Advanced prediction of cardiovascular-kidney-metabolic syndrome using eight machine learning models and 24 composite indices

Abstract

Aim

This study aimed to construct and validate a machine learning classifier for cross-sectionally stratifying existing cardiovascular-kidney-metabolic (CKM) syndrome stages using routine composite inflammatory, metabolic and anthropometric indices, and to interpret core driving biomarkers via SHAP analysis.

Methods and results

We analyzed data from 12,106 participants from the National Health and Nutrition Examination Survey (NHANES). Among 24 initial biomarkers, 10 were selected after addressing multicollinearity and applying feature selection. Several machine learning algorithms were evaluated, with the LightGBM model demonstrating the highest performance (ROC AUC: 0.88). External validation using the China Health and Retirement Longitudinal Study (CHARLS) dataset confirmed the model’s generalizability (ROC AUC: 0.84). SHapley Additive exPlanations (SHAP) analysis revealed that metabolic markers—specifically eGDR, METS_VF, TyG, and TyG_BMI—were consistently the strongest predictors across CKM stages, whereas inflammatory indicators showed more limited utility.

Conclusion

Key metabolic composite indices exhibit significant associations with CKM staging and may serve as practical, clinically feasible tools for risk stratification. The model’s robust performance across distinct populations (U.S. and China) supports its potential clinical utility. Further validation in diverse populations and prospective studies is needed to confirm their predictive value and translational potential.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12872-026-06318-2.

Lay summary

This study developed an easy-to-use clinical stratification tool that categorizes patients into distinct CKM syndrome severity stages using routine blood biomarkers and anthropometric measurements, assisting clinicians in evaluating concurrent cardiometabolic, renal, and cardiovascular burden.

Key Findings:

• Metabolic markers (like eGDR and TyG) were the strongest predictors of CKM stages, while inflammatory markers (like PLR and NHHR) played a smaller role in early disease stages.

• A machine learning model accurately identified high-risk patients (88% accuracy) using just 10 key measurements, which can be easily obtained during routine check-ups.

This tool could help doctors detect CKM stages earlier and personalize prevention strategies.

Introduction

Cardiovascular-Kidney-Metabolic Syndrome (CKM), proposed by the American Heart Association (AHA), is a systemic and progressive health disorder characterized by the coexistence of obesity, diabetes mellitus (DM), chronic kidney disease (CKD), and cardiovascular disease (CVD). It has now become a major global public health challenge [1]. A cross-sectional study on the prevalence of CKM staging among U.S. adults revealed that nearly 90% of American adults meet the criteria for CKM syndrome (stage 1 or higher), with 15% meeting the criteria for advanced stages. From 2011 to 2020, there was no improvement in either of these conditions [2]. The multisystem interplay characteristic of CKM syndrome - encompassing metabolic dysregulation, chronic inflammation, and vascular-renal impairments - renders its clinical management particularly challenging [1]. Composite biomarkers have garnered widespread attention due to their ability to integrate multidimensional pathological information and have demonstrated strong performance in CKM staging. However, given the multisystem nature of CKM, these individual biomarkers are limited in risk assessment, primarily predicting cardiovascular events in patients at CKM stages 0–3, and fail to comprehensively reflect the systemic pathophysiological changes of the disease [3–5]. Therefore, developing predictive tools that integrate multidimensional indicators—such as metabolic, inflammatory, and body composition markers—holds significant importance for achieving precise CKM staging and early intervention.

Machine learning (ML) provides a transformative approach to address current limitations in CKM syndrome stratification by uncovering complex clinical patterns that traditional statistical models often miss. Unlike conventional regression-based methods, ML algorithms excel at analyzing high-dimensional datasets and capturing intricate nonlinear relationships - capabilities that prove essential for predicting multifactorial diseases like CKM where metabolic, inflammatory, and body composition indicators interact through complex pathways [6, 7].

Building upon these considerations, our study leveraged machine learning approaches to develop a predictive model for cardiovascular-kidney-metabolic (CKM) syndrome staging using cross-sectional data from the National Health and Nutrition Examination Survey (NHANES) spanning 2005 to 2018. The investigation comprehensively evaluated 24 composite biomarkers representing three fundamental pathophysiological domains: metabolic dysfunction, chronic inflammation, and altered body composition. The metabolic indicators included clinically validated measures such as the Atherogenic Index of Plasma (AIP), Non-HDL to HDL Cholesterol Ratio (NHHR), and various insulin resistance indices including the Triglyceride-Glucose Index (TyG) and its anthropometric-adjusted derivatives (TyG-WC, TyG-BMI, TyG-WHtR). These were complemented by inflammatory markers like the Systemic Immune-Inflammation Index (SII) and Neutrophil-to-Lymphocyte Ratio (NLR), along with body composition parameters such as the Visceral Adiposity Index (VAI) and Waist-to-Height Ratio (WHtR), all of which have demonstrated significant associations with cardiometabolic diseases in prior research [8–19].

