DNA Methylation Mediates the Association Between Cardiometabolic Risk Factors and Cognition: Findings From the Health and Retirement Study

Abstract

The association between cardiometabolic risk factors and cognitive function has been well documented, but the underlying mechanisms are not fully understood. This longitudinal study aimed to investigate the potential mediating role of DNA methylation in this association. We conducted the analyses in 3 708 participants (mean [standard deviation {SD}] age: 67.3 [9.5], women: 57.9%) from the Health and Retirement Study who were assessed in the 2014–2020 waves, had Infinium Methylation EPIC BeadChip methylation assays from the 2016 Venous Blood Study, and had cognitive assessment between 2016 and 2020. Causal mediation analyses were used to test the mediation role of DNA methylation in the associations between cardiometabolic risk factors and cognition, adjusting for demographic, socioeconomic, and lifestyle factors. Hypertension (−0.061 in composite cognitive z-score; 95% confidence interval [CI: −0.119, −0.004]) and diabetes (−0.134; 95% CI: [−0.198, −0.071]) were significantly associated with worse cognitive function while abnormal body weight and hypercholesterolemia were not. An increased number of cardiometabolic risk factors was associated with worse cognitive function (p = .002). DNA methylation significantly mediated the association of hypertension (mediated effect on composite cognitive z-score: −0.023; 95% CI: −0.033, −0.014), diabetes (−0.022; 95% CI: −0.032, −0.014), and obesity (−0.021; 95% CI: −0.033, −0.011) with cognitive function, whereas the mediation effect was not observed for having hypercholesterolemia. The estimated proportions mediated were 37.4% for hypertension and 16.7% for diabetes. DNA methylation may be an important mediator linking cardiometabolic risk factors to worse cognition and might even provide a potential target for dementia prevention.

Cardiometabolic risk factors have been identified as important risk factors for cognitive impairment and dementia (1–5). Specifically, hypertension and diabetes are associated with lower cognitive function, whereas results are less consistent for hypercholesterolemia and obesity (1–5). Both hypertension and diabetes are associated with vascular cognitive impairment and Alzheimer’s disease (6,7). Despite differences in the pathophysiologies underlying these associations, similar mechanisms such as neuroinflammation-related cognitive impairment have been proposed (6,7).

Epigenetic modifications, as mediators between various extracellular factors and cellular gene expression, may elucidate a potential mechanism for the association between cardiometabolic risk factors and cognition. DNA methylation is a major type of epigenetic modification in which the 5’-position of cytosine residues in DNA sequences are methylated or demethylated, resulting in changes in the corresponding gene expression (8). We hypothesize that cardiometabolic risk factors put stress on the arteries and target organs, leading to an accumulation of small insults over time. These recurrent injury-repair processes could alter the pattern and level of DNA methylation through mechanisms like chronic low-grade inflammation (9,10). Studies have identified significant associations between faster biological aging, as measured by DNA methylation changes, and worse cognition, though the associations vary across different epigenetic age measures and study populations (11–13). Furthermore, studies have found that DNA methylation could alter neuron plasticity and impact learning and memory formation in the adult brain (14,15). It is thus plausible that cardiometabolic risk factors contribute to worse cognition partly through DNA methylation, but this relationship remains undetermined. Such data would inform mechanistic links between cardiometabolic risk factors and cognitive impairment and might even provide a potential therapeutic target for individuals at a high risk of cognitive impairment due to cardiometabolic risk factors.

Therefore, in the current study, we examined the extent to which DNA methylation mediates the effect of cardiometabolic risk factors on cognition using data from a nationally representative longitudinal cohort study. We hypothesized that cardiometabolic risk factors are associated with worse cognition through the acceleration of epigenetic aging, as measured by the differential patterns of DNA methylation.

Method

Study Population

The Health and Retirement Study (HRS) is a nationally representative longitudinal study of adults aged 51 and older in the United States (16). The HRS oversampled African-American and Hispanic households to recruit and retain participants from racial and ethnic minority groups (17). As a result, the HRS cohort allows analyses within an older population that provide adequate representation of minority ethnic groups. The study began in 1992 and has since followed a cohort of over 20 000 participants, collecting data biennially through in-person interviews, phone surveys, and self-administered questionnaires (16). Further details of the HRS design have been described elsewhere (16).

