What this is
- This research evaluates 16 markers of biological aging in 1083 participants from the Berlin Aging Study II (BASE-II).
- The study analyzes both cross-sectional and longitudinal data to assess associations between these markers and various age-related health outcomes.
- Findings reveal that and are the most consistent markers, significantly improving predictions of health issues over a 7.4-year follow-up.
Essence
- and show strong associations with health outcomes in older adults. They improve prediction models for conditions like Metabolic Syndrome and frailty by up to 24 percentage points over 7.4 years.
Key takeaways
- and consistently correlate with various age-related health issues. They outperform other markers in predicting conditions such as Metabolic Syndrome and frailty.
- Markers of aging demonstrate varying strengths in predicting health outcomes, highlighting the need for tailored applications in clinical settings. This study provides insights into which markers may be most useful.
- The study emphasizes the importance of longitudinal analyses, as markers were more predictive of health issues at follow-up than at baseline, suggesting their potential for early risk stratification.
Caveats
- Participants were generally healthier than the average population, which may limit the generalizability of the findings to broader age groups. This could lead to an underestimation of effect sizes.
- The study's design does not establish causation, and future research is needed to explore the mechanisms linking these markers to health outcomes.
- Multiple testing corrections may obscure some true associations, as the conservative approach could lead to missed significant relationships.
Definitions
- Allostatic Load Index: A composite score reflecting the cumulative burden of chronic stress on the body, incorporating various health indicators.
- DunedinPACE: A marker of biological aging that estimates the rate of aging based on multiple biological indicators.
AI simplified
Introduction
Advances in healthcare, hygiene, lifestyle, and housing have led to an increase in the average lifespan in many countries over the past decades1. However, this expansion in lifespan was not matched by the increase in healthspan, the time spent before the onset of chronic disease or age-associated impairments2–4. At the same time, individual aging trajectories vary, and this heterogeneity increases with advancing age. While some individuals maintain good physical and mental health into old age, others show an early onset of chronic disease and impairment5. The geroscience hypothesis states that interventions targeting the biological process of aging may result in an increase of healthspan and prevent or at least postpone the onset of chronic disease6. Therefore, interventions that slow down or even reverse biological aging processes are of high individual but also societal interest7 but traditionally require long follow-up periods3,8. To test interventions in a cost- and time-effective manner, the validation of markers of aging that “either alone or in a composite predict biological age”3 is needed. Although extensive efforts have been made to identify and develop markers that can quantify inter-individual differences in biological age, no consensus has been reached5,9–11. This might be partly due to the differences in the conceptual meaning and the underlying framework of aging they aim to quantify and the methodological approach used. For example, aging clocks based on “-omics” data, like the proteomic clock or the first-generation epigenetic clocks, were trained to predict chronological age, and the difference between the predicted and the actual chronological age was shown to bear biological meaning12. In contrast, more recent versions of the epigenetic clocks are trained to predict age- and mortality-dependent biological changes or the so-called Pace of Aging13,14, and thus differ fundamentally in their underlying conceptualization. Other markers, such as BioAge15 and Allostatic Load16, are markers that aggregate laboratory measures associated with multiple organ systems into one composite biomarker of aging. Consistent with findings from our and other groups is that different indicators of aging only show a low to moderate association with each other5,9,17–25. This may result from the fact that the indicators were derived from different aging domains using varying methods. Whilst existing indicators of aging are individually well-evaluated, comparisons of different markers within the same study population and their relation to age-associated phenotypes are scarce17,23. Further, by focusing on selected health outcomes or high-risk populations, these studies limit our understanding of the broader predictive validity of aging markers in the general population.
To close this gap, we evaluate 16 markers of aging by investigating their relationship to a wide range of physical and mental age-associated health phenotypes in 1083 participants of the Berlin Aging Study II (BASE-II), based on both cross-sectional and longitudinal data. To assess their potential for clinical application, we additionally investigate the ability of each indicator to predict functional, physical and cognitive impairment after a mean follow-up period of 7.4 years. By doing so, we aim to provide a comprehensive overview of their individual strengths and weaknesses, in order to sharpen each biomarker’s individual profile, and contribute to a roadmap for their effective and targeted use in research and ultimately in clinical practice.
