A roadmap to build a phenotypic metric of ageing: insights from the Baltimore Longitudinal Study of Aging

Abstract

Over the past three decades, considerable effort has been dedicated to quantifying the pace of ageing yet identifying the most essential metrics of ageing remains challenging due to lack of comprehensive measurements and heterogeneity of the ageing processes. Most of the previously proposed metrics of ageing have been emerged from cross-sectional associations with chronological age and predictive accuracy of mortality, thus lacking a conceptual model of functional or phenotypic domains. Further, such models may be biased by selective attrition and are unable to address underlying biological constructs contributing to functional markers of age-related decline. Using longitudinal data from the Baltimore Longitudinal Study of Aging (BLSA), we propose a conceptual framework to identify metrics of ageing that may capture the hierarchical and temporal relationships between functional ageing, phenotypic ageing and biological ageing based on four hypothesized domains: body composition, energy regulation, homeostatic mechanisms and neurodegeneration/neuroplasticity. We explored the longitudinal trajectories of key variables within these phenotypes using linear mixed-effects models and more than 10 years of data. Understanding the longitudinal trajectories across these domains in the BLSA provides a reference for researchers, informs future refinement of the phenotypic ageing framework and establishes a solid foundation for future models of biological ageing.

Introduction

Life expectancy has increased steadily over the past century contributing to the ageing of the population worldwide [1]. Over the past three decades, our understanding of the biological mechanisms that affect longevity and healthspan has grown exponentially. A fundamental aspect of this work is the search for an overall metric of human ageing derived from a combination of multidimensional assessments that distinguish and differentiate individuals who are ageing ‘faster’ or ‘slower’ than the population average [2–7]. A corollary of this work is that individuals who appear to age ‘faster’ are closer to transitioning from a state of health to a state of disease, disability or death and, in general, experience a shorter healthspan and lifespan. Characterization of these ‘at risk’ individuals may facilitate discovery of targeted and effective treatments aimed at maximizing healthspan and reducing global susceptibility to diseases typical of old age [4, 8, 9]. Indeed, establishing robust and universal metrics of the pace of ageing would provide powerful clinical tools for prevention and cure. Importantly, we often hear of interventions that can slow ageing processes, but in the absence of a valid metric of ageing, such claims cannot be substantiated.

It is generally accepted that acute and chronic diseases derive from specific pathophysiological processes, and therefore, prevention should be disease specific as well [10–15]. Simultaneously, overwhelming scientific evidence supports the global health benefits of several behaviours, such as maintaining a healthy weight and engaging in regular physical activity in reducing risk of development and progression of multiple chronic diseases [16]. Consistent with this view, geriatric medicine has recognized that the health of older persons is best captured by complex functional measures, such as comprehensive geriatric assessment (CGA), rather than disease-specific measures [17, 18]. However, CGA cannot guide prevention and healthspan extension because a detectable decline in physical and cognitive function occurs when the compensatory mechanism of human physiology is exhausted [18]. Therefore, such functional measures reveal the severity of damage and ‘age acceleration’ too late for implementing preventative interventions that are maximally effective [17, 18]. Alternatively, measures that can stratify risk in younger, healthier populations by detecting deficits earlier in the disease/decline process are essential for optimizing prevention and assessing the effectiveness of interventions initially in clinical trials and, eventually, in clinical applications [19–21].

Composites of phenotypic and biological ageing have been previously developed and their validity has been demonstrated [22–26]. However, most of these indices have selected variables correlated with ageing or associated with mortality from those available in large epidemiological studies and used a weighted average to construct a unique index [22–28]. These tools are limited by the cross-sectional nature of the assessments used and absence of a theoretical construct about the domains of health and function being assessed (Table S1). The wide utilization of ‘omics’ in epidemiological studies has been followed by the appearance in the literature of different indices of ‘biological’ ageing, with those based on DNA methylation the most successful, by far [26, 29–32]. Further development and refinement of these measures would highly benefit from the availability of indices that comprehensively summarize the longitudinal anatomical, physiological and pathological changes that typically occur with ageing, therefore providing a robust measure of phenotypic ageing. Ideally, these measures should be constructed using parameters that capture essential characteristics of ageing, collected longitudinally to avoid the biasing effect of secular trends and selective attrition and organized according to the major domains that mediate the relationships between ageing/diseases and physical and cognitive limitations. Recently, it has been recognized that although risk factors for development of specific medical conditions, clinical progression and mortality can be reliably identified from cross-sectional studies, biomarkers that convey unbiased information on the pace of ageing are best elaborated using longitudinal data [33]. Furthermore, given that most age-related changes begin to manifest in early-to-mid life and accelerate in later decades, observations should cover a substantial proportion of the lifespan rather than being limited to the final stages of life.

