Abstract
Background
Klemera–Doubal’s method (KDM) is an advanced and widely applied algorithm for estimating biological age (BA), but it has no uniform paradigm for biomarker processing. This article proposed all subsets of biomarkers for estimating BAs and assessed their association with mortality to determine the most predictive subset and BA.
Methods
Clinical biomarkers, including those from physical examinations and blood assays, were assessed in the China Health and Nutrition Survey (CHNS) 2009 wave. Those correlated with chronological age (CA) were combined to produce complete subsets, and BA was estimated by KDM from each subset of biomarkers. A Cox proportional hazards regression model was used to examine and compare each BA’s effect size and predictive capacity for all-cause mortality. Validation analysis was performed in the Chinese Longitudinal Healthy Longevity Survey (CLHLS) and National Health and Nutrition Examination Survey (NHANES). KD-BA and Levine’s BA were compared in all cohorts.
Results
A total of 130 918 panels of BAs were estimated from complete subsets comprising 3–17 biomarkers, whose Pearson coefficients with CA varied from 0.39 to 1. The most predictive subset consisted of 5 biomarkers, whose estimated KD-BA had the most predictive accuracy for all-cause mortality. Compared with Levine’s BA, the accuracy of the best-fitting KD-BA in predicting death varied among specific populations.
Conclusion
All-subset analysis could effectively reduce the number of redundant biomarkers and significantly improve the accuracy of KD-BA in predicting all-cause mortality.
Keywords: All-subset, Biological age, All-cause mortality
Aging is an unavoidable, asynchronous, and highly heterogeneous biological process characterized by a progressive decline in tissue and organ function and an increased risk of mortality (1). Individuals of the same chronological age (CA) might have quite different health statuses, for instance, someone aged 60 years old who must walk with a crutch versus someone in their 80s who can run a marathon. Biological aging, which is quantified by several well-developed measures, including but not limited to telomere length (2,3), frailty index (FI) (3,4), homeostatic dysregulation (5,6), allostatic load (7), the Pace of Aging (8), and biological age (BA) (7–35), can explain the heterogeneity of peers and evaluate their aging status in this inevitable process. Although it has been widely used as a biological aging measure in aging evaluation research, BA has not yet been given a precise verbal definition. Its superiority over other biological aging measures is because BA shares the same unit as CA, thus facilitating the interpretation and evaluation of aging. A growing body of literature has shown that BA can be estimated from epigenetic modifications (2,3,9), transcriptomics (2,10), proteomics (2,11–13), metabonomics (2,11), glycomics (14), neuroimaging (15), facial-image (16), blood biomarkers, and physical examination, even including ECG data (also classified as clinical biomarkers) (7,8,17–35). Generally, BA estimated from clinical biomarkers rather than multiomics data is more feasible to understand and interpret.
Several algorithms have been developed to estimate BA from clinical biomarkers, including Klemera–Doubal’s method (7,8,11,18–30), Levine’s method (17,35), Hochschild’s method (18), principal component analysis (PCA) (18–22), multiple linear regression (MLR) (18–21), and machine learning (ML) (31,32). Several previous studies have compared the association between multiple biological aging measures and mortality or other health outcomes (3,18–23,33,34). Among these, BA estimated by KDM (KD-BA) has been proven to be a better predictor of mortality than BA estimated by PCA, MLR, or ML (19,21,22), and to perform at least as well as PhenoAge (also referred to as Levine’s BA) (33,34), although not as well as the FI (3,23). The diverse performance of KD-BA may be attributed to the composition of the aging biomarker sets and cohort heterogeneity. The identification and selection of aging biomarkers have always been a particularly critical and complex process, especially because there are no uniform criteria for estimating BA. In addition to proposing KDM, Klemera and Doubal argued that any measurable property of a human organism that changed systematically with CA might be affected by the individual degree of aging and then used as biomarkers (36). When estimating BA by using KDM, several research teams adopted different biomarker processing strategies, including organ function representation (19,29), stepwise regression (21), domain specificity (27), PCA (28), and association with mortality (30). One drawback of these strategies for biomarker selection was that it could not ensure that the BA thus estimated was of the best performance. For example, based on an organ function representation strategy, diastolic blood pressure is usually discarded due to a much smaller correlation with CA than systolic blood pressure. However, it is unknown whether this strategy is optimal. Another limitation was that more biomarkers were needed to construct the BA equation, therefore complicating the BA equation and reducing its feasibility in clinical practice.
The aim of this study was to identify the BA that most accurately predicts mortality using all subsets of biomarker selection strategies. We estimated BAs from full biomarker subsets in the China Health and Nutrition Survey (CHNS) population older than 20 years and younger than 80 years by using the KDM algorithm and checked the accuracy of the mortality predictions using the Cox hazards proportion model. We validated the BA calculated from KDM parameters in another independent, nationally representative cohort of Chinese older adults. We also compared our best-fitting KD-BA to a version of Levine’s BA developed in the CHNS cohort comprising the same biomarkers as KD-BA and repeated and validated our strategy for screening the most predictive subset of biomarkers and KD-BA in National Health and Nutrition Examination Survey (NHANES).
Method
Data Source and Study Design
CHNS, an ongoing open cohort of an international collaborative project, was designed to examine the effects of the health, nutrition, and family planning policies and programs implemented by national and local governments and to see how the social and economic transformation of Chinese society is affecting the health and nutritional status of its population. The survey used a multistage, random cluster process to draw a cumulative sample of approximately 11 130 households with 42 829 individuals in 10 waves from 1989 to 2015 in 15 provinces that vary substantially in geography, economic development, public resources, and health indicators. Detailed information about the cohort profile has been published previously (37,38). All household members provided individual data on demographics, health-related behaviors (eg, smoking, beverage consumption), activities of daily living (ADL, for elderly participants), health history and physical examinations that included blood pressure (for adults), measurement of weight, height, arm/waist/hip circumference, and triceps skin fold. The blood biomarker data collected in the 2009 wave involved 26 fasting tests on individuals aged 7 and older for the first time. The CHNS was approved by the institutional review boards at the University of North Carolina at Chapel Hill and the National Institute of Nutrition and Food Safety. All participants provided informed consent before the surveys.
The Chinese Longitudinal Healthy Longevity Survey (CLHLS) is a prospective, longitudinal, population-based study that has collected extensive data from a large sample of the oldest-old adults aged over 80 years, including sufficient numbers of centenarians, nonagenarians, and octogenarians, with comparison groups of younger elderly adults aged 65–79 since 1998. The CLHLS is conducted in a random selection of half of the counties and cities in 22 of China’s 31 provinces, and the overwhelmingly predominant population in the 22 provinces is Han Chinese, whose age reporting is reasonably good. To date, 8 waves of data were collected in 1998, 2000, 2002, 2005, 2008–2009, 2011–2012, 2014, and 2018. Blood and urine specimens were collected in the 2008–2009, 2011–2012, and 2014 waves for approximately one-third of the participants. Detailed information about the cohort profile has been published elsewhere (39,40). The Research Ethics Committees of Peking University (IRB00001052-13074) granted approval for the CLHLS. All survey respondents or their proxy respondents provided informed consent before participation.
