinal trend characteristic (slope) can be understood as the direction and rate of change of the subjects' indicators as they increase, decrease or remain stable over time.

[0027] Step S5: Construct a feature vector based on the vertical horizontal features of CHG, the vertical trend features of CHG, the vertical horizontal features of log SII, and the vertical trend features of log SII.

[0028] The joint longitudinal feature vector V can be expressed as: V = (CHG intercept, CHG slope, log SII intercept, log SII slope).

[0029] Step S6: Perform unsupervised clustering on the feature vectors based on the Gaussian mixture model to obtain the clustering model and its corresponding trajectory subtypes; and optimize the number of trajectory subtypes.

[0030] The clustering model based on Gaussian mixture model can output the posterior probability of a subject belonging to each trajectory subtype. The trajectory subtype to which the subject belongs is obtained by maximizing the posterior probability Pmax. The posterior probability can be obtained by using the distance between the subject's clinical test data and the corresponding clustering model of the trajectory subtype, such as calculating the average distance between the clinical test data and each point in the clustering model. The calculation method is existing technology and will not be elaborated here.

[0031] The method for predicting CKM includes the following steps: Step S7: Construct a prediction model based on the trajectory subtype and CKM progression outcome.

[0032] Predictive models can be trained using machine learning methods, such as survival curves (Kaplan-Meier, KM curves), Cox proportional hazards models, etc., but are not limited to these.

[0033] Step S8: Predict the clinical test data of the subjects using a predictive model to obtain the risk of CKM progression and / or survival probability.

[0034] Step S9: Model stability evaluation.

[0035] Using clinical test data, CHG (critical blood glucose and lipid metabolism) reflecting glucose and lipid metabolic burden and SII (systemic immune inflammation) reflecting systemic immune inflammation status were calculated. A linear mixed-effects model (LMM) was used to extract individual longitudinal levels and trends of CHG and SII. A Gaussian mixture model (GMM) was employed to identify different longitudinal trajectory subtypes of CKM. Based on the trajectory subtype-based prediction model, CKM progression outcomes were further correlated and predicted, providing auxiliary references for CKM progression risk stratification and clinical management. This also provides a reference for clinical follow-up, risk monitoring, and individualized management.

[0036] Example 2 provides a specific CKM prediction scheme.

[0037] In step S1, longitudinal clinical data were collected from 1270 CKM-related subjects. Each subject had routine biochemical and complete blood count (CBC) data at multiple time points. Routine biochemical data included total cholesterol (TC), high-density lipoprotein cholesterol (HDL-C), and fasting blood glucose (FBG) for calculating CHG. CBC data included platelet count (PLT), neutrophil count (NEUT), and lymphocyte count (LYMPH) for calculating SII. Progression was defined as CKM progression to stage 4 clinical cardiovascular disease or the occurrence of a related clinical cardiovascular event.

[0038] Follow-up time was calculated from baseline time; for those without progression, the last follow-up time was used as the censoring time. Multi-timepoint test records were organized by subject ID and sorted from earliest to latest test time. The time when each subject first met the inclusion criteria or first entered the analysis cohort was used as the baseline time, and the follow-up time for each subsequent test was calculated relative to the baseline.

[0039] Data preprocessing: Units for the detected indicators are standardized. The indicators required for CHG calculation include TC, HDL_C, and FBG. If the raw units for TC and HDL_C are mmol / L, they can be converted as 1 mmol / L = 38.67 mg / dL. If the raw unit for FBG is mmol / L, it can be converted as 1 mmol / L = 18.00 mg / dL. The indicators required for SII calculation include PLT, NEUT, and LYMPH; the recommended unit is ×10⁻⁶. 9 / L. NEUT and LYMPH should use absolute counts, not percentages. If only percentages are provided in the raw data, they should be converted to absolute counts using white blood cell counts before calculating SII. A consistent unit system should be used throughout the training and application of the same model. Records with missing key indicators, zero denominators, or those clearly outside the physiological range should be removed. A follow-up time should be established for each subject relative to baseline, and this follow-up time should be centered.

[0040] In step S2, log SII is further standardized to obtain the standardized value of the natural logarithm of the systemic immune inflammation index, log SII_z; CHG is standardized to obtain CHG_z; and the follow-up time is centered, denoted as time_c. The structure of the first linear mixed-effects model is as follows: .

[0041] The structural representation of the second linear mixed-effects model is as follows: .

