CardioMetAge estimates cardiometabolic aging and predicts disease outcomes

Abstract

Background

Existing aging clocks, designed to quantify biological aging, primarily capture systemic changes and may overlook alterations crucial for cardiometabolic diseases (CMDs).

Methods

In this study, we developed the CardioMetAge model, an aging clock tailored to predict CMD-related outcomes. Trained in the NHANES-III, the model was applied to the continuous NHANES and UK Biobank. Its associations with cardiometabolic mortality, disease incidence, and transitions between disease states were examined, and its performance in predicting 10-year CMD incidence was also evaluated. We further investigated associations of proteomic pathways, lifestyle factors, and socioeconomic status with CardioMetAge, as well as the impact of caloric restriction intervention on its change.

Results

The final CardioMetAge was constructed as a linear combination of chronological age and 12 common clinical biomarkers. Its age deviation (CardioMetAgeDev) showed stronger associations with CMD mortality (HR per SD [95% CI]: 1.87 [1.83, 1.91]), CMD incidence (1.35 [1.33, 1.37]), and disease progression, including transitions from no CMD to first CMD (1.34 [1.32, 1.35]) and from first CMD to cardiometabolic multimorbidity (1.25 [1.21, 1.30]), compared with deviations of PhenoAge and other traditional biological age models. CardioMetAge also consistently outperformed these models in predicting 10-year CMD incidence. Our findings also highlighted the biological determinants of cardiometabolic aging, with proteomic analyses linking CardioMetAgeDev to inflammatory activation and metabolic disorders. Analysis of modifiable factors revealed that lifestyle and socioeconomic status were associated with CMD risks, partly via CardioMetAgeDev (mediation proportions: 34.5% and 10.7%, respectively). Additionally, two-year caloric restriction slowed the progression of CardioMetAge by 1.23 years (95% CI: [0.61, 1.84]) relative to the ad libitum control.

Conclusions

CardioMetAge outperformed existing aging clocks in ease of use and in predicting CMD-related outcomes. It provides valuable insights into the mechanisms of cardiometabolic aging and holds potential for clinical monitoring and evaluating the effectiveness of interventions.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12916-026-04621-5.

Background

Aging is a complex biological process and a leading risk factor for chronic diseases and mortality [1]. Among these, cardiometabolic diseases (CMDs), such as ischemic heart disease (IHD), stroke, and diabetes, contribute substantially to global morbidity and mortality, imposing significant social and economic burdens [2, 3]. Recent advances in aging research have highlighted the importance of biological age (BA), a conceptual measure that reflects how an individual is aging internally, externally, and functionally [4]. Although BA itself cannot be directly quantified [5, 6], a growing array of molecular and clinical aging biomarkers has emerged as empirical indicators of aspects of biological aging. Among these, aging clocks integrating imaging techniques [7, 8], molecular [9, 10], and clinical [11, 12] data have been developed to approximate the cumulative impact of aging processes. These clocks capture the heterogeneity of biological aging among individuals of the same CA, revealing physiological differences beyond traditional biomarkers.

Modeling of existing aging clocks primarily focuses on CA prediction [10, 12] using machine learning methods, or all-cause mortality [13, 14] prediction, such as PhenoAge [14], with limited consideration of organ- or system-specific disease outcomes. While previous models utilized proteins associated with cognition to construct brain age and improve predictions for Alzheimer’s disease [10], gaps remain for CMD-related mortality and morbidity.

Aging clocks have been associated with molecular profiles [13, 15], lifestyle and socioeconomic status (SES) [9, 12] factors, yet these relationships remain insufficiently studied in the context of cardiometabolic aging. Additionally, caloric restriction (CR) represents a prototypical nutritional strategy with well-documented benefits for cardiometabolic health in humans [16]. However, whether CR can decelerate cardiometabolic aging as captured by an organ-specific biological aging clock remains largely unknown.

In this study, we present the CardioMetAge, a clinical aging clock designed to predict CMD outcomes and uncover factors contributing to cardiometabolic aging. To achieve this, we: (1) Developed the CardioMetAge model using clinical data from the National Health and Nutrition Examination Survey III (NHANES-III) and tested its predictive performance in both the continuous NHANES and the UK Biobank (UKB), designating UKB as the primary cohort for analysis. Evaluation of CardioMetAge focused on disease outcomes, including mortality, chronic disease incidence, and multistate transitions (e.g., from first CMD [FCMD] to cardiometabolic multimorbidity [CMM] and mortality) (2) Examined the associations of proteomic pathways, lifestyle, and SES factors with CardioMetAge, and evaluated the impact of CR as a potential intervention. This comprehensive approach aims to establish CardioMetAge as a robust tool for predicting CMD risks, advancing our understanding of aging mechanisms, and informing strategies for early prevention and personalized interventions.

Methods

Study population

Participants were recruited from three sources: the NHANES surveys in the United States [17, 18], the UKB [19], and the Comprehensive Assessment of Long-term Effects (CALERIE) trial [20]. The NHANES program was designed to assess the health and nutritional status of the U.S. population. Specifically, NHANES-III (1988–1994) [17] provided a nationally representative cross-sectional dataset, while the continuous NHANES [18] is an ongoing survey program; for this study, we used data collected between 1999 and 2018. The UKB [19] is a large-scale prospective cohort established to investigate genetic, environmental, and lifestyle determinants of disease in middle-aged adults, enrolling over 500,000 participants between 2006 and 2010 (Instance 0). A subset of participants was re-assessed during 2012–2013 (Instance 1). The CALERIE trial [20] is a randomized controlled study designed to evaluate the effects of CR on human health.

We developed the CardioMetAge model using the NHANES-III data as the training cohort, tested it using data from the continuous NHANES, and validated it in an independent external cohort, the UKB. The NHANES-III was chosen as the training dataset for several reasons: (1) It is widely used in aging research [14, 21, 22], as exemplified by the classic PhenoAge algorithm [14], which was also developed from this data, making it a well-established dataset for developing aging-related models. (2) It includes a broad age range, enhancing the model’s applicability across different age groups. The UKB was selected as the external validation cohort because it provides detailed longitudinal data on disease incidence and mortality, which allows for an in-depth investigation into CMD. For our primary analysis, the UKB was utilized as the principal dataset. Additionally, the CALERIE trial was incorporated to examine the effects of CR as an intervention on cardiometabolic aging.

In the two NHANES datasets, individuals who were pregnant and under age 20 were excluded (Additional file 1: Fig. S1). For the NHANES-III dataset, participants aged 90 years or older were excluded (Additional file 1: Fig. S1A). Concurrently, for the continuous NHANES dataset, age exclusion criteria were set to 85 years for data collected between 1999 and 2006, and to 80 years for data collected from 2007 onwards (Additional file 1: Fig. S1B). This decision was made because NHANES recorded the data for individuals exceeding certain age thresholds, obscuring their true ages and making the data non-representative and unsuitable for our age-specific analysis. We additionally excluded participants with missing survival status, deaths from unknown causes or accidents, or missing values in clinical factors required for constructing CardioMetAge and PhenoAge [14] (introduced in the following section). The final sample size included 13,262 participants in the NHANES-III cohort and 31,745 participants in the continuous NHANES cohort, serving as the training and testing cohorts, respectively (Additional file 1: Fig. S1; Additional file 1: Tables S1, S2).

In the UKB datasets, individuals who were pregnant, had missing values used for constructing CardioMetAge, and died from an accident in the follow-up were excluded, resulting in a pool of 418,118 participants with complete CardioMetAge data, which served as the basis for subsequent analyses (Additional file 1: Fig. S2; Additional file 1: Table S3).

The CALERIE Phase 2 [23] was a multi-center, randomized controlled trial assessing the effects of 25% caloric restriction over two years in healthy adults with a BMI range of 22 to 28 kg/m². A total of 220 participants were randomized in a 2:1 ratio to either a CR intervention group or an ad libitum (AL) control group. Among those who started intervention and had S_CardioMetAge (a substitute version retrained for CALERIE, introduced in the following section) data available at baseline and at least one follow-up, the analytic sample comprised 121 individuals for the CR group and 65 for the AL group (Additional file 1: Fig. S3; Additional file 1: Table S4).

