The C-reactive protein-triglyceride glucose index (CTI) predicts mortality in cardiovascular-kidney-metabolic syndrome: a dual-cohort study with machine learning validation

Abstract

Background:

Cardiova scular-kidney-metabolic (CKM) syndrome urgently requires accessible biomarkers for stratification of death risk. This study validated the predictive value of a novel inflammatory metabolic biomarker, the C-reactive protein-triglyceride-glucose index (CTI), for all-cause and cardiovascular mortality in dual U.S. and Chinese cohorts and developed a survival analysis machine learning (ML) model.

Methods:

We integrated data from the National Health and Nutrition Examination Survey (NHANES, n = 8784) and China Health and Retirement Longitudinal Study (CHARLS, n = 7745). Multivariate Cox regression was used to evaluate the associations between CTI (formula: 0.412 × Ln(C-reactive protein) + Ln[triglycerides × fasting blood glucose/2]) and mortality. Seven ML models were built using the NHANES data, with CHARLS as the external validation set. SHapley Additive exPlanations (SHAP) clarified the prediction mechanisms.

Results:

Per 1-standard deviation increase in CTI, all-cause mortality risk increased significantly (NHANES: hazard ratios (HRs) = 1.31, 95% confidence interval (CI): 1.19–1.44; CHARLS: HR = 1.67, 95% CI: 1.44–1.93), and cardiovascular mortality increased by 35% in NHANES (HR = 1.35, P < 0.001). The Random Survival Forest (RSF) model performed best: internal validation area under the curve (AUC) = 0.866 (NHANES) with the highest time-dependent Concordance Index, and external validation in CHARLS yielded AUCs of 0.811 (3-year), 0.804 (5-year), and 0.775 (9/12-year), outperforming other models. SHAP analysis identified age (42.2% contribution) and CTI (10.1%) as key predictors, with age, CTI, and systolic blood pressure acting via independent main effects, whereas estimated glomerular filtration rate exerted an influence primarily through synergistic interactions.

Conclusion:

CTI, a novel inflammatory metabolic biomarker, reliably predicts all-cause and cardiovascular mortality in CKM syndrome, with consistent validation across NHANES and CHARLS. The NHANES-derived RSF model (AUC > 0.86) combines high accuracy and clinical utility, and is supported by stable external validation in CHARLS and sensitivity analyses. SHAP-based mechanistic insights further enable personalized risk assessments.

HIGHLIGHTS

• This study is the first to confirm in dual U.S. and Chinese cohorts (NHANES and CHARLS) that the novel inflammatory metabolic biomarker CTI (C-reactive protein-triglyceride-glucose index) can reliably predict all-cause and cardiovascular mortality in patients with cardiovascular-kidney-metabolic (CKM) syndrome, with consistent validation results across different populations.

• The Random Survival Forest (RSF) model developed based on the NHANES data showed the best performance, with an internal validation area under the curve (AUC) of 0.866. In the external validation using the CHARLS cohort, the AUC values for predicting 3-, 5-, 9-, and 12-year all-cause mortality were 0.811, 0.804, 0.775, and 0.775, respectively, demonstrating both high accuracy and cross-population applicability.

• SHAP analysis revealed the heterogeneous mechanisms of action of risk factors: age (42.2% contribution) and CTI (10.1%) affect mortality primarily through independent main effects, while estimated glomerular filtration rate exerts its influence mainly through synergistic interactions with other factors, providing a mechanistic basis for personalized risk assessment.

• The dual-cohort design (8784 participants in the U.S. and 7745 in China) ensures the generalizability of the results. For each 1-standard deviation increase in CTI, the risk of all-cause mortality increased by 31% and cardiovascular mortality by 35% in the NHANES cohort, and the risk of all-cause mortality increased by 67% in the CHARLS cohort, confirming its stability as a risk stratification tool for CKM.

• The calculation of CTI only requires routine measurements of C-reactive protein, triglycerides, and fasting blood glucose, with a simple method and high clinical accessibility. Combined with the high-performance RSF model, it provides a practical tool for precision risk stratification and formulation of intervention strategies for CKM syndrome.

Introduction

Cardiovascular-kidney-metabolic (CKM) syndrome is a complex pathological condition characterized by the coexistence of cardiovascular disease (CVD), chronic kidney disease (CKD), and metabolic disorders (e.g., obesity, insulin resistance (IR), and diabetes)^[1,2]. Its core mechanisms involve chronic inflammatory responses and IR induced by metabolic abnormalities, which accelerate multisystem damage through oxidative stress, endothelial dysfunction, and interorgan interactions, significantly increasing the risk of all-cause and cardiovascular mortality^[3–5]. With its global prevalence continuing to rise^[6–8], CKM syndrome has become a major public health challenge. Studies have indicated that CKM stage 3 has particularly high mortality rates, with all-cause and cardiovascular mortality rates reaching 25.3% and 45.3%, respectively^[9].

Although current guidelines emphasize multiorgan collaborative management, existing risk stratification tools rely primarily on single-disease indicators (e.g., blood glucose^[10] or estimated glomerular filtration rate [eGFR]^[11,12]), making it difficult to comprehensively assess the interplay between “metabolic dysfunction and chronic inflammation.” This limitation results in the inadequate identification of high-risk populations. The development of artificial intelligence (AI) and machine learning (ML), which enable high-dimensional data analysis and capture the associations between risk factors, is increasingly applied in clinical risk stratification^[13,14]. Unlike traditional regression models, ML algorithms (e.g., random survival forests [RSFs]) integrate overlapping risk pathways (inflammation, metabolism, renal function) to improve prediction accuracy; SHapley Additive explanation (SHAP) interpretability analysis further addresses ML’s “black box” limitation.

Recent research has focused on composite biomarkers that integrate the metabolic and inflammatory pathways^[15]. While the traditional triglyceride-glucose (TyG) index reflects IR^[16–18], it overlooks inflammatory mechanisms. The novel C-reactive protein-triglyceride glucose index (CTI)^[19,20] which combines C-reactive protein (CRP) with metabolic markers (triglycerides and glucose), offers a new perspective for CKM risk stratification.