Through the systematic application of diverse machine learning algorithms, we rigorously evaluated the ability of these integrated biomarkers to discriminate between CKM stages. Our goal was to develop a clinically practical yet robust risk classification tool that could identify high-risk individuals in early disease stages, which may support timely interventions and personalized management strategies. This approach addresses critical gaps in current diagnostic paradigms by capturing the complex interplay among metabolic, inflammatory, and structural components in the progression of CKM syndrome.

Methods

Study population in NHANES

The data used in the current analysis are publicly available through the NHANES database (https://www.cdc.gov/nchs/nhanes/index.htm). The protocols for the NHANES study were authorized by the NCHS Research Ethics Review Board. Informed consent was obtained from all NHANES participants. This study did not require institutional review board approval as it utilized de-identified, publicly available data. This study encompasses data from seven NHANES cycles between 2005 and 2018, encompassing a total of 70,190 subjects. We first excluded participants with missing cardiovascular outcome events (N = 30,422), those who did not meet the PREVENT equation criteria (N = 13,523), and those with missing raw data required for composite metrics (N = 16,462). Ultimately, data from 12,106 participants were included in the analysis.

The China Health and Retirement Longitudinal Study (CHARLS) collects high-quality microdata from Chinese households and individuals aged 45 and above to study population aging. To date, four waves of data have been released: the baseline (2011) and three follow-ups (2013, 2015, 2018). Detailed study information is available in prior publications¹⁹, and data can be accessed at http://charls.pku.edu.cn/en. The study was approved by the Peking University Biomedical Ethics Committee, and all participants provided informed consent. For the external test set, data from the 2015 wave (Wave 3; initial N = 21,097) were used. After applying PREVENT equation criteria and removing cases with missing raw data required for composite metrics, 10,540 participants were included.

Definition of CKM syndrome staging outcome

In this study, cardiovascular disease (CVD) was defined based on self-reported conditions in NHANES participants. Covariates included diabetes mellitus (DM), hypertension (HBP), and chronic kidney disease (CKD), defined by clinical history, medication use, or laboratory criteria. According to the American Heart Association (AHA) criteria [20], we stratified cardiovascular-kidney-metabolic (CKM) syndrome into stages 0–4 based on available NHANES data, incorporating risk factors, metabolic abnormalities, chronic kidney disease (CKD), and presence of cardiovascular damage or events. The detailed diagnostic definitions and grouping criteria are provided in Supplementary Table S1.

Calculation of 24 composite biomarkers

In this study, we focused on 24 clinically practical composite metrics related to lipid metabolism, glucose metabolism, body composition assessment and inflammation/immunity. All composite metrics were directly calculated using biomarkers from the NHANES database, including white blood cell count, neutrophil count, lymphocyte count, monocyte count, platelet count, neutrophil percentage, albumin, globulin, total cholesterol, triglycerides, HDL cholesterol, serum creatinine, glycated hemoglobin A1c (HbA1c), uric acid, and urine albumin-to-creatinine ratio. Detailed formulas are provided in Supplementary Table S2 [8–19].

Data preprocessing, statistical analysis and machine learning pipeline

All data preprocessing and statistical analyses were conducted using R statistical software (version 4.4.3). The raw missing proportion of each composite biomarker in the NHANES cohort ranged from 0.8% to 11.7%, as summarized in Supplementary Table S2. Missing data were imputed using the missRanger package’s random forest algorithm, which first initialized missing values with variable-specific means or modes, then iteratively predicted missing values using other observed variables through chained random forests, finally applying predictive mean matching (PMM) to continuous variables to preserve their empirical distributions [21].

The dataset for this study initially consisted of 24 features. To reduce multicollinearity among all features, correlation coefficients were calculated, and features with coefficients exceeding 0.9 were removed. Synthetic Minority Over-sampling Technique (SMOTE) with k-nearest neighbor set to 5 was solely applied to the training subset to avoid data leakage, generating synthetic samples through linear interpolation within minority class neighborhoods [22]. Extreme values were treated using Winsorization [23]. Processed data were randomly split into training (80%) and testing (20%) sets. These procedures were implemented using dplyr and tableone packages. Recursive feature elimination (RFE) was performed as a two-step feature screening workflow. First, preliminary sensitivity tests across feature subsets (5, 10, 15, full collinearity-excluded features) were conducted to compare model accuracy and Kappa values. The 5-feature subset represented the minimal biomarker panel with acceptable baseline performance, yet the 10-feature panel yielded a significant + 9.6% accuracy improvement with only marginal additional feature redundancy (< 1.4% performance gain if further adding variables). We therefore selected the 10-feature subset as the primary final model for all internal discrimination, SHAP interpretation, and performance benchmarking. The 5-feature minimal panel was only retained for auxiliary sensitivity analysis and reported in Supplementary Table S6 for reference, and all main results (AUC, calibration, SHAP, internal test set performance) presented in the manuscript correspond exclusively to the 10-feature optimized model [24].