In the present study, we included HRS participants who met all the following 3 criteria: (1) took part in the 2014 (our study baseline) to 2020 (Waves 12 through 15) core surveys; (2) had complete epigenetic clock data from the 2016 Venous Blood Study (VBS; n = 4018); and (3) completed all relevant cognitive assessments (word recall test, serial sevens subtraction test, and counting backwards test) at least once between 2016 and 2020 (18–20). Participants were excluded if they had missing baseline covariates or did not have at least 1 valid cognitive assessment. Of the remaining participants, we further excluded individuals classified as having dementia by the Langa–Weir classification of cognitive function at our study baseline (21). The ultimate analytic sample consisted of 3 708 participants (Supplementary Figure 1). Participants who were excluded due to missing cognitive assessment data from 2016 to 2020 had older chronological age, older DNA methylation age, lower education level, and lower physical activity level (Supplementary Table 1).

Epigenetic Age Assessment

Epigenetic age was used as a composite measure of the level and pattern of DNA methylation across multiple CpG sites. Thirteen epigenetic clocks were constructed using the DNA methylation data from the 2016 VBS (18). Full details of the construction of these epigenetic clocks have been previously published (18,22). We selected the GrimAge DNAm Age as the measure for epigenetic age in the present study because it has been shown to perform robustly in predicting age-related clinical phenotypes including frailty and worse cognitive function (23). Briefly, GrimAge DNAm Age was trained in a 2-stage process (24). In the first stage, each of the 88 plasma protein variables and smoking pack-years were regressed on chronological age, sex, and methylation levels at 485k CpGs to create DNAm-based surrogate measures for 12 plasma proteins and smoking pack-years using an elastic net regression model (24). In the second stage, time-to-death was regressed against these DNAm-based surrogate measures using elastic net Cox regression to select a linear combination of covariates, which include age, sex, and DNAm-based surrogate measures for smoking pack-years, adrenomedullin (ADM), beta-2-microglobulim (B2M), Cystatin C, GDF-15, Leptin, PAI-1, and tissue inhibitor metalloproteinases 1 (TIMP-1) (24). The final estimate of GrimAge DNAm Age was based on DNA methylation levels at 1 030 unique CpGs with the mean and variance of the model’s prediction being forced to match that of chronological age (24).

Cognitive Assessment

A battery of neurocognitive tests was administered at each HRS wave. An immediate and delayed 10-noun free recall test was administered to measure memory, with scores ranging from 0 to 20; a serial sevens subtraction test was administered to measure working memory with scores ranging from 0 to 5; a counting backward test was administered to measure processing speed with scores ranging from 0 to 2. Individual participants’ scores for each test from all available waves from 2016 to 2020 were averaged to obtain a more stable measure of cognition over the 4-year period (Supplementary Figure 2). z-Scores were obtained for each participant by taking the z transformation of their average scores on each test. Details about these tests have been described elsewhere (25).

Cardiometabolic Risk Factors Assessment

Cardiometabolic risk factors, the exposures of interest, include hypertension, diabetes, body mass index (BMI), and hypercholesterolemia. Hypertension (yes/no), diabetes (yes/no), and hypercholesterolemia (yes/no) statuses were obtained from self-reported interview surveys at baseline as part of the 2014 core HRS interview. BMI was calculated using weight reported at baseline (2 014) and height measured at cohort entry (Supplementary Figure 2). BMI values for participants were then calculated and categorized into 4 categories according to Centers for Disease Control and Prevention (CDC) guidelines for underweight (BMI <18.5 kg/m²), healthy weight (BMI within 18.5–24.9), overweight (BMI within 25.0–29.9), and obesity (BMI ≥30) (26). A count variable was also created by counting cardiometabolic risk factors for each participant. For BMI, obesity or underweight was counted as a cardiometabolic risk factor.