Identification of key clinical factors

Candidate clinical factors in NHANES-III were established and classified into three categories. Skewness was assessed using the skewness() function from the e1071 R package, and variables with skewness > 2 were considered severely right-skewed [24] and subsequently log-transformed (Additional file 1: Fig. S4). Variables undergoing log-transformation were transformed using a log(x + 1) operation. Candidate biomarkers included:

1. Body composition metrics: body mass index (BMI), waist circumference (WC), waist-to-height ratio.

2. Blood biochemistry factors: glucose (log-transformed), glycohemoglobin (HbA1c), albumin, alkaline phosphatase (ALP), C-reactive protein (CRP, log-transformed), creatinine, blood urea nitrogen (BUN), uric acid (UA), Lymphocyte percent, mean corpuscular volume (MCV), white blood cell count (WBC), red cell distribution width (RDW), aspartate aminotransferase (log-transformed), alanine transaminase (log-transformed), high-density lipoprotein, triglycerides (log-transformed), total cholesterol, and triglyceride to high-density lipoprotein ratio (log-transformed).

3. Blood pressure and pulse rate: systolic blood pressure (SBP), diastolic blood pressure, pulse pressure (PP), mean blood pressure, pulse rate.

Based on these factors and CA, a least absolute shrinkage and selection operator (LASSO) Cox-penalized regression model with fivefold cross-validation was applied to predict mortality due to heart diseases, cerebrovascular diseases, and diabetes. The most regularized model was chosen such that its error was within one standard error of the minimum, ensuring robustness while maintaining relatively low model complexity by limiting the number of included variables. To further enhance model stability, a bootstrap procedure (50 resamples) was conducted, retaining variables that appeared in at least 30 iterations. The identified factors, in addition to CA, were HbA1c, RDW, SBP, creatinine, lymphocyte percent, MCV, pulse rate, PP, UA, CRP, WC, and BUN (Additional file 1: Fig. S5A).

Given the absence of HbA1c and RDW in the CALERIE dataset, these markers were excluded from the candidate biomarker pool. Consequently, the identified predictors for the S_CardioMetAge model, in addition to CA, were creatinine, SBP, WBC, pulse rate, glucose, PP, UA, lymphocyte percent, CRP, WC, ALP, and BUN (Additional file 1: Fig. S5B).

Construction of the CardioMetAge model

The CardioMetAge model was constructed by adapting the calculation method from the classic BA algorithm — PhenoAge [14]. Two Gompertz models were utilized in estimating the risk of mortality due to heart diseases, cerebrovascular diseases, and diabetes. The first model only used CA as a tool for prediction; the second one used CA and clinical factors. By equating the 10-year risk from both models, we then worked backward to solve for CA, thus deriving the predicted age. The detailed procedure for deriving the raw predicted age of the CardioMetAge formula was as follows:

1. Formulate the Gompertz cumulative distribution function (CDF) to estimate the 10-year cause-specific mortality risk, using CA as the predictor.

   $$CDF(10,CA)=1-exp(-exp(CA\ast {\beta }_0+c_0){{\gamma }_0}^{-1}(exp({\gamma }_{0}t)-1))$$ 1

2. Construct the Gompertz CDF for the 10-year cause-specific mortality risk, using the risk score (xb) as the predictor. Meanwhile, ‘xb’ designated the linear combination of CA and other risk factors.

   $$CDF(10,xb)=1-exp(-exp(xb){{\gamma }_1}^{-1}(exp({\gamma }_{1}t)-1))$$ 2
   $${xb=\beta }_1\ast CA+b\ast clinicalfactors+c_1$$

3. By setting (1) = (2), we were equating the Gompertz CDF for age to that for the xb at 10 years:

   $$\begin{array}{cc}& exp(CA\ast {\beta }_0+c_0){{\gamma }_0}^{-1}(exp({\gamma }_{0}t)-1)\hfill \\ \hfill & =exp(xb){{\gamma }_1}^{-1}(exp({\gamma }_{1}t)-1)\hfill \end{array}$$

4. Further transformation:

   $$exp\left(C,A,,,\ast ,,{\beta }_0,+,c_0,-,x,b\right)=\frac{{\gamma }_1^{-1}\left(e,x,p,\left({\gamma }_1,t\right),-,1\right)}{{\gamma }_0^{-1}\left(e,x,p,\left({\gamma }_0,t\right),-,1\right)}$$

5. Take the logarithm and solve for it:

   $$CA=\frac{1}{{\beta }_0}ln\frac{{{\gamma }_1}^{-1}(exp({\gamma }_{1}t)-1)}{{{\gamma }_0}^{-1}(exp({\gamma }_{0}t)-1)}-\frac{c_0}{{\beta }_0}+\frac{\mathit{xb}}{{\beta }_0}$$ 3

6. In (3), the ‘CA’ specifically represented the predicted age that we sought. Essentially, the raw predicted age of the CardioMetAge model was a linear combination of CA and risk factors.

For practical use, the primary model was fitted using biomarkers in their original measurement units. To present the relative importance of biomarkers, we additionally fitted a standardized model in which continuous biomarkers were transformed to z-scores (mean = 0, standard deviation [SD] = 1) and the same modelling pipeline was applied. The standardized model’s coefficients represent the effect per one SD change and are therefore comparable across biomarkers; this model is intended solely for interpretability, while the original-unit model is used for prediction (Fig. 1A).

Fig. 1.

The description and mortality associations of the CardioMetAge model. A Original coefficients were estimated using biomarkers in their original measurement units and are used for practical application. Scaled coefficients were obtained from a parallel model in which biomarkers were standardized (z-scores, mean = 0, SD = 1) prior to modeling; they are shown solely to illustrate the relative importance of biomarkers rather than for application. Log-transformed variables were computed as log(x + 1). B Scatter density plots illustrated the association between the raw predicted age of the CardioMetAge model and CA within both training (NHANES-III), test (continuous NHANES), and validation (UKB) cohorts. Each panel displayed the fit of the model (indicated by the black dotted line), annotated by Pearson correlation coefficients (r) and mean absolute errors (MAE). C The distribution of CardioMetAgeDev grouped by men and women across three cohorts. D-G Associations of age deviations with mortality risk. These forest plots illustrated the hazard ratios of various death causes associated with a one SD increase in CardioMetAgeDev or other metrics. D Comparison between CardioMetAgeDev and PhenoAgeDev and other age deviations derived from traditional CA-prediction models in the NHANES-III (n = 13,262). E Comparison of age deviations in the continuous NHANES (n = 31,745). F Comparison of CardioMetAgeDev with epigenetic age deviations of methylation-based PhenoAge (PhenoAgeMethyDev) and GrimAge2 (GrimAge2Dev), and the Dunedin pace of aging (DunedinPoAm) in the continuous NHANES (n = 2,062). G Comparison of age deviations in the UKB (n = 380,212). Abbreviation: HbA1c, glycated hemoglobin; RDW, red blood cell distribution width; SBP, systolic blood pressure; MCV, mean corpuscular volume; PP, pulse pressure; UA, uric acid; CRP, C-reactive protein; Waist Circ., waist circumference; BUN, blood urea nitrogen

Following the recommendations of previous research [25], we adopted ‘aging deviation’, also known as ‘age acceleration’. The CardioMetAgeDev, which indicated the aging deviation compared to individuals of the same CA, was calculated from the residual of the linear regression of the raw predicted age against CA. The CardioMetAge was calculated by adding CardioMetAgeDev to the CA. This procedure, which is commonly applied in aging clock models, effectively removes age-related bias, preventing the raw predicted age from being systematically over- or under-estimated due to its dependence on CA [26, 27]. The formulas could be shown as follows:

$$\widehat{Predictedage}=\widehat{\alpha }+\widehat{\beta }·CA,$$

$$CardioMetAgeDev=Predictedage-\widehat{Predictedage},$$

$$CardioMetAge=CA+CardioMetAgeDev,$$

where $\widehat{\beta }$ represented the parameters obtained from linear regression of the raw predicted age on CA, and $\widehat{\alpha }$ represented the intercept. All the age deviation metrics for their respective BAs mentioned in this study were calculated using the same procedure.

