Comparing Biological Age Estimates Using Domain-Specific Measures From the Canadian Longitudinal Study on Aging

Abstract

Many studies have shown that estimates of biological age (BA) can predict health-related outcomes in older adults. Often, researchers employ multiple measures belonging to a variety of biological/physiological systems, and assess the validity of BA estimates by how well they approximate chronological age (CA). However, it is not clear whether this is the best approach for judging a BA estimate, or whether certain groups of measures are more informative to this end. Using data from the Canadian Longitudinal Study on Aging, we composed panels of biological measures based on the physiological systems/domains they belong to (blood, organ function, physical/cognitive performance), and also composed a panel of measures that optimized the association of BA with CA. We then compared BA estimates for each according to their association with CA and health-related outcomes, including frailty, multimorbidity, chronic condition domains, disability, and health care utilization. Although BA estimated using all 40 measures (r = 0.74) or our age-optimized panel (r = 0.77) most closely approximated CA, the strength of associations to health-related outcomes was comparable or weaker than that of our panel composed only of physical performance measures (CA r = 0.59). All BA estimates were significantly associated to the outcomes considered, with exception to the neurological and musculoskeletal disease domains, and only varied slightly by sex. In summary, while the approximation of CA is important to consider when estimating BA, the strength of associations to prospective outcomes may be of greater importance. Hence, the context in which BA is estimated should be influenced by an investigator’s specific research goals.

Over the life course, from the perinatal period to old age, individuals are exposed to stressors that can trigger allostatic responses, resulting in damage to system integrity. This damage, also referred to as “wear and tear,” or allostatic load (1), contributes to the breakdown of multiple physiological processes, which leads to the development of adverse health conditions and eventually, mortality. Measurements of age-related wear and tear based on biological and physiological measures have been in use since the 1960s (2); more recently, methods that calibrate measures to approximate chronological age (CA) have been developed. These new measures take on values that represent the age at which an individual’s biological and physiological state would be approximately normal in the population, and are commonly known as “biological age” (BA).

A great deal of evidence exists supporting the utility of BA estimation for the prediction of important health outcomes and the classification of one’s health state. For example, BA estimated based on a mixture of molecular (eg, triglyceride levels in blood) and physiological (eg, blood pressure) measures was recently shown to predict 17-year survival in a Korean cohort of nearly 560,000 adults (3). Similarly, in both European (4) and American (5) cohorts, BA was significantly associated with the hazard of all-cause mortality, as well as cardiovascular- and cancer-specific mortality (5), and the incidence of age-related diseases such as stroke, cancer, and diabetes (4). Biological aging has also been shown to be associated with cognitive decline, both in younger (6) and older adults (7), and is related to prior life events such as adverse childhood experiences (8). Other types of BA estimates or “clocks” have been developed using DNA methylation data and are related to health outcomes to varying degrees (9–11).

While there is much evidence on the relationship between BA and adverse outcomes as well as plenty of literature comparing mathematical approaches to estimating BA, the majority of studies published to date employ relatively small panels of measures and fail to report which measures make the greatest contributions to BA. This is important to understand, as some researchers may have to be selective in the types of measures to obtain from their cohort, and certain measures may be more desirable to include than others; for example, those that are highly standardized, minimally invasive, and/or quick to obtain, such as complete blood counts or other measures derived from clinical instrumentation. Furthermore, measures that optimize BA (ie, minimize the error between BA and CA) may not necessarily improve the prediction of health-related outcomes, as previously shown (11).

In the following study, we sought to compare BA estimates derived from different panels of measures with regards to their correlation to CA and association to health-related outcomes. We defined panels based on measurement type or those measures found to optimize the association of BA with CA, and compared them to a complete panel including all measures considered. Our study not only provides important commentary regarding the contribution of different types of measures to the precision of BA estimated using the Klemera–Doubal approach and the association of BA to health-related variables, it is also a resource for researchers that wish to estimate BA using data from the Canadian Longitudinal Study on Aging (CLSA).