Covariates Assessment

Chronological age, sex, race (White, non-White), ethnicity (Hispanic, non-Hispanic), marital status (married, not currently married), education level (no degree, high school degree/some college/unknown degree, college graduate, advanced degree), activity level (<3 times or ≥3 times of vigorous activities/week), alcohol consumption (<1 or ≥1 drink/day), and smoking status (never smoker, past smoker, current smoker) were obtained from self-reported interview surveys at baseline as part of the 2014 core HRS interview. These variables were included as potential confounders in the analyses (Supplementary Figure 3).

Statistical Analysis

We evaluated the associations of individual cardiometabolic risk factors and the total count of cardiometabolic risk factors (ie, the exposures) with GrimAge DNAm Age (ie, the potential mediator) and composite cognitive score (ie, the outcome) using linear regression models, with adjustment for chronological age, sex, race, ethnicity, marital status, education level, activity level, alcohol intake, and current smoking status at our study baseline (2014). Exposure–mediator interactions were evaluated by additionally including the product term and assessing its statistical significance using the Wald test.

To evaluate how much of the association of cardiometabolic risk factors with cognition was statistically mediated by epigenetic age, we conducted mediation analyses using the CMAverse package in R following our hypothesized causal pathway (Supplementary Figure 3) (27) (R Foundation for Statistical Computing, Vienna, Austria). Specifically, effect estimates were estimated in a regression-based model. The total effect represents the difference in composite cognitive z-score associated with the presence (vs absence) of certain cardiometabolic risk factors. The total natural indirect effect (ie, the mediated effect of interest) represents the difference in composite cognitive z-score associated with the presence of certain cardiometabolic risk factors acting through the corresponding change in the mediator while the pure natural direct effect represents the effect through a pathway independent of the change in the mediator. The significance of exposure–mediator interaction was determined by the Wald test. If the interaction was significant, interaction terms were added to the analysis. Standard errors were estimated through bootstrapping with 1 000 bootstrap samples. Age and GrimAge DNAm Age were centered for all mediation analyses (28,29).

To explore whether the total, direct, and mediated effects of the association between the count of cardiometabolic risk factors and cognition followed a dose–response pattern, we repeated the mediation analysis with the count of cardiometabolic risk factors as the exposure. We visually inspected the residuals plot, QQ plot, and scale-location plot for the association between the exposure (count of cardiometabolic risk factors) and mediator, as well as between the exposure and outcome and verified the linearity, normality, and homoscedasticity assumptions for linear regressions.

In the sensitivity analysis, to reduce the risk of bias due to reverse causation, we further excluded participants classified as Cognitively Impaired Not Demented (CIND) at our study baseline (2014) by the Langa–Weir classification (21). To assess the impact of unmeasured confounding, we calculated the E-values for the mediated effects using the methods outlined by VanderWeele et al. (30). Briefly, mediation effects on the difference scale were first converted into risk ratios by an approximation formula, which were then used to calculate the E-values (30).

Results

The baseline (2014) characteristics of the 3 708 participants are summarized in Table 1. Their mean chronological age was 67.3 years (SD = 9.5), with a similar mean GrimAge DNAm Age of 68.0 years (SD = 8.6). The mean 4-year average composite cognitive score from 2016 to 2020 was 14.7 out of 27 pts (25^th–75th percentile: [12, 17.7]). The study population was mostly White (76.2%) and non-Hispanic (87.3%). The assessment of cardiometabolic risk factors, GrimAge DNAm Age, and cognitive function was conducted in a temporal order to reduce the potential of reverse causation and to allow for the use of a causal mediation analysis framework (Supplementary Figure 2).

Table 1.