Survey-weighted statistical models were employed to characterize participants across CKM stages while accounting for the complex sampling design. Continuous variables are presented as weighted means ± standard errors, with categorical variables expressed as weighted frequencies and percentages. Statistical comparisons between CKM stages incorporated appropriate weighted analyses: weighted chi-square tests for categorical variables, survey-weighted analysis of variance (ANOVA) for normally distributed continuous variables, and weighted Kruskal-Wallis H tests for non-normally distributed continuous variables. Bonferroni correction was applied for all intergroup multiple comparisons to control familywise Type I error.

Eight machine learning algorithms were implemented: Generalized Linear Models: Multinomial logistic regression (multinom), Elastic Net (enet), Tree-based Methods: Decision Tree (dt), Random Forest (rf), XGBoost (xgboost), LightGBM (lightgbm), Neural Networks: Multilayer Perceptron (mlp), Distance-based: k-Nearest Neighbors (knn). All algorithms underwent 5-fold cross-validated grid-search hyperparameter tuning via the caret package; for LightGBM, tuned hyperparameters covered learning rate (0.01–0.1), max tree depth (3–8), number of leaves (10–50), and subsample ratio (0.7–0.9). Performance was comprehensively assessed using 13 metrics: Discrimination: ROC AUC, PR AUC; Calibration: Accuracy, Balanced Accuracy, Kappa, MCC; Clinical Utility: Sensitivity, Specificity, PPV, NPV, Youden’s J-index; Precision-Recall: F1-score, Precision, Recall. Benchmarking procedures adhered to TRIPOD guidelines for predictive model validation [7, 25–27]. Decision curve analysis (DCA) was further performed for the optimal LightGBM model to quantify net clinical benefit across different risk thresholds (Figure S13). The iml package is adopted to conduct partial dependency analysis on the features of the optimal model.

To address the low detection sensitivity of advanced Stage 3 and Stage 4 CKM patients under the default maximum-probability classification rule, we performed supplementary post-hoc stage-specific threshold optimization without refitting the finalized LightGBM model. Based on the class-wise predicted probabilities (.pred_0 to .pred_4) output from the test set, we decomposed the five-class CKM stratification task into five independent one-versus-rest (OvR) binary discrimination sub-tasks. For each individual CKM stage, a standalone ROC curve was constructed, and the optimal probability cutoff was determined by maximizing Youden’s J statistic (sensitivity + specificity − 1). A hierarchical classification logic was adopted to prioritize identification of high-risk patients: samples were sequentially judged against Stage 4, Stage 3, Stage 2, Stage 1, and stage-specific thresholds. Predictive performance metrics (overall accuracy, Kappa coefficient, stage-wise sensitivity/specificity, and confusion matrices) under both the default argmax rule and tailored stage thresholds were calculated and summarized in Supplementary Table S11 and Figure S12, following established multi-class threshold calibration frameworks and TRIPOD-AI reporting recommendations.

To assess the generalizability of our model across diverse populations, we performed an external validation using data from the China Health and Retirement Longitudinal Study (CHARLS) 2015 wave. The distribution of CKM stages in the CHARLS cohort is presented in Figure S11. The original 10-feature full LightGBM model required neutrophil count and serum albumin to calculate NPAR and AGR, which are not measured in the CHARLS 2015 wave dataset. We extracted seven shared biomarkers consistently available across both NHANES and CHARLS cohorts: eGDR, METS_VF, TyG, TyG_BMI, SHR, WHtR, NHHR. We retrained a reduced 7-feature LightGBM model on the NHANES training set first, benchmarked its performance against the original 10-feature model on the held-out NHANES internal test set for direct comparison (Supplementary Table S10), then applied this identical 7-feature architecture to the full CHARLS cohort for external discrimination testing. Supplementary Table S10 reports side-by-side performance metrics of the 10-feature full model (NHANES internal test) and 7-feature simplified model (NHANES internal test + CHARLS external validation), including AUC, accuracy, sensitivity, specificity and F1-score to quantify performance loss caused by feature reduction.

The model interpretation was performed using SHapley Additive exPlanations (SHAP) values computed via TreeExplainer, which provides a unified framework for both global and local interpretability. This approach quantifies global feature importance through mean absolute SHAP values while simultaneously enabling detailed visualization of feature effects through summary plots (beeswarm plots) that reveal value distributions and directional impacts. For clinical translation, we examined individual predictions using force plots to demonstrate how specific features contribute to particular risk stratifications, and identified potential biological interactions through dependence plots that capture nonlinear relationships. This multi-modal interpretation strategy synergistically combines game-theoretic principles with local explanation aggregation, offering comprehensive insights that span from population-level feature importance rankings to prediction-specific mechanistic explanations. The SHAP framework demonstrates superior theoretical foundations compared to conventional feature importance methods by rigorously maintaining both consistency (features reflecting true impact receive appropriate attribution) and local accuracy (explanations precisely match individual model outputs) properties throughout all analytical dimensions [7, 28].