[0042] Where patient_id represents the subject ID. The random intercept represents the difference in indicator loading between the subject and the population average; the random slope represents the difference in the rate of change of the indicator relative to the population average. The optimizer is set to Bobyqa, and the maximum number of iterations is set to 2 × 10⁻⁶. 5 This is used to improve the convergence stability of linear mixed-effects models in longitudinal clinical data. Figure 2 The trajectory diagrams of the first linear mixed-effects model and the second linear mixed-effects model are shown.

[0043] In step S6, the number of candidate trajectory subtypes is set to 2, 3, 4, 5, and 6. Each candidate K value is individually fitted to the GMM model, and the corresponding class assignment results, posterior probabilities of each class, and model evaluation metrics are output. A random seed of 123 is set during model fitting to improve the reproducibility of the results. The final model is determined by combining the Bayesian Information Criterion (BIC), Mean AvePP, Min AvePP, Entropy, Min OCC, Min Prop, and trajectory subtype interpretability. Specifically, the BIC is used to evaluate the balance between model fit and complexity; Entropy is used to evaluate the overall classification clarity; Mean AvePP is used to evaluate the overall classification certainty; Min OCC is used to evaluate the quality of class assignment; and Min Prop is used to avoid classes with too small a sample size and weak interpretability.

[0044] After considering the model fit, class determinism, minimum class proportion, and trajectory subtype interpretability, four longitudinal trajectory subtypes of CKM were finally determined. The sample sizes for the four subtypes are as follows: Subtype 1 (Class 1) has 450 cases; Subtype 2 (Class 2) has 331 cases; Subtype 3 (Class 3) has 346 cases; and Subtype 4 (Class 4) has 143 cases. Referring to Table 1, the mean posterior probability (mean AvePP) of these four models is approximately 0.7536, the mean posterior probability (min AvePP) of the minimum class is greater than 0.70, and the minimum class proportion is approximately 11.26%.

[0045] While the 5-class model can further segment the population, its mean posterior probability of the smallest class decreases to approximately 0.6333, and the proportion of the smallest class decreases to approximately 8.35%. Therefore, the 4-class model achieves a better balance between class certainty, class size, and clinical interpretability.

[0046] The selection of the number of candidate subtypes (K) should not be based on a single indicator, but should comprehensively consider model fit, classification stability, class size, and result interpretability. BIC is used to balance model goodness of fit and model complexity; mean posterior probability is used to evaluate the certainty of an individual being assigned to a particular subtype; generally, a higher mean posterior probability indicates more stable classification; minimum class mean posterior probability can be used to determine if there are subtypes with poor classification quality, and it is generally desirable for it to be no less than 0.70; minimum class proportion is used to avoid generating excessively small classes, and it is generally desirable for the minimum class proportion to be no less than 5%, with no less than 10% being more beneficial for result interpretation and subsequent validation in clinical subtyping studies; Entropy is used to evaluate overall classification clarity, but this indicator is easily affected by the number of classes, and therefore should not be used as the sole selection criterion.

[0047] Table 1 The Cox global test or the KM curve can be used to verify whether the obtained subtypes have the ability to distinguish progression risk, but it should not be used as the sole criterion for determining the K value.

[0048] In step S7, after obtaining the longitudinal trajectory subtypes of CKM with K=4, the trajectory subtype label for each subject is merged and correlated with their CKM progression outcome to construct an analysis dataset. This dataset includes subject number, trajectory subtype, maximum a posteriori probability, baseline time, follow-up time, whether progression to CKM stage 4, and covariates such as age, gender, cardiorenal related indicators, and comorbidities. The trajectory subtype label is determined by the maximum a posteriori probability output by the GMM model.

[0049] First, Kaplan-Meier curves are plotted using trajectory subtypes as grouping variables, such as... Figure 3The probability of non-progression and cumulative risk of progression for different CKM longitudinal trajectory subtypes during the follow-up period were estimated, and the log-rank test was used to compare the differences in progression risk among different subtypes. If the KM curves of different subtypes separated, and the global test results indicated differences between groups, it indicates that the subtypes identified by the combined longitudinal trajectory of CHG and SII have the ability to distinguish CKM progression risk.

