Association of Blood Chemistry Quantifications of Biological Aging With Disability and Mortality in Older Adults

Abstract

Quantification of biological aging has been proposed for population surveillance of age-related decline in system integrity and evaluation of geroprotective therapies. However, methods of quantifying biological aging have been little studied in geriatric populations. We analyzed three clinical-biomarker-algorithm methods to quantify biological aging. Klemera–Doubal method Biological Age and homeostatic dysregulation algorithms were parameterized from analysis of U.S. National Health and Nutrition Examination Surveys (NHANES) data (N = 36,207) based on published methods. Levine method Biological Age was adapted from published analysis of NHANES data. Algorithms were applied to biomarker data from the Duke Established Populations for Epidemiologic Studies of the Elderly (Duke-EPESE) cohort of older adults (N = 1,374, aged 71–102 years, 35% male, 52% African American). We tested associations of biological aging measures with participant reported Activities of daily living (ADL), instrumental activities of daily living (IADL) dependencies, and mortality. We evaluated the sensitivity of results to the demographic composition of reference samples and biomarker sets used to develop biological aging algorithms. African American and white Duke-EPESE participants with more advanced biological aging reported dependence in more ADLs and IADLs and were at increased risk of death over follow-up through 2017. Effect sizes were similar across algorithms, but were strongest for Levine method Biological Age (per-quintile increase in ADL incidence rate ratio = 1.25, 95% confidence interval [1.17–1.37], IADL incidence rate ratio = 1.23 [1.15–1.32], mortality hazard ratio = 1.12 [1.08–1.16]). Results were insensitive to demographic composition of reference samples, but modestly sensitive to the biomarker sets used to develop biological aging algorithms. Blood-chemistry-based quantifications of biological aging show promise for evaluating the effectiveness of interventions to extend healthy life span in older adults.

Biological aging refers to the gradual and progressive decline in system integrity that is thought to mediate aging-related increases in disease and disability (1,2). Measurements of biological aging have potential utility for clinical and population surveillance and for evaluation of therapies intended to delay or prevent aging-related disease and disability (3,4).

Several methods have been proposed to measure biological aging (5). Among these are algorithms that combine information from blood chemistry and related data routinely collected during clinical encounters. Even in adults still young enough to be free of most chronic disease, biological aging measurements derived from such clinical data can predict risk of death, disability, and cognitive decline (6–11). However, as most research has focused on midlife samples or samples of mixed chronological age, more research is needed on the utility of these measurements for research on older adults.

To evaluate the potential utility of biological aging measurements in older adults, we tested three proposed methods to measure biological aging in the Duke Established Populations for Epidemiologic Studies of the Elderly (Duke-EPESE) sample of older black and white adults in the United States (N = 1,374, age range = 71–102, 35% male, 52% African American). We quantified biological aging using the published Klemera–Doubal method (KDM) Biological Age (12), homeostatic dysregulation (9), and Phenotypic Age (13) (hereafter referred to as Levine method [LM] Biological Age) algorithms. We constructed KDM Biological Age and homeostatic dysregulation algorithms from analysis of the U.S. National Health and Nutrition Examination Surveys (NHANES; N = 36,207) and then applied these algorithms along with the published LM Biological Age algorithm (13) to Duke-EPESE data to compute biological aging values for participants. We then tested if participants with more advanced biological aging experienced more disability and were at increased risk for mortality.

Methods

Duke Established Populations for Epidemiologic Studies of the Elderly

The Duke-EPESE sampled persons aged 65–105 (response rate 80%, N = 4,162, 54% African American, 45% white, <1% other race) living in five adjacent Piedmont North Carolina area counties in 1986 (14). Data on social context, functional status, and health were gathered annually over the first 7 years of the study, with a final data collection completed in the 10th study year (15). At the time of the seventh data collection in 1991/1992, when all participants were aged 71 years or older, trained phlebotomists visited participants at home and drew blood samples from consenting participants (N = 1,554) (16). This blood-draw sample formed the basis of the current study. The Duke University Medical Center institutional review board approved this study. Written consent was obtained from all participants.