Results

Characteristics of the features

The study cohort comprised 12,106 participants stratified by CKM stage, with a prevalence of: Stage 0: 6.9% (n = 830), Stage 1: 23.8% (n = 2,884), Stage 2: 40.4% (n = 4,895), Stage 3: 8.4% (n = 1,016), and Stage 4: 20.5% (n = 2,481). The cohort was then randomly split into a training set (n = 9,683) and a test set (n = 2,423). Table S3 presents the prevalence of CKM stages within the training and testing sets. All measured parameters showed significant intergroup differences (P < 0.001). Demographic characteristics revealed an age gradient from Stage 0 (44.9 ± 11.5 years) to Stage 3 (71.4 ± 6.3 years), with a modest decrease in Stage 4 (63.6 ± 10.8 years). Comprehensive baseline characteristics including gender, race/ethnicity, education level, marital status, smoking history, medication use, comorbidities, anthropometric measures, and laboratory findings are detailed in Table S4.

Development and validation of the comorbidity disease prediction model

Analysis of 24 composite metabolic and inflammatory markers demonstrated progressive deterioration across CKM stages (all P < 0.001). Atherogenic profiles worsened significantly, with the atherogenic index of plasma (AIP) increasing from 0.03 ± 0.21 (Stage 0) to 0.42 ± 0.31 (Stage 4) and non-HDL/HDL cholesterol ratio (NHHR) rising from 2.03 ± 0.75 to 3.04 ± 1.35. Insulin resistance markers showed particularly dramatic changes: estimated glucose disposal rate (eGDR) declined from 10.93 ± 0.64 to 5.77 ± 2.19 mg·kg-1·min-1, while triglyceride-glucose index (TyG) and its derivatives (TyG_WC, TyG_BMI, TyG_WHtR) exhibited consistent increases. Visceral adiposity measures demonstrated similar progression, with waist-to-height ratio (WHtR) increasing from 0.49 ± 0.04 to 0.63 ± 0.08, visceral fat metabolic score (METS-VF) from 5.58 ± 0.46 to 6.69 ± 0.37, and visceral adiposity index (VAI) from 1.94 ± 0.95 to 5.77 ± 5.97. Inflammatory markers showed parallel elevation: systemic immune-inflammation index (SII) increased from 485.36 ± 269.10 to 573.25 ± 404.01, systemic inflammation response index (SIRI) from 0.99 ± 0.69 to 1.48 ± 1.21, and pan-immune-inflammation value (PIV) from 235.01 ± 180.88 to 354.56 ± 360.95. Neutrophil-related ratios (NPAR, NLR) and platelet-lymphocyte ratio (PLR) followed comparable upward trajectories (Table S5).

Interparameter correlation analysis of the 24 composite biomarkers revealed significant associations through Spearman’s rank correlation. The heatmap (Fig. 1) visualization, with color intensity representing correlation coefficients (ranging from − 1 to + 1), identified five highly correlated features (r > 0.9): TyG_WC, TyG_WHtR, METS-IR, VAI, and SIRI. These redundant parameters were subsequently excluded from downstream modeling to mitigate multicollinearity effects while retaining their representative counterparts with stronger clinical interpretability. RFE revealed optimal model performance with 10 features (accuracy: 76.8% ± 0.4%; Kappa: 0.709 ± 0.005), showing significant improvement over 5-feature models (accuracy + 9.6%). Further feature additions yielded marginal gains (< 1.4%), indicating redundancy. The final 10-feature set balanced predictive power with parsimony (Table S6) (Fig. 2).

Fig. 1.

Heatmap of Spearman’s rank correlation analysis for 24 composite biomarkers

Fig. 2.

ROC curves across CKM stages (Stage 0–4) for all algorithms

The predictive performance of all eight machine learning models is comprehensively detailed in Table 1. While LightGBM achieved the highest numerical AUC (0.884), its overall advantage over other top-performing models was marginal. The multinomial regression model demonstrated strong and competitive performance (AUC = 0.861), and Random Forest achieved the highest scores in accuracy (0.628) and F1-score (0.580). The calibration assessment results demonstrated that all models exhibited good calibration performance, with LightGBM (Brier score = 0.0997), XGBoost (0.100), and random forest (0.101) showing the best performance. Calibration curves (Figure S2) further confirmed that these models maintained high consistency between predicted probabilities and actual frequencies across all probability ranges, particularly showing optimal calibration in the low-probability range with the largest sample size, indicating that the models can provide reliable probability predictions for clinical decision support.

Table 1.