As a first step in addressing this challenge, we delineate longitudinal trajectories of key biological and physiological measures of ageing that encompass different stages across the lifespan using data from the Baltimore Longitudinal Study of Aging (BLSA) [34]. The BLSA was started in 1958, continuously enrolling volunteers over a wide age range, and following them with visits every 1–4 years depending on age, making it the ideal study for this challenge. The BLSA systematically collects state-of-the-art clinical and functional variables purposely selected to study the ageing process. In previous work, we have organized these measures into major domains that mediate the effect of ageing and diseases on physical and cognitive function: energetics, body composition, homeostatic mechanisms and neuroplasticity [34, 35]. The frequency of follow-up is sufficient to capture possible departure from linearity in these trajectories, a consideration essential for any attempt to develop a robust phenotypic measure of ageing. Extensive tables and figures incorporated throughout the manuscript, including those in the appendix, serve as initial steps in delineating measures most pertinent to understanding trajectories of ageing. In the text, we describe the clinical and functional assessments that, in our view, best characterize changes in ageing phenotypes. Information on additional assessments examined can be found in the appendix.

Materials and methods

Study population

The BLSA is a study of normative human ageing, established in 1958 and conducted by the National Institute on Aging Intramural Research Program [36]. A comprehensive revision of the design was conducted in 2003, focused on establishing objective enrolment criteria for recruitment of healthy participants and a more comprehensive focus on phenotypic domain-based measurements [37, 38]. All participants are community-dwelling volunteers free of major chronic conditions upon enrolment. Detailed inclusion/exclusion criteria are described in Table S2. Enrolled participants are followed up with an age-dependent frequency (<60 every 4 years, 60–79 every 2 years, ≥80 every year), by either clinical visit or home visit.

The sample for the current study consists of 1581 participants who underwent a physical examination, health history assessment, and extensive functional testing, cognitive testing, and imaging during their clinic visits between January 2005 and December 2018. Trained and certified technicians administered all assessments following standardized protocols. The study protocol has been approved by the Internal Review Board of the National Institute of Environmental Health Sciences and participants provided written informed consent at each visit.

Measurements

As mentioned in the introduction, the measures collected in the BLSA are organized around domains that are currently conceptualized as interfaces by which disease and ageing affect physical and cognitive function: body composition, energetics, homeostatic mechanisms and neuroplasticity. In the context of BLSA, these variables are considered ‘phenotypes of ageing’ and their emergence is assumed to be caused by underlying biological processes that currently are not well understood (Fig. 1) [34, 39]. Although strict protocols were followed, given the longitudinal nature of the BLSA, it is inevitable that changes in testers and technology occurred over the study period (e.g. analysers, assays). Extensive efforts were made to control for these changes in the analysis.

Fig. 1.

Conceptual framework illustrating the phenotypic measures of ageing organized by domain. The proposed phenotypic measures of ageing include four domains: body composition, energetics, homeostatic mechanisms, and neurodegeneration/neuroplasticity. Changes in the ageing phenotypes are driven by the biological mechanisms (centre circle) of ageing and manifest as functional changes with ageing (peripheral borders) through the phenotypic domains. Specific variables are included as examples and are not meant to be an exhaustive list.

