2.1 Study population

The National Health and Nutrition Examination Survey (NHANES) is a multi-cycle, cross-sectional survey that integrates population-specific health questionnaires and physical examinations to assess the health and nutritional status of adults and children in the United States. The survey uses a complex, stratified, multistage probability sampling method, ensuring that the selected samples can effectively represent the entire US population.²² This study used data from three survey cycles from 1999 to 2004, involving the recruitment of 31,126 participants. Individuals under 18 years of age (n = 14,065), those who reported pregnancy (n = 715), those missing fasting blood sample data (n = 9890), those with fewer than two metabolic abnormality indicators (n = 3662), those missing cardiac and renal biomarker data (n = 423), and those missing death data (n = 2) were excluded, resulting in a final sample of 2369 participants (Supporting Information: Figure S1). The NHANES program has been approved by the Institutional Review Board of the National Center for Health Statistics (NCHS) and participants provided signed informed consent forms.

2.2 MetS

According to the criteria set forth by the American College of Endocrinology (ACE) and the American Association of Clinical Endocrinologists (AACE), MetS is considered present when three or more of the following five conditions exist²: (1) increased waist circumference: ≥88 cm in women and ≥102 cm in men; (2) elevated triglycerides: ≥150 mg/dL; (3) low high-density lipoprotein cholesterol (HDL-C): <40 mg/dL in men and <50 mg/dL in women; (4) elevated blood pressure: systolic blood pressure (SBP) ≥ 130 mmHg or diastolic blood pressure (DBP) ≥ 85 mmHg, or self-reported use of antihypertensive medication; and (5) elevated fasting plasma glucose (FPG): FPG ≥ 100 mg/dL or use of glucose-lowering drugs.

Clinical data for MetS indicators were obtained at the mobile examination centers (MEC). Blood samples were collected from participants who had fasted for ≥8 h to determine FPG, HDL-C, and serum triglyceride concentrations. Waist circumference and blood pressure were measured using standard methods. After the participant had rested in a quiet sitting position for 5 min, blood pressure was measured on the right arm using a mercury sphygmomanometer, with three consecutive readings taken. If needed, take the fourth reading.

2.3 Cardiorenal biomarkers

During the NHANES 1999–2004 survey cycles, measurements of Hs-cTnT, Hs-cTnI, NT-proBNP, β2M, CysC, and Cr were conducted on stored residual serum samples. The limit of detection (LOD) for Hs-cTnT was 3 ng/L, with a total imprecision expressed as a coefficient of variation between 2.0% and 3.1%. The LOD for Hs-cTnI was 1.6 ng/L, with coefficients of variation of 3.8% (12–28 ng/L), 2.7% (120–280 ng/L), and 2.6% (9000–21,000 ng/L). NT-proBNP (Roche Diagnostics) had detection limits of 5 and 35,000 pg/mL, with a coefficient of variation of 2.7%–3.1%. Serum β2M and CysC were measured on the automated multi-channel analyzer Siemens Dimension Vista 1500 (Siemens Healthcare Diagnostics). The detection limits for β2M were 0.72 and 23.0 mg/L, with coefficients of variation of 3.42%–3.88%. Cystatin C had detection limits of 0.23 and 8.00 mg/L, with coefficients of variation of 3.54%–4.36%. Serum creatinine was determined using the Jaffé rate reaction.

For the biomarkers NT-proBNP, β2M, and CysC, when the analysis results were below the lower LOD, they were imputed using the estimated value of the limit of detection divided by the square root of 2 ($$\text{LOD}/\sqrt{2}$$). For results above the upper LOD, impute with an estimated value equal to the upper LOD value. Overall, NT-proBNP measurements for 87 individuals were imputed as $$\text{LOD}/\sqrt{2}$$, β2M measurements for one individual each were imputed as $$\text{LOD}/\sqrt{2}$$ and upper LOD, respectively, and additionally, CysC measurements for 3 individuals were imputed as $$\text{LOD}/\sqrt{2}$$.