Biomarker data were assayed with the SMAC-24 chemistry panel and complete blood count using standard procedures (16). High-sensitivity C-reactive protein (CRP) was measured using enzyme-linked immunosorbent assays (Mesoscale Discovery Systems, Rockville, MD). Glycated albumin (used to estimate glycated hemoglobin (17); Supplementary Methods) was measured using an enzymatic assay (Diazyme Laboratories, CA).

Biological Aging Measures

Calculating human biological age is a relatively recent enterprise (18), and there is disagreement about methods (19). Our aim was to borrow from established methods suitable for application to the Duke-EPESE sample. We analyzed three different algorithms: KDM Biological Age, homeostatic dysregulation, and LM Biological Age.

We selected the KDM Biological Age (12) to estimate biological age from physiological data collected in Duke-EPESE because this method has a track record of replication for prediction of morbidity, mortality, and indicators of healthspan (6,10,20–24) and has been shown to outperform several alternative methods in prediction of mortality (7). We selected the homeostatic dysregulation algorithm (9) because it has been shown to predict morbidity, mortality, and functional deficits (8,10,13) and, unlike the KDM Biological age, does not assume unidirectionality of aging–biomarker relationships, and does not incorporate chronological age into calculations (19).

KDM Biological Age

The KDM builds a biological age algorithm based on a series of regressions of individual biomarkers on chronological age in a reference population (Supplementary Methods). Following previous work (6,22), we used nonpregnant participants in the U.S. Centers for Disease Control and Prevention’s National Health and Nutrition Examination Surveys who were aged 30–75 years to form this reference population. We included data from NHANES III and continuous NHANES panels collected during 1999–2016 (N = 36,207). An individual’s KDM biological age prediction corresponds to the chronological age at which her/his physiology would be approximately normal in the NHANES.

Homeostatic dysregulation

The homeostatic dysregulation method calculates biological age based on a panel of biomarkers’ Mahalanobis distance (25) from a reference population. Following previous work (22), we used nonobese, nonpregnant NHANES participants aged 20–30 years to form a young, healthy reference population. We included data from NHANES III and continuous NHANES panels collected during 1999–2016 (N = 1,687). An individual’s homeostatic dysregulation value corresponds to how different their physiology is from the young, healthy NHANES sample.

To compose the KDM Biological Age and homeostatic dysregulation algorithms, we first selected biomarkers included in previous KDM analyses (6,22) that were available in Duke-EPESE: albumin, alkaline phosphatase, blood urea nitrogen, creatinine, CRP, glycated hemoglobin (HbA1C), uric acid, and white blood cell count. We did not include total cholesterol and systolic blood pressure in the Duke-EPESE KDM Biological age algorithm because, although levels of these biomarkers tend to rise across most of the life-course, late-life declines are associated with increased risk of mortality (26–29). Forced expiratory volume in 1 second and cytomegalovirus optical density were included in Levine’s 2013 KDM algorithm (6), but were not measured in Duke-EPESE. We supplemented the Duke-EPESE biomarker list with two biomarkers identified in machine-learning analysis of NHANES data to derive an alternative algorithm that was not included in previous KDM papers: lymphocyte percent and mean cell volume (13,30). A third biomarker identified in that analysis, red-cell distribution width, was not measured in Duke-EPESE. Biomarker summary statistics for the NHANES reference sample are summarized in Supplementary Dataset 1.

