Biomarkers help us understand how cellular and systemic aging contribute to mortality: a study utilizing a machine-learning approach in the Health and Retirement Study

Abstract

Background

Research suggests aging is a coordinated physiological decline occurring in multiple systems and at multiple biological levels. However, it is largely unknown how general biological aging and specific systemic aging co-occur and influence one another to affect health outcomes. There is also emerging interest in understanding how social exposures may differentially accelerate decline in individual physiological systems.

Methods

We utilize data from the Health and Retirement Study, a nationally representative sample of about 4000 US adults over age 55. We used eXtreme Gradient Boosting (xgboost) in a training subsample to create system-specific mortality risk scores based on sets of biomarkers representing biological systems (eg, brain and nervous system, adaptive immune system, cardiovascular system, renal system) as well as general multisystem aging.

Results

Results suggest that the effects of most biological systems may be well captured by one or a small number of biomarkers and that female sex appears to be a protective or risk factor depending on specific biological system.

Conclusions

The importance of studying both general and system-specific aging is discussed.

According to the geroscience hypothesis, aging is a coordinated decline that occurs within individuals over time across multiple physiological systems (eg, neurological, cardiovascular, renal) and across different biological scales (eg, cell, tissue, organ).^1,2 This hypothesis suggests that aging itself is at the root of age-related health outcomes, including mortality. Thus, the most effective way to promote healthy longevity and to address sociodemographic differences in mortality may be to focus on slowing the aging process itself. Geroscience has been an active and fruitful field in recent years, as there has been a flurry of research on measuring,^3–8 understanding,^9–11 and intervening on^12,13 aging to promote health and longevity.

However, emerging research suggests that aging does not occur simultaneously across individual cells, tissues, and organ systems. Rather, cells in certain tissues may age faster, leading to uneven decline between organ systems.^14–17 This is a critical insight, as understanding how system-specific aging is associated with mortality could direct interventions to focus on slowing aging in biological systems that are of greater importance for health outcomes. Additionally, an individual may experience more accelerated aging in an individual system compared to their general multi-system aging. For example (as shown in Figure 1A) an individual may have a general baseline aging rate, but experience much more rapid cardiovascular aging—leading to greater risk of cardiovascular-related disease and death—and may experience slower renal aging—reducing their risk of kidney disease, etc. Targeting particular systems that have more accelerated aging may delay mortality related to that system; thus, understanding how general multi-system aging and specific systemic aging co-occur and influence one another to affect health outcomes may have important implications for personalized medicine. In this report, we use general multi-system aging to refer to coordinated decline across systems in an individual person and system-specific aging to refer to decline that occurs within an individual system, independent of this multi-level, multi-system aging process.

Figure 1.

(A) Conceptual model. (B) Analytic plan. (C) Aging-associated biomarkers for each of 12 biological systems and multi-system aging. ALT, alanine aminotransferase; AST, aspartate aminotransferase; BDNF, brain-derived neurotrophic factor; BMI, body mass index; BUN, blood urea nitrogen; CMV, cytomegalovirus; CO2, carbon dioxide; CRP, C-reactive protein; DHEAS, dehydroepiandrosterone sulfate; eGFR, estimated glomerular filtration rate; GDF15, growth/differentiation factor 15; GFAP, glial fibrillary acidic protein; HbA1C, hemoglobin A1c; HDL, high-density lipoprotein; IGF1, insulin-like growth factor 1; IL6, interleukin 6; IL10, interleukin 10; IL1RA, interleukin-1 receptor antagonist; LDL, low-density lipoprotein; mtDNA, mitochondrial deoxyribonucleic acid; NfL, neurofilament light chain; NK, natural killer; NT-proBNP, N-terminal prohormone of brain natriuretic peptide; SuPAR, soluble urokinase Plasminogen Activator Receptor; TGFb, transforming growth factor beta; TNFr1, tumor necrosis factor receptor 1.