2.4 Covariates

Covariates were acquired through standardized questionnaires during interviews, including age, race, sex, marital status, education, poverty income ratio (PIR), smoking status, and drinking status. Marital status was categorized into married or living with partner, widowed or divorced or separated, and never married. Education was classified as less than high school (low), high school or equivalent (medium), and college graduate or above (high). PIR was divided into <1.0 and ≥1.0. Smoking status was categorized into never smoked, former smoker, and current smoker. Drinking status was divided into nondrinker, light drinker, moderate drinker, and heavy drinker. Height and weight were measured at the MEC, with body mass index (BMI) calculated as weight (kg) divided by height (m) squared. The determination of CVD history is made by answering the question, “Ever told had congestive heart failure/coronary heart disease/angina or angina pectoris/heart attack/stroke?” CKD is defined as an estimated glomerular filtration rate (eGFR) < 60 mL/min/1.73 m². The eGFR is calculated using the CKD-EPI 2021 equation based on serum Cr and CysC.²³ Diabetes is defined as plasma HbA1c ≥ 6.5% or FPG ≥ 126 mg/dL, self-reported previous diagnosis of diabetes, or use of insulin or oral hypoglycemic medication.²⁴ The diagnosis of cancer is based on the question, “Have you ever been told you had cancer or any type of malignant tumor?” Missing covariates were imputed using multiple imputation methods.

2.5 Ascertainment of mortality

All individuals included in the study were followed up. Mortality from all causes and CVD was determined by matching with the national death index (NDI) up to December 31, 2019. The primary causes of death were identified based on ICD-10 codes. CVD mortality was defined as deaths caused by heart diseases (I00–I09, I11, I13, I20–I51) and cerebrovascular diseases (I60–I69).²⁵

2.6 Statistical analysis

Baseline characteristics of participants were compared by outcome event occurrence. Continuous variables with a normal distribution are presented as means with standard deviations (SD), whereas those with a nonnormal distribution are presented as medians with interquartile ranges (IQR). Categorical variables are expressed in frequencies and percentages. Between-group differences for continuous variables were assessed using the Student's t-test or Wilcoxon rank–sum test, whereas the χ² test or Fisher's exact test for categorical variables.

Univariate and multivariate weighted Cox proportional hazards models were employed to assess the associations of Hs-cTnT (continuous and quartiles), Hs-cTnI (continuous and quartiles), NT-proBNP (continuous and quartiles), Cr (continuous and quartiles), β2M (continuous and quartiles), and CysC (continuous and quartiles) with mortality from MetS. Considering the skewed distribution of variables, log-transformed biomarkers were used for regression analysis. Three models were formulated: Model 1 without adjustment; Model 2 adjusted for age, sex, and race; Model 3 adjusted for age, sex, race, marital status, PIR, education, drinking status, smoking status, BMI, CVD, CKD, diabetes, and cancer. Hazard ratio (HR) and 95% confidence interval (CI) were used to assess the risk in the models. Additionally, restricted cubic spline (RCS) curves based on multivariate weighted Cox regression were plotted to visualize the linear or nonlinear relationships between biomarkers and all-cause mortality, and CVD mortality. We also performed subgroup analysis by potential confounders.

Five machine learning models, random forest (RF), support vector machine (SVM), categorical boosting (CatBoost), eXtreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM), were employed to predict mortality from MetS using cardiac and renal biomarkers. The data were split into a training set (70%, 1658 individuals) and a test set (30%, 711 individuals). Models performance metrics included area under the curve (AUC), accuracy, precision, recall, and F1 score for both the training and test sets. Based on the best machine learning model from the training and test sets, the relative importance of cardiac and renal biomarkers was determined using Shapley Additive exPlanation (SHAP) values. A beewarm plot is a more complex and informative representation based on SHAP values, which not only indicates the relative importance of features but also reveals their actual relationship with the predicted outcome. The top three biomarkers were selected for incorporating into the final predictive model, combined with confounding variables, to construct the optimal model for predicting the mortality risk from MetS.

All analyses were performed using R software, with a two-tailed test, and a p-value < 0.05 was considered statistically significant.

3.1 Baseline characteristics

In 2369 NHANES study population (weighted population 75,042,131), there were 1157 males and 1229 non-Hispanic Whites (Table 1). Among the MetS patients who experienced outcome events, 602 had elevated blood pressure and were more likely to have underlying conditions. Median and IQR levels of serum Cr by outcome event occurrence were 70.72 [61.90–88.40] µmol/L, 79.56 [61.90–97.24] µmol/L, respectively, with no significant differences (p > 0.05). Significant differences were observed in the median levels of Hs-cTnT, Hs-cTnI, NT-proBNP, CysC, and β2M across different groups (p < 0.001). Furthermore, there were no significant differences in Cr across different survey periods (Supporting Information: Table S1). At the same time, we compare the baseline characteristics of the included and excluded populations (Supporting Information: Table S2). The two groups showed no significant differences in terms of gender, educational level, and drink status (p > 0.05), indicating that the two groups are comparable in these aspects.