Therefore, there is an urgent need for biomarkers to predict mortality in CKM. This study systematically evaluated the prognostic value of CTI in CKM patients with CKM by integrating two major population-based cohorts: National Health and Nutrition Examination Survey (NHANES) and China Health and Retirement Longitudinal Study (CHARLS). Furthermore, we developed ML-based predictive models incorporating CTI and employed SHAP analysis to decipher the interaction mechanisms of risk factors, providing a novel approach for precision risk assessment in CKM syndrome.

Methods

Data source

This study utilized data from two nationally representative prospective cohort studies, the CHARLS and NHANES, conducted in China and the United States. CHARLS is a nationwide cohort study of people aged 45 years and older in China. CHARLS is a national cohort study of people aged 45 years and older in China. Multistage stratified probability proportional sampling was used to recruit participants from 150 counties (representing both urban and rural areas) in 28 provinces. The baseline survey (Wave 1) was conducted from June 2011 to March 2012 with a total of 17 708 participants. Follow-up surveys were conducted in 2013 (Wave 2), 2015 (Wave 3), 2018 (Wave 4) and 2020 (Wave 5). The NHANES used a complex stratified multistage probability sampling design to collect nationally representative health and nutrition data from the civilian population of the United States. The study adhered to the ethical principles set forth in the Declaration of Helsinki and was approved by the Institutional Review Board of the National Center for Health Statistics (NCHS). Ethical approval for the CHARLS study was obtained from the Institutional Review Board of the NCHS (approval number: IRB00001052-11015). Written informed consent was obtained from all the participants. Please refer to the official website (https://www. cdc. gov/nchs/nhanes/) to obtain more information on study design and data. Informed consent was obtained from all the participants and/or their legal guardians before participating in the study.

Participant selection

This study incorporated data from two longitudinal cohorts: NHANES (1999–2010 and 2015–2018 cycles) and CHARLS (2011–2020). The following uniform exclusion criteria were applied: (1) non-CKM syndrome population, (2) missing CTI or mortality data, and (3) Patients with stage 4 CKM or those with pre-existing CVD at baseline. After screening, the final analysis included 8784 NHANES and 7745 CHARLS participants (see Figure 1 for the participant selection flowchart).

Figure 1.

Participant selection flowchart. The study population was derived from the U.S. NHANES (1999–2010, 2015–2018) and China CHARLS (2011–2020) cohorts. After exclusion, 8784 participants from NHANES (initial n = 81 385) and 7745 from CHARLS (initial n = 17 708) were included in the final analysis. Abbreviations: NHANES, National Health and Nutrition Examination Survey; CHARLS, China Health and Retirement Longitudinal Study; CTI, C-reactive protein-triglyceride glucose index; CKM, cardiovascular-kidney-metabolic syndrome.

Definitions of CKM syndrome stages 0–4

CKM syndrome is classified into five progressive stages based on metabolic, renal, and cardiovascular risk characteristics^[21]. Stage 0 denotes individuals without CKM risk factors, characterized by a normal body mass index (BMI <23 kg/m² for Asian ethnicity; <25 kg/m² for others) and waist circumference (WC <80/90 cm for Asian women and men; <88/102 cm for non-Asian women/men). Stage 1 involves excess or dysfunctional adiposity (BMI ≥23 kg/m² for Asians or ≥25 kg/m² for others; WC ≥88 cm for women or ≥102 cm for men) without metabolic abnormalities, CKD, or CVD. Stage 2 includes metabolic risk factors (e.g., elevated triglycerides [≥135 mg/dL], hypertension, diabetes, or metabolic syndrome) and moderate-to-high-risk CKD (defined by eGFR and albuminuria). Stage 3 identifies subclinical CVD, defined as a 10-year CVD risk ≥20% (calculated via American Heart Association Prediction Model for Cardiovascular Disease Risk [AHA-PREVENT] equations) or very high-risk CKD (KDIGO criteria). Stage 4 encompasses clinical CVD including coronary heart disease, heart failure, and cerebrovascular accidents (Supplemental Digital Content Table S1, Available at, http://links.lww.com/JS9/F207).

Definitions of CTIs

In this study, the CTI was calculated using the following formula:

CTI = 0.412 × Ln(CRP, mg/L) + Ln [triglycerides (mg/dL) × Fasting blood glucose (mg/dL)/2],

In a preliminary analysis, a continuous CTI variable was considered. Subsequently, they were categorized into quartiles for further analysis.

Covariates of interest

This study analyzed covariates from both NHANES (1999–2010 and 2015–2018) and CHARLS (2011–2020) cohorts to adjust for confounders^[22]. Common variables included demographics (age and sex), socioeconomic status (education and marital status), lifestyle factors (smoking, alcohol use, and BMI), and medical history (hypertension, diabetes, and cancer). To ensure comparability in cross-cohort analysis, categorical variables with differing classifications were harmonized as follows: For education, CHARLS’ original 4-level classification (“Primary or below,” “Junior School,” “High School,” “University or above”) was collapsed into three levels to match NHANES: “Primary or below” + “Junior School” → “Less than High School; “High School” → “High school”; “University or above” → “Above high school.” For marital status, NHANES’ detailed categories (“Married/Living with partner,” “Widowed/Divorced/Separated/Never married”) were aligned with the CHARLS “Yes’ (married) versus “No” (unmarried) grouping.

Definition of clinical outcomes

Mortality data in the NHANES cohort

As of 31 December 2019, mortality data for NHANES were obtained from its publicly linked mortality file, which was probabilistically matched with the National Death Index maintained by the NCHS. The causes of death were coded according to the International Classification of Diseases, Tenth Revision (ICD-10) and further categorized as (1) all-cause mortality (ICD-10 codes: U01–U03, 010) and (2) cardiovascular mortality, including heart disease (054–068) and cerebrovascular disease (070). The mean follow-up duration in the NHANES cohort was 11.54 years.