Calculation of other BA metrics

The PhenoAge model [14], constructed based on the previous study, based on biomarkers including albumin, creatinine, glucose, CRP, lymphocyte percent, MCV, RDW, ALP, and WBC. PhenoAge corresponds to the CA in the reference population that matches an individual’s 10-year risk of all-cause mortality.

The other three CA-prediction based models were using machine learning methods: Elastic Net, Random Forest, and XGBoost, which were frequently employed in BA estimations [11, 22, 28]. These models, developed on the NHANES-III cohort, utilized the same biomarkers as those in the CardioMetAge model.

The epigenetic metrics based on DNA methylation (DNAm) data from 1999–2002 were obtained from the NHANES website: https://wwwn.cdc.gov/nchs/nhanes/dnam/. These metrics included second-generation clocks: PhenoAge-methylation (PhenoAgeMethy) [14], GrimAge2 [29] and Dunedin Pace of Aging methylation (DunedinPoAm) [30]. PhenoAgeMethy is calculated by using DNAm sites to predict PhenoAge; GrimAge2 is a composite biomarker calculated as a weighted linear combination of ten DNAm proxies for human mortality risk. DunedinPoAm measures the pace of aging by analyzing DNAm sites linked to longitudinal biomarkers in individuals of the same CA.

Framingham risk score and cardiometabolic index

The Framingham Risk Score (FRS) is a classic and widely used algorithm for estimating an individual’s 10-year risk of developing cardiovascular disease based on traditional risk factors [31]. The 10-year FRS was derived by incorporating sex, age, triglycerides, high-density lipoprotein, SBP, use of blood pressure medication, smoking status, and presence of diabetes, following the methodology outlined in the previous study [31]. Cardiometabolic index (CMI) was computed as the product of the triglyceride-to-high-density lipoprotein ratio and the waist-to-height ratio, serving as a composite indicator of cardiometabolic risk [32].

Ascertainment of mortality

The survival data of NHANES-III and continuous NHANES were extracted from the NCHS Data 2019 Public-Use Linked Mortality Files at https://www.cdc.gov/nchs/data-linkage/mortality-public.htm. In these public-use files, causes of death are provided in a coarsely categorized form (e.g., heart diseases, cerebrovascular diseases, diabetes) rather than detailed ICD-10 codes. The corresponding definitions and ICD code ranges for each cause-specific mortality category are summarized in Additional file 1: Table S5. For the UKB, date and cause of death were obtained from death certificates held by the National Health Service Information Centre (England and Wales) and the National Health Service Central Register Scotland (Scotland). Detailed information about the linkage procedure is available at https://digital.nhs.uk/services. The definitions of cause-specific mortality used in this study, along with the corresponding ICD code ranges, are provided in Additional file 1: Table S6. The follow-up time was determined as the time from the baseline assessment to either the date of death or the censoring date (December 31, 2019, for the NHANES; October 31, 2022, for England; July 31, 2021, for Scotland; and February 28, 2018, for Wales), whichever occurred first. Participants who died from accidents were excluded from the survival analysis.

Ascertainment of CMD and other chronic diseases

In the NHANES-III and continuous NHANES, chronic diseases, including diabetes, hypertension, cardiovascular diseases (including angina, congestive heart failure, coronary heart disease, heart attack, and stroke), and kidney diseases were self-reported by the study participants. Disease-free participants referred to individuals who did not suffer from any of the aforementioned diseases at baseline.

In the UKB, a spectrum of chronic disease categories was identified by leveraging self-reported data and hospital records, including Parkinsonism, multiple sclerosis, stroke, dementia, depression, bipolar disorder, schizophrenia, IHD, hypertensive diseases, chronic obstructive pulmonary disease, chronic kidney disease, diabetes, cirrhosis, osteoarthritis, osteoporosis, and cancer. These categories encompass all causes and sub-types, with definitions adhering to the previous study [12], and corresponding ICD codes are detailed in Additional file 2: Table S7. For this study, CMD was defined as the presence of any of three major conditions, namely IHD, stroke, and diabetes, consistent with previous publications [33, 34]. Person-time was calculated from baseline to the occurrence of a certain disease, censor events (death), or the end of follow-up (October 30, 2022 [England], August 31, 2022 [Scotland], and May 31, 2022 [Wales]), whichever came first. Disease-free participants also referred to individuals who were free from all these diseases at baseline.

Assessment of covariates

Based on previous studies, covariates included in this study were age, sex (man, woman), ethnicity (white, non-white), education (college or university degree; A levels/AS levels or equivalent; O levels/GCSEs or equivalent; CSEs or equivalent; NVQ or HND or HNC or equivalent; other professional qualifications; other), income (less than £18,000; £18,000–£30,999; £31,000–£51,999; £52,000–£100,000; greater than £100,000), employment status (employed, not employed), physical activity (high, moderate, low), smoking status (current, previous, never), alcohol consumption status(current, previous, never), diet score (healthy, unhealthy; calculation seen in the Additional file 1: Tables S8, S9), BMI, use of cholesterol-lowering and blood pressure medications (yes, no), and family history of CMD (yes, no; whether any parent or sibling had diabetes, stroke, or heart diseases). The missing rates for ethnicity, education, smoking, alcohol, cholesterol-lowering medication, blood pressure medication, and BMI were all below 1%, and these missing values were imputed using the mode or median as appropriate. For physical activity, income, diet score, and family disease history, the missing rates ranged from 9 to 20%, with missing values being assigned an independent category.

CMD progression trajectories

We used continuous-time multistate models to characterize the progression of CMD and its subcomponents over time. Two complementary trajectories were constructed to capture the natural history of cardiometabolic aging, consistent with previous publications [33, 34]. The first, a four-state CMD trajectory, comprised four mutually exclusive states representing sequential disease progression: (1) non-CMD, defined as being free of IHD, stroke, and diabetes at baseline; (2) FCMD, defined as the first occurrence of any CMD (IHD, stroke, or diabetes); (3) CMM, defined as the coexistence of two or more of these conditions; and (4) all-cause death, considered an absorbing terminal state. Participants were allowed to transition along five possible pathways: from non-CMD to FCMD, FCMD to CMM, non-CMD to death, FCMD to death, and CMM to death. The second trajectory, a disease-specific six-state model, was developed to delineate transitions related to individual CMD components. It included six distinct states: (1) non-CMD, (2) IHD, (3) stroke, (4) diabetes, (5) CMM, and (6) all-cause death. Possible transitions included non-CMD to IHD, stroke, or diabetes (representing FCMD onset), IHD to CMM or death, stroke to CMM or death, diabetes to CMM or death, and CMM to death. Both models were fitted using the mstate package in R, with time since baseline as the underlying time scale.

Plasma proteins and enrichment analysis

Proteins were measured using the Olink Explore 3072 platform, covering 2,923 unique proteins across eight panels [35]. After excluding samples with 1000 missing values and proteins with > 10% missing values, 2,916 proteins remained. Missing proteins were imputed using k-nearest neighbors, resulting in 37,942 samples with complete CardioMetAgeDev data. The z-score transformation was applied to the protein data.

Gene set enrichment analysis (GSEA) was performed to evaluate the enrichment of predefined protein sets based on proteins ranked by their partial correlation coefficients with CardioMetAgeDev, adjusted for age and sex. Enrichment Scores were calculated using the adaptive multilevel splitting Monte Carlo method implemented in the R package fgsea, with normalized enrichment scores (NES) representing the enrichment strength normalized for gene set size. Corresponding P values were calculated using permutation tests. The gene sets were obtained from MSigDB at https://www.gsea-msigdb.org/gsea/msigdb/collections.jsp, including collections from Gene Ontology: Biological Process, KEGG, and Reactome.