NHANES is a nationally representative survey of noninstitutionalized civilian residents in the United States using a stratified, multistage, and probability sampling design. Data from NHANES III (1988–1994) and NHANES 1999–2010 were included in this study. The design details, protocols, and other information about the NHANES are available from the official website (https://www.cdc.gov/nchs/nhanes/index.htm).
The flowcharts of the CHNS, CLHLS, and NHANES analysis procedures are presented in Supplementary Figure 1. In brief, 8 009 participants without missing candidate biomarkers and between 20 and 80 years old were set as the reference population to screen the most predictive subset of biomarkers and KD-BA in the CHNS, and 8 441 participants with data for the most predictive subset biomarkers and older than 20 years were included in further analysis. In the CLHLS, 2 439 participants in the 2012 wave and 1 086 new recruits in the 2014 wave were merged, resulting in 3 161 participants after excluding those with missing data included in analysis-1 to validate the association of BA with mortality. A total of 1 287 participants with full biomarkers in both the 2012 and 2014 waves were entered into analysis-2 to validate the association of BA change with mortality. Similarly, 7 142 female and 6 345 male participants were set as the reference population from NHANES III to develop the most predictive sex-specific KD-BA, and 7 616 female and 6 796 male participants were included in further analysis. A total of 12 061 female and 11 310 male participants in the NHANES 1999–2010 cycle were used in the validation analysis.
Biomarker Selection and Combination
Similar to a previous study (29), we added body measurements focusing on height, waist circumference (WC), and hip circumference, which are easily available and have been proven to be associated with mortality (41,42), an acknowledged standard of aging biomarkers (43). In total, 36 clinical biomarkers, including those from physical examinations and blood assays, were extracted from the 2009 wave of the CHNS. Three new variables were constructed from physical examinations, namely, pulse pressure (PP), body mass index (BMI), and waist–hip ratio (WHR). Twenty-three biomarkers that were significantly correlated with CA (|r| ≥ .1) were selected as candidate biomarkers (Supplementary Table S2). To reduce the amount of computation, we first compared the contribution of biomarkers with collinearity (r ≥ .4) to the mortality predictive performance of KD-BA. After discarding biomarkers that were less predictive of death in the comparison subset analysis, 17 candidate biomarkers entered the all-subset analysis. Three to 17 of these candidate biomarkers were extracted in an orderly manner and combined completely to produce 130 918 subsets of biomarkers. The most predictive subset of biomarkers was selected according to the derived BA that had the most accurate prediction of mortality, which was quantified by the concordance index (C-index) of the Cox hazards proportion model (details in the following section of Methods).
Considering that the NHANES contains a rich variety of biomarkers and accessibility of biomarkers in clinical practice, 51 candidate biomarkers were selected from blood pressures, body measurements, complete blood counts, standard biochemistry profiles, blood lipids, and blood glucose profiles (Supplementary Table S2). The selection of biomarkers, all-subset analysis, and comparative analysis in the NHANES were within sex-specific strata. Other analysis procedures were consistent with those in the CHNS and CLHLS.
Biological Age Estimation and Aging Status
KD-BA of comparison subsets and all-subset analysis in the CHNS or NHANES III reference population were calculated by the advanced algorithms described in Klemera and Doubal’s work (36). A Chinese version of the PhenoAge (c-Levine’s BA) algorithm composed of the most predictive subset of biomarkers of KD-BA was developed in the CHNS, following the method described in Belsky’s work (44). For participants included in further analysis of the CHNS cohort and validation analysis of the CLHLS population, KD-BA and c-Levine’s BA were calculated by estimated parameters of KDM and c-Levine’s BA algorithm developed in the CHNS full population with no missing for the most predictive subset of biomarkers. For participants in the validation analysis of NHANES 1999–2010, KD-BA was calculated by estimated parameters of KDM developed in the NHANES III. Levine’s BA in NHANES III and 1999–2010 were calculated according to the equation provided by previous work (17,35).
To compare the predictive capability of different BAs, we used ΔAge, defined as the difference between BA and CA, namely, BA minus CA, to characterize an individual’s aging status. Distinct from previous concepts (45), we redefined the “aging rate” as the rate of change in BA along with CA, namely, ΔBA/ΔCA, to characterize an individual’s aging acceleration. The aging rate was calculated for individuals who had repeated measures in the CLHLS 2012 and 2014 waves. It was generally accepted that a positive value of individual ΔAge indicated that the individual was biologically old or undergoing advanced aging and that a greater ΔAge meant biologically older, whereas a ΔAge less than zero suggested that the individual was biologically younger than expected or experiencing delayed aging. Similarly, a greater aging rate means faster biological aging than others longitudinally.
Mortality
In the CHNS, the date of survey and date of death were integrated into the “MAST_pub_12” data set. We extracted survival information from the 2009, 2011, and 2015 waves. We calculated survival time as the time from the survey date in the 2009 wave to the date of death. If respondents were interviewed in the 2011 (including those lost to follow-up in the 2015 wave) or 2015 wave, their survival time was right censored as the interval between the survey date in 2009 wave and the date of the latest survey in which they were interviewed but not lost to follow-up. Those who were not interviewed in both the 2011 and 2015 waves were classified as “lost to follow-up.”
In the CLHLS, participants’ date of survey, date of death, age at death, or survival status were recorded in each longitudinal data set. In detail, we calculated survival time as the time from the survey date in the 2012 or 2014 wave to the date of death. If participants were alive in the 2014 (including those lost to follow-up in the 2018 wave) or 2018 wave, their survival time was right censored as the interval between the survey date in the 2012 or 2014 wave and the date of the latest survey in which they were still alive but not lost to follow-up. Those who were not interviewed in both the 2014 and 2018 waves were classified as “lost to follow-up.”
In NHANES III and 1999–2010, mortality data were linked to National Death Index records, a centralized database compiled from state vital statistics offices (https://www.cdc.gov/nchs/data-linkage/mortality-public.htm). The records provided information on mortality status and follow-up time from the NHANES interview to death or the end of the follow-up period (December 31, 2015).