A question in biological aging research concerns sensitivity of algorithm-based biological aging measures to the specific set of biomarkers included in the algorithm (31). To address this question, we compared KDM Biological Age algorithms constructed from the biomarkers listed above with algorithms developed using three additional sets of biomarkers matched to the sets used in previous studies: (i) The set of biomarkers included in the original KDM Biological Age algorithm published by Levine (6) that were available in the Duke-EPESE database (albumin, alkaline phosphatase, blood urea nitrogen, creatinine, CRP, HbA1C, total cholesterol, and systolic blood pressure); (ii) the set used in our previous analysis of the CALERIE Trial (22) (albumin, alkaline phosphatase, blood urea nitrogen, creatinine, CRP, HbA1C, uric acid, and white blood cell count); and (iii) the set of biomarkers selected in the recent machine-learning analysis by Levine and colleagues (13) that were available in the Duke-EPESE database (albumin, creatinine, CRP, glucose, lymphocyte percent, mean cell volume, white blood cell count). Our primary biomarker set and the three additional sets are summarized in Supplementary Table 1. Analysis of correlation among chronological age and KDM Biological Age measures based on the different biomarker sets is reported in Supplementary Table 2. In the Duke-EPESE, KDM Biological Age values computed from algorithms including slightly different sets of biomarkers were similarly correlated with chronological age (r = .23–.27) and were highly correlated with one another (r > .58).

A second question in biological aging research concerns whether the demographic composition of the reference sample used to derive a measurement algorithm should be matched to the individual/sample that is the target for biological age measurement (32,33). To address this question, we conducted sensitivity analyses of biological age algorithms developed using reference populations demographically matched to subgroups of the Duke-EPESE. Specifically, we conducted analyses to develop versions of the KDM Biological Age algorithm in African American, white, and older-adult (aged ≥ 65) subsamples of the NHANES database. Biomarker summary statistics in the NHANES subsamples are reported in Supplementary Dataset 1. Analysis of correlation among chronological age and KDM Biological Age measures derived from different training samples is reported in Supplementary Table 3. In the Duke-EPESE, biological age values computed from algorithms trained in the full NHANES dataset were highly correlated with biological-age values computed from algorithms trained in African American and non-Hispanic white subsamples (r > .97). Correlations were somewhat less strong with biological age values computed from the algorithm trained in older NHANES participants (r > .55).

Parameters for each of the KDM Biological Age and homeostatic dysregulation algorithms are reported in Supplementary Datasets 2 and 3.

LM Biological Age

Finally, we computed a measure of biological aging using the algorithm recently published by Levine and colleagues (13,30). This algorithm is referred to in the original publications as “phenotypic age.” For consistency with other naming within this article, we refer to the algorithm as “Levine Method (LM) Biological Age.” The LM Biological Age algorithm was constructed from elastic-net regression of mortality on a panel of 42 biomarkers in the NHANES III data set. The analysis selected 9 biomarkers for inclusion in the final algorithm: albumin, creatinine, CRP, glucose, white blood cell count, lymphocyte percent, mean cell volume, and red-cell distribution width. Within the algorithm, mortality hazards predicted from these nine biomarkers are converted to age values. The conversion is made by comparing the biomarker-model-estimated hazard with hazards estimated from a univariate model including only chronological age. The chronological age producing a univariate hazard equivalent to the person’s biomarker-model-hazard is assigned as that person’s LM Biological Age. Thus, a person’s LM Biological Age corresponds to the chronological age with the equivalent expected mortality risk based on the NHANES III reference sample used to derive the algorithm.

One of the nine biomarkers included in the LM Biological Age algorithm, red-cell distribution width, was not measured in Duke-EPESE. To compute the algorithm, we imputed Duke-EPESE participants’ red-cell distribution width values based on their sex, age, and values for the other biomarkers. Imputation was conducted based on regression analysis in the NHANES database.

Criterion Measures

We conducted tests of criterion validity of biological aging algorithms using two end points, disability and mortality.