Method

Sample

We analyzed data from the CLSA baseline collection (2012–2015). The CLSA is a 20-year longitudinal study including over 50,000 community-dwelling participants between the ages of 45 and 85 at baseline (12). Canadian Longitudinal Study on Aging excludes individuals residing in nursing homes, on First Nations reserves, or in the Canadian territories, full-time members of the armed forces, those unable to respond in English or French, and those with significant cognitive impairment precluding participation without a proxy respondent (13). Our study was based on the comprehensive cohort (Comprehensive Baseline Dataset version 4.0), which included 30,097 community-dwelling adults aged 45–85 (average age: 63 ± 10.3 years; 15,320 women, 14,777 men) who provided data through interviews in their homes and at 1 of 11 data collection sites nationwide; the biological measures employed in this study were only obtained at a data collection site.

Measures of Biological Aging

Biological age estimation using the Klemera–Doubal method

Biological age was estimated from algorithms developed according to the method proposed by Klemera and Doubal (14), which has been shown to perform well as compared to other published approaches (15–17). Specifically, we employed equation 25 ($B_E$), as referenced in the original publication:

$${\displaystyle \begin{array}{l}{\displaystyle \frac{{\displaystyle \sum _{j=1}^m}(x_j-q_j)\frac{k_j}{s_j^2}}{{\displaystyle \sum _{j=1}^m}{\left(\frac{k_j}{s_j}\right)}^2}}\end{array}}$$

Here, for a panel of $m$ measures, CA was first regressed against each of the $j$ measures, and the intercept ($q_j$), slope ($k_j$), and the mean square error ($s_j^2$) were estimated. Biological age was then estimated for each participant by substituting in their $x_j$ measure value. The regression parameters $q_j$, $k_j$, and $s_j^2$ were estimated separately for males and females in a random training sample representing 80% of participants (Supplementary Table 1); accordingly, BA estimates were also calculated in a sex-specific manner in the remaining 20% of participants (test sample). The biological age difference (ΔBA) represents the residual from BA regressed on CA.

Selection of biomarkers used to compose BA estimates

A total of 52 measures were initially considered for analysis (Supplementary Table 2). These measures were drawn from four categories: blood test, physiological, physical performance, and cognitive performance data.

Blood test data included complete blood counts and clinical chemistry. Blood counts were obtained using a Coulter Ac·T diff Analyzer (Beckman Coulter) and included measures of white blood cells, monocytes, lymphocytes, granulocytes, hemoglobin, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin and concentration, red blood cell distribution width, platelets, and mean platelet volume. Clinical chemistry was performed using a Cobas 8000 modular analyzer (Roche Diagnostics) and included measures of 25-hydroxyvitamin D, albumin, alanine aminotransferase, creatinine, ferritin, free thyroxine, high-sensitivity C-reactive protein, cholesterol, high-density lipoprotein, triglycerides, and thyroid-stimulating hormone.

Physiological data were derived from sphygmomanometry (BpTRU Blood Pressure Monitor) to measure systolic and diastolic blood pressure, and pulse rate; spirometry (TruFlow Easy-On Spirometer) to measure forced expiratory volume in one second (FEV1), forced vital capacity (FVC), and the FEV1/FVC ratio; body composition analysis (Hologic DXA) to measure total and appendage lean mass; and anthropometry (by measuring tape and stadiometer) to measure waist circumference, waist to hip ratio, and body mass index.

Physical performance data included tests of gait speed, standing balance, timed get up and go, chair raise, and grip strength (Tracker Freedom Wireless Grip dynamometer).

Cognitive performance data included tests of memory (immediate and delayed Rey Auditory Verbal Learning Test, time and event-based Prospective Memory Test), verbal fluency (Animal Fluency Test, Controlled Oral Word Association Test), executive function (Victoria Stroop Test (18,19), Mental Alternation Test (20), and Choice Reaction Time Test (motor speed). All measures were analyzed as T-scores adjusted for language of administration (French or English) (21).