Statistical Analysis
We presented the correlation matrix for CA and candidate biomarkers in the CHNS and NHANES III. We calculated the Pearson correlation coefficient between BAs and CA to assess the influence of biomarker combinations on BA estimation. We also plotted the correlation between the best-fitting KD-BA, c-Levine’s BA and CA in the CHNS and CLHLS. Cox proportional hazards regression analysis was applied to check the effect size and accuracy of predicted mortality estimated by BA. The concordance index (C-index), which represents the global assessment of the model discrimination power (46), was used to compare the predictive accuracy of Cox models for mortality. The most predictive subset of biomarkers was selected according to the derived KD-BA with the maximum C-index in the sex and CA (CHNS) or CA (NHANES III) adjusted Cox model. For further analysis in the CHNS/NHANES III and validation analysis in the CLHLS/NHANES 1999–2010, we plotted Kaplan–Meier curves for ΔAge and aging rate (only in the CLHLS cohort) grouped by quartile and ran a further Cox model adjusted for behavior, socioeconomic, and health-related covariates.
All analyses were performed in R version 4.0.3 (2020-10-10). The correlation matrix and Pearson coefficient were calculated by the corr.test function in the psych package; the hazard ratio (HR), Z-score, and C-index were provided by the coxph function in the survival package; the difference in the C-index between Cox models was tested by the cindex.comp and concordance.index functions in the survcomp package; the survival curve was generated by the survfit function in the survival package and the ggsurvplot function in the survminer package.
Results
Comparison of Biomarkers with Collinearity
Five groups of biomarkers (WC/WHR, blood pressures, lipid profiles, blood glucose/glycosylated hemoglobin [GH], and creatinine [lnCRE]/urea) with collinearity in the CHNS were compared by evaluating their contribution to the mortality predictive performance of KD-BA (Supplementary Figure 2). Generally, WHR (r with CA = 0.23) could enhance mortality predictive performance, but WC (r with CA = 0.20, with WHR 0.71) did the opposite. Similarly, subsets with DBP (r with CA = 0.23, with SBP = 0.71, with PP = 0.15) performed better than those without, but subsets with SBP (r with CA = 0.44, with PP = 0.80) or PP (r with CA = 0.42) performed worse than those without. Surprisingly, subsets with TC (r with CA = 0.22, with LDL_C = 0.79, with APO_B = 0.84), LDL-C (r with CA = 0.21, with APO_B = 0.81), or APO_B (r with CA = 0.23) all decreased the performance of KD-BA, but lnTG (r with CA = 0.10, with TC = 0.37, with APO_B = 0.40) enhanced it. The presence of lnCRE (r with CA = 0.19, with urea = 0.39) or urea (r with CA = 0.24) or both in the subsets contributed to a better performance than their absence, and lnCRE performed the best. Subsets including lnGLU (r with CA = 0.22, with GH = 0.61) or GH (r with CA = 0.22) performed similarly. As expected, subsets including 2 or more biomarkers with collinearity decreased the performance of KD-BA in predicting death. The findings of the biomarker comparative analysis in NHANES III were similar to these results (Supplementary Figures 3 and 4).
Biomarker Composite and BA Performance
A total of 130 918 panels of KD-BA for 8 009 participants in the CHNS 2009 wave were calculated by the KDM algorithm (results were not provided because the file was larger than 7 GB as an object in R). The Pearson coefficient between each KD-BA and CA varied from 0.39 to 1 (precisely 0.999999996 in the 5513th subset, Supplementary Table S3). The subset with the smallest correlation coefficient had 3 biomarkers, while the subset with the largest had 5 biomarkers. We found that the minimum value of the correlation coefficient in subsets with the same numbers of biomarkers gradually increased as the number of biomarkers in the subset increased, while the maximum value gradually decreased as the number of biomarkers increased from 5 (Figure 1A).
Figure 1.
Plot of Pearson correlation coefficient and C-index for each panel of KD-BA in CHNS. Y-axis, upper (A): Pearson correlation coefficient, lower (B): C-index; x-axis, upper: number of biomarkers in subsets, lower: serial number of panels of subset.
A total of 130 918 panels of ΔAge for 8 009 participants were calculated (results were not provided because the file was larger than 7 GB as an object in R), and 130 918 panels of C-index were extracted from the sex- and CA-adjusted ΔAge Cox models (Supplementary Table S3). The C-index of each model varied from 0.806 to 0.830 (Figure 1B). The subset with the smallest C-index had 3 biomarkers, while the subset with the largest had 5 biomarkers. As Figure 1 shows, the trend of extreme values of the C-index in subsets with the same numbers of biomarkers was similar to that of the correlation coefficient, but they did not correspond to each other (Supplementary Figure 5A and B). Similar trends were found in the all-subset analysis of NHANES III (Supplementary Figures 6 and 7).
Most Predictive Subset of the CHNS and NHANES
As Table 1 shows, the C-index of the base (sex + CA) model was 0.806, while the top 10 C-index values of the ΔAge model adjusted by sex and CA varied from 0.829 to 0.830, with an FDR-adjusted p-value (base + Δage vs base) less than .01, compared with the C-index of the base model (0.806). Detailed results of all 130 918 ΔAge Cox models and the related biomarker composite subsets are presented in Supplementary Table S3. The most predictive subset of biomarkers in the CHNS consisted of height, albumin, lnCRE and lnTG and lnCRP. Figure 2A and B show the correlation plots for CA and best-fitting KD-BA and c-Levine’s BA incorporating the most predictive subset of biomarkers. Detailed results of the all-subset analysis for NHANES III are presented in Supplementary Tables S4 and S5.
Table 1.
Top 10 Models Ranked by C-Index and Corresponding Biomarkers in Subsets in CHNS
| Model | C-index | FDR p-value |
Biomarkers Composite of Subsets | |||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| N | HEI | WHR | SBP | DBP | RBC | PLT | ALB | URE | CRE | lnCRE | TC | TG | LPA | GH | CRP | FET | TRF | |||
| Base | 0.80644 | |||||||||||||||||||
| Base + ΔAge1 | 0.83004 | .00049 | 5 | |||||||||||||||||
| Base + ΔAge2 | 0.83000 | .00049 | 7 | |||||||||||||||||
| Base + ΔAge3 | 0.82986 | .00103 | 5 | |||||||||||||||||
| Base + ΔAge4 | 0.82965 | .00049 | 6 | |||||||||||||||||
| Base + ΔAge5 | 0.82960 | .00049 | 6 | |||||||||||||||||
| Base + ΔAge6 | 0.82957 | .00051 | 8 | |||||||||||||||||
| Base + ΔAge7 | 0.82953 | .00049 | 4 | |||||||||||||||||
| Base + ΔAge8 | 0.82940 | .00097 | 7 | |||||||||||||||||
| Base + ΔAge9 | 0.82939 | .00102 | 4 | |||||||||||||||||
| Base + ΔAge10 | 0.82928 | .00049 | 7 | |||||||||||||||||
| Base + ΔAge** | 0.81581 | .03510 | 12 |
Notes: N = number of biomarkers in subsets; HEI = height; WHR = waist–hip ratio; SBP = systolic blood pressure; DBP = diastolic blood pressure; RBC = red blood cell count; PLT = platelet count; ALB = albumin; URE = urea; CRE = creatinine; lnCRE = log-transformed creatinine; TC = total cholesterol; TG = log-transformed total triglyceride; LPA = log-transformed lipoprotein(a); GH = glycosylated hemoglobin; CRP = log-transformed C reaction protein; FET = logged-ferritin; TRF = transferrin. Base model, gender + CA; Base + ΔAge**, model developed by Liu29, ranked at 44 776th in 130 918 models; FDR p-value, difference between C-index of Base + ΔAge model and Base model.