Body composition domain

Changes in body composition are evident across the lifespan. Traditional body size measures are collected in BLSA, including waist circumference and weight and height, used to estimate body mass index (BMI) [40]. Total lean mass, appendicular lean mass and total fat mass were assessed using total body dual-energy X-ray absorptiometry (DEXA; Prodigy Scanner, GE, Madison, WI) with Encore Software. While total lean mass is composed mainly of both muscle and visceral organs, appendicular lean mass (both arms and legs) mostly represent muscle mass [41]. DEXA measures are complemented with computerized tomography cross-sectional images (10 mm) at the mid-thigh area (10 mm, Somatom Sensation 10, Siemens, Malvern PA) quantified using the GEANIE 2.1 software (BonAlyse Oy, Jyvaskla, Finland) and TIBIA ESTIMATION TOOL (TibEsT v.1.4, Sokratis, NIH).[42] Muscle strength was assessed as grip strength using a Jamar Hydraulic hand dynamometer. The maximum force (kg) measured from three trials on each hand was used in the analysis [41, 43].

Energetic domain

Both parameters of energy availability and energy consumption change with ageing. Several energetics measures (oxygen consumption (VO₂)) are collected in BLSA from resting to maximal exertion. Resting metabolic rate (kcals day⁻¹), the minimal amount of energy required for living, was assessed by indirect calorimetry [44, 45], using a Cosmed k4b2 portable metabolic analyser (Cosmed, Rome, Italy) after awakening in the morning in a quiet, thermo-neutral environment, in a fasted, rested state [46]. Peak VO₂ (mL kg⁻¹ min⁻¹) was assessed during a modified Balke protocol maximal treadmill test as a proxy measure of maximal energy availability (VO₂ max) [47]. The balance between energy availability and demand for physical functioning was estimated by a ratio of the energy cost of slow walking to peak walking capacity (‘cost–capacity ratio’) [48–50]. The energetic cost of slow walking (mL kg⁻¹ min⁻¹) was assessed via indirect calorimetry (Medical Graphics Corp, St Paul, MN, USA) during 5 min of slow treadmill walking at 0.67 m s⁻¹ (1.5 mph), zero percent slope [48]. Peak walking capacity was assessed during a 400-meter walk test performed in an uncarpeted corridor with the participant wearing a portable metabolic analyser, the Cosmed K₄b² (Cosmed) [46, 48]. Forced vital capacity (FVC) and the forced expiratory volume in the first second (FEV₁), indicators of respiratory capacity and functions, were measured using a MedGraphics Gas Exchange System (Medical Graphics Corp) through closed-circuit breath collection [51].

Homeostatic mechanisms domain

The integrity of the ‘self’ and a stable homeostatic equilibrium is essential for healthy life and is maintained by many homeostatic mechanisms, few of them can be evaluated clinically and are particularly relevant to ageing and are assessed in the BLSA. They include chronic inflammation, insulin sensitivity/resistance, cardiovascular parameters and circulating lipids. As biomarkers of chronic inflammation, we considered interleukin-6 (IL-6), C-reactive protein, albumin and haemoglobin. IL-6 was measured by commercial ELISA kits (R&D System, Minneapolis, MN, USA), and C-reactive protein (CRP), by ELISA (ALPCO Diagnostics, Salem NH, or Alpha Diagnostic International, San Antonio, TX, USA or Immundiagnostik AG) [42, 52]. Albumin was measured using dye binding BCG [53–55]. Haemoglobin, red blood cell width (RDW) and absolute neutrophil counts, measured by Sysmex’s multiple methods, have been linked with inflammation and immunosenescence [56]. Fasting plasma glucose was measured by the glucose oxidase method in the morning after at least 10 h overnight fast [42, 57]. Total cholesterol was measured with enzymatic methods, HDL and LDL was measured with dextran magnetic, and triglycerides with colorimetric methods, all also after 10 h overnight fast [53]. Cardiovascular parameters included blood pressures and pulse-wave velocity. Blood pressures were measured in the supine position, three times on each arm, alternating right and left with 1 min between each measurement [58]. Carotid-femoral pulse wave velocity was measured using either Transcutaneous Doppler probes (model 810A, 9 to 10- Mhz probes, Parks Medical Electronics, Inc., Aloha, OR, USA) or Complior SP device (Artech Medical, Paris, France) or SphygmoCor system (AtCor Medical, Sydney, NSW, Australia) by well-trained technicians [58]. Renal function was estimated by calculating body surface area (BSA) adjusted 24-h urine creatinine according to standard methods. Standard formulas for the estimation of glomerular filtration rate were avoided because they include an ‘age’ parameter. Creatinine concentrations in serum and urine were measured either by the enzymatic Vitros CREA method performed on the Ortho Fusion 5.1 Analyzer (Ortho-Clinical Diagnostics, Rochester, NY, USA), or the isotope dilution mass spectrometry (IDMS)-traceable serum creatinine assay [59].