Our analyses reveal individual strengths and weaknesses of the investigated markers. DunedinPACE and Allostatic Load Index show the overall strongest and most consistent associations with the investigated outcomes. Both markers improve the discrimination of a basic clinical model developed to predict incident cases of Metabolic Syndrome, high cardiovascular risk (LS7), and frailty after a 7.4-year follow-up period by up to 24 percentage points.
Methods
Study population
BASE-II is an observational longitudinal study aiming at the identification of factors that predict and shape healthy aging trajectories26. Participants were recruited through the Max Planck Institute for Human Development’s participant pool in Berlin, as well as through advertisements in local newspapers and public transportation networks. As a convenience sample, baseline BASE-II participants were characterized by higher education and better self-reported health status than the general population of Berlin and Germany26. An above-average objective health status of BASE-II participants has been well documented, e.g., for diabetes mellitus type 227, cardiovascular health28 and frailty19. Participants between the ages of 60 and 80 years (older cohort) and 20 and 35 years (younger cohort, not analyzed in this study) were eligible for recruitment. Men and women were recruited in equal numbers. The baseline examination (T0) of 1671 older participants (medical part) was conducted between 2009 and 201426. A follow-up assessment (T1) as part of the GendAge study was conducted between 2018 and 202029 on average 7.4 years after baseline (SD: 1.5 years, range: 3.9 to 10.4 years). Only participants who had provided information on outcomes of interest at baseline as well as at follow-up were included in the final sample (n = 1083, Supplementary Fig. 1), to allow comparison of the aging markers in the same cohort of individuals. Of the 588 participants who dropped out between baseline and follow-up, 126 were confirmed to have died. Reasons for dropout among the remaining 462 participants were not systematically evaluated. Differences in the variables analyzed in this study between the main study sample and participants who dropped out are small, and we do not expect that loss to follow-up substantially altered our findings (Supplementary Data 1).
All participants gave written informed consent. All assessments at baseline and follow-up were conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Charité—Universitätsmedizin Berlin (approval numbers EA2/029/09, EA2/144/16, and EA2/224/21) and were registered in the German Clinical Trials Registry as DRKS00009277 (BASE-II) and DRKS00016157 (GendAge). The Ethics Committee of the Max Planck Institute for Human Development approved the procedure, and the Ethics Committee of the German Society for Psychology (DGPs) additionally approved of the MRI protocol. This manuscript was created in accordance with the STROBE guidelines30.
Variables
The Horvath31, Hannum32, PhenoAge33, GrimAge34, and DunedinPACE14 epigenetic clock algorithms were applied to derive DNA methylation age (DNAmAge) from methylation data measured with the “Infinium MethylationEPIC”, version 1, array (Illumina, Inc., USA). Methylation data for the 7-CpG epigenetic clock18 was obtained through Single Nucleotide Primer Extension (SNuPE)18,35. DNAmAge acceleration (DNAmAA) was calculated as the unstandardized residuals of a leukocyte cell-count adjusted linear regression analysis of DNAmAge on chronological age. Proteomics Age was derived from 248 proteins measured by liquid chromatography-mass spectrometry (LC-MS). Proteomics Age acceleration (ProteomicsAA) was calculated similarly to the DNAmAA as residuals from a linear regression of Proteomics Age on chronological age. Telomere length was assessed through quantitative real-time PCR (rLTL36) as well as estimated by the algorithm proposed by Lu and colleagues37 from methylation data (DNAmTL). The average results across three raters evaluating the number of lentigines from photos of the participants’ skin were used to derive SkinAge15. BioAge is a composite score that aggregates 13 routine laboratory parameters that were identified for their association with mortality15. The Allostatic Load Index (ALI) was computed by awarding a point to every participant within the high-risk quartile of selected variables using the approach described by Seeman and colleagues16. It additionally incorporates information about intake of relevant medication to include successfully treated and thereby masked dysregulation in its calculation38,39. Subjective Felt Age (SFA) was calculated as proportional discrepancy score from self-reported “felt age” and chronological age40. Subjective Life Expectancy (SLE) is calculated as the difference between the age participants expect to live to and their chronological age at the time of assessment. Subjective Health Expectancy (SHE) is the difference between the age participants expect to remain healthy and their chronological age at the time of assessment. Both SLE and SHE were adjusted for chronological age41. In addition to these markers, BrainAge was available in a subgroup of n = 255 BASE-II participants who underwent Magnetic Resonance Imaging (MRI) in a 3-Tesla Siemens Magnetom Trio scanner. BrainAge was calculated using a model trained on participants from the UK Biobank42. Except for Proteomics Age, all other indicators of aging were individually investigated in BASE-II before15,18,19,24,28,43–49.