Lifestyle and SES scores

Each lifestyle factor (diet score, smoking, alcohol consumption, and physical activity) and SES variable (household income, education level, and employment status) was assigned a score, where higher values represented greater frequency for lifestyle factors and better status for SES variables. The diet score was calculated following rules derived from a prior study, which summarized multiple diet components based on updated recommendations for cardiovascular health [36] (Additional file 1: Tables S8, S9). Individual food group scores were computed separately, with higher scores indicating greater intake of specific food groups. These food groups were assessed using a 24-h dietary recall. Total diet calorie intake refers to the energy derived from the overall diet, estimated based on the 24-h dietary recall. This value was adopted as the original measure and expressed in kilojoules (kJ). The overall healthy lifestyle score was calculated as the sum of points assigned to each factor, where a healthy level received 1 point and an unhealthy level received 0 points. This score ranged from 0 to 4, with higher scores reflecting healthier lifestyles. The overall SES score was calculated by combining three SES variables, classified into low (1 point), medium (2 points), and high (3 points) (Additional file 1: Table S10), based on the previous study using latent class analysis [36]. Detailed scoring rules are provided in Additional file 1: Tables S8 to S11.

Statistical analysis

In the UKB, repeated measurements of BA indicators were available at two time points (instance 0 and instance 1), with a median interval of 4.5 years between assessments. To approximate the required sample size for detecting a true difference, we calculated the within-individual differences between the two instances and obtained their SD. The standardized effect size was defined as the expected mean change divided by the within-individual SD. Power analysis was then performed using the pwr.t.test() function in the R package pwr, assuming a two-sided paired t-test with a significance level of 0.05 and 80% power. Because the within-individual variability incorporates both technical noise and true biological variation over time, this procedure yields a conservative estimate of the sample size needed to detect meaningful changes beyond measurement noise.

The intraclass correlation coefficient (ICC) for repeated measurements was calculated to quantify the proportion of total variance attributable to between-subject differences. It was estimated using the icc() function from the R package performance, based on a linear mixed-effects model with a random intercept for each participant.

Agreement between two versions of the CardioMetAge metric (CardioMetAgeDev and S_CardioMetAgeDev) was assessed using the ICC computed by the ICC() function in the psych R package. The ICC(3,1) model [37], corresponding to a two-way random-effects model for absolute agreement of single measures, was used to assess the reliability between measurements.

For cause-specific death, we used R package cmprsk to fit cause-specific Cox proportional hazard regression models in considering competing risk, adjusting for age and sex. Cox proportional hazards regression models were fitted using the R package survival to examine the associations of CardioMetAgeDev and other reference BA metrics with all-cause mortality and incident chronic diseases, adjusting for age and sex. For 10-year CMD prediction models, we calculated the area under the receiver operating characteristic curve (AUC) for each model and applied DeLong’s test for pairwise comparisons of AUCs.

A multi-state model using the Markov proportional hazards framework [38], an extension of competing risks survival analysis, was conducted to evaluate the association of BA metrics with CMD progression. Transition-specific Cox models were fitted using the mstate package, allowing the estimation of hazard ratios (HRs) for each phase of CMD progression. Given the computational complexity of multi-state modeling, the analyses adjusted for a comprehensive set of covariates, including age, sex, education, income, physical activity, smoking, alcohol, diet score, BMI, cholesterol-lowering or blood pressure medication, and family history of CMD, while ensuring that model dimensionality remained manageable.

Mediation analysis was conducted by the R package CMAverse. A linear regression model assessed the association between exposure (lifestyle factors, SES factors, calorie intake) and mediators (CardioMetAgeDev), adjusting for covariates, while a Cox regression model evaluated the mediators’ impact on event risk, controlling for exposure and covariates. The total effect (TE) was decomposed into direct (DE) and indirect effects (IDE), with IDE indicating the mediator-explained exposure impact on event risk, and DE reflecting the exposure effect on event. Nonparametric bootstrapping (100 reps) estimated 95% confidence intervals (CIs) and P values, and the proportion mediated was calculated as (DE*(IDE-1)/[TE-1]) [39]. The common covariates included age, sex, ethnicity, BMI, cholesterol-lowering medication, blood pressure medication, and family history of CMD. Other exposure-specific covariates referred to lifestyle and SES factors beyond the exposure factor itself, which were listed in the Additional file 1: Table S12.

Intention-to-treat analysis tested the effect of randomization to treatment groups (CR vs. AL) on change in S_CardioMetAge using repeated-measures analysis of covariance within a mixed-model framework, consistent with previous CALERIE analyses [40, 41]. The model was fit using the R package lmerTest, and treatment effects were assessed via Type III Analysis of Variance with Satterthwaite’s method for estimating degrees of freedom. The mixed-model included terms for treatment group (${\beta }_1$), follow-up time (${\beta }_2$), and their interaction (${\beta }_3$) to assess heterogeneity in treatment effects across time points. Covariates containing baseline age, sex, race, BMI stratum, and study site (${\beta }_i$) were included to control for potential confounding, alongside baseline S_CardioMetAge (${\beta }_4$) to adjust for regression-to-the-mean effects. The model also included the fixed intercept (${\beta }_0$) to represent the average change, random intercepts ($\mathrm{\gamma }$) to account for individual-level variability, and residual errors ($\mathrm{\epsilon }$) for unexplained variance. The model was represented as:

$$\begin{array}{cc}& Changeof\mathrm{S}\_\mathrm{CardioMetAge}={\beta }_0+{\beta }_1\mathit{Group}\hfill \\ \hfill & +{\beta }_2\mathit{Visit}+{\beta }_3\left(\mathit{Group},\times ,\text{Visit}\right)\hfill \\ \hfill & +{\beta }_{4}Baseline\text{S\_CardioMetAge}+\sum {\beta }_i{x}_i+\gamma +\epsilon \hfill \end{array}$$

Results

Development of the CardioMetAge model

In the NHANES-III, 12 common clinical factors (Fig. 1A) were identified as key factors for the risks of death from heart diseases, cerebrovascular diseases, and diabetes. Using these factors, the CardioMetAge model was trained to estimate an individual’s predicted age as the CA at which their predicted 10-year mortality risk for the three diseases matches the average risk in a reference population. The expression of CardioMetAge was as follows, with the units of these variables detailed in Fig. 1A:

$$\begin{array}{cc}\hfill \mathit{CardioMetAgePredict}& =0.831320\times age+19.5734\times \text{log}(HbA1c+1)+1.77394\times RDW+0.0760217\times SBP+6.18803\times creatinine\hfill \\ \hfill & -0.148076\times lymphocytepercent+0.218946\times MCV+0.105980\times pulserate+0.0603608\times PP+0.636711\hfill \\ \hfill & \times UA+2.40001\times \text{log}(CRP+1)+0.0283277\times WC+0.0754119\times BUN-101.454\hfill \end{array}$$

Based on standardized coefficients, HbA1c, RDW, and SBP were identified as the most influential predictors (Fig. 1A). Among 13,262 participants in the NHANES-III (Additional file 1: Fig. S1; Additional file 1: Table S1), 31,745 in the continuous NHANES (Additional file 1: Fig. S1; Additional file 1: Table S2), and 418,118 in the UKB (Additional file 1: Fig. S2; Additional file 1: Table S3) with complete CardioMetAge data, we found strong correlations between the predicted CardioMetAge and CA, with coefficients of 0.970, 0.961, and 0.889, and mean absolute errors (MAEs) of 4.170, 4.070, and 3.705 years, respectively (Fig. 1B). In the NHANES-III, the overall median CardioMetAgeDev was − 0.4 years (inter-quartile range [IQR] = − 3.4 to 2.9), with men exhibiting higher values (median = 1.2 years, IQR = − 1.5 to 4.2) than women (median = − 1.9 years, IQR = − 4.6 to 1.2) (Fig. 1C; Additional file 1: Table S1). This disparity aligned with trends observed across all the cohorts (Fig. 1C; Additional file 1: Fig. S6; Additional file 1: Tables S1 to S3) and supported existing literature on sex differences in biological aging [13, 42].