We would expect aging rates of individual systems and multisystem aging to be strongly related to chronological age, but it is also important to understand how social and health risk exposures may differentially accelerate decline in individual physiological systems. Individuals are not uniformly exposed to health risks, leading to sociodemographic differences in health and longevity.¹⁸ There is evidence that these differential exposures to health risks lead to differences in rates of aging.^4,19–22 In particular, there is strong evidence for educational gradients in both health outcomes and in biological aging.^9,23 Education may affect healthy aging via multiple pathways, including exposure to stressors, differential access to health resources, and health risk behaviors.^24–28

Similarly, there is a long-standing interest in understanding sex differences in mortality. Women tend to live longer than men across time and across a wide variety of social and country contexts.^29–32 Though women generally live longer, they also tend to have greater rates of disability and certain morbidities (eg, arthritis, depressive symptoms, kidney disease).^33–35 Consistent with their longer life spans, emerging evidence suggests that women’s multi-system aging tends to be slower than men’s.^9,32 Examining system-specific aging may help explain why women are paradoxically at greater risk of disability and several morbidities, despite their longer life expectancy and slower multi-system aging.

Biological, social, and demographic data have been collected by the Health and Retirement Study (HRS), including 64 biomarkers, which include measures of general multi-system aging (eg, epigenetic clocks, telomere length) as well as measures of system-specific dysfunction that are potentially indicative of system-specific aging. These biomarker data were derived from blood samples which were assayed from numerous clinical chemistry measures, from DNA methylation, from flow cytometry, and from physical measurement. Specific measures were selected by HRS because they have been linked to age-related health outcomes in prior research. That is, HRS selected markers for inclusion in their study based on their association with age-related health outcomes.³⁶ Many of these measures (eg, eGFR) are strong clinical predictors of disease. We build on the geroscience perspective that these diseases are part of an underlying aging process that may be occurring within (ie, system-specific aging) and between biological systems (ie, general aging). Such data on a large nationally representative sample provide a unique and exciting opportunity to understand how general multi-system and system-specific aging are associated with sociodemographic factors and how they affect health and longevity. To accomplish this, we sorted biomarkers into 12 biological systems and multi-system aging based on prior literature (see online supplement). We utilized eXtreme Gradient Boosting (xgboost), a machine-learning (ML) technique, to estimate general multi-system and system-specific aging based on morality risk. We then assessed how these aging measures are associated with sociodemographic differences and health outcomes. Our analytic plan is outlined in Figure 1B.

In this study, we address the following questions: (1) What biomarkers within each system are of the greatest relative importance for mortality. To assess this, we estimated xgboost models with markers from each biological system and general aging predicting 6-year mortality. By comparing the relative importance of biomarkers within a given system, we can identify the biomarkers of greatest importance. (2) What biological systems are of the greatest importance for mortality? The models that are the outcome of question one can be used to produce algorithms for system-specific mortality risk. To address question 2, we used these algorithms to create system-specific risk scores for each individual. We then trained a new xgboost model predicting mortality using the system-specific risk scores, chronological age, sex, race/ethnicity, and years of education as features. We use the relative importance scores from this model to identify the most important biological systems compared to other systems and sociodemographic variables. (3) Finally, how is system-specific risk associated with sociodemographic factors. As described above, we expect system-specific and multisystem aging to be associated with chronological age. We expect that greater education and being female will be protective against aging. To assess this, we regress each system-specific risk score on age, sex, and years of education.

Methods

Data

As part of data collection in 2016, a subset of consenting HRS participants had blood drawn by a certified phlebotomist. A large number of multiomic assays were conducted, including epigenome-wide DNA methylation (DNAm), transcriptomics, and many common clinical assays in a sample of about 4000 older persons in this nationally representative sample. The HRS has produced epigenetic aging measures using these DNAm data. More detail on the collection, assays, quality control, and reliability associated with these biomarkers is available elsewhere.^36,37 These data are designed to be population-representative of community dwelling older US adults when weighted.