Figure 2.
Correlation plot and Kaplan–Meier curves with 95% confidence intervals for ΔAge and Aging Rate grouped by quartile. Participants whose ΔAge/aging rate > third quartile were defined as biologically older/aging faster, whose ΔAge/aging rate < first quartile as biologically younger/aging slower, between first and third quartiles as biologically normal/aging normal.
Association of the Best-Fitting KD-BA and c-Levine’s BA With All-Cause Mortality in the CHNS
In univariate analysis of KD and c-Levine’s ΔAge by using Kaplan–Meier curves, we found that biologically older participants had starkly worse survival distributions compared with both biologically normal and younger participants over approximately 75 months of follow-up in the full sample (log-rank p < .0001, Figure 2C and D) and older adults (log-rank p < .0001, Supplementary Figure 8A and B). In the full sample of the CHNS (Table 2), both increased KD and c-Levine’s ΔAge values significantly increased the hazard of mortality, and the difference in mortality predictive performance between the best-fitting KD-BA and c-Levine’s BA was not significant (p = .143). For participants aged 60 years or older, similar results are presented in Supplementary Table S6. The accuracy of the best-fitting KD-BA for death prediction was significantly higher than that of the previously developed Liu’s KD-BA in both the full sample and older adults of the CHNS (Supplementary Table S7).
Table 2.
Association of ΔAge and Aging Rate With All-Cause Mortality in CHNS and CLHLS
| Model | Predictors | Best-Fitting KD-BA | c-Levine’s BA | N/Death | ||||
|---|---|---|---|---|---|---|---|---|
| HR (95% CI) | Z-score | C-Index (se) | HR (95% CI) | Z-score | C-Index (se) | |||
| CHNS (full sample) | ||||||||
| Base | CA | 1.098 (1.087, 1.110) | 17.41 | 0.818 (0.013) | 8 441/254 | |||
| Base + ΔAge | CA | 1.096 (1.084, 1.108) | 16.98 | 0.838 (0.012) | 1.100 (1.087, 1.110) | 17.54 | 0.836 (0.013)¶1 | 8 441/254 |
| ΔAge | 1.047 (1.035, 1.060) | 7.68 | 1.090 (1.067, 1.118) | 7.40 | ||||
| Base + ΔAge + Cov | CA | 1.090 (1.076, 1.103) | 13.68 | 0.846 (0.012) | 1.091 (1.078, 1.105) | 13.98 | 0.845 (0.012) | 8 441/252 |
| ΔAge | 1.047 (1.035, 1.060) | 7.42 | 1.095 (1.069, 1.121) | 7.48 | ||||
| CLHLS | ||||||||
| Base | CA | 1.088 (1.081, 1.094) | 27.32 | 0.748 (0.007) | 3 161/1 161 | |||
| Base + ΔAge | CA | 1.087 (1.080, 1.094) | 26.71 | 0.756 (0.007) | 1.087 (1.080, 1.090) | 26.96 | 0.758 (0.007)¶2 | 3 161/1 161 |
| ΔAge | 1.018 (1.013, 1.022) | 7.79 | 1.040 (1.030, 1.050) | 8.42 | ||||
| Base + ΔAge + Cov | CA | 1.083 (1.076, 1.091) | 22.37 | 0.757 (0.007) | 1.083 (1.076, 1.091) | 22.54 | 0.760 (0.007) | 3 161/1 116 |
| ΔAge | 1.018 (1.013, 1.023) | 7.70 | 1.042 (1.032, 1.052) | 8.49 | ||||
| Base | CA | 1.083 (1.072, 1.090) | 16.04 | 0.750 (0.012) | 1 287/373 | |||
| Base + aging rate | CA | 1.080 (1.071, 1.090) | 15.75 | 0.753 (0.012) | 1.082 (1.072, 1.093) | 15.74 | 0.757 (0.012)¶3 | 1 287/373 |
| Aging rate | 1.020 (1.008, 1.040) | 2.92 | 1.069 (1.038, 1.101) | 4.49 | ||||
| Base + aging rate + Cov | CA | 1.080 (1.070, 1.096) | 13.20 | 0.762 (0.012) | 1.080 (1.071, 1.097) | 13.08 | 0.765 (0.012) | 1 287/355 |
| Aging rate | 1.020 (1.006, 1.041) | 2.60 | 1.060 (1.032, 1.096) | 3.98 |
Notes: Base model, gender + CA Cox model; Cov, covariates; Cov in CHNS: urbanization index, nationality, education level, marital status, smoking history, drinking behavior, hypertension, diabetes, other cardiovascular diseases; Cov in CLHLS: residence, marital status, smoking history, drinking behavior, hypertension, diabetes, other cardiovascular diseases. BA, biological age; KD-BA, BA estimated by Klemera–Doubal’s algorithm; N = sample size.
¶1 p-value for difference between C-index of KD-BA and C-index of c-Levine’s BA: .1432.
¶2 p-value for difference between C-index of KD-BA and C-index of c-Levine’s BA: .9973.
¶3 p-value for difference between C-index of KD-BA and C-index of c-Levine’s BA: .008148.
Association of the Best-Fitting KD-BA and c-Levine’s BA With All-Cause Mortality in the CLHLS
As shown in Figure 2E and F, the differences between strata of both KD and c-Levine’s ΔAge were substantial and significant for all-cause mortality (log-rank p < .0001) in survival over 80 months of follow-up. Two years later, those aging faster had a much steeper decline compared with those with a normal or slower aging rate (log-rank p < .0001; Figure 2G and H). In multivariate analysis, both ΔAge and aging rate from 2012 to 2014 measured by best-fitting KD-BA or c-Levine’s BA were significantly associated with increasing mortality risk after adjusting for sex and CA and further adjusted for covariates (Table 2). The difference in mortality predictive performance between KD and c-Levine’s ΔAge was not significant (p = .997), and the KD aging rate was less accurate than c-Levine’s aging rate in mortality prediction (p = .008).
Association of the Best-Fitting KD-BA and Levine’s BA With All-Cause Mortality in the NHANES
In repeated validation analysis conducted in NHANES III and NHANES 1999–2010, KD ΔAge was significantly associated with an increasing mortality risk after CA adjustment, similar to Levine’s ΔAge (Table 3). For female participants, KD and Levine’s ΔAge had similar mortality predictive performance in NHANES III (p = .1965), but KD ΔAge’s prediction of death was not as accurate as that of Levine’s ΔAge in NHANES 1999–2010 (p = .0002). For male participants, KD ΔAge had better accuracy in predicting death than Levine’s ΔAge in both NHANES III (p < .0001) and NHANES 1999–2010 (p = .0097).