Neurodegeneration/Neuroplasticity domain

The central nervous system was assessed by brain volume, and the peripheral nervous system was assessed by nerve conduction velocity. Brain volume was measured using a 3T Philips Achieva Magnetic Resonance Imaging (MRI) system to acquire magnetization-prepared rapid gradient echo (MPRAGE) scans (repetition time = 6.8 ms, echo time = 3.2 ms, flip angle = 8°, image matrix = 256 × 256, 170 slices, pixel size = 1 × 1 mm, slice thickness = 1.2 mm; sagittal acquisition). Anatomical labels and global and regional brain volumes were obtained using Multi-atlas region Segmentation using Ensembles of registration algorithms and parameters (MUSE) [60]. To measure the fibular nerve conduction velocity, a trained technician performed a standard nerve conduction velocity test on the peroneal nerve. To measure the fibular nerve conduction velocity, a trained technician performed a standard nerve conduction velocity test on the peroneal nerve [61].

Other covariates

Height and weight were assessed in light clothing using a stadiometer and calibrated scale, respectively.

Statistical methods

Descriptive statistics are presented as mean (SD) for continuous variable, and number (%) for categorical variables. Student’s t test and chi-squared tests were used for baseline comparisons between men and women for continuous variables and categorical variables, respectively. The cohort structure of BLSA is depicted in Fig. 2.

Fig. 2.

Graphic summary of the BLSA cohort that was used for this analysis. Each line or dot denotes a BLSA participant; A line ending with a solid dot means the participant died, while a line ending with an arrow and a halo dot means the participant is still alive at that age. The upper panel shows the time since enrolment by participant age. The lower panel shows each participant’s age versus calendar time. Men are plotted in blue, and women are plotted in red.

Linear mixed models with random intercepts and random slopes were used to model the longitudinal change of the phenotypic measures by domain. All models included variables for baseline age, male sex, time, baseline age x time, and (male sex) × time [62]. Baseline age was centred at 50 years, and time was scaled at the unit of decade, so the interpretation of the coefficient of time is the estimated change in the next decade for a 50-year-old female. The coefficients for baseline age should be interpreted as the cross-sectional relationships between chronological age and the phenotypic measure and are included in the tables. The coefficient of male sex should be interpreted as the sex difference in the phenotypic measure at age 50, and the coefficient of the interaction term, (male sex) × time, should be interpreted as the sex difference in the estimated change in the next decade starting at 50 years old. All sex difference coefficients are reported in the tables, but only those with significant differences are discussed in the results section. For blood pressure measures, analytical models included three-way interactions between baseline age, sex and time to better describe sex-differences in the age-associated changes in haemodynamics [58, 63]. The comparisons between models and fitted plots with and without the three-way interaction were presented in the supplements.