Direct assessment of technical noise was not feasible, as replicate assays were unavailable in these cohorts. As an alternative, test–retest variability was approximated using repeated phenotypic measurements at two time points (instance 0 and instance 1) with a median follow-up interval of 4.5 years in the UKB (n = 7,933). For CardioMetAgeDev, the ICC was 0.730 (95% CI: 0.720 to 0.739) and the within-individual SD of the differences between the two instances was 2.97 years (Additional file 1: Fig. S7A, S7B). Based on this variability, the estimated sample sizes required to detect mean longitudinal changes of 0.25, 0.5, 1.0, and 2.0 years with 80% power (two-sided α = 0.05) were approximately 1113, 280, 72, and 20 participants, respectively (Additional file 1: Fig. S7C).

Association of CardioMetAge with all-cause and cause-specific mortality

We compared CardioMetAgeDev with other age deviation metrics regarding their associations with all-cause and cause-specific mortality risks across three cohorts. After adjustment for age and sex, CardioMetAgeDev exhibited higher HRs per SD for mortality due to heart diseases, cerebrovascular diseases, diabetes, and nephritis, compared to PhenoAge deviation (PhenoAgeDev) in the NHANES-III and the continuous NHANES (Fig. 1D, 1E; Additional file 2: Table S13). In the continuous NHANES, where DNA methylation data were available, similar results were observed when comparing CardioMetAgeDev with age deviations of epigenetic age (methylation-based PhenoAge [14] [PhenoAgeMethyDev], and GrimAge2 [29] [GrimAge2Dev]) and the DunedinPoAm [43](Fig. 1F; Additional file 2: Table S14). In the UKB, CardioMetAgeDev (HR per SD [95% CI]: 1.866 [1.825, 1.907]) also showed a stronger association compared to PhenoAgeDev (1.659 [1.628, 1.690]) with CMD mortality, including IHD, stroke, and diabetes (Fig. 1G; Additional file 2: Table S13). For non-CMD outcomes, such as respiratory diseases, CardioMetAgeDev did not consistently outperform PhenoAgeDev and epigenetic metrics (Fig. 1D-1G; Additional file 2: Tables S13, S14). However, it generally demonstrated higher HRs with all cause and cause-specific mortality compared to traditional machine-learning models, including Elastic Net, Random Forest, and XGBoost (Fig. 1D, E, G; Additional file 2: Table S13). These results remained robust when analyses were restricted to all included chronic disease-free individuals (Additional file 1: Fig. S8; Additional file 2: Table S15).

Prediction of CMD incidence by CardioMetAge

To evaluate the relative utility of different aging clocks in predicting chronic disease risk, we compared CardioMetAgeDev with the age deviations of other traditional BA models. Analyses were performed for each disease outcome separately, excluding participants with baseline occurrences of each specific disease. After adjustment for age and sex, CardioMetAgeDev (HR per SD [95% CI]: 1.350 [1.334, 1.367]) demonstrated stronger associations with CMD incidence than PhenoAgeDev (1.236 [1.222, 1.251]) (Fig. 2A; Additional file 2: Table S16), supporting its design as a CMD-specific aging metric. The same trend was observed for hypertensive diseases, chronic kidney disease, and cirrhosis, whereas the associations with other conditions, such as chronic obstructive pulmonary disease and depression, were weaker. Compared to models based on Elastic Net, Random Forest, and XGBoost, CardioMetAgeDev demonstrated higher HRs per SD for nearly all considered diseases (Fig. 2A; Additional file 2: Table S16). In the participants free from all included chronic diseases at baseline, a similar trend was observed (Additional file 1: Fig. S9; Additional file 2: Table S16).

Fig. 2.

Associations of age deviations with chronic disease incidence risk and predictive comparison of models for 10-year CMD incidence. A Associations of CardioMetAgeDev and age deviations of other traditional BAs with the incidence risk of chronic diseases in each specific disease-free participant of UKB. B-I Receiver operator characteristic curves demonstrated the model’s predictive performance for the 10-year incidence of IHD (Panels B and F), stroke (Panels C and G), diabetes (Panels D and H), and composite CMD (Panels E and I). The area under the receiver operating characteristic curve (AUC) values compared the predictive power of the model using CardioMetAge versus other indices or risk factors. Traditional risk factors in the panel (Panels F-I) included age, sex, ethnicity, education, income, employment status, physical activity, smoking, alcohol, diet score, BMI, cholesterol-lowering medication, blood pressure medication, and family history of CMD. Abbreviation: CMI, cardiometabolic index; FRS, Framingham risk score

We further conducted 10-year incidence risk predictions for IHD, stroke, diabetes, and composite CMD, comparing CardioMetAge with CA, other traditional BAs, the classic FRS [31], the recently widely adopted CMI [32], and other traditional risk factors. Among all age-based measures, CardioMetAge consistently achieved the highest AUC values, outperforming CA, PhenoAge, and models based on Elastic Net, Random Forest, and XGBoost, across all evaluated outcomes (Fig. 2B-E; Additional file 2: Table S17). For all these diseases, traditional risk factors (age, sex, education, income, employment, physical activity, smoking, alcohol, diet score, BMI, cholesterol-lowering medication, blood pressure medication, family history of CMD) yielded the highest AUCs (IHD: AUC = 0.725; stroke: 0.716; diabetes: 0.804; composite CMD: 0.732), with the inclusion of CardioMetAge alongside these factors led to a notable improvement in predictive accuracy (IHD: AUC = 0.729, Delong test P = 2.8 × 10^–25; stroke: 0.723, P = 3.5 × 10^–12; diabetes: 0.830, P = 2.0 × 10^–132; composite CMD: 0.743, P = 8.4 × 10^–90) (Fig. 2F-I; Additional file 2: Table S17). While CardioMetAge showed lower predictive performance than the FRS for IHD, it significantly outperformed FRS in predicting stroke and diabetes. Compared with the CMI, CardioMetAge demonstrated weaker prediction for diabetes but stronger discrimination for IHD, stroke, and composite CMD.

The associations of CardioMetAge with CMD trajectories

The associations between CardioMetAgeDev and transitions in the CMD progression were further characterized. Among 343,973 participants free of CMD at baseline, 39,440 (11.5%) developed FCMD, and 4,246 (10.8%) progressed to CMM, with a total of 21,576 deaths occurred (Fig. 3A). Disease-specific trajectories revealed heterogeneous patterns of progression: 22,682 (6.6%) participants initially developed IHD, 5,565 (1.6%) developed stroke, and 11,193 (3.3%) developed diabetes. Of these, 2,533 (11.2%), 534 (9.6%), and 1,179 (10.5%) subsequently progressed to CMM, respectively (Fig. 3B). Deaths occurred in 10.0% of those with IHD, 22.8% with stroke, 10.8% with diabetes, and 20.2% with CMM.

Fig. 3.

Associations of CardioMetAge with CMD trajectories. A-B CMD progression was depicted in two complementary trajectories, with the first trajectory describing the overall CMD process (Panel A), and the second focused on disease-specific pathways (Panel B). Numbers (percentages) of participants in the transition pattern were shown for each trajectory. C-D Hazard ratios per standard deviation (SD) increase for transitions between CMD states. Transition-specific Cox models were adjusted for age, sex, ethnicity, education, income, physical activity, smoking, alcohol, diet score, BMI, cholesterol-lowering/blood pressure medication, and family history of CMD

In the multi-state models, compared with PhenoAgeDev and other metrics, the associations of CardioMetAgeDev with both first disease onset and multimorbidity progression were consistently stronger. Each 1-SD increase (3.7 years) in CardioMetAgeDev was associated with a 33.8% higher risk of transitioning from non-CMD to FCMD (HR per SD [95% CI]: 1.338 [1.324, 1.353]) and a 25.2% higher risk of progressing from FCMD to CMM (1.252 [1.209, 1.298]), after adjusting for traditional risk factors (Fig. 3C; Additional file 2: Table S18). However, its associations with later transitions, such as from CMM to death, were relatively attenuated (1.155 [1.078, 1.238]). When specific CMDs were modeled separately, CardioMetAgeDev was most strongly related to the onset of diabetes (1.746 [1.713, 1.780]), followed by stroke (1.253 [1.217, 1.291]) and IHD (1.177 [1.159, 1.194]) (Fig. 3D; Additional file 2: Table S18). Furthermore, it remained robust associations with progression from IHD (1.354 [1.288, 1.425]), stroke (1.200 [1.097, 1.312]), and diabetes (1.098 [1.031, 1.170]) to CMM.