We transformed values of measures to approximate normal distributions. We applied log transformations to blood and physiological measures, with the exception of the FEV1/FVC ratio, which was cubed. Physical performance measures were log-transformed, with the exception of grip strength. Of the cognitive performance measures, choice reaction time was inverse-transformed to minimize skewness (22) and the Stroop test was log-transformed as in previous work (19). Spirometry measures represent the average of the three highest-ranked trials.

Many of the measures we analyzed were correlated with one another. To select a subset with maximum information, we computed variance inflation factors (VIFs) for each pair of measures and removed one measure from each pair with a VIF > 10. The measure considered to be less informative biologically (eg, derived measures) from each collinear pair was dropped; for pairs of measures deemed equally informative, the measure with the stronger linear relationship with CA was selected. The VIFs and rationale for removing biological measures are described in Supplementary Table 3. The final measure count in each panel are as follows: full panel (FULL), 40 measures; blood test panel (BLD), 19 measures; physiological panel (PHYS), 7 measures; performance panel (PERF), 5 measures; cognitive panel (COG), 9 measures. The five panels (FULL, BLD, PHYS, PERF, and COG) and their constituent measures are listed in Table 1.

Table 1.

The Panel of 40 Measures Employed to Estimate Biological Age

Panel	Name	Acronym
Blood (BLD)	Lymphocytes (absolute number)	Lyc
	Monocytes (absolute number)	MOc
	Granulocytes (absolute number)	GRc
	Hemoglobin	HGB
	Mean corpuscular volume	MCV*
	Red blood cell distribution width	RDW
	Platelets	PLT*
	Mean platelet volume	MPV
	25-Hydroxyvitamin D	VITD*
	Albumin	ALB
	Alanine aminotransferase	ALT*
	Creatinine	CREAT*
	Ferritin	FERR*
	Free thyroxine	T4
	High-sensitivity C-reactive protein	CRP
	Cholesterol	CHOL
	High-density lipoprotein	HDL
	Triglycerides	TRIG
	Thyroid-stimulating hormone	TSH
Physiological (PHYS)	Systolic blood pressure	SYSTOLIC*
	Diastolic blood pressure	DIASTOLIC*
	Pulse	PULSE
	Ratio of forced expiratory volume after 1 s to forced vital capacity	RAT*
	Forced vital capacity	FVC*
	Appendage lean mass	ApLean*
	Waist to hip ratio	WHR
Performance (PERF)	4-m walk test	WALK
	Timed get up and go test	TUG
	Chair rise test	CR
	Grip strength	GS
	Single leg balance test	BAL*
Cognitive (COG)	Animal Fluency Test	AFT
	Rey Auditory Verbal Learning Test (immediate recall)	REYI
	Rey Auditory Verbal Learning Test (delayed recall)	REYII*
	Mental Alternation Test	MAT
	Event-based Prospective Memory Test	PMT*
	Time-based Prospective Memory Test	TMT
	Victoria Stroop Test	STRP*
	Choice Reaction Time Test	CRT*
	Controlled Oral Word Association Test	FAS

Note: The measure panel FULL includes all measures listed above. Measures with an asterisk were included in the age-optimized panel (OPT).

To identify an age-optimized set of measures (OPT), we applied an iterative approach based on the minimization of the mean squared error (MSE; the average of squared residuals). In this process, we estimated BA for each individual measure and estimates were generated separately in males and females. For each estimate, we computed the MSE of CA regressed on BA, then, beginning with the measure with the smallest MSE, we calculated a new set of BA estimates based on that measure plus each of the other measures in turn. Again, the pair with the smallest MSE was retained and the procedure was repeated. We repeated this procedure until all measures were retained in the estimate of BA. To clarify this approach, a detailed example is provided in the Supplementary Methods, along with a flowchart describing each step. The final optimized set included those measures for which the MSE was reduced by at least 0.5% as compared to the previous iteration in both sexes (Figure 1); hence, OPT included 16 measures: FVC, BAL, STRP, ApLean, FERR, RAT, VITD, SYSTOLIC, DIASTOLIC, PMT, CREAT, CRT, REYII, ALT, MCV, and PLT (Table 1; Figure 1).