Outcome measures investigated in this study include two measures of frailty. The Fried Frailty Phenotype (Fried FI) includes information on unintended weight loss, exhaustion, weakness, slow walking speed, and low physical activity50. The SPRINT-BASEed frailty index (SP FI) used in this study is an adapted version51 of the deficit-based measure developed by Pajewsky and colleagues52. Finger Floor Distance (FFD) was measured in centimeters. The Tinetti Mobility Test assesses mobility and balance through a series of simple tasks that are rated by the test supervisor53. Mini Mental State Examination (MMSE) is a well-established interviewer-administered instrument that was used to assess cognitive impairment54. Additionally, cognitive performance, specifically processing speed, was measured by the Digit Symbol Substitution Test (DSST), for which participants were asked to match symbols to numbers according to a given key55. Depressive symptoms were assessed using the Center for Epidemiologic Studies Depression Scale (CES-D)56. Independence during everyday life and the ability to perform tasks without help was measured using the Activities of Daily Living questionnaire (ADL, “Barthel Index”)57. The nutritional status was assessed using results from a short questionnaire and the measured circumference of the upper arm and the calf (Mini Nutritional Assessment, MNA58). Type 2 Diabetes (T2D) was diagnosed based on the criteria defined by the American Diabetes Association (ADA) guidelines59. Diabetes-associated complications were quantified using the composite score developed by Young and colleagues27,60. The Morbidity Index (MI) was calculated to assess the overall morbidity burden of the BASE-II participants by adapting the approach first described by Charlson and colleagues61. The Systematic Coronary Risk Evaluation assessment (SCORE262) and the SCORE2-OP (for participants >70 years, hereafter referred to as SCORE263) were calculated according to the recommendations in the respective publications to assess the risk for cardiovascular events. An adapted version28,64 of the Life’s Simple 7 (LS765) was calculated to quantify modifiable cardiovascular risk factors. Metabolic Syndrome (MetS) was diagnosed using the definition suggested by the American Heart Association/International Diabetes Federation/National Heart, Lung, and Blood Institute criteria 200966. The outcome variables described above were chosen because they represent specific health aspects that we analyze as downstream effects of the biological aging process quantified by the investigated markers of aging. However, in other contexts, some of the outcome variables themselves could also be investigated as aging markers and vice versa.
Confounding variables included in the regression models were chronological age (years), sex, alcohol consumption (g/d), and nicotine consumption (packyears), which were assessed during 1:1 interviews with trained study personnel. Body Mass Index (BMI) was calculated as kg/m2 using measurements from an electronic measuring station (seca 763, SECA, Germany). Genetic ancestry was quantified using the first four components from a principal component analysis on genome-wide single-nucleotide polymorphism genotyping data67. Details on the variables investigated in this study are described in the Supplementary Material.