Mortality data in the CHARLS cohort

Participants enrolled in the CHARLS baseline survey (2011–2012) were followed across four subsequent waves (2013, 2015, 2018, and 2020). Death events were verified through (1) official death certificates, (2) medical records, and (3) interviews with the next of kin during follow-up surveys. The survival time was calculated from the baseline survey date to the date of death (censored at the last follow-up for survivors). Unlike NHANES, CHARLS only recorded all-cause mortality, without further cause-specific classification. The mean follow-up duration was 8.90 years.

Statistical analysis

All analyses incorporated sample weights, clustering, and stratification to ensure an accurate variance estimation and national representativeness. The normality of continuous variables was assessed visually using histograms^[23] with bell-shaped curves indicating a normal distribution^[24]. Except for age and poverty–income ratio (PIR) in NHANES (non-normal distributions), all other continuous variables followed normal distributions (Supplemental Digital Content Figure S1, Available at, http://links.lww.com/JS9/F207 displays the CTI distribution). Skewed variables were expressed as medians (Q1, Q3) and analyzed using the Kruskal–Wallis test. Normally distributed variables were reported as means (standard errors) and compared using analysis of Variance. Categorical variables are presented as weighted counts and percentages (n [weighted %]) and were analyzed using the chi-square or Fisher’s exact test. Statistical significance was set at P < 0.05. Python (Version 3.9) and R (Version 4.4.2) software were used to analyze the data of the entire study.

Cox proportional hazards models were used to evaluate CTI–mortality associations, and the results were expressed as hazard ratios (HRs) and 95% confidence intervals (CIs). For comparability, the CTI was standardized as Z-scores^[25] (effect size per 1-standard deviation [SD] increase). The variance inflation factors for all covariates were <5 (Supplemental Digital Content Table S2, Available at, http://links.lww.com/JS9/F207), indicating no significant multicollinearity^[26]. Covariates were adjusted based on clinical relevance and previous studies^[27,28]. Restricted cubic splines (RCS) have been used to visualize nonlinear mortality relationships. Kaplan–Meier curves with log-rank tests were used to compare survival differences. Stratified and interaction analyses were used to examine effect modifications by sex, marital status, smoking, alcohol use, diabetes, hypertension, and CKM stages 0–3. All analyses adhered to the STROCSS 2025 guidelines^[29].

The NHANES dataset was randomly divided into a training set (N = 6148) and an internal validation set (N = 2636) at a 7:3 ratio, with the CHARLS dataset used as the external validation set. The research workflow was as follows. The models were first developed using the training set, and their performance was evaluated using internal and external validation sets. Feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) regression method^[30] and seven survival analysis-specific ML models were constructed: (1) Light Gradient Boosting Machine (LightGBM), (2) RSF, (3) LASSO regression, (4) Cox Proportional Hazards Boosting (CoxBoost), (5) Extreme Gradient Boosting (XGBoost), (6) Supervised Principal Component Analysis (Superpc), and (7) Partial Least Squares Regression for Cox Models (plsRcox). The hyperparameters of the seven ML models are listed in Supplemental Digital Content Table S9, Available at, http://links.lww.com/JS9/F207.

1. Model performance was comprehensively evaluated using the following methods: (1) Time-dependent receiver operating characteristic (ROC) curves and area under the curve (AUC) based on 150-month (12.5-year) follow-up data; (2) Time-dependent Concordance index (C-index) to assess discriminatory ability during follow-up (a value closer to 1 indicates better predictive consistency)^[31]; (3) Decision Curve Analysis (DCA) to evaluate clinical utility; and (4) External validation at multiple time points (3, 5, 9, and 12 years).

To address the “black-box nature of machine-learning models^[32], we employed the SHAP methodology^[33] based on cooperative game theory. SHAP values quantify each feature’s marginal contribution to individual predictions by decomposing model outputs into additive feature effects: (1) Main Effects (independent feature contributions) and (2) Interaction Effects (synergistic/counteractive effects from feature combinations). In our study, (1) Positive SHAP values indicate increased mortality risk, (2) negative values denote protective effects, and (3) absolute magnitudes reflect effect strengths. Furthermore, SHAP interaction values quantitatively revealed significant feature interdependencies by measuring the average marginal contribution changes across different feature combinations^[34].

Result

Baseline characteristics

The NHANES data revealed a strong positive association between elevated CTI levels and adverse metabolic profiles (all P < 0.001), with the highest CTI quartile (Q4) demonstrating significantly worse metabolic parameters (BMI 32.42 vs 24.12 kg/m²; SBP 127.20 vs 118.07 mmHg; eGFR 100.96 vs 106.36 mL/min/1.73 m²) and higher prevalence of diabetes (27.29% vs 4.18%), hyperlipidemia (95.11% vs 45.36%), and metabolic syndrome (54.35% vs 3.35%) compared to Q1. This metabolic deterioration corresponded with progressive CKM staging (87.08% stages 2 and 3 in Q4 vs 30.28% stage 0 in Q1). These findings were validated in the CHARLS cohort, with the increase of CTI (from Q1 to Q4), there were statistical differences in renal function-related indicators (blood urea nitrogen (BUN) and creatinine (CREA)) and hemoglobin levels among the groups (all P < 0.001), and increased prevalence of diabetes (31.49% vs 4.03%), hypertension (57.41% vs 32.85%), and more advanced CKM stages (65.10% vs 50.00% stage 3 in Q4 vs Q1) (Tables 1 and Supplemental Digital Content Table S3, Available at, http://links.lww.com/JS9/F207).

Table 1.