Table 3.
Association of ΔAge With All-Cause Mortality in NHANES III and NHANES 1999–2010
| Group | Model | Predictor | Best-Fitting KD-BA | Levine’s BA | N/Death | ||||
|---|---|---|---|---|---|---|---|---|---|
| HR (95% CI) | Z-score | C-Index (SE) | HR (95% CI) | Z-score | C-Index (SE) | ||||
| NHANES III | |||||||||
| Female | CA | CA | 1.087 (1.085, 1.088) | 59.6 | 0.840 (0.004) | 7 616/2 547 | |||
| CA + ΔAge | CA | 1.092 (1.090, 1.093) | 59.8 | 0.850 (0.004) | 1.089 (1.088, 1.091) | 60.2 | 0.851 (0.004)¶1 | 7 616/2 547 | |
| ΔAge | 1.040 (1.038, 1.042) | 18.5 | 1.042 (1.040, 1.044) | 21.4 | |||||
| Male | CA | CA | 1.080 (1.078, 1.081) | 58.4 | 0.814 (0.004) | 6 796/2 868 | |||
| CA + ΔAge | CA | 1.084 (1.082, 1.085) | 60.2 | 0.830 (0.004) | 1.078 (1.076, 1.079) | 57.2 | 0.828 (0.004)¶2 | 6 796/2 868 | |
| ΔAge | 1.040 (1.038, 1.042) | 22.8 | 1.046 (1.044, 1.048) | 22.9 | |||||
| NHANES 1999–2010 | |||||||||
| Female | CA | CA | 1.079 (1.076, 1.081) | 31.7 | 0.797 (0.007) | 12 061/1 069 | |||
| CA + ΔAge | CA | 1. 079 (1.076, 1.081) | 31.5 | 0.804 (0.007) | 1.081 (1.078, 1.083) | 32.3 | 0.820 (0.007)¶3 | 12 061/1 069 | |
| ΔAge | 1.031 (1.028, 1.034) | 10.1 | 1.056 (1.053, 1.059) | 20.2 | |||||
| Male | CA | CA | 1.073 (1.071, 1.075) | 35.7 | 0.770 (0.006) | 11 310/1 548 | |||
| CA +ΔAge | CA | 1.078 (1.076, 1.080) | 36.4 | 0.797 (0.006) | 1.070 (1.068, 1.072) | 33.6 | 0.795 (0.006)¶4 | 11 310/1 548 | |
| ΔAge | 1.049 (1.046, 1.051) | 20.6 | 1.048 (1.046, 1.050) | 22.2 |
Notes: BA, biological age; KD-BA, BA estimated by Klemera–Doubal’s algorithm; N, sample size.
¶1 p-value for difference between C-index of KD-BA and C-index of Levine’s BA: .1965.
¶2 p-value for difference between C-index of KD-BA and C-index of Levine’s BA: .0002427.
¶3 p-value for difference between C-index of KD-BA and C-index of Levine’s BA: 1.379e-18.
¶4 p-value for difference between C-index of KD-BA and C-index of Levine’s BA: .0097.
Discussion
In summary, we proposed a novel strategy to deal with aging biomarkers with high collinearity in using KDM to estimate BA. This strategy was mortality-oriented and used all-subset analysis to find the most predictive combination of biomarkers for estimating BA. We validated the best-fitting KD-BA developed in the CHNS in another independent cohort that included older adults, the CLHLS, and compared it with another highly accepted, mortality-oriented algorithm, Levine’s BA, also known as PhenoAge. Finally, we repeated all the analyses in NHANES III and 1999–2010 to further confirm the validity and superiority of our strategy.
Death is the ultimate end of biological aging. The association of BA, especially BA based on clinical biomarkers and the KDM algorithm, with mortality has been confirmed by several independent research teams (15,19,21,23,24,28–30,33,35). However, the predictive accuracy of BA for mortality did not receive enough attention in these studies. From bench to bedside, the predictive accuracy of BA is important. In some other biomarker processing methods, mortality has been used as a major health outcome to screen aging biomarkers. Kim et al. used machine learning to select health items that were highly predictive of survival/mortality (30). Levine et al. used Cox penalized regression of mortality on 42 clinical markers and CA to select variables for inclusion in their phenotypic age score (17). In our research, we used the C-index of the Cox model to evaluate the predictive performance of KD-BAs for all-cause mortality. Considering that the BAs and CA were usually highly correlated (0.39–1), which might increase the instability of the Cox model, we used the difference between BA and CA (ΔAge) instead of BA alone as an aging predictor of mortality in the sex- and CA-adjusted Cox model. Unsurprisingly, univariate (K–M curve) and multivariate analyses both confirmed the strong association of ΔAge with all-cause mortality in the CHNS (full sample and older adults), CLHLS, NHANES III, and NHANES 1999–2010. In general, adding ΔAge to the sex- and CA-adjusted Cox models enhanced the accuracy of mortality prediction for the majority of BA panels, that is, it improved the C-index values (Figure 1B, Supplementary Table S3). We also used aging rate as an extra predictor to assess the association of BA longitudinal change with mortality. In the CLHLS cohort, although the population was much older, the aging rate was also significantly associated with mortality, which was validated in both univariate and multivariate analyses. This finding indicated that BA could be applied not only in a cross-sectional context to assess aging status for individuals but also in longitudinal studies to monitor their aging trajectories. Additionally, BA was found to be a potential indicator for evaluating the effectiveness of antiaging interventions in clinical trials, as shown in previous work (47), since the aging biomarkers involved in the construction of BA were readily available. Previous studies defined BA divided by CA (BA/CA) as the “aging rate” to evaluate an individual’s aging status (45). We thought that BA change along with CA (ΔBA/ΔCA) was appropriate for measuring the real aging rate, since it reflected the aging process but not one’s stationary status characterized by ΔAge (BA-CA) or previous “aging rate” (BA/CA) (45).