Statistics and reproducibility
| Variables | Mean (SD) | Min | Max |
|---|---|---|---|
| Chronological age | 68.27 (3.49) | 60.16 | 84.63 |
| 7-CpG DNAmAA | −0.02 (6.94) | −22.93 | 22.33 |
| Horvath DNAmAA | 0.08 (4.22) | −11.64 | 14.45 |
| Hannum DNAmAA | −0.04 (3.43) | −10.9 | 19.15 |
| PhenoAge DNAmAA | 0.08 (4.65) | −16.88 | 14.28 |
| GrimAge DNAmAA | 0.07 (3.12) | −8.52 | 10.2 |
| DunedinPACE | 1.01 (0.11) | 0.58 | 1.42 |
| DNAmTL | 6.97 (0.18) | 6.26 | 7.44 |
| rLTL | 1.15 (0.23) | 0.17 | 1.92 |
| SkinAge | 1.71 (0.90) | 0 | 3 |
| ProteomicsAA | −0.02 (1.34) | −4.18 | 5.11 |
| BioAge | 0.17 (5.91) | −16.37 | 20.21 |
| Allostatic Load | 3.98 (2.40) | 0 | 12 |
| Subj. agea | 0.09 (0.08) | −0.13 | 0.41 |
| Subj. life expectancy | 15.89 (7.68) | 0 | 47 |
| Subj. health expectancy | 11.47 (7.06) | 0 | 46 |
| BrainAgeb | −0.06 (3.23) | −9.03 | 8.03 |
| T0 | T1 | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Variable | Mean,n | SD, % | Min | Max | Mean,n | SD, % | Min | Max | Mean Diff. | SMD | -valuep |
| Chronological age | 68.272 | 3.489 | 60.2 | 84.6 | 75.62 | 3.768 | 64.9 | 94.1 | 7.348 | 2.024 | <2.2e-16 |
| Fried frailty | 0.364 | 0.61 | 0 | 3 | 0.763 | 0.878 | 0 | 4 | 0.399 | 0.528 | <2.2e-16 |
| Frailty SPRINT-BASEed | 0.132 | 0.063 | 0.008 | 0.44 | 0.16 | 0.075 | 0 | 0.523 | 0.028 | 0.411 | <2.2e-16 |
| Finger floor distance | 9.746 | 11.839 | 0 | 52 | 10.576 | 12.673 | 0 | 159 | 0.829 | 0.068 | 0.005 |
| Tinetti score | 27.629 | 1.691 | 0 | 28 | 26.572 | 2.668 | 7 | 28 | −1.056 | −0.473 | <2.2e-16 |
| Falls (past 12 months) | 336 | 0.31 | 295 | 0.272 | 0.043 | ||||||
| MMSE | 28.59 | 1.335 | 22 | 30 | 28.578 | 1.524 | 18 | 30 | −0.012 | −0.008 | 0.827 |
| DSST | 45.001 | 8.301 | 5 | 74 | 40.801 | 8.674 | 2 | 73 | −4.2 | −0.495 | <2.2e-16 |
| CES-D | 7.285 | 6.613 | 0 | 36 | 7.17 | 6.165 | 0 | 36 | −0.115 | −0.018 | 0.526 |
| ADL | 99.003 | 3.72 | 5 | 100 | 98.698 | 4.086 | 15 | 100 | −0.305 | −0.078 | 0.046 |
| MNA | 27.383 | 1.756 | 19 | 30 | 26.595 | 2.158 | 16 | 30 | −0.788 | −0.401 | <2.2e-16 |
| DCSI | 0.777 | 1.163 | 0 | 7 | 1.264 | 1.486 | 0 | 8 | 0.488 | 0.365 | <2.2e-16 |
| Morbidity index | 0.975 | 1.208 | 0 | 7 | 1.432 | 1.542 | 0 | 9 | 0.457 | 0.33 | <2.2e-16 |
| SCORE2 | 10.809 | 4.222 | 4 | 32 | 16.073 | 6.16 | 5 | 47 | 5.264 | 0.997 | <2.2e-16 |
| LS7 | 8.609 | 1.978 | 2 | 14 | 8.44 | 1.86 | 3 | 14 | −0.17 | −0.088 | 0.001 |
| T2D (diagnosed) | 127 | 0.117 | 185 | 0.171 | 1.874e-9 | ||||||
| Metabolic syndrome (diagnosed) | 388 | 0.358 | 491 | 0.453 | 5.018e-11 | ||||||
Results
Participants
In this study, 1083 BASE-II participants who provided information at baseline and follow-up on average 7.4 years later were analyzed with respect to cross-sectional and longitudinal associations of 16 markers of aging (Supplementary Fig. 4) with a range of age-associated outcomes. Mean chronological age at baseline was 68.3 years (SD: 3.5 years), and 52% were women (Table 1, Supplementary Data 1). Frequency of impairment in the analyzed variables is shown in Supplementary Data 4. All age-associated outcome variables differed statistically significantly between examinations with the exception of FFD, Falls, MMSE, ADL, CES-D, and LS7 (Table 2). As described in BASE-II before24, correlations between markers of aging derived from different data sources and domains were moderate to low with r ≤ |0.31| while higher correlations were found between markers which were calculated based on the same data (e.g., epigenetic clocks, r ≤ 0.49) or which were closely related to each other (SHE and SLE, r = 0.73, Supplementary Fig. 2).