Disability

We analyzed disability in Duke-EPESE using activities of daily living (ADL) and instrumental activities of daily living (IADL) scales. ADLs were measured using Katz (34), Rosow-Breslau (35), and Nagi (36,37) scales. Each scale measured a count of activities the individual was unable to do without assistance. Katz activities are walking across a small room, bathing, personal grooming, dressing, feeding oneself, transferring from a bed to chair, and toileting. Rosow-Breslau activities are carrying a bag of groceries, climbing one flight of stairs, and walking one-half mile. Nagi activities are moving large objects, stooping or kneeling, carrying weights over 10 lbs, reaching arms above shoulders, and writing or fingering small objects. We computed an overall ADL count as the number of activities for which participants reported impairments across all three scales. The three scales included a total of 15 items. The observed range of ADLs was 0–13 (M = 2, SD = 3). IADLs were measured using the Older Americans Resources and Services IADL scale (38). The IADL scale measured a count of activities the participant was unable to do without assistance. IADL activities were using a telephone, driving a car or traveling alone by bus or taxi, going shopping for household needs, preparing meals, doing housework, taking medicines, and handling money (e.g. paying bills). We computed the IADL count as the number of activities for which the participant reported requiring assistance (range 0–7, M = 1, SD = 2).

Mortality

Participants’ death dates were obtained from linkage with the National Death Index (39). Linkages were performed most recently in 2017, by which time 97% of participants had been linked to a date of death (16).

Statistical Analysis

We analyzed KDM biological age, homeostatic dysregulation, and LM Biological Age among N = 1,374 (35% male) African American (52%) and white (48%) participants in Duke-EPESE with data on biological aging measures and disability and mortality outcomes; 11 individuals who reported other race/ethnicity were classified as white for analysis. Sample characteristics are reported in Table 1. Biomarker summary statistics for the Duke-EPESE sample are reported in Supplementary Dataset 1. Comparison of biomarker distributions in the Duke-EPESE to the NHANES training sample is shown in Supplementary Figure 1.

We tested associations of biological aging measures with chronological age, disability, and mortality in the full Duke-EPESE sample and in subsamples of African American and white participants. We analyzed data from Duke-EPESE participants for whom KDM Biological Age, homeostatic dysregulation, and LM Biological Age could be measured and for whom disability and mortality data were available (N = 1,374; 35% male; 52% African American). We analyzed associations among measures of biological aging using Pearson correlations. We analyzed associations between measures of biological aging and chronological age using linear regression to estimate standardized beta coefficients (interpretable as Pearson r). We tested associations between measures of biological aging and counts of ADLs and IADLs using negative binomial regression to estimate incidence rate ratios (IRRs). We tested associations between biological aging measures and mortality using survival analysis. We used Cox proportional hazard models to estimate hazard ratios (HRs). Subjects entered survival analysis on the date of their blood draw and exited on the date of their death.

For analysis of disability and mortality, we analyzed biological age measures as follows: First, we computed the difference between participants’ biological and chronological ages. For KDM Biological Age and LM Biological Age, we subtracted chronological age from estimated biological age; for homeostatic dysregulation, we fitted a regression of homeostatic dysregulation on chronological age and predicted residual values. Hereafter, we refer to these values as measures of biological age advancement. Second, we divided biological age advancement values into quintiles, with the first quintile identifying individuals with the least biological age advancement and the fifth quintile identifying individuals with most biological age advancement. Third, we fitted regressions with these quintiles as the predictor variables.

In the analysis of the pooled Duke-EPESE sample, models included covariates for chronological age, African American/white race/ethnicity, and sex. All statistical analyses were conducted in R and Stata. R packages used in the analysis included compareGroups, ggplot2, ggpubr, ggstatsplot, MASS, survival, and survminer (citations in the Supplementary Materials).

Results

Consistent with findings from studies of midlife and mixed-age adults (10,40), Duke-EPESE participants’ KDM Biological Age, homeostatic dysregulation, and LM Biological Age values were correlated. Participants measured to have more advanced biological age by one algorithm also tended to have more advanced biological age on the others (Pearson r = .48–.75; Figure 1). We first report results for the analysis of KDM Biological Age and then report results from parallel analysis of homeostatic dysregulation and LM Biological Age.