[0050] Subsequently, a first model based on Cox proportional hazards was constructed, incorporating trajectory subtype, age, and gender as categorical independent variables. Using a pre-defined reference subtype (subtype 3) as a benchmark, the hazard ratio (HR), 95% confidence interval (CI), and p-value for other subtypes relative to the reference subtype were calculated. The second model further adjusted for covariates such as age, gender, cardiorenal related indicators, and comorbidities based on the first model to evaluate the independent association between trajectory subtype and the risk of CKM progression.

[0051] When evaluating the risk discrimination capability within different time windows, a model containing only trajectory subtypes can be constructed based on the trajectory subtype labels output by the second model, or an extended model containing clinical covariates and trajectory subtypes can be constructed, and the area under the time-dependent ROC curve at a specified follow-up time point can be calculated. By comparing the basic model and the extended model with the addition of trajectory subtypes, the incremental value of the longitudinal joint trajectory subtype analysis of CHG and SII for stratifying the risk of CKM progression can be evaluated.

[0052] Referring to Table 2, a basic model without trajectory subtypes was also constructed, and the area under the time-dependent ROC curve, the net reclassification improvement index at a specific time point, and the net benefit of the decision curve were compared between the basic model and the first model. This was used to determine whether longitudinally combining trajectory subtypes of CHG and SII could further improve the risk stratification ability for CKM progression based on routine clinical information.

[0053] Table 2 Kaplan-Meier survival analysis showed a significant separation in the cumulative progression curves of the four longitudinal trajectory subtypes of CKM, with an overall log-rank test P < 0.0001, indicating that different trajectory subtypes have different CKM progression risks. The first model based on multivariate Cox regression showed that the subtypes had hazard ratio levels in the same direction as the survival curves (see Table 2). Figure 4Time-dependent ROC analysis showed that the genotyping results based on the combined CHG-SII longitudinal trajectory had good discriminative ability for CKM progression risk, with an AUC of 0.770 at 3 years, 0.711 at 5 years, and 0.676 at 8 years. Therefore, this invention, through the two-step LMM+GMM method, can identify four CKM longitudinal trajectory subtypes from 1270 CKM-related subjects, and the obtained subtypes can distinguish different CKM progression risks.

[0054] Step S9 employs two evaluation methods: Bootstrap resampling stability verification and high posterior probability subset stability verification.

[0055] Bootstrap resampling stability verification. Since GMM is an unsupervised clustering model, its results may be affected by the sample structure. Therefore, this example uses the Bootstrap resampling method to evaluate whether the LMM+GMM master model can still obtain similar trajectory structures and risk rankings under sample perturbation conditions.

[0056] Bootstrap resampling is performed on a per-subject basis, rather than on a per-record basis. Each time a subject is selected, all longitudinal records for that subject are retained to avoid disrupting the time-series structure within the same subject.

[0057] Specific steps: Using subjects as the sampling unit, sampling with replacement was performed from 1270 samples to generate a Bootstrap resampling dataset. In each Bootstrap resampling dataset, CHG and log SII were recalculated according to the method in Example 1, and follow-up time construction, time centering, and standardization were completed. LMMs for CHG and log SII were reconstructed separately, and the CHG random intercept, CHG random slope, log SII random intercept, and log SII random slope for each subject were extracted. The four longitudinal features were input into the GMM, and trajectory subtype identification was performed according to the master model settings. The above process was repeated, using 1000 Bootstrap resamplings as a stability verification setting. In this embodiment, K=4, determined by the master model, was used as the target subtype number. Four trajectory subtypes were obtained from the Bootstrap sample. The results obtained from the Bootstrap resampling were compared with the results of the second model to evaluate whether the trajectory structure and category ratio were stable. The trajectory subtype labels of each subject in the Bootstrap sample were merged with the CKM progression outcome to reconstruct the outcome analysis dataset. Subsequently, the same method was used to construct Kaplan-Meier analysis and a Cox proportional hazards model to evaluate the relationship between different trajectory subtypes and the risk of CKM progression in the Bootstrap sample. The risk ranking of different trajectory subtypes, the separation trend of the KM curves, and whether the direction of the Cox hazard ratio remained consistent with the master model were observed after resampling.

[0058] Bootstrap resampling verification shows that, under sample perturbation conditions, the LMM+GMM two-step method of the present invention can still repeatedly obtain CHG-SII longitudinal trajectory structures similar to the second model. The risk ranking of the main trajectory subtypes is generally consistent with the results of the second model, indicating that the trajectory subtypes obtained by the present invention do not completely depend on a single fixed sample division.