Aging biomarker processing is an extremely important part of BA construction, but it has not been highlighted in previous studies. Our work bridges this gap. The major difference from previous work was that we included biomarkers with high collinearity in our study, including WC and WHR; glucose and GH; lnCRE and urea; SBP, DBP and PP; and TC, LDL_C, TG and APO-B. First, the presence of biomarkers with collinearity in a subset might reduce the accuracy of BA in mortality prediction, even lnCRE and urea, whose correlation was only 0.39. In previous studies, the criteria for determining whether there was collinearity between biomarkers were not uniform across researchers, varying from 0.4 to 0.7 (7,29). If the criteria are set too loosely, redundant biomarkers might be included, which might hinder the use of the BA equation and reduce the performance of BA. For trade-offs between biomarkers with collinearity, previous articles generally left biomarkers with greater correlation with CA or biomarkers with fewer missing values (7,21,29). Our results did not support this trade-off approach. For instance, compared to urea, lnCRE had smaller correlation coefficients with CA, but BAs estimated by subsets with creatinine performed better than BAs estimated by subsets with urea in predicting mortality. This phenomenon was also observed for blood pressures, and DBP, but not SBP or PP, enhanced BA performance for predicting death in CHNS. Rather unexpectedly, TC, LDL_C and APO_B, which all strongly correlated with CA, significantly attenuated the performance of BA. Most of the articles related to the construction of BA with clinical biomarkers used SBP and blood lipid profiles as key biomarkers. Liu’s work is a typical example of these studies (29). Liu’s work and our study both used the data from the CHNS 2009 wave to construct the BA equation, and the differences are listed as follows. First, we included biomarkers with collinearity to find the most predictive subset. Second, we obtained body measurements, and the final results confirmed that their participation improved the performance of BA. Third, Liu’s BA was composed of 12 biomarkers. In the all-subset analysis, the C-index of Liu’s model ranked 44 776th of 130 918 models, and our best-fitting KD-BA far surpassed Liu’s BA in terms of the accuracy of mortality prediction. Furthermore, the superiority of our best-fitting KD-BA was verified in another independent cohort, the CLHLS, but Liu’s team did not observe a significant association of their KD-BA with mortality in the same cohort (48). More importantly, we had only five biomarkers in the best-fitting KD-BA equation, which was more concise and practical. Inappropriate biomarker selection might diminish the mortality predictive performance of KD-BA. This might be why Li et al. reported that physiological age (based on KDM algorithms) had a weaker effect on mortality risk than other aging measures, as physiological age is composed of biomarkers with high collinearity, such as SBP, DBP, TC and APO_B (3).
We also compared our best-fitting KD-BA with the currently most popular method, Levine’s BA in terms of accuracy in predicting death. Due to the specificity of biomarkers and ethnic differences between cohorts, we developed a Chinese version of Levine’s BA (c-Levine’s BA) composed of the most predictive subsets of biomarkers of KD-BA in the CHNS. Although the correlation between the best-fitting KD-BA and CA was not as strong as that between c-Levine’s BA and CA, the best-fitting KD-BA and c-Levine’s BA had comparable accuracy in predicting death, both in the CHNS and CLHLS. Surprisingly, for male participants, the accuracy of the best-fitting KD-BA for predicting mortality was better than that of Levine’s BA in both the reference population and NHANES 1999–2010. For women, the poor performance of the KD-BA in NHANES 1999–2010 might be due to the absence of WHR, one of the biomarkers in the most predictive subset in the reference population.
Much previous work has reported the correlation between BA and CA or the scatter plot of both (18,20,21,27,29), and some researchers even applied the correlation to evaluate or compare the performance of obtained BAs (27). This did not make much sense because CA was involved in BA construction as an aging biomarker based on the KDM algorithm (36). In fact, we found that almost all estimated BAs highly correlated with CA (r > .99) had much poorer performance for predicting death (Supplementary Table S3). In other words, BA, which was almost equal to CA, could not provide more information about individual biological aging beyond CA (49). A total of 130 918 panels of coefficients also did not have a one-to-one correlation with the C-index. Therefore, the correlation coefficient was not a suitable indicator for evaluating the performance of BA.
Considering the relatively low mortality rate in the CHNS, we did not stratify by sex to estimate BA, as done in previous studies (19,21,28), to avoid low test power. The repeat and validation analyses in NHANES stratified by sex made up for this limitation. As specific causes of death were not available in the CHNS and CLHLS, using all-cause mortality as the evaluation indicator might weaken the predictive performance of KD-BA for aging-related mortality. As we have just conducted a preliminary validation analysis in the NHANES, for technical reasons we chose all-cause mortality rather than a better choice—aging-related mortality—as the indicator for KD-BA. This may partly explain the superiority of the best-fitting KD-BA over Levine’s BA for all-cause mortality prediction in men. Another important question that needs to be addressed in further research is decisive reasons for the selection of biomarkers with high collinearity into the most predictive subset. In addition, the all-subset approach required the support of complex programming coding and strong computational power.
Overall, we found the most predictive subset composed of fewer biomarkers, and the estimated KD-BA had the most accurate predictive value for all-cause mortality. Further studies should focus on the association of the best-fitting KD-BA and other aging-related outcomes, such as diseases and cognitive phenotypes.
Supplementary Material
Acknowledgments
The extensive data calculation was supported by the Medical Research Data Center of Fudan University. This research uses data from China Health and Nutrition Survey (CHNS) and Chinese Longitudinal Healthy Longevity Survey (CLHLS). We are grateful to research grant funding from the National Institute for Health (NIH), the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) for R01 HD30880, National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) for R01DK104371 and R01HL108427, the NIH Fogarty grant D43 TW009077 for financial support for the CHNS data collection and analysis files since 1989, and the China-Japan Friendship Hospital, Ministry of Health for support for CHNS 2009, Chinese National Human Genome Center at Shanghai since 2009, and Beijing Municipal Center for Disease Prevention and Control since 2011. We thank the National Institute for Nutrition and Health, China Center for Disease Control and Prevention, Beijing Municipal Center for Disease Control and Prevention, and the Chinese National Human Genome Center at Shanghai. We also appreciated the Center for Healthy Aging and Development Studies, Peking University.
Contributor Information
Kai Wei, Department of Laboratory Medicine, Huashan Hospital, Fudan University, Shanghai, China.
Shanshan Peng, Department of Laboratory Medicine, Huashan Hospital, Fudan University, Shanghai, China.
Na Liu, Department of Laboratory Medicine, Huashan Hospital, Fudan University, Shanghai, China.
Guyanan Li, Department of Clinical Laboratory Medicine, Fifth People’s Hospital of Shanghai Fudan University, Shanghai, China.
Jiangjing Wang, Shanghai Advanced Institute of Finance, Shanghai Jiao Tong University, Shanghai, China.
Xiaotong Chen, Department of Clinical Laboratory, Central Laboratory, Jing’an District Central Hospital of Shanghai, Fudan University, Shanghai, China.
Leqi He, Department of Clinical Laboratory Medicine, Fifth People’s Hospital of Shanghai Fudan University, Shanghai, China.
Qiudan Chen, Department of Clinical Laboratory, Central Laboratory, Jing’an District Central Hospital of Shanghai, Fudan University, Shanghai, China.
Yuan Lv, Department of Laboratory Medicine, Huashan Hospital, Fudan University, Shanghai, China; National Clinical Research Center for Aging and Medicine, Huashan Hospital, Fudan University, Shanghai, China.