Performance Metrics across different feature subset sizes (5-feature minimal panel vs. 10-feature primary final model; all main text results rely on the 10-feature model)

Model	Accuracy	Kappa	Sensitivity	Specificity	PPV	NPV	MCC	J-index	Balanced Accuracy	Precision	Recall	F1-score	ROC AUC	standard error	ECE	Brier score
multinom	0.546	0.4	0.566	0.883	0.494	0.88	0.406	0.45	0.725	0.2	0.494	0.566	0.855	0.0008918597	0.0612	0.114
enet	0.549	0.403	0.566	0.884	0.494	0.881	0.409	0.45	0.725	0.2	0.494	0.566	0.855	0.0008999775	0.0617	0.115
dt	0.59	0.448	0.578	0.893	0.533	0.892	0.454	0.47	0.735	0.2	0.533	0.578	0.853	0.0014547045	0.0609	0.110
rf	0.599	0.463	0.6	0.896	0.539	0.893	0.467	0.496	0.748	0.2	0.539	0.6	0.869	0.0011224484	0.0745	0.110
xgboost	0.61	0.477	0.605	0.899	0.544	0.897	0.482	0.504	0.752	0.2	0.544	0.605	0.877	0.0012559987	0.0553	0.102
mlp	0.606	0.475	0.615	0.899	0.541	0.896	0.48	0.514	0.757	0.2	0.541	0.615	0.876	0.0011927299	0.153	0.134
lightgbm	0.609	0.478	0.609	0.9	0.546	0.897	0.482	0.509	0.754	0.2	0.546	0.609	0.878	0.0007650322	0.0537	0.101
knn	0.584	0.452	0.593	0.895	0.527	0.89	0.458	0.488	0.744	0.2	0.527	0.593	0.848	0.0014307562	0.0716	0.111
psvm	0.521	0.376	0.56	0.879	0.507	0.876	0.391	0.44	0.72	0.2	0.507	0.56	0.851	0.0008594538	0.0783	0.119

PPV Positive Predictive Value, NPV Negative Predictive Value, MCC Matthews Correlation Coefficient, ECE Expected Calibration Error

Our selection of LightGBM for subsequent interpretability analysis was therefore not based solely on a negligible AUC improvement. Instead, the decision was driven by its well-balanced and robust performance across the majority of key metrics (e.g., high specificity of 0.901, balanced accuracy of 0.759, and competitive F1-score of 0.566), combined with its superior computational efficiency. This combination of high predictive performance, metric balance, and practical efficiency made it the most suitable candidate for applying resource-intensive explainability techniques like SHAP analysis to our large dataset. Given that the LightGBM model did not demonstrate statistically significant superiority over other models, we present the complete chi-square values for all variables across all CKM stages from the multinomial logistic regression model in Table S7 to enhance the credibility and robustness of our research findings.

The external validation results on the CHARLS cohort are summarized in Table S10. Despite the significant differences in population characteristics and the reduction in features, the adapted model maintained strong discriminatory power, achieving a macro ROC AUC of 0.843 (95% CI: 0.827–0.859). This represents only a slight decrease from its performance on the NHANES test set (0.882), confirming the model’s robustness. Notably, several key metrics, including Accuracy (0.648 vs. 0.615) and Specificity (0.912 vs. 0.900), showed modest improvements on the external dataset, suggesting that the model’s performance in identifying true negatives may be particularly robust across populations. Other metrics, such as macro Sensitivity (0.574) and F1-score (0.552), remained consistent with the internal validation results, indicating a balanced performance across all CKM stages.

The permutation-based feature importance analysis revealed substantial heterogeneity in the predictive contributions of different variables for stage classification. As visually represented in Figure S3, METS_VF emerged as the most critical predictor with an importance score of 1.813 (95% CI: 1.799–1.829), closely followed by eGDR at 1.811 (95% CI: 1.804–1.819). The dot plot clearly illustrates the superior predictive power of these metabolic indicators, which substantially outperformed other features as reflected in their permutation errors of 0.543 for both variables. The visualization further demonstrates that TyG-related indices, including TyG_BMI (1.651, 95% CI: 1.630–1.655) and TyG itself (1.579, 95% CI: 1.578–1.601), formed a distinct secondary cluster of predictors. Figure S3 effectively captures the gradual decline in importance scores, showing WHtR (1.258, 95% CI: 1.251–1.262) occupying an intermediate position, while inflammatory markers such as PLR (1.047) and AGR (1.037) appear grouped in the lower importance range. The graphical representation confirms the hierarchical organization of predictor importance, with NPAR (1.054), SHR (1.037), and NHHR (1.027) forming the least influential cluster, while maintaining non-overlapping confidence intervals that underscore the statistical significance of these observed differences. FigureS4-8 illustrates the partial dependence patterns of all clinical features in relation to CKM staging. Each subplot depicts how variations in a specific feature influence the predicted probability of each CKM stage while marginalizing over the distributions of other features. All primary predictive results, SHAP interpretability analyses, and internal validation metrics reported in the main text were generated from the 10-feature optimized LightGBM model; the 5-feature subset was only analyzed for auxiliary sensitivity comparison and does not constitute the core prediction tool of this study.

SHapley additive exPlanations analysis

To gain deeper insights into the predictive capabilities of the LightGBM model, we employed SHAP values to elucidate the contribution of each feature to CKM stage prediction, thereby identifying the most critical factors influencing model decisions. The importance of each feature and its impact on individual predictions were visualized through stage-specific SHAP summary plots (Fig. 3A-E) and feature importance plots (Figure S9). Furthermore, the mean absolute SHAP values, which provide a clear measure of global feature importance in the LightGBM model, are presented in Supplementary Table S8.