Duke-EPESE participants’ chronological ages were correlated with their KDM Biological Ages (r = .26, p < .001). This correlation is lower than that has been reported from studies of samples with broader age distributions (6,20). However, sensitivity analyses testing the algorithm in hold-out samples drawn from NHANES III indicated that this attenuated correlation is characteristic of samples with more restricted age ranges, especially samples of older adults. For example, the correlation between KDM Biological Age and chronological age in NHANES III participants aged 70 and older was r = .35 (Supplementary Results, Supplementary Figure 2D). This attenuated correlation partly reflects increased total variance with advancing chronological age (NHANES III σ ²_BA = 256.75 in participants aged ≥ 65 compared with 164.21 in participants aged 40–60). These findings parallel published reports of reduced correlation between DNA methylation clocks and chronological age in restricted age range samples compared with broader age range samples (23,41,42).

African American Duke-EPESE Participants Had More Advanced KDM Biological Age When Compared With White Participants of the Same Chronological Age

Duke-EPESE participants’ KDM Biological Ages were, on average, 1.90 (SD = 20.0) years older than their chronological ages. This difference was greater in African Americans when compared with whites (b = 5.96 years, 95% confidence interval [CI] [3.86, 8.06]; p < .001; Supplementary Figure 3A), implying more advanced biological aging among the African American when compared with white EPESE participants. Stratified by sex, African American women had more advanced KDM Biological Age compared with white women (b = 8.44 years, 95% CI [5.91, 10.97]; p < .001), but there was no difference in African American men compared with white men (b = 1.32 years, 95% CI [−2.36, 5.01]; p = .480, Supplementary Figure 3).

Men in Duke-EPESE Had Similarly Advanced KDM Biological Age Compared With Women of the Same Chronological Age

Women in Duke-EPESE had somewhat younger KDM Biological Ages when compared with men, but this difference was not statistically significant at the α = .05 threshold (b = −1.69 years 95% CI [−3.91, 0.53]; p = .135). In studies in younger populations, women tend to have younger KDM Biological Ages when compared with men of the same chronological age (6,7). The statistically weaker findings in Duke-EPESE may reflect processes of selective mortality affecting the construction of the sample. Life expectancy at birth for the youngest participants in our Duke-EPESE sample (born in 1920) was 54.4 years for white men and 55.6 years for white women (43). Life expectancy was about a decade less for nonwhite Americans (data not reported specifically for African Americans). Thus, all participants in our analysis had outlived the expectation for their birth cohorts by a substantial degree. This remained true when considering life expectancies calculated for the 1920 birth cohort at age 20 years, which added about a decade of life to the estimates calculated at birth (44).

Duke-EPESE Participants With More Advanced KDM Biological Age Reported Dependence in More ADLs and IADLS When Compared With Participants With Less Advanced KDM Biological Age

We next turned from absolute differences in biological aging to consideration of relative differences. Specifically, we tested the hypothesis that older-adult participants in Duke-EPESE who exhibited signs of more advanced biological aging were at increased risk for disability and mortality. In these models, the predictor was quintile of biological age advancement, the difference between biological age and chronological age. At the time of biological aging measurement, 64% of EPESE participants reported limitation on at least one ADL and 38% reported limitation on at least one IADL. For both ADLs and IADLs, participants with more advanced KDM Biological Age reported greater dependence (for ADLs, per-quintile IRR = 1.19 [1.12, 1.27]; p < .001; for IADLs, IRR = 1.18 [1.10, 1.26]; p < .001; Figure 2).