For analyses of the energetics domain, variables that convey information on body size, such as height and composition (fat and lean mass), were included as covariates. To account for batch effects in the analyses of the homeostatic domain, indicators variables for the different methods used over time as well as time of measurement were included as covariates. For analyses of the body composition domain, body height was included to provide an index of body size. For the neurodegeneration/neuroplasticity domain, covariates for intracranial volume (ICV) were included for brain volume measurements, and the analysis was limited to those ≥40 years old (due to limited data below this age) with no history of Alzheimer’s disease, stroke, traumatic brain injury or Parkinson’s disease. Different technicians and technology used over time were accounted for by including appropriate variables as confounders in nerve conduction velocity analysis. Models with and without baseline age squared were performed to assess the cross-sectional nonlinear age trend. The final models presented are based on the Akaike information criterion and predefined hypothesis for each model. The covariates included in the final models are listed in the table. These models were subsequently also used to plot the longitudinal change in each phenotypic measure by sex to demonstrate the rate of change. In these plots, the starting points for each line are based on the population average at the starting point and might be influenced by cohort effects as well by selective survival and study attrition [62, 64]. Since the main goal of the plots is to demonstrate the change over a decade at different ages, the interpretation and visual perception of these plots should be focused on the slope more than on the intercepts [40, 62].

Results

Overall characteristics

A conceptual description of the phenotypic domains included in the BLSA design and a dynamic view of the BLSA cohort are illustrated in Figs 1 and 2, respectively. Variables selected within each domain are generally considered dimensions that change with ageing; a comparison of the BLSA phenotypic variables with some of previously published papers in this area is summarized in Table S1. Among the 1581 participants aged 22–96 at baseline included in this report, 51.4% were female, 68.7% were white, and average years of education was 17 (Table S3). Given the evolution of the BLSA visit structure and measures, not all measurements were available across all participants. For example, only participants without potential or obvious pathological changes of the central nervous system were selected to receive brain imaging examination at subsequent visits [65, 66]. The sample sizes for different analyses ranged from n = 935 (brain volumes) to n = 1575 (IL-6 measures). Missing values were due to eligibility requirements for specific tests, or a result of technical problems. Detailed information about the number of participants for each variable, number of data points and length of follow-up are described in Tables S4 and S5.

Body composition domain

The results concerning variables in the body composition domain are summarized in Table 1. BMI (Fig. 3a) increased in early-to-mid life, reaching a peak around the age of 65, at which point it began to decline. These changes in BMI reflect the complex make-up of this composite variable and reflect changes in other body composition measures. For instance, total fat mass (Fig. 3b) increased over time (β = 1.70 kg per decade, P < 0.001) and the slope of increase was greater for men than for women (β = 1.72 kg per decade, P < 0.001), but the rate of change reversed over time, showing a decline in fat mass starting in the mid-sixties for women and at around age eighty for men. Interestingly, waist circumference (Fig. 3c), a biomarker of visceral obesity, increased progressively in both men and women, showing some minimal decline only after the age of 90. Total lean mass (Fig. 3d) increased up to the mid-forties in men and mid-fifties in women and declined thereafter. Appendicular lean mass (Fig. 3e) showed a similar trend as total lean mass, with declines beginning in the fifties for men and the sixties for women. Similarly, the mid-thigh muscle area measured by CT showed an accelerated decline after the age of 50, and rates of decline were steeper in men (β = 1073 mm² per decade, P < 0.001) than in women (β = 475 mm² per decade, P < 0.001) after the age of 60 (Fig. S1, and Table S6). Grip strength (Fig. 3f) declined an average rate of 2.44 kg per decade at age 50 (P < 0.001), with a rate of decline that accelerated with ageing (P < 0.001) and was steeper in men (β = −3.73 kg per decade, P < 0.001) than in women (β = −2.15 kg per decade, P < 0.001).

Table 1.

Results of the mixed model analyses for variables from the body composition domain^d