Figure 1.

Iterative procedure to identify those measures that minimize the mean squared error (MSE) of chronological age regressed on biological age. The measures included in the final age-optimized panel (OPT) are those that reduced the MSE at least 0.5% relative to the previous iteration.

Thus, our final set of biological aging variables included six BA estimates: an estimate based on the full panel of 40 measures (BA_FULL), estimates based on each of the four domain-specific panels (BA_BLD, BA_PHYS, BA_PERF, BA_COG), and an estimate based on biomarkers selected to optimize the correlation between BA and CA (BA_OPT). Each estimate was generated using the Klemera–Doubal approach.

Health-Related Variables and Outcomes

To compare the performance of BA estimates derived from different panels of measures, we tested associations with aging-related health variables and outcomes: a frailty index, a count of chronic health conditions (multimorbidity), a count of chronic condition domains, disability, emergency department (ED) visits, and overnight hospitalization (Table 2).

Table 2.

Summary Statistics of Health-Related Variables and Outcomes for Participants Included in the Test Sample (n = 5,314)

		Male	Female
Frailty	Mean ± SD	0.11 ± 0.06	0.13 ± 0.07
	Min/max	0/0.45	0.003/0.54
	Missing	8 (0.3%)	6 (0.2%)
Multimorbidity	Median [25th–75th]	2 [1–4]	3 [1–5]
	Min/max	2/12	3/19
	Missing	3 (0.1%)	0 (0%)
Mental health	No	2,204 (83%)	1,998 (75%)
	Yes	438 (16%)	648 (24%)
	Missing	13 (0%)	13 (0%)
Cardiometabolic	No	1,733 (65%)	1,925 (72%)
	Yes	862 (32%)	684 (26%)
	Missing	60 (2%)	50 (2%)
Neurological	No	2,411 (91%)	2,088 (79%)
	Yes	229 (9%)	556 (21%)
	Missing	15 (1%)	15 (1%)
Gastrointestinal	No	2,225 (84%)	1,932 (73%)
	Yes	413 (16%)	703 (26%)
	Missing	17 (1%)	24 (1%)
Musculoskeletal	No	1,507 (57%)	1,145 (43%)
	Yes	1,080 (41%)	1,391 (52%)
	Missing	68 (3%)	123 (5%)
Disability	Independent	2,392 (90%)	2,114 (80%)
	Disabled	259 (10%)	538 (20%)
	Missing	4 (0%)	7 (0%)
ED visit* (past year)	No	2,125 (80%)	2,066 (78%)
	Yes	425 (16%)	481 (18%)
	Missing	105 (4%)	112 (4%)
Hospitalization* (past year)	No	2,357 (89%)	2,357 (89%)
	Yes	195 (7%)	195 (7%)
	Missing	103 (4%)	107 (4%)

Note: *Emergency department (ED) visit and overnight hospitalization in the past year were obtained from a maintaining contact questionnaire delivered 3.3–39.4 months after the baseline questionnaire.

Frailty

The frailty index (23) was generated using 65 deficits related to chronic conditions, activities of daily living, depression, perceptions of health, satisfaction with life, body mass, and social participation. These deficits were selected based on previous CLSA studies on frailty (24,25), and do not overlap with any of the biological measures employed in this study (Supplementary Table 4). Frailty was scaled to have mean = 0, SD = 1 prior to analysis to facilitate interpretation of regression coefficients.

Multimorbidity

We measured multimorbidity as the number of chronic conditions participants reported having been diagnosed with from a list of 32 chronic conditions recorded by the CLSA. We excluded hypertension and chronic obstructive pulmonary disease because these conditions are diagnosed based on blood pressure and spirometry measures included in the estimation of BA.

Count of chronic condition domains

Individual chronic conditions were grouped according to physiological domain—mental health, cardiometabolic, neurological, gastrointestinal, and musculoskeletal; the individual conditions included in each domain can be found in the Supplementary Methods. We computed the count as the number of domains in which participants reported at least one chronic health condition.