Duke-EPESE Participants With More Advanced KDM Biological Age Were at Increased Mortality Risk Across Three Decades of Follow-up

By latest follow-up in National Death Index data in 2017, 97% of participants were matched to a death record (mean years of survival = 7.93, SD = 4.99). Duke-EPESE participants with more advanced KDM Biological Age were at increased risk of death (per-quintile HR = 1.09, [1.06, 1.13]; p < .001; Figure 3).

Effect-Size Comparisons Between African American and White Duke-EPESE Participants

Effect sizes for KDM Biological Age associations with disability and mortality were somewhat larger in white when compared with African American Duke-EPESE participants: for African Americans ADL IRR = 1.18, 95% CI [1.08–1.28], IADL IRR = 1.12 [1.03–1.21], mortality HR = 1.06 [1.01–1.12]; for whites ADL IRR = 1.23 95% CI [1.11–1.37], IADL IRR = 1.29 [1.15–1.45], mortality HR = 1.14 [1.08–1.20] (Supplementary Table 4). We tested African American-white effect-size differences using regression models with product terms testing interactions between biological aging and race/ethnicity. Interaction terms were not statistically different from zero at the α = .05 threshold for any analysis (for ADLs: p = .519; for IADLs: p = .086; for mortality: p = .117).

Results Were Similar in Analyses of Homeostatic Dysregulation

We computed homeostatic dysregulation using the same biomarkers used to compute KDM Biological Age. Compared with the young, healthy NHANES reference, Duke-EPESE participants showed substantial homeostatic dysregulation (M = 1.43, SD = 0.24; Table 1). Similar to results for KDM Biological Age, correlation between chronological age and homeostatic dysregulation was positive, but weaker than reported in studies of samples with younger and wider age distributions (r = .10; p < .001; Supplementary Results, Supplementary Figure 4A) (8,22). Homeostatic dysregulation was greater in African American compared with white Duke-EPESE participants (b = 0.10 log dysregulation units 95% CI [0.07, 0.12]; p < .001), but there was no difference in men compared with women (b = −0.02 log dysregulation units 95% CI [−0.05, 0.00]; p = .101). Duke-EPESE participants with greater homeostatic dysregulation reported dependence in more ADLs (per-quintile IRR = 1.17 [1.10, 1.25]; p < .001) and IADLs (IRR = 1.17 [1.10, 1.26]; p < .001; Supplementary Results, Supplementary Table 5 and Supplementary Figure 5B) and were at increased risk of death (per-quintile HR = 1.08 [1.04, 1.12]; p < .001; Supplementary Results, Supplementary Table 5 and Supplementary Figure 6B). Effect sizes for associations of homeostatic dysregulation with disability were somewhat larger for white when compared with African American Duke-EPESE participants, but these differences were not statistically significant at the α = .05 threshold (for ADLs: p = .468; for IADLs: p = .144). In African American Duke-EPESE participants, accelerated biological aging as measured by homeostatic dysregulation quintile was not associated with the hazard of mortality (HR = 1.03 [0.97, 1.08]; p = .350). Results from stratified analyses are reported in Supplementary Results, Supplementary Table 5.

Results Were Similar in Analyses of LM Biological Age

Duke-EPESE participants’ LM Biological Ages were, on average, 1.61 (SD = 8.63) years older than their chronological ages. The correlation between LM Biological Age and chronological age was stronger than for the KDM Biological Age (r = .49 when compared with r = .26 for KDM Biological Age), but still weaker than was reported for a mixed-age sample (13). Analysis of differences between African American when compared with white, and male when compared with female Duke-EPESE participants, yielded similar results to analyses of KDM Biological Age, but neither difference was statistically significant at the α = .05 threshold (for African American when compared with white participants, b = 0.85 years 95% CI [−0.06, 1.76]; p = .067; for female when compared with male participants, b = −1.69 years 95% CI [−3.91, 0.53]; p = .135). Again, similar to KDM Biological Age analyses, Duke-EPESE participants with more advanced LM Biological Age reported dependence in more ADLs (IRR = 1.25 [1.17, 1.33]; p < .001) and IADLs (IRR = 1.23 [1.15, 1.32]; p < .001) and were at increased risk of death (HR = 1.12 [1.08, 1.16]; p < .001; Supplementary Results, Supplementary Table 4; Supplementary Figures 5C and 6C). Effect sizes for associations of LM Biological Age with disability and mortality were somewhat larger for white when compared with African American Duke-EPESE participants; differences for ADLs and mortality were not statistically significant at the α = .05 threshold (for ADLs: p = .667; for mortality: p = .412). For IADLs, LM Biological Age was more predictive of disability in whites compared with African Americans (p < .001). Results from stratified analyses are reported in Supplementary Results, Supplementary Table 4; Supplementary Figures 4C and 5C.