Baseline Characteristics of the Study Population

	(N = 3 708)
Age, years	67.3 (9.5)
Female	2 146 (57.9%)
Race
White	2 825 (76.2%)
Black	607 (16.4%)
Other	276 (7.4%)
Hispanic	472 (12.7%)
Married	2 188 (59.0%)
Education level
No degree	534 (14.4%)
High school degree/some college/unknown degree	2 001 (54.0%)
College graduate	793 (21.4%)
Advanced degree	380 (10.2%)
Vigorous activity >3 times/week	1 035 (27.9%)
Daily drinker	223 (6.0%)
Smoking status
Never smoker	1 638 (44.2%)
Former smoker	1 613 (43.5%)
Current smoker	457 (12.3%)
Hypertension	2 289 (61.7%)
Diabetes	924 (24.9%)
Body mass index (BMI)
Healthy weight (<18.5)	1 030 (27.8%)
Underweight (18.5–24.9)	52 (1.4%)
Overweight (25–29.9)	1 335 (36.0%)
Obesity (≥30)	1 291 (34.8%)
High cholesterol	1 908 (51.5%)
Grimage DNA methylation Age*	68.0 (8.6)
Mean composite cognition score†	14.7 (4.0)

Notes: Values are means (SD) or number of participants (percentages). SD = standard deviation.

^*GrimAge DNAm Age was measured in 2016.

^†Mean composite cognitive score was the average score of all available assessments from 2016 to 2020; total possible score was 27.

Association Between Chronological Age and GrimAge DNAm Age

Figure 1 shows the linear association between chronological age and GrimAge DNAm Age with a high correlation (R² = 0.693). The slope of 0.922 indicates that, on average, GrimAge DNAm Age roughly mirrors chronological age in the study population. However, the difference between individual participant’s age and GrimAge showed considerable variation.

Figure 1.

Visualization of the correlation between chronological age and GrimAge DNAm Age. GrimAge DNAm Age is based on DNA methylation levels at 1 030 unique CpGs, which provide DNAm-based surrogate measures for seven plasma proteins and smoking pack-years (see text) (24).

Association of Cardiometabolic Risk Factors With GrimAge DNAm Age and Cognition

Table 2 presents the associations of cardiometabolic risk factors with GrimAge DNAm Age and cognition, adjusting for chronological age, demographic variables, and lifestyle variables. We observed consistently older GrimAge DNAm Age in participants with any of the cardiometabolic risk factors than in those without, though the differences were only statistically significant for hypertension (β = 0.963; 95% confidence interval [CI]: 0.711, 1.214), diabetes (β = 0.978; 95% CI: 0.698, 1.257), and obesity (β = 1.076; 95% CI: 0.759, 1.393). Hypertension (β = −0.061; 95% CI: −0.119, −0.004) and diabetes (β = −0.134; 95% CI: −0.198, −0.071) were associated with significantly lower composite cognitive z-scores while having other cardiometabolic risk factors were not associated with any significant differences in the composite cognitive z-score (Table 2). When the count of cardiometabolic risk factors was further assessed, a higher count of cardiometabolic risk factors was associated with older GrimAge DNAm Age (p-for-trend < .001) and worse composite cognitive z-scores (p-for-trend = .002), both in a dose–response manner. The assumptions of the dose–response test were assessed and met (Supplementary Figures 4 and 5). The observed associations were consistent with our expectations and provided the basis for causal mediation analyses.

Table 2.

Association of Cardiometabolic Risk Factors With GrimAge DNAm Age and Cognition

	Estimated Difference in GrimAge	Estimated Difference in Cognitive Function z-Score (total effect)
Hypertension (Yes vs No)	0.963 (0.711, 1.214)	−0.061 (−0.119, −0.004)
Diabetes (Yes vs No)	0.978 (0.698, 1.257)	−0.134 (−0.198, −0.071)
Hypercholesterolemia (Yes vs No)	0.177 (−0.063, 0.417)	−0.095 (−0.078, 0.031)
BMI category (vs healthy weight)
Underweight	0.816 (−0.212, 1.844)	−0.056 (−0.290, 0.178)
Healthy weight	0 (Ref)	0 (Ref)
Overweight	0.241 (−0.064, 0.546)	0.067 (−0.002, 0.137)
Obesity	1.076 (0.759, 1.393)	0.024 (−0.056, 0.088)
Count of cardiometabolic risk factors (vs 0)
0	0 (Ref)	0 (Ref)
1	0.379 (−0.011, 0.746)	0.002 (−0.082, 0.086)
2	0.841 (0.475, 1.207)	−0.057 (−0.141, 0.026)
3	1.343 (0.936, 1.749)	−0.070 (−0.162, 0.023)
4	1.946 (1.441, 2.450)	−0.164 (−0.279, −0.049)