Disability

A participant was considered disabled if they were unable to perform one or more of the 14 basic and instrumental activities of daily living without any assistance.

Emergency department visits and hospitalization

As part of a maintaining contact questionnaire that was administered over the phone between 3.3 and 39.4 months (average, 16.6 months) following the baseline questionnaire, participants were asked if they were seen in an ED and if they were a patient in a hospital for at least one night anytime in the previous year. Responses were categorized as yes or no.

Statistical Analysis

We excluded participants that were missing more than half of the measures from at least one of the four measurement panels (BLD, PHYS, PERF, COG); the remaining sample included N = 26,566 individuals. Within this analysis sample, we imputed missing data for biological measures by predictive mean matching in the R package “mice” (26). We generated 50 data sets (maximum iterations = 25), which were merged into a single data set by taking the mean or mode of imputed variables.

All mathematical and statistical procedures were performed using R version 3.6. Measures of normality and identification of high-leverage values were determined using quantile-quantile and Cook’s distance plots, respectively, while VIF was calculated using the package “car.” For frailty, associations with ΔBA were determined by ordinary least squares regression; for the number of chronic conditions, negative binomial regression; and for individual chronic condition groups, disability and ED visit and hospitalization, binomial logistic regression. For all regression models, ΔBA was scaled (mean = 0, SD = 1; separately in males and females) to facilitate comparability across measure panels, age was included as a covariate, and 95% confidence intervals (CIs) were estimated using a profile likelihood-based approach. Regression models were performed separately in men and women, and where applicable, coefficients were compared by z-test, where: $z=({\beta }_{Male}-{\beta }_{Female})/\sqrt{S{E}_{Male}^2+S{E}_{Female}^2}$.

Results

Calculating BA in the CLSA

We computed six different measures of BA based on different biomarker panels: BA_OPT, BA_FULL, BA_PHYS, BA_PERF, BA_COG, and BA_BLD. All BA estimates were correlated with CA, although the magnitude of correlations varied: BA_OPT (r = 0.77) was the highest, followed by BA_FULL (r = 0.74), BA_PHYS (r = 0.65), BA_PERF (r = 0.59), BA_COG (r = 0.57), and BA_BLD (r = 0.47) (Figure 2A). The MAE for each estimate followed a similar order (Figure 2B). Correlations were similar in men and women, with the exception of BA_BLD, which was more strongly correlated with CA in men (r = 0.51 for men vs r = 0.44 for women; Supplementary Table 5). As with the comparisons above, the MAE for each ΔBA was similar between men and women with the exception of ΔBA_BLD, in which men were notably lower (13.6 vs 16.7 years) (Supplementary Table 5).

Figure 2.

Comparison of biological age (BA) estimates for each measure panel. (A) Biological age regressed on chronological age (CA). Statistics in the upper left of each figure panel summarize the respective BA estimate, and include: r, Pearson’s correlation with CA; lm, the linear regression model describing the relationship between BA and CA (x). (B) The difference between BA and CA (ΔBA). MAE represents the mean absolute error, or the average absolute difference between BA and CA.

Relative to ΔBA_FULL and ΔBA_OPT, the strength of correlation with other panels ranked very similarly: for ΔBA_FULL, correlations with ΔBA_COG and ΔBA_PERF were almost identical (r = 0.74), followed by ΔBA_PHYS (r = 0.55) and ΔBA_BLD (r = 0.33); ΔBA_OPT correlated with ΔBA_COG the strongest (r = 0.67), followed by ΔBA_PHYS (r = 0.62), ΔBA_PERF (r = 0.51), and ΔBA_BLD (r = 0.30). The correlation between ΔBA_FULL and ΔBA_OPT was 0.86. For the remaining measures, bivariate correlations were relatively low: for ΔBA_COG, correlations with ΔBA_PERF, ΔBA_PHYS, and ΔBA_BLD were 0.28, 0.16, and 0.04, respectively; for ΔBA_PERF, correlations with ΔBA_PHYS and ΔBA_BLD were 0.28 and 0.13, respectively; the correlation between ΔBA_PHYS and ΔBA_BLD was 0.13.