Measures

Sixty-four blood-based and other physiological biomarkers sorted into 12 biological systems and multi-system aging (ie, general aging that occurs throughout an organism as often described by the geroscience hypothesis) are utilized in this study. These markers include many common blood-based measures regularly assessed by physicians for screening and diagnostic purposes (eg, calcium, HbA1C, blood lipids), more innovative markers thought to be involved in aging and disease processes (eg, pTau181, NT-proBNP, CD4⁺ and CD8⁺ RA45⁺ effector memory cells), and physiological biomarkers of functioning (eg, peak expiratory flow, blood pressure). Because lung functioning is highly related to height and sex/gender, we separately regressed peak expiratory flow on height in men and women and standardized the residuals of these regressions to use as peak expiratory flow in all analyses. These markers were selected for inclusion in HRS because they are thought to be associated with age-related health outcomes and the process of aging. Information about collection, assays, and processing is available elsewhere.^36,37 To make relative importance values comparable across markers, all biomarkers were scaled to have a mean of 0 and variance of 1. Markers and their systems are shown in Figure 1C.

Mortality was assessed using individuals known to be dead to HRS over a period of about 6 years. As of this writing, information about mortality was last released in January 2024 and includes deaths up to May of 2021. We note that this is all-cause mortality.

Sociodemographic variables We focus on several sociodemographic variables, including chronological age, sex/gender (female coded as 1), and years of education. We control for race/ethnicity (non-Hispanic Black, Hispanic, non-Hispanic other race, and non-Hispanic White as the reference category) in models addressing questions 2 and 3 described above. We also control for wealth using the sum of all wealth components less all debt. We then added a constant to all observations of the smallest wealth value plus .001. We then natural log transformed wealth. If participants were missing wealth information in 2016, 2014 wealth was used.

Analytic plan

We sorted these biomarker data into 12 biological systems and multi-system aging. Biomarkers in their systems are shown in Figure 1C. For each system, we used eXtreme Gradient Boosting (xgboost) to estimate mortality risk associated with the biomarkers in each system (excluding endocrine which had only one marker). These xgboost algorithms thus represent one way of assessing system-specific aging. We focus on 3177 HRS participants with mortality information and who were in the innovative subsample with DNAm data. We randomly split the HRS data 50:50 into training and testing sets (Training Sample A (N = 1589) and Testing Sample A (N = 1588) in Figure 1B) and trained xgboost models in the training data. Hyperparameters, including maximum depth, learning rate (η), minimum child weight, and number of rounds were selected using a grid search procedure with 5-fold cross validation.

Xgboost models produce information about the relative importance of features based on the gain in predictive accuracy of branches a given feature is on. We multiply the importance values from each model by the total R² for that model to create a weighted gain value. To address question one—what biomarkers are of the greatest relative importance for all-cause mortality within each system—we compare relative importance within each system.

We next used the algorithms trained in the training data set to estimate general and system-specific risk measures based on mortality in Testing Sample A. We further split Testing Sample A 50:50 to produce Training Sample B (N = 789) and Testing Sample B (N = 788) (see Figure 1B). We then estimated an xgboost model predicting all-cause mortality using the system-specific and general aging measures, as well as age, sex/gender, years of education, race/ethnicity, and wealth (as statistical controls), as potential features. That is, xgboost algorithms used to answer question 1 were used to generate system-specific scores in Training Sample B that were then included together, along with demographic covariates, in a single xgboost model predicting all-cause mortality. For the endocrine system, only DHEAS was available, so that biomarker (multiplied by −1 to indicate more risk) was used to represent that system. To assess question 2—what biological systems are of the greatest importance for each health outcome—we compared relative importance measures across systems. We do not weight the relative importance for this model because we are not comparing across models.

Finally, to address question 3—how is system-specific risk associated with sociodemographic factors—we estimated system-specific risk scores using the algorithms developed in models addressing question 1 in Testing Sample A. We regressed these risk scores on age, sex/gender (female coded as 1), and years of education, controlling for race/ethnicity.