Notes: Count was based on the sum of yeses for hypertension, diabetes, hypercholesterolemia, and obesity/underweight. Models were adjusted for age, sex, race (White, non-White), ethnicity (Hispanic, non-Hispanic), marital status (married, not currently married), education level (no degree, high school degree/some college/unknown degree, college graduate, advanced degree), activity level (<3 times or ≥3 times of vigorous activities/week), alcohol consumption (<1 or ≥1 drink/day), and smoking status as of 2014 (never smoker, past smoker, current smoker). BMI = body mass index.

The Mediation Effect of GrimAge DNAm Age

The mediation analyses tested our hypothesis that DNA methylation mediates the association between cardiometabolic risk factors and worse cognition. Table 3 presents the results from the mediation analyses using individual cardiometabolic risk factors as exposures. BMI as an exposure exhibited significant exposure–mediator interaction with GrimAge DNA methylation age while hypertension, diabetes, and hypercholesterolemia did not. The total effect, pure natural direct effect, and total natural indirect effect on composite cognitive z-score were negative (indicating worse cognitive function) for all cardiometabolic risk factors except for the BMI categories. The mediated effects were significant for hypertension (estimate = −0.023 in the composite cognitive z-score, p < .001), diabetes (−0.022, p < .001), and obesity (-0.021, p < .001), but not for hypercholesterolemia, underweight, or overweight. Being in nonhealthy weight categories had negative total natural indirect effects but statistically nonsignificant positive total effects and pure natural direct effects. The estimated proportions mediated were 37.4% for hypertension and 16.7% for diabetes.

Table 3.

The Mediation Effect of GrimAge DNAm Age on the Association Between Individual Cardiometabolic Risk Factors and Composite Cognitive Score

	Total Effect (on composite cognitive z-score)	Pure Natural Direct Effect	Total Natural Indirect Effect (mediated effect)	Proportion Mediated
Hypertension (Yes vs No)	−0.061 (−0.124, −0.003)	−0.038 (−0.103, 0.018)	−0.023 (−0.033, −0.014)*	37.4%
Diabetes (Yes vs No)	−0.134 (−0.196, −0.072)	−0.112 (−0.176, −0.050)	−0.022 (−0.032, −0.014)*	16.7%
BMI category (vs healthy weight)†
Underweight	0.044 (−0.225, 0.314)	0.086 (−0.198, 0.377)	−0.042 (−0.142, 0.014)	—
Healthy weight	0 (Ref)	0 (Ref)	0 (Ref)	—
Overweight	0.058 (−0.012, 0.127)	0.063 (−0.005, 0.133)	−0.006 (−0.013, 0.001)	—
Obesity	0.017 (−0.055, 0.090)	0.038 (−0.034, 0.110)	−0.021 (−0.033, −0.011)*	—
Hypercholesterolemia (Yes vs No)	−0.024 (−0.077, 0.031)	−0.019 (−0.074, 0.033)	−0.004 (−0.011, 0.001)	18.2%

Notes: Models were adjusted for age, sex, race (White, non-White), ethnicity (Hispanic, non-Hispanic), marital status (married, not currently married), education level (no degree, high school degree/some college/unknown degree, college graduate, advanced degree), activity level (<3 times or ≥3 times of vigorous activities/week), alcohol consumption (<1 or ≥1 drink/day), and smoking status as of 2014 (never smoker, past smoker, current smoker). Proportion mediated was not calculated if the total effect had an opposite sign from the mediated effect. BMI = body mass index.

^†Exposure–mediator interaction was considered in the mediation analysis since the interaction between BMI Category and GrimAge DNAm Age was significant as determined by the Wald Test (p = .015).

^* p < .001.