Cross-sectional association between markers of aging and age-associated outcomes at baseline
As a sensitivity analysis, we report on the results of logistic regression models of dichotomized continuously scaled variables on the markers of aging (Supplementary Data) and sex-stratified subgroup analyses (Supplementary Dataand). All analyses were repeated in the subsample of participants for which BrainAge was available (Supplementary Dataand) to allow direct comparison of the results of BrainAge with the other markers in the same individuals. 6 5 6 7 8
Standardized regression coefficients of unadjusted linear regression analyses of outcome variables on markers of aging in 1083 participants from the BASE-II at baseline. BrainAge was available in a subgroup of= 255 participants. *Bonferroni-corrected statistical significance (< 0.0001). LS7 Life’s Simple Seven, SCORE2 systematic coronary risk evaluation 2, MI Charlson’s morbidity index, DCSI diabetes complications severity index, MNA mini nutritional assessment, DSST digit symbol substitution test, Tinetti Tinetti test, FFD finger floor distance, SP FI SPRINT-BASEed frailty phenotype, Fried FI Fried’s frailty phenotype, MetS metabolic syndrome, T2DM type 2 diabetes mellitus. n p
Longitudinal associations between markers of aging at baseline and age-associated outcomes after 7.4 years of follow-up
Standardized regression coefficients of unadjusted longitudinal linear regression analyses of outcome variables at T1 on markers of aging at T0 in 1083 participants from the BASE-II. BrainAge was available in a subgroup of= 255 participants. *Bonferroni-corrected statistical significance (< 0.0001). LS7 Life’s Simple Seven, SCORE2 systematic coronary risk evaluation 2, MI Charlson’s morbidity index, DCSI diabetes complications severity index, MNA mini nutritional assessment, DSST digit symbol substitution test, Tinetti Tinetti test, FFD finger floor distance, SP FI SPRINT-BASEed frailty phenotype, Fried FI Fried’s frailty phenotype, MetS metabolic syndrome, T2DM type 2 diabetes mellitus. n p
Prediction of incident cases in outcome variables over a 7.4-year follow-up period
To promote translation into clinical practice and to evaluate the markers in a possible use-case scenario, we evaluated the markers’ ability to predict the incidence of impairment in the analyzed assessments, as well as incident cases of the investigated diseases and frailty over the average follow-up time of 7.4 years. Each marker-outcome-combination was evaluated in two scenarios. First, we simulated the case in which no other clinical information is available and assessed the markers raw predictive value. In a second step, we investigated the additional value that each marker adds to a basic prediction model.