Baseline characteristics of the CKM stage 0–3 stratified by survival status, weighted for representativeness (NHANES)

Variable	Total (n = 8784)	Q1 (n = 1924)	Q2 (n = 2125)	Q3 (n = 2293)	Q4 (n = 2442)	P
Age (years)	46.00 (33.00, 58.00)	39.00 (27.00, 52.00)	46.00 (33.00, 59.00)	47.00 (36.00, 60.00)	48.00 (37.00, 59.00)	<0.001
PIR	3.09 (1.63, 5.00)	3.36 (1.70, 5.00)	3.16 (1.69, 5.00)	3.07 (1.66, 5.00)	2.77 (1.41, 4.83)	<0.001
CTI	8.01 (0.02)	6.77 (0.02)	7.73 (0.01)	8.35 (0.00)	9.19 (0.01)	<0.001
eGFR, ml/min/1.73 m2	103.15 (0.48)	106.36 (0.80)	103.34 (0.62)	101.93 (0.74)	100.96 (0.73)	<0.001
BMI, kg/m2	28.63 (0.12)	24.12 (0.15)	27.58 (0.16)	30.41 (0.18)	32.42 (0.19)	<0.001
SBP, mm Hg	123.66 (0.27)	118.07 (0.50)	123.93 (0.46)	125.42 (0.47)	127.20 (0.46)	<0.001
DBP, mm Hg	71.52 (0.21)	68.91 (0.40)	71.47 (0.33)	72.43 (0.33)	73.27 (0.37)	<0.001
Cancer, n (%)						0.009
Yes	748 (8.27)	140 (7.09)	178 (7.60)	202 (8.09)	228 (10.29)
No	8036 (91.73)	1784 (92.91)	1947 (92.40)	2091 (91.91)	2214 (89.71)
Gender, n (%)						<0.001
Male	4406 (50.87)	942 (46.59)	1170 (55.88)	1182 (52.73)	1112 (48.27)
Female	4378 (49.13)	982 (53.41)	955 (44.12)	1111 (47.27)	1330 (51.73)
Race, n (%)						<0.001
Mexican American	1775 (7.72)	245 (5.53)	376 (7.16)	497 (8.52)	657 (9.68)
Other Hispanic	664 (4.59)	132 (3.70)	154 (4.20)	192 (5.60)	186 (4.87)
Non-Hispanic White	4315 (71.85)	910 (69.36)	1086 (72.87)	1112 (71.38)	1207 (73.80)
Non-Hispanic Black	1525 (9.83)	469 (13.28)	389 (10.06)	368 (8.88)	299 (7.08)
Other Race	505 (6.01)	168 (8.13)	120 (5.71)	124 (5.61)	93 (4.57)
Education, n (%)						<0.001
Less than high school	2381 (17.09)	382 (13.02)	546 (17.11)	659 (17.51)	794 (20.71)
High school	2095 (24.90)	410 (19.99)	517 (25.27)	567 (27.48)	601 (26.88)
Above high school	4308 (58.01)	1132 (66.99)	1062 (57.62)	1067 (55.01)	1047 (52.41)
Marital status, n (%)						0.001
Married/Living with partner	5606 (66.38)	1128 (62.91)	1339 (65.47)	1529 (69.32)	1610 (67.80)
Widowed/Divorced/Separated/Never married	3178 (33.62)	796 (37.09)	786 (34.53)	764 (30.68)	832 (32.20)
Smoking status, n (%)						<0.001
No	4638 (51.88)	1150 (59.46)	1124 (52.38)	1188 (50.47)	1176 (45.23)
Yes	4146 (48.12)	774 (40.54)	1001 (47.62)	1105 (49.53)	1266 (54.77)
Alcohol intake, n (%)						0.002
No	2554 (24.34)	530 (22.88)	548 (22.29)	681 (24.58)	795 (27.60)
Yes	6230 (75.66)	1394 (77.12)	1577 (77.71)	1612 (75.42)	1647 (72.40)
Hypertension, n (%)						<0.001
No	4998 (61.88)	1357 (76.94)	1191 (61.04)	1220 (56.13)	1230 (53.40)
Yes	3786 (38.12)	567 (23.06)	934 (38.96)	1073 (43.87)	1212 (46.60)
Diabetes, n (%)						<0.001
No	7242 (86.56)	1801 (95.82)	1878 (91.15)	1934 (86.57)	1629 (72.71)
Yes	1542 (13.44)	123 (4.18)	247 (8.85)	359 (13.43)	813 (27.29)
Hyperlipidemia, n (%)						<0.001
No	2092 (25.62)	1034 (54.64)	599 (28.48)	336 (14.47)	123 (4.89)
Yes	6692 (74.38)	890 (45.36)	1526 (71.52)	1957 (85.53)	2319 (95.11)
METS, n (%)						<0.001
No	6524 (74.56)	1854 (96.65)	1850 (87.05)	1616 (68.91)	1204 (45.65)
Yes	2260 (25.44)	70 (3.35)	275 (12.95)	677 (31.09)	1238 (54.35)
CKM syndrome, n (%)						<0.001
Stage 0	673 (10.08)	496 (30.28)	142 (8.49)	32 (1.49)	3 (0.06)
Stage 1	1152 (14.63)	454 (25.27)	409 (20.38)	224 (9.90)	65 (2.98)
Stage 2	6365 (68.46)	904 (41.72)	1441 (64.62)	1874 (80.42)	2146 (87.08)
Stage 3	594 (6.83)	70 (2.73)	133 (6.51)	163 (8.19)	228 (9.89)

Skewed data: median (Q1, Q3; Kruskal–Wallis test); normal data: mean (SE) (ANOVA); categorical data: n (weighted %; chi-square/Fisher’s test). P < 0.05 was significant. Abbreviations: ANOVA, analysis of variance; BMI, body mass index; BUN, blood urea nitrogen; CKM syndrome, cardiovascular-kidney-metabolic syndrome; CVD, cardiovascular disease; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; METS: metabolic syndrome; SBP, systolic blood pressure; SE, standard error; PIR, poverty–income ratio; CTI, C-reactive protein triglyceride glucose index.