Associations of BA With Current Health Status

Across all biomarker panels, participants with higher ΔBA tended to exhibit higher levels of frailty and multimorbidity (Figure 3A). For frailty, ΔBA_PERF (standardized regression coefficient: male = 0.32 [0.28–0.35], female = 0.36 [0.32–0.39]) and ΔBA_FULL (male = 0.32 [0.29–0.36], female = 0.34 [0.30–0.37]) exhibited the strongest associations, followed by ΔBA_OPT (male = 0.24 [0.20–0.28], female = 0.22 [0.18–0.26]) and lastly, ΔBA_BLD, ΔBA_COG, and ΔBA_PHYS (all below 0.19 for men and women). For multimorbidity, a similar ranking was observed for ΔBA_PERF (incident rate ratio: male = 1.15 [1.12–1.18], female = 1.20 [1.17–1.23], p = .04) and ΔBA_FULL (male = 1.16 [1.13–1.20], female = 1.18 [1.15–1.21]), while ΔBA_COG exhibited the weakest associations (male = 1.07 [1.03–1.10], female = 1.05 [1.02–1.08]).

Figure 3.

Associations of the biological age difference (ΔBA) with current health status in males and females. (A) Estimates for frailty (regression coefficient) and multimorbidity (incident rate ratio) were calculated using ordinary least squares regression, and negative binomial regression, respectively. (B) Associations with prevalent chronic condition domains (mental, mental health; neuro, neurological; musc, musculoskeletal; cardio, cardiometabolic; gastro, gastrointestinal; ADL, disability) were estimated by logistic regression, with odds ratios presented; the dotted line represents an odds ratio of 1 (no difference between cases and controls). All models were adjusted for age and error bars represent the 95% confidence interval. Frailty was scaled to have mean = 0, SD = 1 prior to analysis to facilitate interpretation of regression coefficients.

We also compared binomial associations to prevalent chronic condition domains and disability. The strongest associations with ΔBA were most apparent for disability and cardiometabolic disease in both men and women (Figure 3B). For both sexes, ΔBA_FULL and ΔBA_PERF tended to exhibit the strongest associations across all health-related variables, followed by ΔBA_OPT, ΔBA_PHYS, ΔBA_COG, and ΔBA_BLD, although this ranking was dependant on both sex and domain considered. While sex differences in the associations between ΔBA and these health-related variables were for the most part not obvious, ΔBA_PERF estimates tended to be higher in women, and ΔBA_OPT estimates higher in men (Figure 3B).

Associations of BA With Health Care Utilization

We also investigated how different BA estimates were related to whether participants visited the ED or were hospitalized overnight in the past year (both reported as yes/no). For this analysis, nearly all BA measures, in both men and women, were significantly associated with both outcomes (Figure 4). Specifically, ΔBA_FULL tended to exhibit the strongest associations with ED visit (odds ratio [95% CI], M = 1.24 [1.12–1.38], F = 1.33 [1.20–1.47]) and hospitalization (M = 1.66 [1.44–1.90], F = 1.46 [1.27–1.68]), and ΔBA_PERF (ED visit: M = 1.18 [1.07–1.31], F = 1.30 [1.18–1.43]; hospitalization: M = 1.45 [1.27–1.66], F = 1.45 [1.26–1.66]) and ΔBA_OPT (ED visit: M = 1.16 [1.05–1.29], F = 1.27 [1.15–1.40]; hospitalization: M = 1.46 [1.27–1.68], F = 1.32 [1.14–1.51]) were next strongest. Interestingly, although the estimates for men and women were very similar for nearly all measures and outcomes, the association between ΔBA_BLD and hospitalization for men was dramatically higher than for women, and the association for women was not significant (M = 1.36 [1.18–1.56], F = 1.01 [0.87–1.17], p = .004) (Figure 4).