[0059] Stability verification of high posterior probability subsets. The GMM outputs the posterior probability of each subject belonging to each trajectory subtype. If the posterior probabilities of some subjects are close across different subtypes, then that subject belongs to the boundary sample. To verify whether the results of the second model are mainly driven by boundary samples, the stability analysis is performed only on subjects with high classification certainty by increasing the maximum posterior probability threshold.

[0060] The genotyping results of 1270 CKM-related subjects were used. Each subject had a maximum posterior probability Pmax, which is the highest posterior probability that they were assigned to the current trajectory subtype.

[0061] Specific steps: Based on the GMM model, calculate the posterior probability of each subject belonging to each CKM longitudinal trajectory subtype. Extract the maximum posterior probability Pmax for each subject. Set two high-posterior probability thresholds: Pmax ≥ 0.70 or 0.80, retain subjects meeting these conditions, forming two high-posterior probability deterministic subsets. In the two high-posterior probability subsets, retain the assigned trajectory subtype labels and reconstruct the CKM progression outcome analysis dataset. Subsequently, use the same Kaplan-Meier analysis and Cox proportional hazards model to evaluate the relationship between different trajectory subtypes and CKM progression risk in the high-posterior probability subsets, and whether the CHG-SII longitudinal characteristics of each subtype are consistent. Compare the CKM progression risk ranking of each trajectory subtype in the two high-deterministic subsets and compare it with the full-sample master model results.

[0062] Results: In the high deterministic subset with Pmax ≥ 0.70, such as Figure 6 , Figure 8 and Figure 10 Of the 758 subjects retained, the four CKM longitudinal trajectory subtypes were still preserved, and the main CHG-SII longitudinal features of each subtype remained consistent with the full-sample master model results. Figure 7 , Figure 9 and Figure 11 In the more stringent high-determinism subset with Pmax ≥ 0.80, 520 subjects were retained, and the four CKM longitudinal trajectory subtypes were still preserved, indicating that even after removing boundary samples with high classification uncertainty, the obtained trajectory structure did not disappear. Further comparison of the progression risk results shows that the main risk ranking in the high posterior probability subset is basically consistent with the results of the full-sample master model. This result indicates that the CKM longitudinal trajectory subtypes obtained in this invention are not driven by low-determinism boundary samples, but still have a stable trajectory structure and risk stratification trend in populations with high classification certainty.

[0063] Example 3 provides a prediction system for implementing the above prediction method, such as... Figure 12 As shown, it includes an input module 11, a subtype recognition module 12, a prediction module 13, and an output module 14; The input module 11 is used to collect the clinical test data of the subjects to obtain the first dataset; The subtype identification module 12 is used to calculate the posterior probability of the subject belonging to each trajectory subtype based on the first dataset and the clustering model based on the Gaussian mixture model; and to obtain the trajectory subtype to which the subject belongs by maximizing the posterior probability Pmax.

[0064] The prediction module 13 is used to predict the clinical test data of the subjects based on the prediction model to obtain the risk of CKM progression and / or survival probability.

[0065] Output module 14 is used to output or display the posterior probability, the trajectory subtype to which the subject belongs, the risk of CKM progression, and / or the survival probability.

[0066] The prediction system also includes a first training module and a second training module; The first training module is used to perform unsupervised clustering of feature vectors based on Gaussian mixture model to obtain clustering model and corresponding trajectory subtype; the second training module is used to train the training set with trajectory subtype as variable to obtain prediction model.

[0067] Example 4 provides a prediction device, including a processor and a memory. The memory stores code for implementing the above-described subtype construction method and prediction method. When the code is processed by the processor, the above-described subtype construction method and prediction method are implemented.

[0068] This invention transforms multi-timepoint detection data of CHG and SII into longitudinal horizontal features and longitudinal trend features, and further identifies longitudinal trajectory subtypes of CKM, achieving stratification of the dynamic heterogeneity of metabolism-inflammation in CKM-related populations. Compared with existing methods based on static metabolic phenotype clustering or static cardiorenal metabolic index risk prediction, this invention has the following advantages.