Association between CTI and mortality

The NHANES cohort results (Table 2) demonstrated significant associations between CTI and all-cause and cardiovascular mortality. In the fully adjusted model (Model III), each 1-SD increase in CTI was associated with a 31% higher risk of all-cause mortality (HR: 1.31, 95% CI: 1.19–1.44, P < 0.001) and a 35% higher risk of cardiovascular mortality (HR = 1.35, 95% CI: 1.13–1.61, P = 0.001). When stratified by CTI quartiles, the highest quartile (Q4) showed a 74% increased risk of all-cause mortality (HR = 1.74; 95% CI: 1.33–2.28, P < 0.001). Although the association between Q4 and cardiovascular mortality did not reach statistical significance (HR = 1.51, 95% CI: 0.98–2.31, P = 0.059), it approached the significance threshold and there was a significant dose-response trend (P-trend = 0.008), suggesting that an elevated CTI may still be associated with an increased risk of cardiovascular mortality. These findings were corroborated in the CHARLS cohort (Supplemental Digital Content Table S4, Available at, http://links.lww.com/JS9/F207), where each 1-SD increase in CTI corresponded to a 67% higher risk of all-cause mortality in the fully adjusted model (Model III).

Table 2.

Weighted multivariate Cox regression models for the association between CTI index and mortality risk (NHANES)

	Model Ia		Model IIb		Model IIIc
Variables	HR (95% CI)	P	HR (95% CI	P	HR (95% CI)	P
All-cause mortality
Per 1 SD	1.43 (1.33–1.53)	<0.001	1.31 (1.20–1.42)	<0.001	1.31 (1.19–1.44)	<0.001
CTI Index
Q1	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Q2	2.06 (1.62–2.61)	<0.001	1.23 (0.95–1.58)	0.115	1.28 (0.99–1.64)	0.06
Q3	2.32 (1.83–2.95)	<0.001	1.37 (1.08–1.74)	0.009	1.43 (1.09–1.86)	0.009
Q4	2.76 (2.18–3.51)	<0.001	1.74 (1.37–2.21)	<0.001	1.74 (1.33–2.28)	<0.001
P trend		<0.001		<0.001		<0.001
CVD mortality
Per 1 SD	1.53 (1.36–1.73)	<0.001	1.45 (1.24–1.70)	<0.001	1.35 (1.13–1.61)	<0.001
CTI Index
Q1	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Q2	1.99 (1.33–2.96)	<0.001	1.10 (0.74–1.62)	0.647	1.03 (0.68–1.56)	0.892
Q3	3.04 (2.02–4.57)	<0.001	1.70 (1.14–2.54)	0.009	1.53 (1.01–2.35)	0.05
Q4	3.03 (2.03–4.53)	<0.001	1.88 (1.29–2.73)	<0.001	1.51 (0.98–2.31)	0.059
P trend		<0.001		<0.001		0.008

Bold font indicates statistically significant differences (P < 0.05).

^aModel I: Crude

^bModel II: adjusted for gender, age, and race.

^cModel III: adjusted for age, cancer, gender, race, education, marital, smoking, alcohol, hypertension, diabetes, hyperlipidemia, METs, eGFR, BMI, SBP, DBP, and PIR.

CTI, C-reactive protein triglyceride glucose index; CKM syndrome, cardiovascular-kidney-metabolic syndrome; CVD, cardiovascular diseases; 95% CI, 95% confidence interval; HR, hazard ratio.

RCS and survival analysis

Nonlinear relationship analysis using RCS revealed a significant positive linear association between CTI levels and mortality in both NHANES and CHARLS cohorts (all nonlinearity P-values >0.05; Figure 2A,C, and Supplemental Digital Content Figure S2, Available at, http://links.lww.com/JS9/F207). Survival analysis demonstrated progressively declining survival probabilities with increasing CTI levels, as evidenced by Kaplan–Meier curves (log-rank P < 0.001; Figure 2B,D, and Supplemental Digital Content Figure S3, Available at, http://links.lww.com/JS9/F207).

Figure 2.

Restricted cubic spline (RCS) and Kaplan–Meier survival analyses. Panels A and C show RCS curves depicting dose-response relationships between CTI and mortality risk (all-cause [A] and cardiovascular [C]). Panels B and D display Kaplan–Meier survival curves stratified by CTI quartiles, with log-rank tests comparing intergroup differences. Abbreviations: CTI, C-reactive protein-triglyceride glucose index; 95% CI, 95% confidence interval; HR, hazard ratio.

ML Model Performance Comparison

First, ML models were developed in the NHANES cohort, with the dataset split into training and internal validation sets at a 7:3 ratio. LASSO regression was used for feature selection from all the candidate variables (Fig. 3A-B). Based on the optimal penalty coefficient (λ.1se) criterion, 13 predictive features were identified: CTI, age, sex, race, marital status, PIR, systolic blood pressure, diastolic blood pressure, BMI, eGFR, diabetes, cancer history, and smoking status.

Figure 3.

LASSO regression for variable selection and performance evaluation of seven machine learning models for all-cause mortality prediction in the NHANES cohort. (A) Coefficient path plot illustrating variable shrinkage across penalty parameters (λ). (B) Ten-fold cross-validation curve identifying the optimal λ (λ.1se), which selected 13 predictive variables while balancing model complexity and accuracy. (C–D) Receiver operating characteristic (ROC) curves comparing the performance of models between training (C) and internal validation (D) sets. (E–F) Time-dependent C-index curves comparing the performance of models between training (E) and internal validation (F) sets. (G–H) Decision curve analysis (DCA) assessing the clinical utility of the models between training (G) and internal validation (H) sets. The models included LightGBM, random survival forest (RSF), Lasso-Cox, CoxBoost, XGBoost, Superpc, and plsRcox.

ROC curve analysis (Fig. 3C,D) showed the validation set AUC values of the models as follows: LightGBM (0.857), RSF (0.866), Lasso-Cox (0.867), CoxBoost (0.865), XGBoost (0.866), Superpc (0.823), and plsRcox (0.866). Evaluation using the time-dependent C-index demonstrated that the RSF exhibited the highest predictive consistency in both the training and validation sets (Fig. 3E,F). DCA further confirmed that RSF yielded the optimal net clinical benefit (Fig. 3G,H), thus establishing it as the final predictive model.