Table 4 shows that the association of cardiometabolic risk factors with worse cognition increased as the count of such factors increased, with an overall increasing trend in the magnitude of the total effect, pure natural direct effect, and total natural indirect effect (Figure 2). The count of cardiometabolic risk factors, as an exposure, also had a significant exposure–mediator interaction with GrimAge DNA methylation age. Specifically, for both the total effect and the pure natural direct effect on cognition, the magnitude of the associations was larger with more cardiometabolic risk factors, with the biggest difference seen when the count of cardiometabolic risk factors reached 4. In contrast, the total natural indirect effect (ie, the mediated effect through GrimAge DNAm Age) increased when the count of cardiometabolic risk factors increased from 0 to 2 with a plateau afterward.

Table 4.

The Mediation Effect of GrimAge DNAm Age on the Association Between the Count of Cardiometabolic Risk Factors and Composite Cognitive Score

	Total Effect	Pure Natural Direct Effect	Total Natural Indirect Effect (mediated effect)	Proportion Mediated
Count of cardiometabolic risk factors†
0	0 (Ref)	0 (Ref)	0 (Ref)	—
1	0.013 (−0.071, 0.103)	0.021 (−0.063, 0.112)	−0.008 (−0.019, 0.000)	—
2	−0.045 (−0.135, 0.042)	−0.022 (−0.114, 0.063)	−0.022 (−0.036, −0.011)*	50.3%
3	−0.060 (−0.161, 0.035)	−0.032 (−0.133, 0.060)	−0.027 (−0.044, −0.012)*	45.9%
4	−0.145 (−0.254, −0.026)	−0.125 (−0.237, −0.003)	−0.020 (−0.048, 0.004)	13.8%

Notes: Count was based on the sum of yeses for hypertension, diabetes, hypercholesterolemia, and obesity/underweight. Models were adjusted for age, sex, race (White, non-White), ethnicity (Hispanic, non-Hispanic), marital status (married, not currently married), education level (no degree, high school degree/some college/unknown degree, college graduate, advanced degree), activity level (<3 times or ≥3 times of vigorous activities/week), alcohol consumption (<1 or ≥1 drink/day), and smoking status as of 2014 (never smoker, past smoker, current smoker). Proportion mediated was not calculated if the total effect had an opposite sign as the mediated effect.

^†Exposure–mediator interaction was considered in the mediation analysis since the interaction between the count of cardiometabolic risk factors and GrimAge DNAm Age was significant as determined by the Wald test (p = 0.049).

^* p < .001.

Figure 2:

Changes in the total effect, pure natural direct effect, and total natural indirect effect on composite cognitive score by the count of cardiometabolic risk factors. Error bars represent the 95% confidence interval for each effect estimate.

Sensitivity Analysis

In our sensitivity analysis among those who were classified as cognitively normal at the baseline (ie, excluding those with CIND classification) in the mediation analyses, the results showed similar or more pronounced association estimates than those observed in the primary analysis (Supplementary Table 2). This analysis suggests the robustness of the findings to potential reverse causation. Supplementary Table 3 shows the estimated E-values for the mediation effects to evaluate the potential impact of unmeasured confounders on our results. The E-values indicate that, for cardiometabolic factors with statistically significant mediated effects, unmeasured confounders associated with changes in both epigenetic aging and cognition with approximate risk ratios of more than 1.16- to 1.17-fold each could completely explain away the observed mediated effects, controlling for the measured covariates.

Discussion

In this nationally representative cohort study among adults over age 50 years, we identified consistently positive associations between cardiometabolic risk factors and epigenetic age. Hypertension or diabetes was significantly associated with worse cognition. Accelerated epigenetic aging mediated the association between cardiometabolic risk factors and worse cognition, and the estimated proportion mediated was 37.4% for the hypertension–cognition association and 16.7% for the diabetes–cognition association.