	Estimate	SE	P-value
Waist (cm)a
Baseline age	0.278	0.03	<0.001
Male	10.291	0.777	<0.001
Time	3.051	0.421	<0.001
Baseline age × time	−0.129	0.019	<0.001
Male × time	−0.111	0.455	0.807
Waist-to-height ratio
Baseline age	0.001	0.0001	<0.001
Male	0.034	0.003	<0.001
Time	0.022	0.003	<0.001
Baseline age × time	−0.001	0.0001	<0.001
Male × time	−0.002	0.003	0.495
Body mass index (kg m−2)b
Baseline age	0.065	0.013	<0.001
Male	0.999	0.234	<0.001
Time	1.349	0.149	<0.001
Baseline age × time	−0.056	0.007	<0.001
Male × time	0.107	0.169	0.527
Lean mass (kg)c
Baseline age	−0.087	0.009	<0.001
Male	10.698	0.355	<0.001
Time	0.282	0.168	0.093
Baseline age × time	−0.09	0.008	<0.001
Male × time	−1.595	0.189	<0.001
Appendicular lean mass (kg)a
Baseline age	−0.057	0.007	<0.001
Male	5.693	0.194	<0.001
Time	0.850	0.128	<0.001
Baseline age × time	−0.045	0.006	<0.001
Male × time	−1.263	0.145	<0.001
Fat mass (kg)a
Baseline age	0.189	0.028	<0.001
Male	−6.603	0.746	<0.001
Time	1.701	0.346	<0.001
Baseline age × time	−0.134	0.016	<0.001
Male × time	1.721	0.391	<0.001
Mid-thigh area (mm2)a
Baseline age	−72.695	5.532	<0.001
Male	3704.241	142.322	<0.001
Time	−105.534	96.429	0.274
Baseline age × time	−36.944	4.258	<0.001
Male × time	−597.801	101.608	<0.001
Grip strength (kg)a
Baseline age	−0.187	0.018	<0.001
Male	12.011	0.466	<0.001
Time	−2.153	0.329	<0.001
Baseline age × time	−0.099	0.015	<0.001
Male × time	−1.58	0.364	<0.001

^aAdjusted for baseline height, baseline age².

^bAdjusted for baseline age².

^cAdjusted for baseline height.

^dBaseline age is centred at 50 years old, and time is scaled at unit of a decade.

Fig. 3.

Longitudinal fitted changes in selected variables in the body composition domain. The above plot shows the fitted longitudinal changes in body mass index, fat mass, waist circumference, total lean mass, appendicular lean mass and grip strength by decade.

Energetics domain

Results from variables in the energetic domain are summarized in Table 2. Peak VO₂ (Fig. 4a), which approximates energy availability, declined with age at an average of 1.4 mL kg⁻¹ min⁻¹ per decade for women. The rate of decline in peak VO₂ increased with advancing age and was greater for men than women (P < 0.001). Resting metabolic rate (Fig. 4b) declined linearly over time, averaging 160.1 kcals day⁻¹ per decade for women. This decline was steeper for men (β = −118.4 kcals day⁻¹ decade⁻¹, P < 0.001). Energy reserves, as measured by the cost–capacity ratio (Fig. 4c) decreased an average of 6.5% per decade (P < 0.001), commencing in the early thirties, and accelerating with increasing age (P < 0.001). FEV₁ and FVC (Fig. 4d,e) decreased by 0.19 L per decade and 0.26 L per decade, respectively, for women (P < 0.001 for both). The rates of decline increased over time (P < 0.001) and these declines were greater for men than women (P < 0.001). Though the ratio of FEV₁/FVC was negatively associated with age, the change in the ratio of FEV₁/FVC was not obvious until the age of 50 and decreased later in life.

Table 2.

Results of the mixed model analyses for variables from the energetics domain^e

	Estimate	SE	P-value
Resting metabolic rate (kcal day−1)a
Baseline age	−3.591	0.579	<0.001
Male	76.205	25.515	0.003
Time	−160.726	27.245	<0.001
Baseline age × time	−2.188	1.189	0.066
Male × time	−118.435	28.495	<0.001
Peak VO2 (treadmill) (mL kg−1 min−1)b
Baseline age	−0.224	0.013	<0.001
Male	1.175	0.346	0.001
Time	−1.427	0.308	<0.001
Baseline age × time	−0.036	0.014	0.010
Male × time	−1.216	0.340	<0.001
Cost–capacity ratioc
Baseline age	0.002	0.001	<0.001
Male	−0.061	0.012	<0.001
Time	0.065	0.019	0.001
Baseline age × time	0.004	0.001	<0.001
Male × time	0.027	0.019	0.151
FEV1 (L)d
Baseline age	−0.029	0.001	<0.001
Male	0.486	0.035	<0.001
Time	−0.189	0.026	<0.001
Baseline age × time	−0.005	0.001	<0.001
Male × time	−0.112	0.028	<0.001
FVC (L)d
Baseline age	−0.031	0.002	<0.001
Male	0.616	0.047	<0.001
Time	−0.255	0.033	<0.001
Baseline age × time	−0.005	0.002	0.001
Male × time	−0.103	0.036	0.004
FEV1/FVCd
Baseline age	−0.001	0.00017	<0.001
Male	0.003	0.005	0.531
Time	0.002	0.004	0.697
Baseline age × time	−0.001	0.0002	0.006
Male × time	−0.010	0.005	0.035