Radar plots showing the results of unadjusted logistic regression analyses of incident cases of analyzed outcome variables at T1 on markers at T0. Participants with prevalent cases at T0 were excluded from this analysis. Due to the multiple imputation procedure, the sample sizes between each imputed dataset differ slightly. The range of sample sizes is indicated in Supplementary Data. Due to very strong effect sizes observed for Allostatic Load, the scale of this plot was adjusted to increase readability of the other plots. Therefore, the individual axis limits are displayed for each plot individually. Variable names of associations that were statistically significant after Bonferroni correction for multiple testing (< 0.0001) are displayed within a black box. LS7 Life’s Simple Seven, SCORE2 systematic coronary risk evaluation 2, MI Charlson’s morbidity index, DCSI diabetes complications severity index, MNA mini nutritional assessment, DSST digit symbol substitution test, Tinetti Tinetti test, FFD finger floor distance, SP FI SPRINT-BASEed frailty phenotype, Fried FI Fried’s frailty phenotype, MetS metabolic syndrome, T2DM type 2 diabetes mellitus. 18 p
![ROC curves illustrating the sensitivity and specificity of selected logistic regression models predicting impairment in age-associated phenotypes on average 7.4 years after marker assessment. The predictive performance of a basic clinical model including age and sex (blue) is shown compared to a prediction model that is extended by the respective markers of aging (red). The marker-outcome combinations presented here were selected to highlight the most compelling results from the three most promising markers (ALI, DunedinPACE, and GrimAge). A complete list of all AUC values can be found in Supplementary Data.-values for the difference between prediction models were calculated using the approach described by DeLongas part of the roc.test function (pROC package). After Bonferroni correction, the level of statistical significance was defined at< 0.0001. 18 P p [72]](https://europepmc.org/articles/PMC13031708/bin/43856_2025_1233_Fig4_HTML.jpg.jpg "Click to view full size")
ROC curves illustrating the sensitivity and specificity of selected logistic regression models predicting impairment in age-associated phenotypes on average 7.4 years after marker assessment. The predictive performance of a basic clinical model including age and sex (blue) is shown compared to a prediction model that is extended by the respective markers of aging (red). The marker-outcome combinations presented here were selected to highlight the most compelling results from the three most promising markers (ALI, DunedinPACE, and GrimAge). A complete list of all AUC values can be found in Supplementary Data.-values for the difference between prediction models were calculated using the approach described by DeLongas part of the roc.test function (pROC package). After Bonferroni correction, the level of statistical significance was defined at< 0.0001. 18 P p [72]
Discussion
In this study, we investigated 16 markers of aging in 1083 older participants from the Berlin Aging Study II (BASE-II). All markers were cross-sectionally and longitudinally analyzed in the context of a wide range of age-associated outcome variables representing different aspects of aging, including frailty, mobility, cognitive function, depressive symptoms, autonomy, nutrition, overall health, and chronic disease. Additionally, we investigated the markers’ ability to predict incident impairment in the age-associated outcomes over the average follow-up period of 7.4 years.
Our analyses showed that GrimAge DNAmAA and DunedinPACE performed best with respect to their association with cardiovascular health and cognitive capacity. Allostatic Load seems to represent frailty and overall morbidity-associated variables (MI, T2D and MetS). Longitudinally, the subjective psychological markers were associated with depressive symptoms (CES-D) and cognitive functioning (DSST, except SFA). Additionally, they were associated with the SPRINT-BASEed frailty index (cross-sectionally and longitudinally) and Fried’s frailty index (longitudinally). The strongest effect sizes in cross-sectional and longitudinal analyses were found for ALI with respect to diagnosed T2D, MetS, LS7, MI and SPRINT-BASEed frailty index. This presumably results from the way it is calculated, as many of the variables included in the ALI variable are closely related to health and are, in some instances, also part of the diagnostic criteria for T2D, MetS, LS7, and SCORE2. However, sensitivity analyses with modified ALI versions that excluded all variables that were also part of the outcomes in question (MetS, T2D, LS7, SCORE2) remained in almost all cases statistically significant (Supplementary Data–). This increases confidence that ALI is indeed a robust marker for biological age. 14 16