Sensitivity Analysis of Biological-Age-Algorithm Reference-Sample Composition

We tested if developing KDM Biological Age and homeostatic dysregulation algorithm parameters from samples demographically matched to the Duke-EPESE would yield measures with improved criterion validity, that is, stronger associations with disability and mortality. We conducted analyses to develop the KDM Biological Age algorithm in older adults (aged ≥ 65) and in subsets of African American and white NHANES participants. We then used these algorithms to compute biological ages for EPESE participants. KDM Biological Age values computed from race/ethnic-matched training samples were nearly identical to the original KDM Biological Age values (r > .95). KDM Biological Age values computed from the older-adult training sample were also correlated with the original values, but the correlation was more modest (r = .56). Correlations among biological aging algorithms trained in different NHANES subsamples are reported in Supplementary Table 3. Finally, we computed effect sizes for associations between biological age measured by the new algorithms and measures of disability and mortality and compared these effect sizes to effect sizes for the original biological aging measure. Effect sizes were similar (Supplementary Results, Supplementary Tables 2–6). We repeated this analysis for the homeostatic dysregulation measure. Results were congruent with results for KDM Biological age (Supplementary Results, Supplementary Table 5).

Sensitivity Analysis of Biological-Age-Algorithm Biomarker Composition

We tested if composing biological age algorithms from biomarker lists used in previous studies affected results for analysis of disability and mortality (Supplementary Results, Supplementary Table 1). Algorithms trained using biomarker sets from previous studies yielded similar correlations with chronological age, were strongly correlated with one another, and produced similar results in criterion validity tests (Supplementary Results, Supplementary Tables 2 and 7). However, our primary biomarker set and the biomarker set matched to the LM Biological Age tended to produce modestly larger effect sizes for disability when compared with the biomarker sets from the 2013 Levine article and our analysis of the CALERIE Trial. Results for analysis of mortality were similar for all biomarker sets. Details are reported in Supplementary Results, Supplementary Table 7.

Discussion

We analyzed three methods proposed to quantify biological aging from blood chemistry data in a cohort of older African American and white adults, the Duke-EPESE (mean age 78 years at baseline). We computed KDM Biological Age, homeostatic dysregulation, and LM Biological Age for Duke-EPESE participants based on published algorithms and new analysis of data from the U.S. National Health and Nutrition Examination Surveys (NHANES). Across the three methods, more advanced biological age was associated with increased risk of disability and mortality. Effect sizes were in line with previous studies of younger populations.

Biological aging measures are receiving increased attention in gerontology as potential tools for risk stratification, etiological studies, and as surrogate end points for clinical trials of therapies to slow aging. Among proposed approaches to quantify biological aging, algorithms that combine information from blood chemistries and organ-system function tests are appealing because they can be implemented using already-existing data in many studies. Our results lend credence to the application of these clinical-data-based algorithms in studies of geriatric populations. Importantly, our findings replicate results from studies of younger and predominantly white samples in African Americans and white older adults in their 70s and 80s.

