Accelerated biological ageing flags chronic disease and mortality risk in US adults

A single blood-panel-derived number that estimates how fast a body is ageing has been linked to both the weight of chronic disease a person carries and their risk of dying from any cause. In a retrospective analysis of 7,624 US adults drawn from the National Health and Nutrition Examination Survey (NHANES), researchers calculated each participant’s phenotypic age from chronological age and nine clinical biomarkers, then asked how closely that biological age tracked with diagnosed illness and later mortality. The findings, published on 22 September 2026 in Cardiovascular Diabetology – Endocrinology Reports, suggest that phenotypic age metrics could add a layer of risk stratification that chronological age, body mass index or simple disease counts do not capture on their own.

The team worked with two related but distinct measures. Phenotypic age itself is a composite biomarker-based estimate of biological ageing, originally developed in NHANES III to capture mortality-related physiological dysregulation. The phenotypic age gap, or PAG, was defined simply as phenotypic age minus chronological age, and treated as a continuous variable: positive values mean the biological profile looks older than the birth certificate, zero or negative values indicate a maintained or younger biological age. From PAG the researchers derived a binary classification, PhenoAgeAccel, labelling participants with PAG greater than zero as “accelerated” and those at or below zero as “maintained.” The nine biomarkers feeding the calculation were albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage, mean cell volume, red blood cell distribution width, alkaline phosphatase and white blood cell count, combined through a published formula.

The analytic sample came from NHANES cycles spanning 2005 to 2010 and 2015 to 2018. The 2011 to 2014 period was excluded because C-reactive protein data needed for the phenotypic age calculation were unavailable. Starting from 48,069 recruited participants, the authors sequentially removed anyone missing data for the phenotypic age calculation (16,459 people), missing disease-diagnosis information (8,460), missing demographic covariates (2,341) or missing lifestyle covariates (13,163), leaving 7,624 adults in a complete-case analysis. Because phenotypic age was built from laboratory biomarkers collected during the mobile examination centre visit, the researchers applied MEC examination weights, dividing the two-year weights by the number of included survey cycles, and accounted for survey strata and primary sampling units throughout the descriptive and regression analyses.

Disease history came primarily from self-reported physician diagnoses in the NHANES Medical Conditions Questionnaire. A participant counted as having a condition, asthma for example, only if they answered yes to whether a doctor or other health professional had ever told them they had it; refusals and “don’t know” responses were excluded under the complete-case strategy. Cardiometabolic multimorbidity, abbreviated CMM, was defined as the presence of two or three conditions among diabetes, ischaemic heart disease and stroke. The cross-sectional component used multivariable logistic regression to estimate odds ratios for associations between each chronic disease and accelerated ageing status, with demographic factors such as sex, age, race/ethnicity, education, marital status and poverty-income ratio, lifestyle factors including BMI category, smoking and alcohol use, and the full list of chronic conditions included as covariates.

After multivariable adjustment, several common conditions stood out. Obesity, hypertension, arthritis, diabetes and malignancy were among the diseases most strongly associated with positive PhenoAgeAccel. Female sex was associated with lower odds of accelerated ageing, while widowed and divorced or separated participants had higher odds than married participants. The logistic model showed good discrimination, with an area under the receiver operating characteristic curve of 0.868 on an independent test set. Model diagnostics were broadly reassuring: variance inflation factors stayed below five, the Hosmer–Lemeshow test assessed calibration, and Schoenfeld residual tests found no significant overall violation of the proportional hazards assumption, although race showed a covariate-level violation that sensitivity analyses indicated did not materially affect the main exposure effect.

To explore which variables carried the most predictive information, the researchers trained ten classification algorithms, including LightGBM, XGBoost, CatBoost, Random Forest, Support Vector Machine and Decision Tree, on an 80:20 train-test split with grid search and ten-fold cross-validation for hyperparameter tuning. LightGBM emerged as the best performer, combining high test-set AUC, competitive accuracy, stable cross-validation results and rapid training at roughly 1.25 seconds, whereas the Decision Tree showed the lowest test performance and greater overfitting. Shapley Additive Explanations analysis of the LightGBM model ranked obesity, hypertension, arthritis, diabetes and malignancy as the most influential features driving higher predicted probabilities of accelerated ageing, with obesity and hypertension exerting consistently strong influences across samples. The authors stress that these SHAP rankings are model-based importance measures, not estimates of independent or causal disease effects.