Proteins and enriched pathways linked to CardioMetAge

To elucidate the biological mechanisms underlying cardiometabolic aging, we explored the associations between CardioMetAgeDev and circulating proteins. In 37,942 participants with Olink plasma proteins available, out of 2,916 proteins analyzed, 2,183 exhibited significant positive correlations (P < 0.05, false discovery rate [FDR] corrected) with CardioMetAgeDev, while 303 showed significant negative correlations (Fig. 4A; Additional file 2: Table S19). Among positively correlated proteins, HGF (Hepatocyte Growth Factor) is known for its role in maintaining cardiac homeostasis, mitigating oxidative stress in normal cardiomyocytes, and providing cardioprotection in injured hearts [44]. FABP4 (Fatty Acid-Binding Protein 4), associated with lipid metabolism and insulin resistance [45], also showed a strong positive correlation. Likewise, IL1RN (Interleukin-1 Receptor Antagonist) was positively associated, reflecting compensatory anti-inflammatory activity in response to chronic metabolic stress [46]. Conversely, negatively correlated proteins include PON3 (Serum Paraoxonase 3) [47], recognized for its antioxidant properties, and SHBG (Sex Hormone-Binding Globulin) [48], which regulates hormone bioavailability and glucose metabolism. These findings underscore the relevance of inflammation, lipid metabolism, and hormonal balance in influencing CardioMetAgeDev.

Fig. 4.

Associations of plasma proteins and pathways with CardioMetAgeDev. A Top 10 plasma proteins with the strongest positive (red) and negative (blue) partial correlations with CardioMetAgeDev, adjusted for age and sex (n = 37,942). B-C Top 15 significantly (two-sided permutation test; P-value after false discovery rate correction, P_adj < 0.05) upregulated (NES > 0) (Panel B) and all 8 significantly downregulated (NES < 0) (Panel C) pathways identified by Gene set enrichment analysis using partial correlation between plasma proteins and CardioMetAgeDev (n = 37,942). Statistical significance was denoted as follows: * P < 0.05, ** P < 0.01, *** P < 0.001. Abbreviation: GOBP, gene ontology biological process

GSEA analysis revealed upregulation of 61 pathways and downregulation of 8 pathways (P < 0.05, FDR corrected) (Fig. 4B, C; Additional file 2: Table S20). The most significantly enriched pathways with positive NES were primarily involved in innate and humoral immunity, host defense, and inflammation. Key hits included ‘neutrophil degranulation’ (NES = 1.53, P = 5.8 × 10⁻⁵), ‘defense response’ (NES = 1.33, P = 5.8 × 10⁻⁵), and ‘innate immune system’ (NES = 1.36, P = 7.4 × 10⁻⁴), indicating enhanced neutrophil activity and innate signaling. Additional enrichment in ‘humoral immune response’, ‘response to bacterium’, and ‘positive regulation of leukocyte migration’ reflects upregulated systemic inflammation and host–pathogen interactions. Notably, ‘regulation of insulin-like growth factor transport and uptake by IGF binding proteins’ (NES = 1.56, P = 5.3 × 10⁻³) may reflect crosstalk between immune activation and metabolic signaling. Processes related to vascular remodeling, such as ‘blood vessel morphogenesis’ (NES = 1.35, P = 7.8 × 10⁻³), were also significantly enriched. Conversely, pathways with negative NES were mainly related to extracellular matrix (ECM) organization and metabolic homeostasis. The most downregulated pathways included ‘chondroitin sulfate biosynthesis’, ‘dermatan sulfate biosynthesis’, and defective glycosyltransferase-related processes such as ‘defective CHST14 causes EDS musculocontractural type’, ‘defective CHST3 causes SEDCJD’, and ‘defective CHSY1 causes TPBS’ (each NES = –2.49, P = 4.5 × 10⁻³), all critical for glycosaminoglycan biosynthesis and ECM integrity. Metabolic and ionic regulation pathways, including ‘glucocorticoid metabolic process’ (NES = –2.39, P = 1.3 × 10⁻²) and ‘potassium ion homeostasis’ (NES = –2.33, P = 2.2 × 10⁻²), were also negatively enriched, indicating potential endocrine and ionic dysregulation.

Associated lifestyle and SES factors for CardioMetAge

We further examined how lifestyle and SES factors relate to CardioMetAgeDev, aiming to identify potential modifiable determinants of cardiometabolic aging. In 255,250 non-CMD participants with complete lifestyle and SES factors, both lifestyle score (r = − 0.109, P < 2.23 × 10⁻³⁰⁸) and SES score (r = − 0.108, P < 2.23 × 10⁻³⁰⁸) were significantly negatively associated with CardioMetAgeDev (adjusted for age and sex, FDR corrected) (Fig. 5A; Additional file 1: Table S21). Specifically, individuals with a healthier diet (r = − 0.105, P < 2.23 × 10⁻³⁰⁸) exhibited lower CardioMetAgeDev. Among dietary components, higher intakes of fruits, vegetables, whole grains, fish, and dairy were associated with slower aging, whereas higher consumption of oils, refined grains, and processed/unprocessed meats was linked to accelerated aging. Additionally, greater total diet calorie intake was linked to increased CardioMetAgeDev (r = 0.018, P = 4.3 × 10^–4, n = 38,302). Smoking (r = 0.050, P = 2.9 × 10⁻¹⁴¹) and alcohol (r = 0.026, P = 8.2 × 10⁻³⁶) were positively correlated with CardioMetAgeDev, while physical activity (r = − 0.072, P = 7.1 × 10⁻²⁸⁸) was inversely correlated with it. Regarding SES factors, higher levels of education (r = − 0.104, P < 2.23 × 10⁻³⁰⁸), household income (r = − 0.102, P < 2.23 × 10⁻³⁰⁸), and being employed (r = − 0.048, P = 9.7 × 10⁻¹²⁸) were all associated with lower CardioMetAgeDev.

Fig. 5.

Associated lifestyle and SES factors and their mediation through CardioMetAgeDev in CMD outcomes. A The bubble plot visualized the partial correlation coefficients between exposure factors (two-sided t test; P < 0.05, after false discovery rate correction) and CardioMetAgeDev in non-CMD participants, adjusted for age and sex (n = 38,302 for calorie intake; n = 255,250 for all other factors). The lifestyle score combined diet score, smoking, alcohol consumption, and physical activity, while the socioeconomic status (SES) score encompassed income, education level, and employment status. B The bar chart illustrates the proportion mediated (PM; all nonparametric bootstrapping procedures yielded P values < 2 × 10^–16) by CardioMetAgeDev when serving as a mediator between exposure factors and CMD outcomes, including CMD incidence and mortality. These models were adjusted for common covariates, including age, sex, ethnicity, BMI, cholesterol-lowering medication, blood pressure medication, family history of CMD, as well as other lifestyle and SES factors beyond the exposure factors themselves

Mediation analysis was further conducted to explore the relationships among CardioMetAgeDev-associated factors, CardioMetAgeDev itself, and CMD outcomes, which encompassed the incidence and mortality of CMD. It revealed that a healthier lifestyle and better SES were associated with lower CMD outcomes risk, with 34.5% and 10.7% of the respective effects mediated through changes in CardioMetAgeDev, after adjusting for other risk factors (Fig. 5B; Additional file 2: Table S22). Specifically, CardioMetAgeDev mediated 26.9% of the effect of adequate physical activity and 32.8% of the effect of a healthier diet on risk reduction, while smoking and alcohol consumption accelerated CardioMetAgeDev, mediating 6.0% and 3.1% of their effects on higher risk, respectively. Additionally, CardioMetAgeDev explained 7.7% ~ 12.7% of the protective effect of the SES score and its components against CMD outcomes risk.