[0069] (1) It can identify CKM longitudinal trajectory subtypes with a certain degree of stability. Using the multi-time point detection results of CHG and SII, the long-term level and trend of individual-level changes are extracted through longitudinal modeling. It can not only reflect the metabolic or inflammatory state at a certain time point, but also reflect the process of the co-evolution of glucose and lipid metabolic load and systemic immune inflammatory state over time. In the research and development validation, based on the longitudinal data of 1270 CKM-related subjects, the number of candidate trajectory subtypes was set to K=2, 3, 4, 5, and 6, for a total of 5 candidate models. For each candidate K value, Gaussian mixture model was refitted based on the longitudinal features of CHG and SII, and BIC, Entropy, mean posterior probability, minimum class mean posterior probability, minimum correct classification advantage ratio, minimum class proportion, and the global association results between trajectory subtype and CKM progression outcome were calculated.

[0070] Statistical results of model fitting show that, under the adopted BIC output rule, the BIC is -11849.77 when K=4, which is the optimal value among the candidate models from K=2 to K=6. Compared with K=3, the BIC of K=4 is further improved; when the number of subtypes continues to increase to K=5 or K=6, the BIC does not improve further, and the proportion of the minimum category decreases to below 10%, and the mean posterior probability of the minimum category is also below 0.70, suggesting that further increasing the number of subtypes may lead to an undersized sample size for some categories and classification instability. Although the K=2 and K=3 models have a high category proportion and posterior probability, the genotyping granularity is relatively coarse and cannot fully reflect the heterogeneity in the combined longitudinal evolution of CHG and SII. The K=4 model achieves a better balance between BIC, mean posterior probability, mean posterior probability of the minimum category, minimum category proportion, and clinical interpretability. Finally, K=4 was determined as the number of CKM longitudinal trajectory subtypes, resulting in 4 CKM longitudinal trajectory subtypes. The sample sizes for the four subtypes were 450, 331, 346, and 143, respectively, with the smallest class proportion being approximately 11.26%. The mean posterior probability was approximately 0.7536, and the mean posterior probability for the smallest class was 0.7050, meeting the preset classification quality requirements (see Table 1).

[0071] In further progress risk stratification analysis, the trajectory subtypes obtained based on K=4 can distinguish different CKM progress risk levels through Kaplan-Meier analysis and Cox proportional hazards model, indicating that the CKM longitudinal trajectory subtypes obtained in this invention not only have statistical classification stability, but also have application value for CKM progress risk stratification.

[0072] (2) It can distinguish the progression risk of different CKM longitudinal trajectory subtypes: There is a correspondence between the identified CKM longitudinal trajectory subtypes and the subsequent CKM progression risk. Through survival analysis and risk model evaluation, different trajectory subtypes can show different cumulative progression risks. It can not only achieve population typing, but also be used for progression risk stratification. It is different from the method that only outputs static cluster labels, and also different from the traditional risk prediction method that only outputs risk probabilities based on baseline variables. In the research and development validation, the progression of CKM to the clinical cardiovascular disease stage or the occurrence of related clinical cardiovascular events was used as the outcome, and the time from baseline to the occurrence of the event or the last follow-up was used as the follow-up time. Kaplan-Meier survival analysis was used to compare the cumulative progression risk of different trajectory subtypes. The results showed that the cumulative progression curves of the four CKM longitudinal trajectory subtypes were significantly separated, and the overall log-rank test P<0.0001.

[0073] (3) Compared to single CHG or SII indicators, it has better risk discrimination ability: A single CHG or SII can only reflect the metabolic or inflammatory state at a certain detection time point, and is easily affected by the detection time point, short-term fluctuations, and intra-individual changes. This invention extracts longitudinal trajectory features from multi-time point data, which can retain information on the changes of CHG and SII over time. In the research and development verification, the same study population, the same definition of CKM progression outcome, and the same follow-up time window were used to compare the discrimination ability of the CHG-SII combined longitudinal trajectory subtype with single cross-sectional CHG and single cross-sectional SII for CKM progression risk. Time-dependent ROC analysis shows that, Figure 4 The genotyping results based on the combined CHG-SII longitudinal trajectory showed good discriminative ability in the 3-year, 5-year and 8-year follow-up time windows, with AUC of 0.770, 0.711 and 0.676 in 3-year, 5-year and 8-year respectively, all of which were higher than the single cross-sectional CHG or SII index.