The longitudinal component linked participants to National Death Index mortality follow-up data and used survey-weighted Cox proportional hazards models. Positive PhenoAgeAccel was associated with markedly increased all-cause mortality, with a hazard ratio of 5.52 and a 95 percent confidence interval of 4.26 to 7.15. Kaplan–Meier curves by PAG quartile showed a graded pattern: participants in the highest quartile, representing the greatest biological-age burden relative to chronological age, had substantially lower survival probabilities than those in lower quartiles, and cumulative hazard curves showed the fourth quartile accumulating risk fastest over follow-up. Continuous modelling of PAG quantified the average change in mortality risk per unit increase, while the quartile analysis served mainly descriptive and risk-stratification purposes.

The joint analysis of biological ageing and cardiometabolic multimorbidity produced one of the study’s most striking observations. Participants were grouped by the combination of PhenoAgeAccel status and CMM status, with people having neither exposure serving as the reference. Those with cardiometabolic multimorbidity and positive PAG had the poorest survival outcomes of any group, and within the CMM population, Kaplan–Meier analyses showed a clear separation in survival probabilities between PAG-positive and PAG-negative individuals. Multiplicative interaction analyses, however, found no statistically significant interactions between PAG and CMM or other examined covariates, with all P values for interaction above 0.05, so the authors interpret these results as subgroup-specific differences in risk rather than evidence of biological synergy between the two exposures.

The authors are careful about what the study can and cannot show. Because the cross-sectional component cannot establish temporal ordering, it cannot support causal inference, and the associations between PhenoAgeAccel and specific diseases may partly reflect shared biological pathways, since phenotypic age itself incorporates biomarkers tied to inflammation, metabolism, haematologic status and organ function. Chronic disease history rested on self-report and may be subject to recall bias or misclassification; complete-case analysis may have introduced selection bias; and the team lacked information on disease severity, duration, treatment and temporal changes in disease status. The machine-learning models were trained on unweighted data, so the feature-importance results are exploratory rather than nationally representative estimates. The study was reported in line with the STROCSS 2025 guidelines.

Within those limits, the authors argue that phenotypic age metrics may earn a practical role in risk assessment, particularly for multimorbid populations. As a continuous measure, PAG preserves the full gradient of biological-age deviation and can capture graded differences in risk, while the binary PhenoAgeAccel classification offers a convenient way to flag high-risk subgroups in clinical and public health settings. Used together, the two measures may help distinguish people with a similar number of chronic conditions but different levels of underlying physiological vulnerability, informing more tailored follow-up and prevention strategies. The researchers suggest this framework could be relevant to patterns such as cardiovascular-kidney-metabolic syndrome, metabolically unhealthy obesity and non-obese diabetes, where conventional indicators like BMI or disease counts may not fully reflect systemic physiological burden. They conclude that integrating phenotypic ageing into risk models may improve early identification of vulnerable populations and refine conventional risk stratification, while emphasising that the metrics should be understood as markers of biological ageing and systemic vulnerability rather than surrogates for any single disease.

Subject of Research: Associations between phenotypic age acceleration, chronic disease burden and all-cause mortality in US adults

Article Title: Phenotypic age metrics for risk stratification of chronic disease burden and all-cause mortality in adults

Article References: Wang, Z., Zhong, X., Cai, Y., Wang, J., Lin, B., Liu, L., Li, S., He, H., Yao, D., Yan, C., Zou, Y., Fu, Y., Sun, J., & Zhang, Q. (2026). Phenotypic age metrics for risk stratification of chronic disease burden and all-cause mortality in adults. Cardiovascular Diabetology – Endocrinology Reports, 12(1), Article 62. https://doi.org/10.1186/s40842-026-00331-2

Image Credits: AI Generated

DOI: 10.1186/s40842-026-00331-2

Keywords: phenotypic age, biological ageing, NHANES, all-cause mortality, chronic disease, cardiometabolic multimorbidity, obesity, hypertension, diabetes, machine learning, LightGBM, risk stratification