Huan Guo, Department of Occupational and Environmental Health, State Key Laboratory of Environmental Health (Incubating), School of Public Health, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China.
Yong Lin, Department of Laboratory Medicine, Huashan Hospital, Fudan University, Shanghai, China; Department of Clinical Laboratory Medicine, Fifth People’s Hospital of Shanghai Fudan University, Shanghai, China; Department of Clinical Laboratory, Central Laboratory, Jing’an District Central Hospital of Shanghai, Fudan University, Shanghai, China; National Clinical Research Center for Aging and Medicine, Huashan Hospital, Fudan University, Shanghai, China.
Funding
This work was supported by the National Key Research and Development Program of China (2018YFC2000200).
Conflict of Interest
None declared.
Author Contributions
K.W. designed the analysis, analyzed the data, conducted, and drafted the manuscript. H.G. and Y.L. critically revised the analysis plan and manuscript. All authors had commented on the report drafts and approved the final submitted version.
References
- 1. Rando TA, Wyss-Coray T. Asynchronous, contagious and digital aging. Nat Aging. 2021;1(1):29–35. doi: 10.1038/s43587-020-00015-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Jansen R, Han LK, Verhoeven JE, et al. An integrative study of five biological clocks in somatic and mental health. Elife. 2021;10:e59479. doi: 10.7554/eLife.59479 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Li X, Ploner A, Wang Y, et al. Longitudinal trajectories, correlations and mortality associations of nine biological ages across 20-years follow-up. Elife. 2020;9:e51507. doi: 10.7554/eLife.51507 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Williams DM, Jylhävä J, Pedersen NL, Hägg S. A frailty index for UK Biobank participants. J Gerontol A Biol Sci Med Sci. 2019;74(4):582–587. doi: 10.1093/gerona/gly094 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Cohen AA, Milot E, Yong J, et al. A novel statistical approach shows evidence for multi-system physiological dysregulation during aging. Mech Ageing Dev. 2013;134(3–4):110–117. doi: 10.1016/j.mad.2013.01.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Li Q, Wang S, Milot E, et al. Homeostatic dysregulation proceeds in parallel in multiple physiological systems. Aging Cell. 2015;14(6):1103–1112. doi: 10.1111/acel.12402 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Hastings WJ, Almeida DM, Shalev I. Conceptual and analytical overlap between allostatic load and systemic biological aging measures: analyses from the National Survey of Midlife Development in the United States. J Gerontol A Biol Sci Med Sci. 2022;77(6):1179–1188. doi: 10.1093/gerona/glab187 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Belsky DW, Caspi A, Houts R, et al. Quantification of biological aging in young adults. Proc Natl Acad Sci USA. 2015;112(30):E4104–E4110. doi: 10.1073/pnas.1506264112 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Horvath S, Raj K. DNA methylation-based biomarkers and the epigenetic clock theory of ageing. Nat Rev Genet. 2018;19(6):371–384. doi: 10.1038/s41576-018-0004-3 [DOI] [PubMed] [Google Scholar]
- 10. Peters MJ, Joehanes R, Pilling LC, et al. The transcriptional landscape of age in human peripheral blood. Nat Commun. 2015;6:8570. doi: 10.1038/ncomms9570 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Earls JC, Rappaport N, Heath L, et al. Multi-omic biological age estimation and its correlation with wellness and disease phenotypes: a longitudinal study of 3,558 individuals. J Gerontol A Biol Sci Med Sci. 2019;74(Suppl 1):S52–S60. doi: 10.1093/gerona/glz220 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Johnson AA, Shokhirev MN, Wyss-Coray T, Lehallier B. Systematic review and analysis of human proteomics aging studies unveils a novel proteomic aging clock and identifies key processes that change with age. Ageing Res Rev. 2020;60:101070. doi: 10.1016/j.arr.2020.101070 [DOI] [PubMed] [Google Scholar]
- 13. Lehallier B, Shokhirev MN, Wyss-Coray T, Johnson AA. Data mining of human plasma proteins generates a multitude of highly predictive aging clocks that reflect different aspects of aging. Aging Cell. 2020;19(11):e13256. doi: 10.1111/acel.13256 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Krištić J, Vučković F, Menni C, et al. Glycans are a novel biomarker of chronological and biological ages. J Gerontol A Biol Sci Med Sci. 2014;69(7):779–789. doi: 10.1093/gerona/glt190 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Cole JH, Ritchie SJ, Bastin ME, et al. Brain age predicts mortality. Mol Psychiatry. 2018;23(5):1385–1392. doi: 10.1038/mp.2017.62 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Xia X, Chen X, Wu G, et al. Three-dimensional facial-image analysis to predict heterogeneity of the human ageing rate and the impact of lifestyle. Nat Metab. 2020;2(9):946–957. doi: 10.1038/s42255-020-00270-x [DOI] [PubMed] [Google Scholar]
- 17. Levine ME, Lu AT, Quach A, et al. An epigenetic biomarker of aging for lifespan and healthspan. Aging (Albany NY). 2018;10(4):573–591. doi: 10.18632/aging.101414 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Cho IH, Park KS, Lim CJ. An empirical comparative study on biological age estimation algorithms with an application of Work Ability Index (WAI). Mech Ageing Dev. 2010;131(2):69–78. doi: 10.1016/j.mad.2009.12.001 [DOI] [PubMed] [Google Scholar]
- 19. Levine ME. Modeling the rate of senescence: can estimated biological age predict mortality more accurately than chronological age? J Gerontol A Biol Sci Med Sci. 2013;68(6):667–674. doi: 10.1093/gerona/gls233 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Jee H, Park J. Selection of an optimal set of biomarkers and comparative analyses of biological age estimation models in Korean females. Arch Gerontol Geriatr. 2017;70:84–91. doi: 10.1016/j.archger.2017.01.005 [DOI] [PubMed] [Google Scholar]
- 21. Zhong X, Lu Y, Gao Q, et al. Estimating biological age in the Singapore Longitudinal Aging Study. J Gerontol A Biol Sci Med Sci. 2020;75(10):1913–1920. doi: 10.1093/gerona/glz146 [DOI] [PubMed] [Google Scholar]
- 22. Hirota N, Suzuki S, Arita T, Yagi N, Otsuka T, Yamashita T. Prediction of biological age and all-cause mortality by 12-lead electrocardiogram in patients without structural heart disease. BMC Geriatr. 2021;21(1):460. doi: 10.1186/s12877-021-02391-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Mitnitski A, Howlett SE, Rockwood K. Heterogeneity of human aging and its assessment. J Gerontol A Biol Sci Med Sci. 2017;72(7):877–884. doi: 10.1093/gerona/glw089 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Murabito JM, Zhao Q, Larson MG, et al. Measures of biologic age in a community sample predict mortality and age-related disease: the Framingham Offspring Study. J Gerontol A Biol Sci Med Sci. 2018;73(6):757–762. doi: 10.1093/gerona/glx144 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Waziry R, Gras L, Sedaghat S, et al. Quantification of biological age as a determinant of age-related diseases in the Rotterdam Study: a structural equation modeling approach. Eur J Epidemiol. 2019;34(8):793–799. doi: 10.1007/s10654-019-00497-3 [DOI] [PubMed] [Google Scholar]