Effect of caloric restriction intervention on CardioMetAge

A hallmark of effective aging clocks is their responsiveness to intervention effects. Leveraging this insight, we investigated the CardioMetAge model’s reaction to CR as an intervention strategy using data from the CALERIE phase 2 [49]. The final expression of S_CardioMetAge, the substitute version retrained for CALERIE, was as follows, with the units of these variables detailed in Fig. 6A:

Fig. 6.

S_CardioMetAge model in the CALERIE trial. A Original coefficients for S_CardioMetAge were estimated using biomarkers in their original measurement units and are used for practical application. Scaled coefficients were obtained from a parallel model in which biomarkers were standardized (z-scores, mean = 0, SD = 1) prior to modeling; they are shown solely to illustrate the relative importance of biomarkers rather than for application. Log-transformed variables were computed as log(x + 1). B It displayed the predicted age of the S_CardioMetAge model at three visits (baseline, 12 months, 24 months) for the caloric restriction (CR, n = 121) and ad libitum (AL, n = 65) control groups in the CALERIE trial, with each individual's measurements connected by a dashed line. C Trajectories of CardioMetAge for the CR and AL groups over three visits. It illustrated the mean changes from baseline values, along with their respective 95% confidence intervals, as estimated by the linear mixed model, which adjusted for baseline CA, baseline CardioMetAge, sex, BMI stratum, site, and race. P value for the group effect was obtained from a Type III analysis of variance in the mixed-effects model (two-sided). Abbreviation: ALP, alkaline phosphatase; BUN, blood urea nitrogen; CALERIE, Comprehensive Assessment of Long-term Effects of Reducing Intake of Energy; cor, Pearson correlation coefficient; CRP, C-reactive protein; MAE, mean absolute error; n, number of participants; PP, pulse pressure; SBP, systolic blood pressure; UA, uric acid; Waist Circ., waist circumference; WBC, white blood cell count

$$\begin{array}{cc}\hfill \text{S\_CardioMetAgePredict}& =0.871849\times age+7.47627\times creatinine+0.0768106\times SBP+0.559944\times WBC\hfill \\ \hfill & +0.107509\times pulserate+5.61992\times \text{log}(glucose+1)+0.0511339\times PP+0.457352\hfill \\ \hfill & \times UA-0.0734294\times lymphocytepercent+2.64485\times \text{log}(CRP+1)+0.0388157\hfill \\ \hfill & \times WC+0.0180186\times ALP+0.0926833\times BUN-58.5502\hfill \end{array}$$

where creatinine, glucose, and PP were identified as the most influential predictors (Fig. 6A). The Bland–Altman analysis confirmed agreement between the original and S_CardioMetAgeDev (bias: 5.03 × 10^–16, 95% CI: − 4.22 × 10⁻² to 4.22 × 10⁻²), with an ICC of 0.92. While the association of S_CardioMetAgeDev with CMD-related mortality and incidence was weaker than that of the original CardioMetAgeDev, it remained stronger than that of PhenoAgeDev (Additional file 1: Fig. S10; Additional file 2: Table S23).

Among CALERIE participants with S_CardioMetAge available from baseline and at least one follow-up (12 months and 24 months), the sample sizes were 121 for the CR group and 65 for the AL group (Fig. 6B; Additional file 1: Fig. S3; Additional file 1: Table S4). The predicted age from the S_CardioMetAge model closely aligned with CA, with a correlation of 0.919 and a MAE of 6.485 (Fig. 6B). The ICC across three time points was 0.957 (95% CI: 0.940 to 0.966) for S_CardioMetAge, and 0.670 (95% CI: 0.596 to 0.719) for S_CardioMetAgeDev, reflecting moderate consistency as expected for age-deviation indices [50]. As expected, participants’ S_CardioMetAge tended to increase over time (Fig. 6C). Notably, the increase was significantly greater in the AL group than in the CR group (ANOVA, F[1, 175.5] = 23.0, P = 3.41 × 10⁻⁶) (Additional file 1: Table S24). The least-squares mean difference between groups was 1.32 (95% CI: 0.70 to 1.94) years at 12 months and 1.23 (95% CI: 0.61 to 1.84) years at 24 months (Fig. 6C; Additional file 1: Table S25), indicating a consistently slower progression in the CR group.

Discussion

In this work, we developed a clinical CardioMetAge model tailored for predicting CMD-related outcomes and understanding cardiometabolic aging. With its concise form and easily available indicators, the CardioMetAge model underwent training, testing, and validation across a total of three cohorts and one intervention. CardioMetAgeDev exhibited stronger associations with CMD-related mortality, incidence, and multi-state transitions, compared to PhenoAgeDev and other traditional metrics. Its biological and lifestyle determinants were also characterized. Proteomic profiling indicated activation of immune and inflammatory pathways alongside downregulation of ECM organization and metabolic homeostasis pathways. Additionally, the associations of a healthier lifestyle and higher SES with lower CMD risk were partly mediated by CardioMetAgeDev. The potential for interventions was further highlighted by the CR trial, which slowed the progression of CardioMetAge compared to an AL diet.

The CardioMetAge model consolidates multiple risk measures into an age scale, providing a specific numerical value that indicates an individual’s cardiometabolic aging status relative to their peers of the same CA. Unlike many traditional risk or BA models that often lack interpretability, CardioMetAge is transparent and user-friendly, clearly illustrating how each variable contributes to BA calculation. For instance, the positive coefficient of SBP aligns with its established association with increased cardiometabolic risk, enhancing clinical intuitiveness. By combining robust predictive performance with a concise and interpretable structure, CardioMetAge provides a distinct perspective on cardiometabolic health within the BA spectrum, complementing traditional risk models like the FRS. Together, these indicators serve as complementary tools for assessing cardiometabolic health, with CardioMetAge excelling in reflecting aging status and offering unique advantages in identifying targets for clinical intervention [25]. Its intuitive design also encourages individuals to adopt healthier lifestyle choices, highlighting its potential for widespread application in clinical practice.

Early BA models were primarily designed as supervised predictions of CA, but as their precision in fitting CA improved, the ability to capture meaningful information beyond CA diminished [51, 52]. This limitation led to the emergence of second-generation BA models focusing on predicting future health outcomes, predominantly mortality, with enhanced accuracy in disease risk prediction [53]. Despite advancements, most second-generation clocks, such as PhenoAge and epigenetic BAs, focus on overall aging and fail to adequately capture organ- or system-specific changes. In this study, the modeling of CardioMetAge specifically targets CMD mortality, thus demonstrating superior predictive performance for CMD outcomes compared to other BAs. This modeling approach could be generalized to other studies focusing on organ- or system-specific outcomes, providing a domain-specific framework for understanding and predicting targeted health risks.

The present analysis provides an approximation of the test–retest reliability of CardioMetAge, given the lack of true technical replicates in the available datasets. Nevertheless, the observed consistency across repeated measurements supports its potential utility as a longitudinal biomarker of cardiometabolic aging. Previously optimized principal component–based methylation clocks derived from repeated measurements over periods of up to 20 years have shown ICCs ranging from 0.5 to 0.8, which are comparable to our estimates [50]. The moderate-to-high ICCs observed across repeated measurements in the present study suggest that CardioMetAge can detect meaningful longitudinal changes with feasible sample sizes. Although these estimates likely capture both technical and biological components of variability, they provide a conservative yet practical approximation of measurement noise under real-world conditions. Overall, these findings demonstrate that CardioMetAge exhibits sufficient temporal stability to capture biologically relevant shifts in cardiometabolic aging, even over relatively short intervention periods. Future research incorporating true technical replicates would further refine the assessment of its test–retest reliability.

Previous studies have debated the idea of incorporating aging as a disease in clinical practice [54–56]. Cardiometabolic aging shares many pathophysiological features with CMD, such as increased arterial stiffness, oxidative stress, and inflammation, all of which affect heart function and structure [56, 57]. Chronic, low-grade inflammation, in particular, is known to accelerate aging and contribute to cardiovascular diseases²⁸. Our proteomic analysis further supported the role of immune activation and metabolic disorders in driving cardiometabolic aging. The identified proteins and pathways reinforce the biological validity of CardioMetAge and enhance understanding of its molecular underpinnings. Nevertheless, causal evidence is still needed to clarify whether these protein alterations represent drivers or consequences of accelerated cardiometabolic aging. Several of these proteins may also serve as promising therapeutic or monitoring targets for future research.