Analyses were conducted in R 4.4.0 “Puppy Cup”³⁸ using the tidyverse,³⁹ survey,⁴⁰ caret,⁴¹ and xgboost⁴² packages.

Results

Estimation of multi-system and system-specific aging measures

We began by splitting the HRS sample randomly into a training and testing sample (Training Sample A and Testing Sample A in Figure 1B). Descriptive statistics for these samples are shown in Table S1 (see online supplementary material). We estimated xgboost models predicting mortality for each system (excluding endocrine which had only one marker). Each system pairing thus has a set of hyperparameters and fit statistics. These values are shown in Table S2 (see online supplementary material).

Model fit—as measured by R² and area under the curve (AUC)—varied substantially across these models. The AUCs for the neurological, innate immune, and cardiovascular system models all surpass 0.75 (a traditional cut-off value⁴³), and renal and respiratory system models surpass 0.70.

Relative importance of biomarkers—Question 1

Relative importance for explaining all-cause mortality relative to other features in the same system in each model is shown in Figure 2A and in Table S3. Darker squares indicate more important variables relative to others in the same model. For the cardiovascular, innate immune, neurological, renal, and respiratory systems, one feature in each of these systems is the strongest predictor of mortality (NT-proBNP, GDF15, NfL, eGFR, and peak expiratory flow, respectively). Thus, for these systems, there appears to be a biomarker in that system that is most important in explaining all-cause mortality risk.

Figure 2.

(A) Relative importance for mortality of biomarkers from xgboost models for multi-system and system-specific aging measures weighted by R² of each model; note, DHEAS is the only endocrine marker, and is therefore not included in the figure. (B) Relative importance for mortality of multi-system and system-specific aging measures and controls from xgboost model. BDNF, brain-derived neurotrophic factor; BMI, body mass index; BUN, blood urea nitrogen; CO2, carbond dioxide; CRP, C-reactive protein; DHEAS, dehydroepiandrosterone sulfate; GDF15, growth/differentiation factor 15; GFAP, glial fibrillary acidic protein; HbA1C, hemoglobin A1c; HDL, high-density lipoprotein; HRS, Health and Retirement Study; IGF1, insulin-like growth factor 1; IL6, interleukin 6; IL10, interleukin 10; IL1RA, interleukin-1 receptor antagonist; LDL, low-density lipoprotein; mtDNA, mitochondrial deoxyribonucleic acid; NfL, neurofilament light chain; NK, natural killer; NT-proBNP, N-terminal prohormone of brain natriuretic peptide; SuPAR, soluble urokinase Plasminogen Activator Receptor; TGFb, transforming growth factor beta; TNFr1, tumor necrosis factor receptor 1.

Relative importance of systems—Question 2

We next used the algorithms trained in the Training Sample A to estimate general and system-specific risk measures based on mortality in Testing Sample A. We then split Testing Sample A into Training Sample B and Testing Sample B. We estimated xgboost models on Training Sample B predicting mortality using the risk measures developed in Training Sample A. We then evaluated these models in Testing Sample B. This model fit the data quite well (AUC = 0.823, R² = 0.229).

Relative importance for explaining all-cause mortality of each system-specific aging measure, general aging, chronological age, sex/gender, race/ethnicity, years of education, and wealth are shown in Figure 2B and in Table S4. Darker squares indicate more important variables relative to others. The most important factor for mortality was chronological age, followed closely by risk related to the neurological system. Innate immune and cardiovascular system-related risk were also important, with importances greater than 0.1. Multisystem aging, respiratory, and renal system-related risk all had relative importances greater than 0.05. That is, about 20% of the gain in explanatory power of this model to predict mortality comes from neurological system-related risk, about 18% from innate immune, 11% cardiovascular, and 9% general, multisystem aging-related risk.