Subsequently, seven survival prediction models were constructed using 10 variables common to both the NHANES and CHARLS cohorts (CTI, age, sex, marital status, systolic blood pressure, diastolic blood pressure, BMI, diabetes, cancer history, and smoking status) based on the NHANES cohort. The CHARLS cohort was used as an external validation set to evaluate the performance of each model in predicting all-cause mortality at 3, 5, 9, and 12 years of follow-up.

Time-dependent receiver operating characteristic (tdROC) curve analysis revealed that in the CHARLS validation set, the RSF model achieved the highest external validation set AUC among all models for all-cause mortality prediction at different follow-up durations: AUC = 0.811 for 3-year prediction, 0.804 for 5-year prediction, and 0.775 for both 9-year and 12-year predictions (Fig. 4A-H). These results indicate that RSF outperformed other models (including LightGBM and Lasso-Cox) and exhibited superior predictive efficacy in multi-time point all-cause mortality prediction in the CHARLS validation set.

Figure 4.

External validation in CHARLS of NHANES-developed survival models for all-cause mortality using time-dependent ROC curves across follow-up durations. (A, E) Three-year all-cause mortality prediction in the NHANES training set (A) and CHARLS external validation set (E); (B, F) 5-year all-cause mortality prediction in the NHANES training set (B) and CHARLS external validation set (F); (C, G) 9-year all-cause mortality prediction in the NHANES training set (C) and CHARLS external validation set (G); (D, H) 12-year all-cause mortality prediction in the NHANES training set (D) and CHARLS external validation set (H).

SHAP Analysis of Feature Importance and Interaction Patterns

SHAP analysis of the RSF model (Fig. 5A) identified age (SHAP = 1.246), CTI (0.297), and smoking (0.267) as top positive mortality predictors, whereas sex (−0.152) and BMI (−0.130) showed protective effects. The feature importance ranking (Fig. 5B) highlighted age (42.2%) and CTI (10.1%) as dominant predictors.

Figure 5.

SHAP analysis of the RSF model. (A) Beeswarm plot: dots represent individual samples, colors indicate feature values, and horizontal displacement reflects SHAP values (direction and magnitude of risk contribution). (B) Feature importance ranking (mean absolute SHAP values). (C) Main and interaction effects analysis, highlighting how key variables (e.g., age, CTI) independently or jointly influence mortality risk.

Analysis of variable interaction patterns (Fig. 5C) revealed significant heterogeneity; age exhibited a main effect-to-interaction effect ratio of 21:1, indicating that its risk impact was almost entirely driven by independent effects. For CTI (ratio = 1.84), SBP (ratio = 1.58), and smoking (ratio = 2.4), the main effects predominated. Although synergistic interactions existed, independent risk contributions remained the primary mode. In contrast, eGFR (ratio = 0.3:1) and sex were dominated by interaction effects; eGFR’s influence of eGFR was largely mediated by synergistic interactions with age, while sex exerted effects primarily through interactions with factors including eGFR.

Subgroup and sensitivity analysis

Subgroup analyses demonstrated robust positive associations between CTI and all-cause mortality in both NHANES and CHARLS cohorts (all HR > 1, P < 0.05). Interaction analyses revealed no significant effect modifications by marital status, smoking, alcohol use, hypertension, or diabetes (all interactions, P > 0.05; Supplemental Digital Content Table S5, Available at, http://links.lww.com/JS9/F207). Notably, a significant gender interaction was observed in the NHANES (P = 0.017), indicating a higher risk in males, although this finding was not replicated in CHARLS (Supplemental Digital Content Table S6, Available at, http://links.lww.com/JS9/F207). Collectively, these results support the robustness of CTI as a mortality risk predictor.

To verify the generalizability of the results, a sensitivity analysis including CKM stage 4 was supplemented (Supplemental Digital Content Table S7, Available at, http://links.lww.com/JS9/F207), which showed that the association between CTI and all-cause mortality remained stable in the expanded population (per 1-SD increase, HR = 1.33, 95% CI: 1.21–1.47, P < 0.001), suggesting that its predictive value may be applicable to a broader CKM spectrum.

Cumulative CTI (cuCTI) was calculated to integrate dynamic information, and the sensitivity analysis (Supplemental Digital Content Table S8, Available at, http://links.lww.com/JS9/F207) showed that in the fully adjusted model, each 1-SD increase in cuCTI was associated with a 28% increase in all-cause mortality risk (HR = 1.28, 95% CI: 1.07–1.52, P = 0.006), whereas the mortality risk in the highest quartile group was 2.06 times that in the lowest quartile group (95% CI: 1.27–3.32, P = 0.003).

The above results support the reliability of CTI as a predictor of mortality risk.

Discussion

This dual-cohort study is the first to establish the CTI as a novel core biomarker that integrates the inflammatory and metabolic dimensions. It has demonstrated robust predictive efficacy for all-cause and cardiovascular mortality in individuals with CKM syndrome across ethnically diverse populations, including the U.S. (NHANES) and Chinese (CHARLS) cohorts. In the NHANES cohort, each 1-SD increase in CTI was independently associated with a 31% higher risk of all-cause mortality and a 35% higher risk of cardiovascular mortality. In the CHARLS cohort, the risk of all-cause mortality was significantly increased by 67%. More innovatively, the CTI-based RSF model we developed exhibited both high accuracy (internal validation AUC = 0.866) and cross-population generalizability (external validation AUC 0.775–0.811). Additionally, SHAP interpretability analysis revealed the heterogeneous mechanisms of multi-factor action: CTI, age, and systolic blood pressure primarily drive mortality risk through “independent main effects,” while renal function (eGFR) exerts its influence mainly through “synergistic interactions with other metabolic factors.” These findings not only fill the critical gap in CKM management regarding “comprehensive biomarkers and interpretable prediction tools validated across populations” but also provide a framework with clinical accessibility and biological rationality for personalized risk stratification and the formulation of precise prevention strategies through the analysis of “main effect-interaction effect” mechanisms.