Prior studies have focused on the association between cardiometabolic risk factors and epigenetic age as well as the association between cardiometabolic risk factors and cognition (1–4,7,31–35). Our study added new mechanistic insights into the potential mediating role of epigenetic age in the association between cardiometabolic risk factors and cognition among a nationally representative sample of older adults. Our results confirm the positive associations between cardiometabolic risk factors and epigenetic age reported in prior studies (33–35). Notably, 2 of these studies used longitudinal data with the study by Lundgren et al. using data from a twin cohort (33,34). Consistent with our results, several studies have also shown that hypertension and diabetes seem to have a stronger overall association with worse cognition than hypercholesterolemia and BMI (2–4). The weaker association of hypercholesterolemia with cognition may be attributable to the late-life measurement and lack of distinction between LDL- and HDL-cholesterol (5). We examined BMI as a categorical variable because we expected its association with cognition to be non-linear, which was consistent with our observations. In addition, we observed two potentially important trends. First, individual cardiometabolic risk factors that were more strongly associated with accelerated GrimAge DNAm Aging were also more strongly associated with worse cognition. Second, having more cardiometabolic risk factors simultaneously was associated with accelerated GrimAge DNAm Aging and worse cognition. These observations support the hypothesis that DNA methylation mediates the association of cardiometabolic risk factors with worse cognition. Another report using the HRS 2016 VBS dataset showed that faster DNA methylation aging is associated with worse cognitive outcomes in cross-sectional analyses stratified by sex (11). Additionally, when matched by DNA methylation aging rate, women outperformed men in verbal learning and memory cognitive domains (11). The association between the DNA methylation aging rate and cognitive outcomes while controlling for sex is consistent with our results. The sex difference in the cognitive outcomes when controlling for DNA methylation aging rate was an interesting observation not expanded upon in the current study.

In mediation analyses, we observed that the mediated effects for hypertension, diabetes, and obesity through DNA methylation were roughly the same despite different total effects on cognition, suggesting that the cumulative damage caused by these factors through DNA methylation may be similar. This observation also implies that the varied total effects observed for different cardiometabolic risk factors are explained to a larger extent by direct effects that are not mediated by DNA methylation. Mediation analyses also revealed interesting differences in how the total and mediated effects changed with the count of cardiometabolic risk factors. The total effect size increased modestly from 0 to 3 risk factors and substantially from 3 to 4. On the other hand, the mediated effect size increased gradually until the number of cardiometabolic risk factors reached 2, after which a plateau was observed. Further, the mediated effect for the count of risk factors plateaued at about the same level as the mediated effect for hypertension, diabetes, and obesity considered alone. A possible explanation for this observation is that the effect of DNA methylation on cognition plateaus after a certain level while the effect on cognition through other pathways rises with the presence of more cardiometabolic risk factors, which may represent the homeostatic dysregulation of the physiologic system (36).

There are several potential overarching explanations for the observations in the current study. The hypothesized pathway between cardiometabolic risk factors and cognition through DNA methylation likely accounted for a considerable proportion of the effects observed. Our hypothesis is supported by recent discoveries on the underlying biological mechanisms. An animal study by Wielscher et al. showed that altered CpG methylation is consequential of high blood C-reactive protein (CRP) levels, suggesting that chronic low-grade inflammation caused by cardiometabolic risk factors could result in changes in the DNA methylation patterns (9). In addition, the relationship between DNA methylation and neural plasticity has been reported by animal studies with a recent cohort study by Ho et al. showing that DNA methylation changes could also affect functional brain connectivity (14,37,38). However, we cannot rule out the possibility of reverse causation. Epigenetic changes may have caused the presence of cardiometabolic risk factors by altering relevant gene expressions whereas cognitive problems may also cause epigenetic modifications. Despite these concerns about reverse causation, the 2-year lead time between our baseline and measurement of DNA methylation should have mitigated this issue. Emerging evidence from recent studies has shown that changes in DNA methylation after certain exposures could take anywhere between hours to months depending on the type of exposure and the CpG sites affected (39,40). The 2-year period should thus allow enough time for the causal effects of cardiometabolic risk factors on DNA methylation to manifest. In addition, unmeasured confounding may still exist in the current study. For example, diet and sleep pattern differences may make an individual more susceptible to developing cardiometabolic risk factors, accelerated epigenetic aging, and worse cognition.