Fig. 3.

SHAP summary plots illustrating the contribution distributions of key features across CKM stages (Stage 0–4) (A) Stage 0; (B) Stage 1; (C) Stage 2; (D) Stage 3; (E) Stage 4

The SHAP summary plots (Figs. 3A-E) revealed that eGDR, METS_VF, TyG, and TyG_BMI exerted substantial influence on model outputs, with wide SHAP value ranges indicating heterogeneous effects across individuals. As shown in Figure S3 (Supplementary Materials), these features also demonstrated the highest mean SHAP values, reflecting their dominant average impact on predictions. However, their predictive contributions varied by disease stage: eGDR was consistently influential across all stages except Stage 3, METS_VF showed stronger effects in Stages 2–4, TyG primarily contributed to Stages 1–2, and TyG_BMI maintained robust importance throughout all stages.

To further dissect individualized predictions, we generated SHAP force plots for each stage (Figs. 4A-E), which visually represent how specific features contributed to single predictions. These plots provided granular insights into the contextual role of each biomarker across diverse samples. Collectively, the SHAP analysis offered a comprehensive understanding of the LightGBM decision-making process, pinpointing key predictive factors critical for model refinement and clinical interpretation.

Fig. 4.

SHAP force plots dissecting individualized prediction drivers across CKM stages (Stage 0–4) (A) Stage 0; (B) Stage 1; (C) Stage 2; (D) Stage 3; (E) Stage 4

The identified feature importance patterns not only validate the model’s biological plausibility but also highlight their potential utility in risk stratification. These findings enable a more holistic health assessment and provide a scientific foundation for personalized prevention strategies.

Stage-specific threshold sensitivity analysis

We conducted one-versus-rest ROC decomposition to derive independent optimal probability cutoffs for each CKM stage to mitigate the underdetection of advanced patients under standard maximum-probability classification. As listed in Supplementary Table S12, the stage-tailored optimal thresholds ranged narrowly from 0.1831 to 0.2015 across all five strata, far lower than the implicit variable cutoffs generated by the default argmax assignment rule. Each stage achieved perfect specificity (1.00) at its optimized threshold, with single-stage sensitivity exceeding 81% for all subgroups (Stage 0: 96.30%, Stage 1: 90.97%, Stage 2: 86.58%, Stage 3: 81.91%, Stage 4: 83.03%).

We further compared global predictive performance and confusion patterns between the default classification strategy and hierarchical stage-threshold judgment (Supplementary Figure S10 and Figure S12). Under the original maximum-probability rule, the overall test-set accuracy reached 61.49%, yet severe misclassification of Stage 3 and Stage 4 patients was observed: only 90 out of 199 true Stage 3 patients and 170 out of 495 true Stage 4 patients were correctly identified, with a large proportion of high-risk subjects misassigned to moderate Stage 2. After applying hierarchical stage-specific thresholds prioritizing severe CKM cases, the number of correctly classified Stage 4 patients rose sharply to 411, and nearly all Stage 3 individuals were flagged as high-risk (classified as Stage 3 or Stage 4). This marked improvement in capturing advanced cardiometabolic patients was accompanied by a moderate decline in overall accuracy to 53.32%. Near-zero Kappa coefficients (− 0.0083 and − 0.0129 for the default and threshold-tailored rules, respectively) were observed for both analytical pipelines, driven by severe inherent class imbalance in the test cohort dominated by Stage 2 participants rather than flawed model architecture or threshold calibration.

Discussion

This study analyzed data from 12,106 participants and revealed significant associations between cardiovascular-kidney-metabolic (CKM) syndrome stages and a panel of metabolic-inflammatory biomarkers (all P < 0.001). After removing highly collinear features and performing recursive feature selection, ten key biomarkers were identified for modeling: eGDR, METS_VF, TyG, TyG_BMI, SHR, NPAR, AGR, PLR, WHtR, and NHHR. These clinically practical indicators can be easily derived from routine clinical and laboratory measures, providing healthcare professionals with an accessible tool for CKM stage discrimination.

A major strength of this study is the external validation of the model in a geographically and ethnically distinct population. Despite substantial demographic and clinical differences from the original NHANES cohort, the model retained high discriminative performance (ROC AUC: 0.843) in the CHARLS cohort, supporting its generalizability. This suggests that the model captures core pathophysiological pathways of CKM syndrome—primarily related to insulin resistance, visceral adiposity, and cardiovascular risk—that are consistent across diverse groups. Although adaptation was required to incorporate seven commonly available clinical features, this adjustment highlights the model’s applicability in resource-limited settings where advanced testing is unavailable. The robust performance of this simplified version confirms that readily obtainable metrics are strongly predictive of CKM staging.