Our investigation identified cardiometabolic aging as a mediator in the links between SES, lifestyle, and CMD, offering new insights into previous studies that focused on their direct associations [36, 58]. This finding suggests that part of the adverse effects of unfavorable socioeconomic and lifestyle conditions on cardiometabolic health may operate through the acceleration of cardiometabolic aging. In this context, CardioMetAge may serve as a useful biomarker to quantify the biological impact of modifiable exposures and to identify individuals at higher cardiometabolic risk who may benefit from early preventive strategies. Moreover, the observed associations highlight the importance of addressing social and behavioral determinants in strategies aimed at promoting healthy cardiometabolic aging and reducing health disparities.

We clarified the relationship between the reduced calorie intake and the slowdown of cardiometabolic aging, addressing both association and intervention perspectives. Our findings complement previous reports on the benefits of CR in reducing cardiometabolic risk factors [59]. Prior research has also explored the effect of CR on epigenetic BAs [40]. While PhenoAgeMethy and GrimAge showed no significant differences between the AL and CR groups, the DunedinPACE measure revealed significant effects. Given the long observational periods required to study human diseases, which pose significant challenges for intervention trials, changes in CardioMetAge have the potential to serve as surrogate endpoints for evaluating the effects of interventions on healthy lifespan [25, 56].

It should also be noted that phenotype-based aging clocks, including PhenoAge and the CardioMetAge proposed here, inherently summarize multidimensional physiological information into a single latent construct. This integrative property enables the models to capture overall system-level aging processes, but it may also obscure specific pathological abnormalities [52]. For instance, an individual with elevated C-reactive protein, indicating low-grade inflammation, may still appear ‘biologically younger’ if other biomarkers (e.g., glucose levels or blood pressure) remain optimal. Therefore, BA should be interpreted as a holistic indicator of cumulative cardiometabolic burden rather than as a diagnostic measure for any single biomarker. In clinical contexts, the estimated BA should therefore be considered alongside individual biomarker profiles and conventional risk factors.

This study also had potential limitations. First, this study was conducted primarily within Western populations, including the racially diverse U.S. NHANES, the predominantly European UKB, and the CALERIE trial comprising mainly U.S. White participants. Although the consistency of results across these datasets supports the robustness of the CardioMetAge framework, further validation in non-Western and ethnically diverse populations is warranted. Second, the modeling solely focused on cardiometabolic mortality as an aging outcome. Future studies could expand to encompass the incidence of cardiometabolic diseases and additional aging phenotypes, such as organ fibrosis [60], for a more comprehensive assessment. Third, our approach primarily leveraged the robustness and accessibility of clinical markers, while acknowledging that metabolomic, proteomic, and imaging-based aging clocks can capture complementary molecular and structural information, thereby enriching the understanding of cardiometabolic aging. Fourth, although biomarker selection was performed using bootstrap iterations of LASSO regression to improve robustness, different feature selection approaches or modeling strategies might identify alternative sets of biomarkers, which should be explored in future studies. Fifth, while our model was designed with a simplified linear structure to enhance clinical utility, this approach inevitably may overlook complex nonlinear relationships. Future studies could explore integrating complex modeling techniques to capture these intricate patterns and further improve predictive accuracy. Sixth, our model was developed based on biomarkers measured at a single time point, and incorporating repeated measurement data in the future can better track the dynamic changes in aging. Finally, because both the exposures and the mediator were measured simultaneously at baseline, the mediation analyses cannot establish temporal or causal relationships. These findings should therefore be considered exploratory, highlighting potential biological pathways for future validation in longitudinal or intervention studies.

Conclusions

We developed and validated the CardioMetAge model, a concise and interpretable indicator for quantifying cardiometabolic aging, outperforming existing general aging clocks in predicting CMD-specific outcomes. By capturing biological pathway signals and integrating lifestyle and SES factors, CardioMetAge offers insights into cardiometabolic aging mechanisms and health management. Its potential for monitoring interventions, such as CR, further highlights its utility in advancing precision medicine and promoting healthy aging.

Abbreviations

AL

Ad libitum

ALP

Alkaline phosphatase

AUC

The area under the receiver operating characteristic curve

BA

Biological age

BMI

Body mass index

BUN

Blood urea nitrogen

CA

Chronological age

CALERIE

Comprehensive Assessment of Long-term Effects of Reducing Intake of Energy

CardioMetAge

Cardiometabolic biological age

CardioMetAgeDev

Cardiometabolic biological age deviation

CDF

Cumulative distribution function

CI

Confidence interval

CMI

Cardiometabolic index

CMD

Cardiometabolic disease

CMM

Cardiometabolic multimorbidity

CR

Calorie restriction

CRP

C-reactive protein

DBP

Diastolic blood pressure

DE

Direct effect

DNAm

DNA methylation

DunedinPoAm

Dunedin Pace of Aging methylation

ECM

Extracellular matrix

FCMD

First cardiometabolic disease

FDR

False discovery rate

FRS

Framingham risk score

GSEA

Gene set enrichment analysis

GrimAge2Dev

Age deviation of GrimAge2

HbA1c

Glycohemoglobin

HR

Hazard ratio

ICC

Intraclass correlation coefficient

IDE

Indirect effect

IHD

Ischemic heart disease

IQR

Inter-quartile range

LASSO

Least absolute shrinkage and selection operator

MAE

Mean absolute errors

MCV

Mean corpuscular volume

NES

Normalized enrichment score

NHANES

National Health and Nutrition Examination Survey

PhenoAgeDev

Age deviations of PhenoAge

PhenoAgeMethyDev

Age deviation of methylation-based PhenoAge

PP

Pulse pressure

RDW

Red cell distribution width

SBP

Systolic blood pressure

S_CardioMetAge

Substitute version of CardioMetAge

S_CardioMetAgeDev

Age deviation of S_CardioMetAge

SD

Standard deviation

SES

Socioeconomic status

TE

Total effect

UA

Uric acid

UKB

UK Biobank

WBC

White blood cell count

WC

Waist circumference

Authors’ contributions

YCL, KLX, and XDC conceived and designed the study. YCL, XMX performed the data analysis. YZ, XYH, and JCW contributed to data collection and preprocessing. ZQL provided statistical expertise and guidance. YCL drafted the manuscript, with significant input from XMX and KLX. YFJ, CS, and TJZ assisted in interpreting the results and contributed to the manuscript’s modifications. XG provided critical feedback on the study design and manuscript. KLX and XDC supervised the overall project. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (grant number: 82304239), Natural Science Foundation of Shanghai, China (grant number: 23ZR1414000), Science and Technology Innovation 2030 Major Projects (grant numbers: 2022ZD0211600, 2023ZD0510000), National Key Research and Development Program of China (grant numbers: 2022YFC3400700), and the Shanghai Municipal Science and Technology Major Project (grant number: 2023SHZDZX02).

Data availability

Data from the NHANES are available at https://www.cdc.gov/nchs/nhanes/index.html. Data from the UKB are available on application at https://www.ukbiobank.ac.uk. Data from the CALERIE are available for academic research purposes from the CALERIE Biorepository: https://calerie.duke.edu.

Declarations

Ethics approval and consent to participate

The NHANES study was conducted by the National Center for Health Statistics (NCHS) with approval from the NCHS Research Ethics Review Board. All participants provided written informed consent.

The UKB, having received initial approval from the North West Multi-Centre Research Ethics Committee as a Research Tissue Bank in 2011, undergoes renewal every five years. All participants have provided their informed consent with signed agreements.

The CALERIE follows the ethical principles of the Declaration of Helsinki and is registered on ClinicalTrials.gov (NCT00427193). All participants provided informed consent and Health Insurance Portability and Accountability Act (HIPAA) authorization before study procedures.

Consent for publication

Not applicable. This study used de-identified data from the publicly available database, which contains no personally identifiable information. Therefore, individual patient consent for publication was not required.

Competing interests

The authors declare no competing interests.