Existing research on biomarkers for CKM syndrome has primarily focused on isolated metabolic and inflammatory markers^[35,36]. The TyG index is a reliable surrogate indicator of IR^[27,37,38] and is associated with cardiovascular events and renal dysfunction in patients with diabetes and metabolic syndrome^[39–41]. A meta-analysis^[42] confirmed that an elevated TyG index is associated with an increased risk of heart failure and coronary artery disease. However, such single-pathway biomarkers often overlook the critical synergistic role of chronic inflammation, which is the core pathological driver of CKM progression^[41,43].

In contrast, the composite CTI integrates TyG (metabolic pathway) and CRP (inflammatory pathway) levels, enabling a more comprehensive reflection of the pathophysiological characteristics of CKM. Unlike traditional indicators that only assess isolated metabolic abnormalities (e.g., fasting blood glucose [FPG]) or renal function impairment (e.g., eGFR), CTI uniquely captures the synergistic interaction between IR and chronic inflammation, which is a core feature of CKM pathophysiology.

Existing evidence highlights significant limitations of single-pathway biomarkers. For example, the atherogenic index of plasma (AIP), a lipid-based marker, can predict cardiovascular mortality^[44], but lacks inflammatory indicators, limiting its application value in overall CKM risk assessment. Similarly, the stress hyperglycemia ratio shows a U-shaped association, but fails to consider chronic inflammatory factors, thereby weakening its predictive ability in the early CKM stages (stages 0–3)^[45]. Crucially, independent biomarkers (such as TyG or AIP) only assess isolated pathways (e.g., metabolic dysfunction or dyslipidemia) and are not involved in the synergistic interaction between IR and inflammation, which is the core mechanism of CKM progression^[46,47].

The synergistic interaction between chronic inflammation and insulin resistance is the core driver of CKM progression, and the advantage of CTI lies in quantifying this bidirectionally regulated vicious cycle. On one hand, is a classic inflammatory marker, CRP, that promotes the release of inflammatory factors such as interleukin [IL]-6 and tumor necrosis factor-α (TNF-α) from adipocytes by activating the NF-κB pathway, inhibiting the phosphorylation of insulin receptor substrate 1 (IRS-1), and directly exacerbating IR^[48]. Conversely, the IR state reflected by the TyG index can induce the accumulation of advanced glycation end products through hyperglycemia, activate the NLRP3 inflammasome, and further stimulate CRP synthesis^[49]. This “inflammation → metabolic disorder → more severe inflammation” cascade reaction cannot be fully captured by a single biomarker: for instance, TyG only reflects metabolic abnormalities, while CRP only indicates the level of inflammation, and both fail to capture the synergistic effect when used alone^[50]. However, the potential of nanocarrier technologies (such as lipid nanoparticles) mentioned in the study by Priyanka et al^[51], which enhances the targeting ability and efficacy of bioactive substances, suggests that in the future, based on CTI risk stratification, precision intervention strategies targeting the regulation of this vicious cycle (e.g., nanocarrier-mediated anti-inflammatory-metabolic synergistic drugs) can be developed for CKM patients with high CTI. Meanwhile, the characteristic of CTI in integrating dual pathways shares a similar logic with mRNA vaccines, which enhances the scope of intervention through multi-target coverage, emphasizing the comprehensive addressing of complex pathological mechanisms^[52]. By integrating the weights of both through the formula (0.412 × Ln(CRP) + Ln(TG × FPG/2)), CTI retains the sensitivity of CRP to vascular endothelial damage and incorporates the assessment of metabolic overload by TyG, thus more accurately reflecting the overall risk in patients with CKM^[53].

A growing body of evidence supports the superiority of combined metabolic-inflammatory biomarkers in CVD risk assessment. In the general population, a study by Sun et al, based on the NHANES data, showed that the CTI had a linear positive correlation with the risk of total CVD (OR: 2.85, 95% CI: 2.32–3.52) and various CVD subtypes (such as congestive heart failure, coronary heart disease, etc.). Its predictive performance significantly outperforms that of CRP or the TyG index alone^[54]. In the general middle-aged and elderly population, further research by Huo et al^[55] confirmed that CTI has a linear positive correlation with stroke risk (HR: 1.19), and this association was more significant in populations with normal glucose regulation (HR: 1.33) and prediabetes (HR: 1.20). ROC curve analysis showed that the AUC 0.587 of CTI in predicting stroke was better than that of CRP alone (0.565) and TyG index alone (0.567), confirming its advantage in integrating the dual pathological pathways of inflammation and IR. In the hypertensive population, through dual assessment of IR (TyG component) and inflammatory status (CRP component), CTI significantly improves the accuracy of stroke risk stratification^[56]. In patients with chronic coronary syndrome, the combined application of TyG index and high-sensitivity CRP significantly enhances the predictive efficacy for major adverse cardiovascular events^[57,58]. Notably, the application value of CTI has been extended to multiple fields, such as oncology, depression, erectile dysfunction, and liver diseases, further highlighting its broad applicability as a cross-disease biomarker^[5].

The RSF model developed in this study exhibited an excellent performance in predicting all-cause mortality in the CKM population. First, the flexibility of the ML algorithm in the RSF model allows for more accurate capture of the complex associations between CKM risk factors. Traditional scores such as PREVENT innovatively incorporate renal function indicators (eGFR) to cover the core dimensions of CKM^[59], but their linear framework struggles to quantify nonlinear relationships (e.g., interactions between blood pressure and metabolic indicators). In contrast, RSF can automatically identify the combined effects of inflammatory-metabolic indicators (such as CTI) and basic health status (such as diabetes and smoking history) by integrating multiple survival trees, which may be the key reason why its AUC (0.866) for predicting 12.5-year all-cause mortality in the NHANES cohort was higher than the median C-statistic (0.757–0.813) of PREVENT for cardiovascular composite outcomes. Second, the stability of RSF across populations highlights its clinical utility. PREVENT shows ethnic differences in the C-index for predicting 10-year cardiovascular risk in multiethnic cohorts (0.65–0.70)^[60], while RSF maintains a high predictive efficacy (0.775–0.866) in both NHANES (multiethnic) and CHARLS (Chinese population). Finally, the long-term prediction advantage of RSF is more in line with the CKM management needs of CKM. The AUC of the Chinese domestic CKM₂S₂-BAG score for predicting 5-year cardiovascular risk was 0.705^[61], while RSF achieved an AUC of 0.775–0.811 for predicting 3- to 12-year all-cause mortality in the CHARLS cohort, which remained stable with extended follow-up. This characteristic matches the “progressive course of CKM syndrome and can provide a more reliable basis for long-term risk stratification and selection of intervention timing.