Figure 4.

Associations of the biological age difference (ΔBA) with an emergency department (ED) visit and overnight hospitalization in the past year, in males and females. The odds ratio of reporting an ED visit or overnight hospitalization was estimated using logistic regression, adjusting for age, and error bars represent the 95% confidence interval. The dotted line represents an odds ratio of 1 (no difference between cases and controls).

Discussion

Although many studies have indicated that estimates of BA based on a variety of biological measures are associated with health-related variables and outcomes, relatively few have scrutinized the composition of panels employed. This is partly due to a lack of data sets that are both large in sample size, and contain enough biological measures to perform such an analysis. In order to fill this knowledge gap, we employed the large and rich database of the CLSA.

Of the six panels of measures we considered, BA estimated using the age-optimized panel (BA_OPT), as expected, was found to correlate most strongly with CA, albeit not by a lot (BA_OPTr = 0.77, BA_FULLr = 0.74). This small gain by BA_OPT is likely due to the exclusion of noninformative measures, which were observed to increase the amount of error (MSE) between BA and CA in linear regression analysis. Regarding the other panels considered, BA estimates for BLD were the worst (r = 0.47), especially so in women, likely due to nonlinear relationships between age and hematological measures, which is most evident around the perimenopausal and early postmenopausal periods; similar trends have been shown elsewhere (27,28). In contrast to our findings, a recent study by Putin and colleagues (29) showed that BA estimated using a deep neural network approach based solely on blood-derived measures can approximate CA with very little error. This suggests that aspects inherent to one’s study design, including the particular combination of measures chosen, the manner in which BA is estimated, and even the cohort employed, may play an important role in the precision of subsequent BA estimates. In the case of the study by Putin and colleagues, the ability of their deep neural network approach to model complex, nonlinear measures across the life span likely contributed greatly to the precision of their estimates.

Interestingly, when comparing the strength of associations between our BA estimates and health-related outcomes, BA estimated using physical performance measures (BA_PERF) tended to associate as well as that of BA_FULL and BA_OPT; associations with frailty, multimorbidity, cardiometabolic disease, disability, and overnight hospitalization were notable. Estimates of BA using blood-derived measures (BA_BLD), although the weakest as far as correlations with CA are concerned, were broadly associated with health-related outcomes, including disability, which we recently reported (30). These relationships were dependant on sex though, as women generally exhibited lower and sometimes nonsignificant associations as compared to men. The major conclusion drawn from these analyses is that the magnitude in which a BA estimate correlates with age does not necessarily dictate whether it will most strongly associate with health-related outcomes. Levine’s phenotypic age approach (31), which does not train its algorithm using CA, supports this, as it was shown to outperform BA with regards to predicting mortality. Similarly, Zhang and colleagues recently showed that as you increase the accuracy of CA estimation using DNA methylation data (ie, epigenetic clock), associations with mortality decrease (11).

The reason that certain age-related biological measures, namely those used to measure physical performance and organ function, tend to result in BA estimates that more strongly associate with health-related outcomes likely has to do with the importance of those measures to the pathophysiology or even diagnosis of those diseases and conditions. For example, disability is likely to be more prevalent in individuals that perform poorly in physical performance tests; hence, BA estimates based on physical performance measures would be expected to associate more strongly to those conditions, even if the observed correlation with CA was relatively poor. This would be similar for cardiometabolic diseases and BA estimates that include blood pressure, pulse, and body composition measures. Estimates such as these that exhibited strong associations to a given condition, as well as our use of cross-sectional data, illustrate the need for caution in our interpretation, due to the potential for reverse causality. Poor physical performance may precede disability, but may also be caused by it; likewise, cardiometabolic diseases may be influenced by, or influence, blood pressure, pulse, and body composition measures. Generally speaking, the health outcomes we studied have at least some potential to influence our metrics of BA. Thus, a proper interpretation is not that BA causes health outcomes, nor that BA is “merely correlated with” health outcomes, but rather that BA and health outcomes are intimately linked through feedback loops that defy traditional notions of causality (32).