Subsequently, we investigated a potential use-case of the markers in a clinical context by predicting incident cases of impairment at follow-up by markers assessed on average 7.4 years earlier at baseline. ALI, GrimAge DNAmAA and DunedinPACE improved our basic prediction model in its prediction of incidence of impairment by up to 24 percentage points (Fig. 4). Although this increase in accuracy of the prediction model is expected to be less substantial when the markers are added to more comprehensive and specific prediction models, these results nevertheless illustrate that the markers hold biological information that can substantially improve the prediction of impairment years before it becomes clinically apparent. To estimate how the predictive value of the markers of aging would change in comparison to more comprehensive clinical prediction models, we calculated a sensitivity analysis that investigated the value added by the markers of aging to an extended clinical model (age, sex, BMI, smoking, and alcohol) (Supplementary Data 19). Although this did not lead to substantial changes in most marker’s interpretation, the added value of ALI with respect to the AUC of the prediction models for MetS, T2D, and LS7 was reduced by up to 17 percentage points compared to our original analysis. This suggests that future studies need to compare individual markers of aging with more comprehensive and specific clinical prediction models to fully understand how the markers of aging can contribute to incident case prediction in a clinical setting. Generally, we found more statistically significant results and stronger effect sizes in the longitudinal analyses compared to the cross-sectional results. This was observable especially for DunedinPACE, Allostatic Load and the psychological aging markers. Markers of aging are expected to depict the underlying aging processes, which are anticipated to happen prior to clinical manifestations. Therefore, it is expected that an accelerated aging process that did not (yet) exceed the individual mechanisms of resilience and therefore did not manifest through clinical phenotypes would be trackable through markers. The results presented here suggest that the respective markers appear to recognize an acceleration in the underlying aging process that was (at least at baseline) not advanced enough or was still compensated by the physiological and cognitive coping mechanisms to not result in clinically observable phenotypes. However, at follow-up examination, the acceleration in the biological aging process, which was already picked up at baseline by the respective markers, potentially resulted in clinical manifestations.
Another main finding of our study is the in part large differences in effect sizes of associations with the examined outcome variables, which suggests that the analyzed markers capture distinct aspects of the biological aging processes9,17,18,25. This interpretation of our findings is particularly plausible due to the different concepts of biological age that underlie the examined markers. For example, the first-generation epigenetic clocks aim at the prediction of chronological age and the residuals of a regression of chronological on epigenetic age are used to derive the analyzed marker (DNAmAA)18,31,32. Second- and third-generation epigenetic clocks, in contrast, are trained on more complex measures of biological aging calculated from several individual variables14,33,34, and DunedinPACE captures the rate of aging comparable to a speedometer14,70. Other markers, such as the psychological aging markers, rely on the subjective self-assessment of the participants. Composite markers, like Allostatic Load and BioAge, on the other hand, incorporate information from numerous systems to quantify aging, while other measures, such as telomere length and BrainAge, aim at the quantification of age-related biological changes. These differences result in unique strengths and weaknesses of each marker, which in turn define how they could potentially be used in the scientific and clinical context.
As sex-specific differences in the aging process are well known and to be in line with recommendations for the validation of markers of aging69, sex-stratified analyses were presented in the Supplementary Material of this manuscript. In some cases, statistically significant findings that were observable for analyses including the whole study sample did not reach statistical significance in the sex-stratified subgroup analyses. This most likely results from the smaller sample sizes of the respective subgroups. On the other hand, in some cases, it can be seen that the association observed in the whole sample is driven mostly by one of the sex-stratified subgroups. For example, the cross-sectional association between ALI and DCSI (logistic regression, Supplementary Data 6) observed for the whole study sample (OR = 1.4, p = 0.00001) seems to be mostly driven by men (OR = 1.5, p = 0.00005) compared to the subgroup of women (OR = 1.3, p = 0.02) in which the association does not reach the Bonferroni adjusted significance level (Supplementary Data 6). Due to the large number of markers and outcomes, as with the main analyses, we refrained from a detailed discussion of all findings from these subgroup analyses.