Notably, feature importance analyses from both the LightGBM model and multinomial logistic regression showed substantial concordance, reinforcing the robustness of our key findings. Both methods consistently identified metabolic indices—eGDR, TyG, METS_VF, and TyG_BMI—as the top-ranking predictors, while inflammatory and ratio-based markers (such as PLR, NHHR, and AGR) exhibited more modest effects. This agreement across fundamentally different modeling approaches provides strong evidence that metabolic dysfunction is a central driver of CKM staging.

In particular, eGDR demonstrated strong predictive performance for CKM progression, corroborating previous reports of its utility in predicting cardiovascular events among individuals in CKM stages 0–3 [29]. A cross-sectional study further supports this, linking higher eGDR levels to reduced risks of cardiovascular disease and mortality, which aligns with the trends observed in our SHAP analysis and partial dependence plots (PDPs). As shown in Fig. 3 and Figures S3–S7, higher eGDR values were associated with a lower likelihood of advanced CKM stages (2–4) and a higher probability of earlier stages (0–1) [30]. It is also noteworthy that for stages 0–1—reflecting normal cardiometabolic status—the predicted probability increased sharply from low to moderate eGDR values but declined slightly beyond a certain high eGDR threshold.

Furthermore, METS-VF demonstrated robust predictive performance, with higher levels strongly associated with advanced CKM stages (3–4), while lower values were primarily linked to earlier stages (1–2). These findings align with previous cross-sectional studies that reported similar associations between METS-VF and cardiometabolic risk stratification [31, 32]. Partial dependence plots further revealed a strong non-linear and inverse relationship between METS-VF and the predicted probability of CKM stage 1. Specifically, the predicted risk decreased sharply across lower METS-VF values, with a pronounced saturation effect observed beyond approximately 6.2, beyond which further increases conferred minimal additional reduction in risk.

TyG showed a predictive trend similar to that of METS-VF. In contrast, the derivative index TyG_BMI exhibited an inverse relationship: higher TyG_BMI levels were associated with CKM stages 1–2, while lower values correlated with stages 3–4. This finding diverges from existing cross-sectional studies, which consistently report a positive association between elevated TyG_BMI and cardiovascular event risk [33].

Inflammatory markers such as SHR, NPAR, AGR, PLR, and NHHR contributed relatively little to early CKM stages prediction in our machine learning models. This may be explained by several factors. In the early CKM stages (0–2), metabolic abnormalities—such as insulin resistance and dyslipidemia—are the dominant drivers, whereas systemic inflammatory responses (e.g., elevated IL-6 or TNF-α) typically become prominent only in advanced stages (3–4) with established organ damage. Consequently, conventional inflammatory markers show limited sensitivity for early screening [34]. Additionally, markers like PLR and NHHR are subject to confounding by acute infections, anemia, or medications [18]. More importantly, machine learning feature selection prioritizes strong, stable predictors directly related to metabolic dysregulation. Metabolic indices not only reflect early pathophysiological changes more sensitively but also demonstrate greater numerical stability compared to more variable inflammatory parameters. Thus, the model inherently assigns lower importance to inflammatory markers in early risk prediction.

This finding suggests that future efforts in developing early CKM stages prediction tools should prioritize core metabolic indicators while also exploring novel, organ-specific inflammatory biomarkers, rather than relying heavily on traditional systemic inflammatory markers.

A major strength of this study is its use of extensive cross-sectional data, the inclusion of multiple composite metrics, and the application of diverse machine learning algorithms. The LightGBM model, in particular, outperformed other models, enabling real-time risk prediction for various CKM stages based on patients’ serological and clinical laboratory data. This cross-sectional stratification tool assists clinicians in identifying patients with prevalent early-to-late CKM for targeted risk management and individualized intervention planning; prospective longitudinal cohorts are required to verify its capacity for forecasting incident cardiometabolic events. Furthermore, the integration of SHAP values improves model interpretability, helping medical practitioners identify major contributing factors and enhancing the model’s acceptability and utility in clinical settings.

Another key observation is the pronounced imbalance in prediction errors across CKM stages. Under the conventional maximum-probability classification rule adopted for primary model reporting, the LightGBM model exhibited substantially reduced sensitivity for advanced Stage 3 and Stage 4 patients, which may hinder clinical screening utility if applied without threshold adjustment. To resolve this limitation, we supplemented a post-hoc one-versus-rest threshold calibration workflow that derived independent low cutoffs (0.18–0.20) for each stratum, with hierarchical judgment prioritizing severe CKM subgroups. As demonstrated in Supplementary Table S9, this adjustment drastically elevated the detection rate of Stage 3 and Stage 4 patients at the cost of an 8.17% drop in overall classification accuracy. This trade-off is clinically acceptable for population-based cardiometabolic screening: missed diagnosis of advanced CKM carries substantial risks of progressive cardiorenal complications, while false-positive labeling of moderate-stage patients only requires follow-up re-evaluation without irreversible organ harm. Near-zero Kappa values across both classification schemes were attributable to extreme uneven sample distribution in the NHANES test set, where Stage 2 constituted over 40% of all participants; this cohort-driven limitation is acknowledged in our study constraints.