Overall, results from the current study indicate that our hypothesized mediation pathway is plausible and could be shared among different cardiometabolic risk factors. Understanding the pathways through which cardiometabolic risk factors affect cognition could help us better understand the underlying pathology linking cardiometabolic risk factors to cognitive decline and dementia. The current study has several clinical implications. First, the primary prevention of cardiometabolic risk factors, such as preventing smoking, promoting a balanced diet, and encouraging exercise, is important public health initiative for preventing cognitive impairment and dementia (41). Second, identifying DNA methylation as an important mediator in the effect of cardiometabolic risk factors on cognition also informs the potential of DNA methylation as an intervention target to mitigate the risk of cognitive decline and dementia. Indeed, it was reported in a randomized clinical trial that diet and lifestyle interventions reversed epigenetic aging (42). Current pharmacological agents like DNMT inhibitors and emerging therapeutic tools like the CRISPR/dCas system could also inform future therapeutic interventions targeting epigenetic aging (43). These interventions may benefit patients with cardiometabolic diseases at a high risk of cognitive impairment.

Strengths and Limitations

Our study has several limitations. First, although the exposure and mediator were measured in temporal order (exposure ascertainment at baseline in 2014 before mediator ascertainment in 2016), they were ascertained at a single time point for each participant without repeated measurements. So we were not able to confirm if the measured DNA methylation occurred before or after the development of cardiometabolic diseases. Therefore, a causal interpretation of the results must be made with caution. Second, cardiometabolic risk factors were assessed using self-reported questionnaires and the majority of the participants did not have continuous measures of the exposures (ie, blood pressure, blood glucose, and cholesterol), which could have introduced misclassification bias. Finally, although we adjusted for important demographic and lifestyle variables, there may still be unmeasured confounding by factors like neighborhood environment and dietary habits. Additionally, though GrimAge DNAm Age appeared to be an adequate measure of DNA methylation, future studies might be able to parse out the mechanism in more detail by looking at DNA methylation levels at specific CpG sites. It may also be of interest to use the presence of cardiometabolic risk factors at early- and mid-life stages to better quantify the cumulative effect over the life course.

Despite these limitations, to our knowledge, the present study is the first to investigate DNA methylation as a mediator in the association between cardiometabolic risk factors and cognition. Using a national representative sample from the HRS study, our findings are generalizable to adults in the United States over 50 years of age. Our assessment of cognitive function was also relatively robust as we constructed the cognitive function z-score based on 3 cognitive domains and used the mean of 3 longitudinal measurements across 4 years. The use of average cognitive score over 4 years allowed us to keep participants in the study base even if only 1 cognitive assessment was done in the 4-year period, which mitigated potential bias due to loss-to-follow-up and measurement errors. Furthermore, the use of the causal mediation framework with adjustments for important confounding structures and potential exposure–mediator interactions also rendered the findings more robust.

Conclusion

Changes in DNA methylation levels and patterns may partly explain the observed association between cardiometabolic risk factors and worse cognition. Future work elucidating how DNA methylation mediates the effect of cardiometabolic risk factors on cognition would enable us to better understand the mechanisms underlying cognitive impairment and dementia, better predict the heterogeneous patterns in late-life cognitive changes, and further inform the potential of DNA methylation as a therapeutic target for preventing cognitive decline and dementia.

Funding

The Health and Retirement Study is sponsored by the National Institute on Aging (NIA, grant number U01AG009740) and is conducted by the University of Michigan. This study was also supported by two grants from the NIA (R00AG071742 to Y.M. and P30AG062421 to D.B.).

Conflict of Interest

None.

Data Availability

The source variables used in this analysis are publicly available at the Health and Retirement Study website (https://hrs.isr.umich.edu).

Author Contributions

L.B.C., L.V., T.M.H., and D.B. contributed to methodology, result interpretation, and reviewing of manuscript drafts. Y.M. contributed to conceptualization, methodology, data analysis, result interpretation, and reviewing of manuscript drafts. Z.W. was the primary contributor to the manuscript’s aspects including conceptualization, study design, data analyses, result interpretation, and editing of the manuscript.