Sociodemographic differences in multi-system and system-specific aging measures—Question 3

Finally, we regressed each system-specific risk score (based on algorithms trained in Training Sample A and estimated in the Testing Sample A, as described above) as well as the total risk score (calculated as part of Question 2) on age, sex/gender, and years of education, controlling for race/ethnicity. Results from these regressions are shown in Figure 3. Each box represents a single regression with the dependent variable listed on the right. Results for chronological age were very highly consistent, as chronological age was positively associated with all risk factors except lipid aging after false discovery rate (FDR) correction treating the 42 p-values shown in the figure as a family.⁴⁴ Educational attainment was relatively consistent. In 7 of the 14 models, educational attainment was protective against systemic/general risk after FDR correction.

Figure 3.

Estimates and 95% confidence intervals from regressions of each system-specific measure on sociodemographic factors. Circles represent significant associations (p < .05) after Benjamini–Hochberg false discovery rate (FDR) correction and Xs represent non-significant associations after FDR correction. Each box represents a separate regression of the dependent variable (shown in gray box on right) on age, sex/gender, and years of education. FDR, false discovery rate.

In contrast, the associations between sex/gender and risk were different for different systems. Sex/gender, or being female, was negatively associated with adaptive immune risk, lipid, and nutrient/electrolyte-based risk after FDR correction, and with general aging before FDR correction. Sex/gender was positively associated with endocrine risk after FDR correction. Notably, sex was not associated with total risk.

We conducted a number of additional sensitivity analyses, including additionally controlling for wealth, health behaviors, and baseline morbidities. We also tested whether increasing the number of folds for cross validation to 10 would affect findings. Results from these analyses were very similar to main results presented above and would not change conclusions drawn. We also estimated preliminary models focused on specific cause of death. We consider these results preliminary because of low statistical power, but they generally support the conclusions drawn from the main analysis (see online supplement).

Discussion

In this study, we utilized an ML method to examine differences in general and system-specific aging based on 6-year all-cause mortality. Our analyses suggest that within several biological systems, there is one or a small set of markers that well-characterize mortality risk. Thus, a small set of markers may be useful for understanding aging in many systems. Though some of these markers are routinely collected by physicians to monitor health status (eg, eGFR), many are atypical or are only collected in specific cases (eg, peak expiratory flow would likely only be collected in a patient with lung-related symptoms). Similarly, many large studies of aging in humans have collected epigenetic data (needed to estimate epigenetic aging), but few have collected GDF15 or NfL. Our results suggest that NT-proBNP, GDF15, NfL, eGFR, and peak expiratory flow are all highly useful for monitoring cardiovascular aging, innate immune aging, neurological aging, renal aging, and respiratory aging, respectively, at least at the population level.

We find that chronological age is still the most important factor for mortality when compared with system-specific risk measures. Though general, multisystem aging is still selected in this model, several other systems (especially neurological and innate immune) are selected as more important. Thus, our results suggest that addressing general aging (a principle goal of geroscience) would reduce mortality risk; however, identifying which specific systems are aging faster than others may be essential for identifying which systems will show dysfunction first, will show disease first, and may ultimately be a cause of mortality. Longitudinal data are needed to show that changing the rate of aging in a specific system affects morbidity and mortality associated with that system.

As expected, chronological age was generally a risk factor and educational attainment was generally protective against general, multisystem and system-specific aging. In our past research, we have used ML to show that indices of aging that include biomarkers from many biological systems do not explain sex differences in dementia and cognitive functioning.⁴⁵ Consistent with this, we do not find evidence that biomarkers combined as a whole (ie, the total risk row in Figure 3) explain sex differences in mortality risk. However, several individual system aging measures had opposite associations with sex/gender. Being female appeared to be protective against adaptive immune risk, lipid, and nutrient/electrolyte-based risk but was a risk factor for endocrine-related risk. Our results help clarify past research on sex and gender differences in health and aging. For example, women tend to live longer than men in the United States (thus have lower general aging-related risk), but experience menopause and childbearing (and thus may have higher endocrine-related risk).⁴⁶ These opposing system-specific effects appear to counteract each other when biomarkers are assessed in aggregate. Thus, analyses that focus solely on general aging risk missing how sex/gender interacts with system-specific aging to generate health differences.