For those BA estimates that associated relatively poorly to the outcomes included in this study (ie, BA_COG, BA_BLD, BA_PHYS), it is tempting to conclude that those component measures must not contribute to the pathogenesis of those outcomes; however, such logic may be fallacious. For example, ΔBA_BLD was not associated to the odds of having a mental health disorder, even though several blood test measures have been shown to correlate with depression and anxiety symptom severity (33). There are at least two possible reasons for this discrepancy: first, as asserted previously, the age-related change in a given measure may not be equivalent to the change that contributes to a chronic condition. Thus, the levels of a given measure relative to the mean of their peers may not be informative in predicting disease. Second, as demonstrated by Nelson and colleagues (34), measures that are causally related to the risk of death would be expected to exhibit a dampened correlation with CA in cross-sectional data. This form of “cohort selection” would also be expected to reduce the contribution of cognitive measures to BA estimates derived from some cross-sectional studies, given that significant cognitive impairment is a common exclusion criterion for participation. In their report, Nelson and colleagues showed that the bias caused by mortality-related cohort selection could be corrected if BA estimates are composed of biomarkers that sufficiently correlate with mortality (34); although this is not possible for the current study, this is a important consideration for studies that have access to longitudinal data.

While our study featured significant strengths, such as a large, nationally generalizable sample of older adults, a rich and diverse set of biological measures, and a rigorous statistical approach, it is not without its drawbacks. Of particular importance is the use of cross-sectional data, which may bias estimates related to the biological measures and outcomes employed. Certain biological measures may be influenced by cohort effects and this could affect the magnitude of correlation with CA. For example, it has been shown that birth weight, which increased steadily in the 20th century (35), is significantly associated with pulmonary function later in life (36). As such, the change in FEV1 or FVC with age that is estimated using cross-sectional data may not accurately reflect the true slope over time. A limitation of basing estimates on cross-sectional data and not incident outcomes is that they may be inflated; hence, depending on the role of reverse causality, they may or may not be good predictors of subsequent risk. However, BA estimates based on the aforementioned measures would be expected to associate more strongly with the incidence of those conditions and diseases since they are also important predictors of both progression and severity. Furthermore, although not a true longitudinal outcome since the date of reported events overlapped with the date in which biological measures were recorded for some participants, our BA estimates almost uniformly associated with ED visit and overnight hospitalization for both men and women. In fact, when performing this analysis on those participants in which there was no overlap between reported events and the date of biological measures (n = 3,722 of 5,314), the association between BA estimates and ED visit or hospitalization were very similar and all were statistically significant (data not shown). It should also be noted that the parameters estimated in this study for use in the Klemera–Doubal approach may specific to older Canadian adults or possibly just those who participated in the CLSA Comprehensive cohort.

In summary, our data show that considerations should be made when evaluating which biological measures are used to estimate BA, as this can significantly influence the association to health-related variables and outcomes. First, certain biological measures will have greater value in the estimation of BA with regards to the approximation of CA. Second, BA estimates that most closely approximate CA will not necessarily associate with health-related outcomes the strongest. Third, although a comprehensive panel of measures will generally be most effective for determining associations to health-related outcomes, understanding the causes and pathophysiology of a given outcome will inform the selection of biological measures and allow for a BA estimate that not only approximates CA but also associates with the outcome of interest reasonably well.

Funding

This work was supported by the Canadian Institutes of Health Research (CIHR) and Canadian Space Agency (CIHR Catalyst Grant, ACD-162980). D.W.B. is partly supported by the U.S. National Institute on Aging (R21AG054846), and L.E.G. is supported by the McLaughlin Foundation Professorship in Population and Public Health. A.A.C. is supported by a CIHR New Investigator Salary Award and is a member of the Fonds de recherche du Québec - Santé (FRQ-S)-supported Centre de recherche sur le vieillissement and Centre de recherche du CHUS.

Conflict of Interest

None declared.