- 26. Crimmins EM, Thyagarajan B, Kim JK, Weir D, Faul J. Quest for a summary measure of biological age: the health and retirement study. Geroscience. 2021;43(1):395–408. doi: 10.1007/s11357-021-00325-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Verschoor CP, Belsky DW, Ma J, Cohen AA, Griffith LE, Raina P. Comparing biological age estimates using domain-specific measures from the Canadian Longitudinal Study on Aging. J Gerontol A Biol Sci Med Sci. 2021;76(2):187–194. doi: 10.1093/gerona/glaa151 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Chan MS, Arnold M, Offer A, et al. A biomarker-based biological age in UK Biobank: composition and prediction of mortality and hospital admissions. J Gerontol A Biol Sci Med Sci. 2021;76(7):1295–1302. doi: 10.1093/gerona/glab069 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Liu Z. Development and validation of 2 composite aging measures using routine clinical biomarkers in the Chinese population: analyses from 2 prospective cohort studies. J Gerontol A Biol Sci Med Sci. 2021;76(9):1627–1632. doi: 10.1093/gerona/glaa238 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Kim S, Fuselier J, Welsh DA, Cherry KE, Myers L, Jazwinski SM. Feature selection algorithms enhance the accuracy of frailty indexes as measures of biological age. J Gerontol A Biol Sci Med Sci. 2021;76(8):1347–1355. doi: 10.1093/gerona/glab018 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Sagers L, Melas-Kyriazi L, Patel CJ, Manrai AK. Prediction of chronological and biological age from laboratory data. Aging (Albany NY). 2020;12(9):7626–7638. doi: 10.18632/aging.102900 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Sayed N, Huang Y, Nguyen K, et al. An inflammatory aging clock (iAge) based on deep learning tracks multimorbidity, immunosenescence, frailty and cardiovascular aging. Nat Aging. 2021;1:598–615. doi: 10.1038/s43587-021-00082-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Parker DC, Bartlett BN, Cohen HJ, et al. Association of blood chemistry quantifications of biological aging with disability and mortality in older adults. J Gerontol A Biol Sci Med Sci. 2020;75(9):1671–1679. doi: 10.1093/gerona/glz219 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Hastings WJ, Shalev I, Belsky DW. Comparability of biological aging measures in the National Health and Nutrition Examination Study, 1999-2002. Psychoneuroendocrinology. 2019;106:171–178. doi: 10.1016/j.psyneuen.2019.03.012 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Liu Z, Kuo PL, Horvath S, Crimmins E, Ferrucci L, Levine M. A new aging measure captures morbidity and mortality risk across diverse subpopulations from NHANES IV: a cohort study. PLoS Med. 2018;15(12):e1002718. doi: 10.1371/journal.pmed.1002718 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Klemera P, Doubal S. A new approach to the concept and computation of biological age. Mech Ageing Dev. 2006;127(3):240–248. doi: 10.1016/j.mad.2005.10.004 [DOI] [PubMed] [Google Scholar]
- 37. Popkin BM, Du S, Zhai F, Zhang B. Cohort profile: The China Health and Nutrition Survey—monitoring and understanding socio-economic and health change in China, 1989-2011. Int J Epidemiol. 2010;39(6):1435–1440. doi: 10.1093/ije/dyp322 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Zhang B, Zhai FY, Du SF, Popkin BM. The China Health and Nutrition Survey, 1989–2011. Obes Rev. 2014;15 (Suppl 1):2–7. doi: 10.1111/obr.12119 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Zeng Y, Feng Q, Hesketh T, Christensen K, Vaupel JW. Survival, disabilities in activities of daily living, and physical and cognitive functioning among the oldest-old in China: a cohort study. Lancet. 2017;389(10079):1619–1629. doi: 10.1016/S0140-6736(17)30548-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Lv YB, Gao X, Yin ZX, et al. Revisiting the association of blood pressure with mortality in oldest old people in China: community based, longitudinal prospective study. BMJ 2018;361:k2158. doi: 10.1136/bmj.k2158 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41. Zhao Y, Zhang M, Liu Y, et al. Adult height and risk of death from all-cause, cardiovascular, and cancer-specific disease: the Rural Chinese Cohort Study. Nutr Metab Cardiovasc Dis. 2019;29(12):1299–1307. doi: 10.1016/j.numecd.2019.05.067 [DOI] [PubMed] [Google Scholar]
- 42. Chen Y, Yang Y, Jiang H, Liang X, Wang Y, Lu W. Associations of BMI and waist circumference with all-cause mortality: a 22-year cohort study. Obesity (Silver Spring). 2019;27(4):662–669. doi: 10.1002/oby.22423 [DOI] [PubMed] [Google Scholar]
- 43. Partridge L, Deelen J, Slagboom PE. Facing up to the global challenges of ageing. Nature. 2018;561(7721):45–56. doi: 10.1038/s41586-018-0457-8 [DOI] [PubMed] [Google Scholar]
- 44. Kwon D, Belsky DW. A toolkit for quantification of biological age from blood chemistry and organ function test data: BioAge. Geroscience. 2021;43(6):2795–2808. doi: 10.1007/s11357-021-00480-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45. Wang C, Guan X, Bai Y, et al. A machine learning-based biological aging prediction and its associations with healthy lifestyles: the Dongfeng-Tongji cohort. Ann N Y Acad Sci. 2022;1507(1):108–120. doi: 10.1111/nyas.14685 [DOI] [PubMed] [Google Scholar]
- 46. Uno H, Cai T, Pencina MJ, D’Agostino RB, Wei LJ. On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Stat Med. 2011;30(10):1105–1117. doi: 10.1002/sim.4154 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47. Belsky DW, Huffman KM, Pieper CF, Shalev I, Kraus WE. Change in the rate of biological aging in response to caloric restriction: CALERIE biobank analysis. J Gerontol A Biol Sci Med Sci. 2017;73(1):4–10. doi: 10.1093/gerona/glx096 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48. Jin S, Li C, Cao X, Chen C, Ye Z, Liu Z. Association of lifestyle with mortality and the mediating role of aging among older adults in China. Arch Gerontol Geriatr. 2022;98:104559. doi: 10.1016/j.archger.2021.104559 [DOI] [PubMed] [Google Scholar]
- 49. Zhang Q, Vallerga CL, Walker RM, et al. Improved precision of epigenetic clock estimates across tissues and its implication for biological ageing. Genome Med. 2019;11(1):54. doi: 10.1186/s13073-019-0667-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.