This study systematically evaluated predictive biomarkers for cardiovascular-kidney-metabolic (CKM) syndrome staging using machine learning and yielded noteworthy findings; however, several limitations warrant careful consideration. First, both the NHANES and CHARLS datasets adopted a cross-sectional design, which brings two critical constraints. The cross-sectional nature only allows us to identify significant cross-sectional associations between metabolic-inflammatory biomarkers and concurrent CKM stages, rather than establishing causal relationships or forecasting future incident cardiometabolic, renal, or cardiovascular events. On one hand, we cannot distinguish whether abnormal biomarkers drive CKM progression or serve as secondary consequences of advanced organ damage; for instance, an elevated TyG index may reflect insulin resistance exacerbating CKM or secondary metabolic dysregulation induced by severe renal impairment. On the other hand, our model only achieves classification of prevalent, existing CKM stages instead of prospective risk prediction for new-onset diseases. Accordingly, we have tempered all clinical claims throughout the manuscript: this tool is not capable of independently predicting future disease onset, and it merely stratifies participants with already-present metabolic, renal, and cardiovascular abnormalities by CKM severity. Rigorous causal inference analyses (e.g., Mendelian randomization) and long-term prospective cohort validations are required to clarify causal pathways and verify the model’s true predictive performance for incident adverse outcomes.

Second, due to the unavailability of certain laboratory measures in the external validation set, we employed a simplified model. Although this adapted version exhibited strong generalizability, it may partially underestimate the full model’s predictive capability. Future studies should prioritize prospective validation in international cohorts with complete biomarker collection. Moreover, when implementing the model in new populations, we recommend local calibration to align output probabilities with local disease prevalence and improve absolute risk estimation.

Additionally, the sensitivity of all models did not exceed 62%, which may reflect constraints in sample size or feature selection. Further validation using larger datasets or transfer learning approaches is needed to enhance generalizability. Although multiple machine learning algorithms were used to assess predictive performance, model efficacy could still be affected by sample size limitations and class imbalance. The supplementary stage-specific threshold optimization strategy we proposed effectively alleviates the low sensitivity of advanced CKM subgroups under unified classification logic; however, the hierarchical judgment rule reduces overall categorical accuracy, indicating that single fixed thresholds cannot simultaneously optimize both global concordance and high-risk detection. Subsequent research should incorporate larger, more balanced datasets, cost-sensitive machine learning loss functions, and advanced multi-class calibration techniques to improve model robustness and reliability, alongside prospective external validation of tailored threshold strategies.

Despite these limitations, our predictive framework offers clinicians a practical, evidence-based tool for early CKM stages stratification using routinely available composite biomarkers. Integration of this machine learning-assisted assessment into clinical workflows may facilitate: Timely identification of high-risk individuals across CKM stages, Tailored interventions based on individualized biomarker profiles, Improved long-term cardiorenal-metabolic outcomes. The ten-feature panel—including eGDR, METS_VF, TyG and related indices—strikes a balance between predictive accuracy (AUC: 0.88) and clinical feasibility, as all components can be derived from standard laboratory and anthropometric measures. Future validation across diverse populations and healthcare settings will help establish its role in precision prevention strategies.

In summary, this study demonstrates the clinical potential of eGDR, METS_VF, TyG, TyG_BMI, SHR, NPAR, AGR, PLR, WHtR, and NHHR in predicting CKM syndrome risk. These composite biomarkers provide new tools for CKM assessment and are expected to see broader clinical application in the future. Further research should validate their utility across diverse ethnic and regional populations and evaluate their performance in longitudinal settings.

Authors’ contributions

Minghao Li, Junyuan Xiang, and Fang Xiao contributed equally to this work as co-first authors. Minghao Li, Junyuan Xiang, and Fang Xiao participated in the conception, design, data acquisition, formal statistical analysis, result interpretation, and full manuscript revision responding to peer reviewers. Qianxiang Wang, Lin Liu, Yujia Huo and Chunyu Zhang assisted with data sorting, supplementary figure and table production, and critical revision of the manuscript. Jian Feng and Li Deng supervised the entire research project, provided key academic guidance throughout the study, and finalized the revised manuscript as corresponding authors. All authors have read and approved the final submitted version, and take full responsibility for the integrity and accuracy of all content in this paper.

Funding

This research was funded by grants from Sichuan Science and Technology Program (2022YFS0610), Xuyong County People’s Hospital—Southwest Medical UniversityScience and Technology Strategic Cooperation Project (2024XYXNYD18), Hejiang People's Hospital - Southwest Medical University Science and Technology Strategic Cooperation Project (2022HJXNYD05), Gulin People's Hospital - Affiliated Hospital of Southwest Medical University Science and Technology strategic Cooperation (2022GLXNYDFY13), 2022-N-01-33 project of China International Medical Foundation.

Data availability

The data used to support the findings of this study are included within the article.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.