Trials of interventions to slow aging are being planned and executed in geriatric populations (45,46). Using existing data from the CALERIE Trial, we showed that moderate caloric restriction over 2 years slowed the rate of biological aging in healthy midlife adults (22). Results from the present study suggest that analogous evaluations can be applied to data collected from trials testing interventions in older adults.

Our results address outstanding questions about methods to quantify biological aging. One question is how sensitive are biological aging measures to the reference population used to develop the algorithm? We compared algorithms developed in a race/ethnicity diverse, mixed-age sample to algorithms developed using data on older adults and algorithms developed using race/ethnicity matched individuals. We found that the diverse, mixed-age sample produced an algorithm that was equally or more predictive of disability and mortality when compared with algorithms developed using samples demographically matched to the test sample. Databases representing the general population, such as the U.S. NHANES, can be used to develop biological aging algorithms for application in a range of population segments.

A second question is how sensitive are biological aging measures composed of multiple blood chemistries and organ function tests to the specific set of measures included in the algorithm? Many data sets will be missing one or another biomarker in a given algorithm. Importantly, our results suggest that algorithms composed of biomarker sets that differ on one or a few biomarkers are robust and yield similar results; measures of biological aging were highly correlated with one another and had similar-magnitude associations with disability and mortality.

A third question is how useful might biological age algorithms be in the clinical care of older adults? Biological aging measures derived from the type of data routinely collected during clinical encounters could offer health care providers a new tool for communication with patients about health risks and behaviors. Our results show promise for application of existing methods to quantify biological aging in older adults, but further testing is needed in clinical settings. Specifically, quantification of biological aging from electronic medical record or clinical chart review data is a critical next step.

Finally, our results suggest that blood-chemistry measures of biological aging may be less predictive of disability and mortality in African American when compared with white older adults. This difference in algorithm performance did not appear to reflect poor representation of African Americans within the NHANES samples used to develop the algorithms. For both KDM and homeostatic dysregulation, developing alternative algorithms within African American-only NHANES subsamples did not affect results. Instead, poorer algorithm performance in African American Duke-EPESE participants could reflect the selectivity of this sample. At the time they participated in the EPESE blood draw, African American Duke-EPESE participants had substantially outlived life expectancies for their birth cohorts. It is possible these participants carried certain robustness characteristics not well measured by the biological aging algorithms we studied. Replication of findings in more recently recruited cohorts of older adults is needed to establish whether blood-chemistry biological aging measures are less informative about disability and mortality in African American when compared with white older adults.

We acknowledge limitations. The Duke-EPESE sample was recruited in the 1980s in North Carolina. It does not represent the U.S. population as a whole. Moreover, given evidence suggesting change in the rate of biological aging across birth cohorts (20), replication is needed in more recently collected geriatric cohorts. However, it is a strength of our analysis that the long follow-up of Duke-EPESE participants allowed us to observe timing of death for nearly all of the study members. To date, Duke-EPESE has not conducted genomic analysis of biospecimens. Therefore, our analyses could not address epigenetic clocks (47) or telomere length, which are popular approaches to quantification of biological aging. The analysis sample size was moderate (N = 1,374). Thus, power to distinguish small differences in effect sizes between race/ethnic groups and different biological aging measures was limited. Finally, some biomarkers included in previously published biological age algorithms were not available in Duke-EPESE. Thus, comparison of biomarker sets from those algorithms is imperfect.

Within the context of these limitations, our results advance gerontology research by providing initial validation evidence for blood-chemistry-based measures of biological aging in a geriatric population. Results encourage future etiological and intervention studies in older adults to utilize clinical data to quantify biological aging.

Funding

This work was supported by Duke University’s Claude D. Pepper Older Americans Independence Center with support from the National Institutes of Health (grant P30AG028716); National Institutes of Health (grants R01AG054840 and R21AG054846). D.P. was supported by National Institutes of Health training grant T32AG000029. D.W.B. is partly supported by a fellowship from the Jacobs Foundation.

Conflict of Interest

None reported.

References