There is an emerging body of research that is producing system-specific aging measures based on both clinical chemistry markers and DNAm.^16,17,47,48 These studies have been highly valuable for understanding how specific systems may age at different rates. This study builds on this past research by estimating system-specific and multi-system, general aging measures at one time using blood-based biomarkers, rather than omic surrogates, in a nationally representative, diverse, US sample. This allows us to more easily compare system-specific aging rates and directly compare these with general, multi-system aging described by the geroscience hypothesis. Unlike many past studies, we trained our aging measures on mortality, rather than chronological age.⁴⁷ Past research has shown that aging measures trained on health outcomes, rather than age, have much greater predictive power. We additionally build on this past research by examining relative importance of system-specific aging, contrasted with general, multi-system aging, for mortality and by examining sociodemographic differences in these aging scores.

This study is not without limitations. Models were only trained and tested in HRS. Lack of an external validation dataset is an important limitation. However, proper validation would require a dataset with the same set of predictors. We are unaware of an existing dataset with this same set of biomarkers. Future research using harmonized data with a partial set of markers (eg, the Longitudinal Study of Aging in India—Diagnostic Assessment of Dementia) would be of great value. This is a US sample of older adults. Populations from other countries and cultures would likely have different patterns of association between sociodemographics and aging measures, given differing patterns of mortality and morbidity across countries. Moving forward, research in non-western populations will be essential for advancing geroscience. Findings would likely be different in younger populations as well. We were interested in studying the aging process, and so utilized a sample of older adults; however, aging is increasingly recognized as beginning at conception. More aging research in middle aged and young samples could help us understand what markers of aging emerge earlier in life and are therefore useful for targeting interventions. We focus on mortality as a critical aging endpoint. However, morbidity, disability, and other health endpoints are critical for understanding healthy longevity. Future research should investigate system-specific aging and its association with these health and quality of life outcomes. We utilize all-cause mortality (though see the online supplement for analyses using non-accidental death) as a general indicator of the endpoint of aging processes. Future research should investigate cause-specific mortality to further refine system-specific aging processes. Blood data were only available at one time point. Longitudinal change in biomarkers is critical for establishing temporal order of processes and for understanding how changes in exposures or interventions might affect system-specific and general aging. Mortality data was only available for 6 years after blood collection. Longer follow-up may identify longer-term aging processes. Many large studies, like the HRS, are collecting, or have very recently collected, longitudinal biological data. These data will be of extraordinary value for aging and geroscience research.

Despite these limitations, this study advances the aging literature by highlighting the importance of aging in individual systems. We find support for the geroscience hypothesis, that there is a measurable, distinct process underlying aging throughout an individual that can be termed “aging.” However, individual systems depart in their speed of aging from this general aging rate. Understanding this system-specific aging is critical for understanding individual disease processes and may help disentangle the complex ways that sex/gender interacts with aging. Sex appears to have no effect on general aging; however, this masks risk and protective effects of female sex on system-specific aging. Future geroscience and aging research could be accelerated by integrating general aging with system-specific aging processes.

Supplementary material

Supplementary material is available at The Journals of Gerontology, Series A: Biological Sciences and Medical Sciences online.

Funding

Research reported in this study was supported by the National Institute on Aging of the National Institutes of Health under Award Numbers T32AG000037, R01AG060110 and P30AG017265. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The HRS (Health and Retirement Study) is sponsored by the National Institute on Aging (grant number NIA U01AG009740) and is conducted by the University of Michigan.

Conflicts of interest

None declared.

Data availability

The data supporting these findings are publicly available at https://hrs.isr.umich.edu/about or are in the process of being made available.

Author contributions

Eric T. Klopack (Conceptualization, Methodology, Formal analysis, Writing—original draft) and Eileen M. Crimmins (Conceptualization, Resources, Data curation, Writing—review & editing, Supervision, Project administration, Funding acquisition)