3.2 Associations between cardiorenal biomarkers and mortality risk

During the follow-up, a total of 774 (32.67%) participants died from various causes, including 260 (10.98%) CVD-related deaths. Table 2 shows the relationship between different cardiorenal biomarkers and all-cause mortality. After multivariable adjustment (Model 3), compared to the control group (Q1), the highest quartile of NT-proBNP, and β2M were associated with an increased risk of all-cause mortality by 106% (HR 2.06, 95% CI 1.50–2.83), 94% (HR 1.94, 95% CI 1.24–3.05), respectively. In addition, a similar association was found between cardiorenal biomarkers and CVD mortality (Table 3). After adjusting for potential confounders (Model 3), the HRs for the highest quartile of Hs-cTnI, NT-proBNP, and β2M compared with the control group (Q1) were 2.31 (95% CI 1.25–4.28), 2.86 (95% CI 1.26–6.48), and 2.80 (95% CI 1.09–7.18), respectively.

3.3 Dose–response relationship between cardiorenal biomarkers and mortality

We used RCS analysis to further explore the potential nonlinear associations between biomarkers and MetS mortality. The associations of biomarkers with all-cause mortality risk is visualized in Figure 1. Hs-cTnT, Hs-cTnI, Cr, and CysC showed significant nonlinear associations with all-cause mortality in MetS (p < 0.05 for nonlinearity), whereas NT-proBNP and β2M exhibited positive linear associations (p > 0.05 for nonlinearity). A nonlinear association was found only between Cr level and CVD mortality (p < 0.05 for nonlinearity, Supporting Information: Figure S2).

3.4 Subgroup analysis

In the subgroup analysis of all-cause mortality in patients with MetS and cardiorenal biomarkers (Supporting Information: Table S4), after multivariable adjustment, the risk of death in women aged ≥60 years increased with higher levels of Hs-cTnT, Hs-cTnI, NT-proBNP, and β2M (p < 0.05, HR > 1). Similar results were observed in the subgroup analysis of cardiovascular mortality (Supporting Information: Table S5). Furthermore, we applied Spearman's rank correlation to estimate the correlation between age and biomarkers (Supporting Information: Figure S3). All six biomarkers were positively correlated with age, with the correlation coefficient for NT-proBNP being 0.64 and for Cr being 0.18. Meanwhile, among patients with MetS without comorbidities, NT-proBNP was a risk factor for both all-cause and CVD mortality, whereas β2M and Hs-cTnI were risk factors for all-cause mortality and CVD mortality, respectively.

3.5 Performance comparison of machine learning models

To determine the optimal machine learning model for predicting mortality in MetS, the 2369 study subjects were randomly divided into a training set (70%, N = 1658) and a testing set (30%, N = 711). Through the comparison of the training set and the testing set, it is found that the baseline characteristics of the two groups of objects are comparable (Supporting Information: Table S3). Six biomarkers were included, Hs-cTnT, Hs-cTnI, NT-proBNP, Cr, CysC, and β2M, to construct five machine learning models, SVM, RF, XGBoost, LightGBM, and CatBoost. We utilized five evaluation metrics—AUC, accuracy, precision, recall, and F1 score—to assess the performance of these machine learning models in the training and testing cohorts. In the training set, the RF model exhibited the best predictive performance, with all evaluation metrics being 1. In the testing set, the CatBoost model demonstrated the highest AUC (0.862), accuracy (0.805), precision (0.832), recall (0.892), and F1 (0.861), followed by the SVM (0.858), RF (0.853), XGBoost (0.847), and LightGBM (0.828) models (Supporting Information: Table S6). Figure 2A shows the AUC value with 95% CI for five machine learning models, and Figure 2B,C shows the receiver operating characteristic (ROC) curves for RF and CatBoost, respectively.

3.6 Global and local explanations of the machine learning models

To explain how machine learning models predict mortality in MetS, we aimed to elucidate the impact of each biomarker on the predictive models through variable importance and SHAP values. Ranking the variable importance based on the RF and CatBoost models, Hs-cTnT, NT-proBNP, and β2M emerged as the top three important variables in both models (Figure 3A–C). Figure 3D–F displays the dependency plots for the top three features determined by average absolute SHAP values, revealing similarities in the relationship between SHAP values and variable values across features.

3.7 CatBoost model for the combined assessment of biomarkers in predicting all-cause mortality

In the original model, the addition of Hs-cTnT, NT-proBNP, and β2M individually resulted in AUC of 0.901, 0.903, and 0.903, respectively (Supporting Information: Figure S4). The combinations of biomarkers, Hs-cTnT+NT-proBNP, Hs-cTnT+β2M, and NT-proBNP+β2M, showed minor improvements in model performance, with AUCs of 0.903, 0.904, and 0.904, respectively. Combining all three biomarkers, the CatBoost model's predictive performance was higher than that of the individual biomarkers and was significantly better than the original model, with an AUC of 0.904.