[0074] (4) Incremental risk stratification information that can be used as a base model: The CKM longitudinal trajectory subtype obtained in this invention can also be used as a supplementary variable in downstream risk assessment to evaluate whether it can provide additional risk stratification information beyond the base model. In the 5-year CKM progress risk assessment, for the first model after adding the CKM longitudinal trajectory subtype, such as Figure 5 Net Reclassification Index (NRI) analysis showed that the first model had improved reclassification ability: non-event NRI was 0.461, 95% CI 0.406–0.527; event NRI was 0.280, 95% CI 0.147–0.411; and overall NRI was 0.741, 95% CI 0.593–0.893. Decision curve analysis showed that, within the low to medium risk threshold range, the model with added trajectory subtypes had a higher net benefit. This incremental value assessment is a downstream effect verification after trajectory subtype formation and does not change the construction process of the LMM+GMM principal subtyping model.

[0075] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing subtypes of CKM, characterized in that, Includes the following steps: Collect clinical test data; preprocess the clinical test data to obtain the first dataset; Based on the first dataset, calculate the natural logarithm of the cholesterol-high-density lipoprotein-glycemic index (CHG) and the systemic immune inflammation index (log SII). A first linear mixed-effects model and a second linear mixed-effects model were constructed based on CHG and log SII, respectively. The longitudinal horizontal features and longitudinal trend features of CHG were extracted from the intercept and slope of the first linear mixed-effects model. The longitudinal horizontal features and longitudinal trend features of log SII are extracted from the intercept and slope of the second linear mixed-effects model; Based on the vertical horizontal features of CHG, the vertical trend features of CHG, the vertical horizontal features of log SII, and the vertical trend features of log SII, a feature vector is constructed. Unsupervised clustering of the feature vectors is performed based on Gaussian mixture model to obtain the clustering model and the corresponding trajectory subtype.

2. The subtype construction method according to claim 1, characterized in that, Methods for obtaining subject trajectory subtypes include: Based on the clustering model and the clinical data of the subjects, the posterior probability of belonging to each trajectory subtype is calculated; By maximizing the posterior probability, the trajectory subtype to which the subject belongs can be obtained.

3. The subtype construction method according to claim 2, characterized in that, The trajectory subtype to which the subject belongs is obtained by maximizing the posterior probability and the posterior probability threshold. The total number of trajectory subtypes is four.

4. The subtype construction method according to claim 2, characterized in that, After standardizing CHG and log SII, the first linear mixed-effects model and the second linear mixed-effects model are constructed respectively.

5. The subtype construction method according to claim 1, characterized in that, The cholesterol-high-density lipoprotein-glycemic index is calculated as follows: CHG=ln[(TC×FBG) / (2×HDL_C)] (1); Among them, TC is total cholesterol, FBG is fasting blood glucose, and HDL_C is high-density lipoprotein cholesterol; The formula for calculating the systemic immune inflammatory index (SII) is as follows: SII=PLT×NEUT / LYMPH (2); Among them, PLT represents platelet count, NEUT represents neutrophil count, and LYMPH represents lymphocyte count.

6. A method for predicting CKM, characterized in that, Includes the following steps: Based on the trajectory subtype constructed by the subtype construction method as described in any one of claims 1-5, and the CKM progression outcome, a prediction model is constructed; Predictive models are used to forecast the clinical test data of subjects to obtain the risk of CKM progression and / or survival probability.

7. The prediction method according to claim 6, characterized in that, The predictive models include KM survival curves and / or a first model based on Cox proportional hazards.

8. The prediction method according to claim 6, characterized in that, The predictive model includes a second model based on Cox proportional hazards, and the covariates or independent variables of the second model also include age, sex, cardiorenal related indicators, and comorbidities.

9. A prediction system, characterized in that, For implementing the prediction method as described in any one of claims 6-8, the prediction system includes an input module, a subtype identification module, and a prediction module; The input module is used to collect the clinical test data of the subjects to obtain the first dataset; The subtype identification module is used to calculate the posterior probability of a subject belonging to each trajectory subtype based on the first dataset and a clustering model based on a Gaussian mixture model; by maximizing the posterior probability, the trajectory subtype to which the subject belongs is obtained. The prediction module is used to predict the risk of CKM progression and / or survival probability based on the clinical test data of the subjects according to the prediction model.

10. The prediction system according to claim 9, characterized in that, It also includes a first training module and a second training module; The first training module is used to perform unsupervised clustering of feature vectors based on Gaussian mixture models to obtain clustering models and corresponding trajectory subtypes; The second training module is used to train the training set with trajectory subtypes as variables to obtain a prediction model.

Images