Our results are indicative that these markers are indeed capable of depicting underlying aging processes and do so with a higher sensitivity than clinical aging measures. Thereby, they could potentially be used to identify participants who are especially prone to future impairment due to an acceleration in one of the biological aging processes long before they show this decline clinically. Furthermore, we want to note that in this study, the markers of aging are analyzed separately with respect to their ability to predict incident cases in the outcome variables. This neglects the possibility of complementary and synergistic effects when combining markers from different domains. In addition, while the results indicate the promising potential of these markers as screening parameters, the study design does not allow to conclude on any clinical recommendations. Further studies that examine the added value of including these markers in screening processes, ideally in a randomized controlled design, with a meaningful endpoint, are needed. While screening programs are often evaluated with mortality as endpoint, in the aging context, it might be important to analyze other variables that are specific to the outcome of interest and are associated with quality of life and overall health. For example, a reduction of diabetes-associated complications might provide substantial benefit for individuals even if the overall lifespan could not be extended. Therefore, future studies with a focus on adverse outcome prevention are needed to further explore the possibility of utilizing these markers of aging as possible screening markers in a clinical context.
In a previous study by Kuo and colleagues, change in epigenetic age over time was associated with mortality71. As a sensitivity analysis, we calculated prediction models investigating the change in the two most promising markers of aging, ALI and DunedinPACE, between T0 and T1 (Supplementary Data 20). Interestingly, DunedinPACE at T0 showed better results compared to the longitudinal change in DunedinPACE, which only remained nominally statistically significant in the unadjusted logistic regression model predicting MI. On the other hand, the longitudinal difference in ALI showed stronger effect sizes when predicting falls, T2D, and MetS, but generally less strong or overall, not statistically significant associations for all other outcome variables. One limitation of our study in this context is that the second timepoint for assessing the longitudinal difference was also used to assess incident cases. Future studies with a third timepoint for incident cases assessment might be needed to evaluate the potential of the aging rate to predict outcomes other than mortality.
We want to point out several limitations to this study. First, the participants of this study are above-average health26. Consequently, a comparatively low prevalence of impairment in the functional and cognitive assessments was observed, which could lead to an underestimation of the true effect sizes, and the generalizability of our results to the underlying source population might be limited. Second, the covariates used in the fully adjusted regression models were chosen because of their frequent use in other studies in the field, as well as their known or suspected association with the independent and dependent variables in our regression models. An individual covariate selection for each marker-outcome-combination was outside the scope of this investigation and would also have compromised comparability of effect sizes between markers within this study. Future studies with a stronger focus on causal associations and the mechanistic relationship between markers and outcomes are needed to deepen our understanding. Third, as with any study that reports on a large number of individual statistical tests, multiple testing is an issue. Here, we adjusted the p-values using the Bonferroni correction. While this correction is robust, it also comes with the risk of missing true effects due to its conservative approach. Therefore, it is possible that true associations did not reach statistical significance in our study. Finally, the comparatively small age range in BASE-II might be a reason for the weak correlation between some markers and chronological age, and limits our conclusions to this age group. Future longitudinal studies with a wider age range and repeated marker measurements are needed to further improve our understanding of longitudinal associations between the markers investigated here and clinical aging phenotypes.
Strengths of these analyses include the comparative analysis of a large number of markers derived from a wide variety of aging domains, including epigenetics, proteomics, telomeres, composite markers, psychological markers, SkinAge, and BrainAge. The availability of these variables, in addition to numerous age-associated outcomes in a large longitudinal sample, allowed a comprehensive comparison of these markers cross-sectionally as well as over time. We provide information on the distinct abilities of these markers, which can inform future studies and sharpen their profiles. As the measurement of numerous markers is costly, these analyses allow a targeted selection of markers that can be used in future studies based on their strength of association with the specific aging domain of interest.
Conclusion
Our comparative analyses of 16 markers of aging in the context of a wide range of age-associated outcome variables highlight the distinction between the investigated domains of aging. Interestingly, markers were more frequently and also more strongly associated with outcomes on average 7.4 years after their assessment compared to cross-sectional analyses, suggesting their sensitivity to biological aging processes that became clinically apparent only much later during the follow-up examination. This finding underscores the potential of these markers for early risk stratification. Namely, ALI and DunedinPACE substantially contributed to the prediction of incident impairment at follow-up examination.
Supplementary information
Supplementary Material Description of Additional Supplementary Files Supplementary Data 1-20 Supplementary Data 21