SHAP interpretability analysis showed that age (contribution, 42.2%) and CTI (10.1%) were the dominant independent predictors (ratio of main effect to interaction effect > 1.8), supporting targeted intervention measures (such as anti-inflammatory therapy). In contrast, eGFR mainly affects risk through interactions, and thus requires comprehensive assessment. From the perspective of clinical intervention, the different modes of action of the CTI and eGFR suggest different management strategies. As a marker of metabolic-inflammatory coupling, an elevated CTI mainly reflects the vicious cycle of IR-driven lipotoxicity and chronic inflammation: high triglycerides promote hepatic fat synthesis through the peroxisome proliferator-activated receptor-α pathway, while CRP inhibits the phosphorylation of IRS-1 through the nuclear factor-κB (NF-κB) pathway, exacerbating IR^[46,49]. Therefore, interventions targeting CTI focus more on breaking the metabolic-inflammatory cycle; in terms of lifestyle adjustments, a low-carbohydrate diet can improve insulin sensitivity^[62], and aerobic exercise reduces the release of inflammatory factors by inhibiting the NF-κB pathway^[63]. In drug therapy, glucagon-like peptide-1 receptor agonists can simultaneously reduce triglyceride levels and CRP expression^[64], and metformin can improve IR and exert mild anti-inflammatory effects by activating the AMP-activated protein kinase pathway^[65]. Additionally, the anti-inflammatory mechanism of mesenchymal stem cells (MSCs) provides a supplementary direction for the intervention of populations with high CTI – MSCs can regulate the inflammatory microenvironment by releasing anti-inflammatory factors (e.g., IL-10) and inhibiting the expression of pro-inflammatory factors (e.g., TNF-α and IL-6) through paracrine effects^[66]. Animal studies have also confirmed that MSCs can alleviate local inflammatory responses and optimize the tissue microenvironment by secreting trophic factors^[67], which holds potential reference value for populations with high CTI who show limited responses to traditional interventions.

In contrast, eGFR and gender mainly affect risk through interactions: eGFR synergistically amplifies risk with age, which is consistent with the research conclusion that “age-related decline in renal perfusion exacerbates metabolic damage”;^[68,69] the interaction between gender and eGFR may be related to the renal protective effect of estrogen or androgen-driven activation of the renin–angiotensin system^[70]. In clinical practice, sodium-glucose cotransporter 2 (SGLT2) inhibitors (such as dapagliflozin) can interfere with the interaction between eGFR and age by improving renal hemodynamics^[64]. This mechanism-based stratification strategy enhances the precision of CKM syndrome management.

This study has several advantages. First, it innovatively combined two nationally representative cohorts (NHANES and CHARLS) to validate the prognostic value of CTI for all-cause and cardiovascular mortality in CKM stages 0–3, with consistent results obtained in both the U.S. and Chinese populations. Second, the simple calculation method of CTI (using routine measurements of CRP, triglycerides, and FPG) ensures its clinical utility. Third, the study adopts multidimensional analytical approaches ranging from traditional Cox regression to ML (such as RSF) and SHAP interpretability analysis, ensuring the robustness and generalizability of the results.

This study had several limitations. First, despite extensive covariate adjustment, unmeasured confounding factors (such as dietary habits and genetic factors) may still exist. Second, owing to the lack of specific cause-of-death data for CHARLS, this study was unable to validate the predictive efficacy of CTI for cardiovascular mortality in this cohort. Moreover, this data discrepancy may have affected the direct comparison of the research conclusions with those from the NHANES cohort. Future studies should rely on Chinese cohorts with detailed cause-of-death information to further clarify the specific value of the CTI in the prognostic assessment of specific causes of death. Finally, limited by the relatively short follow-up period (2015–2020) and the small number of death events during this period, this study failed to construct a prediction model for dynamic changes in cuCTI. Future studies should rely on cohorts with longer follow-up periods and larger sample sizes, systematically collect multi-time point CTI data, and further clarify whether dynamic CTI can improve the efficacy of death risk prediction by constructing and validating prediction models incorporating dynamic change parameters, thereby providing more precise tools for clinical risk stratification.

Conclusion

CTI, a novel inflammatory metabolic biomarker, reliably predicts all-cause and cardiovascular mortality in CKM syndrome, with consistent validation across NHANES and CHARLS. The NHANES-derived RSF model (AUC > 0.86) combines high accuracy and clinical utility, and is supported by stable external validation in CHARLS and sensitivity analyses. SHAP-based mechanistic insights further enable personalized risk assessments.

Supplementary Material

js9-112-1340-001.doc ^{(2.8MB, doc)}

Ethical approval

NHANES was approved by the Institutional Review Board of the National Center for Health Statistics, U.S.A. The CHARLS study received ethical approval from the Institutional Review Board of Peking University (approval no. IRB00001052-11015). Written informed consent was obtained from all the participants.

Consent

All participants provided written informed consent.

Sources of funding

This study was funded by Yunnan Province Science and Technology Department-Kunming Medical University Joint Project (No. 202201AY070001-294).

Author contributions

Conception and design: GS; Administrative support: GS; Provision of study materials or patients: GS; Collection and assembly of data: GS; Data analysis and interpretation: GS; Manuscript writing: GS; Final approval of the manuscript: GS.

Conflicts of interest disclosure

The author declares that there are no conflicts of interests.

Guarantor

Gao Song.

Research registration unique identifying number (UIN)

Not required, as de-identified CHARLS data were used.

Provenance and peer review

Not commissioned, externally peer-reviewed.

Data availability statement

The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request.