FEV₁, forced expiratory volume in the first second; FVC, forced vital capacity; VO₂, oxygen consumption.

^aAdjusted for height, fat mass, lean mass.

^bAdjusted for height, fat-to-lean ratio, baseline age².

^cAdjusted for height, weight, baseline age².

^dAdjusted for height, baseline age².

^eBaseline age is centred at 50 years old, and time is scaled at unit of a decade.

Fig. 4.

Longitudinal fitted changes in selected variables in the energetics domain. The above plot showed the fitted longitudinal changes in peak VO₂, resting metabolic rate, cost–capacity ratio, FEV₁ and FVC by decade.

Homeostatic mechanisms domain

Results concerning variables in the homeostatic domain are summarized in Table 3. IL-6 (Fig. 5a) increased linearly over time, averaging 0.937 pg mL⁻¹ per decade (P < 0.001). CRP (Fig. 5b) showed a nonsignificant increasing trend of longitudinal change (β = 0.13 mg L⁻¹ decade⁻¹, P = 0.384). Albumin (Fig. 6a) declined linearly over time, averaging 0.11 g dL⁻¹ per decade. Haemoglobin (Fig. 6b) declined linearly over time, averaging at 0.238 g dL⁻¹ per decade (P = 0.01) for women, and 0.249 g dL⁻¹ per decade among men (P < 0.001). Red blood cell distribution width (Fig. S2A) increased linearly over time, averaging 0.496 % per decade, and was lower in men than women at baseline (β = −0.152 %, P = 0.002). Absolute neutrophil counts (Fig. S2B) increased an average of 407.03 cells μL⁻¹ per decade, and the rate of increase was progressively steeper at older ages (β = 6.96 cells μL⁻¹ decade⁻¹, P = 0.024). Fasting glucose (Fig. 6c) increased an average of 5.5 mg dL⁻¹ per decade, and the rate of increase was faster in male than in females (β = 1.31 mg dL⁻¹ decade⁻¹, P = 0.049) slowed with advancing age (β = −0.16 mg dL⁻¹ decade⁻¹, P < 0.001), and started declining after about age 80. For the blood pressure measures, because of the inclusion of a three-way interaction, the effect of sex on time trajectories varied across different ages. To facilitate the interpretation, in addition to the longitudinal plots shown in Fig. 6d,e, we calculated for the estimated rate of change at aged 50 and 80 among men and women. Systolic blood pressure (Fig. 6d) increased in women averaging 7.4 mmHg per decade (P < 0.001) at age 50 and slowing to 6.4 mmHg per decade (P < 0.001) at age 80, while it increased at decelerating rates in men averaging 4.9 mmHg per decade (P = 0.002) at age 50 and 2.5 mmHg per decade (P = 0.098) at age 80 (at 50, male sex–time interaction β = −2.5 mmHg per decade, P = 0.179; at 80, male sex–time interaction β = −3.9 mmHg per decade, P = 0.031). Thus, overall, the rate of increase in SBP declined over time, and men experienced faster decline. Diastolic blood pressure (Fig. 6e) declined inwomen averaging −0.7 (P = 0.369) mmHg per decade at age 50 to −1.1 (P = 0.183) mmHg per decade at age 80 and demonstrated accelerated decline averaging −0.873 mmHg per decade (P = 0.315) at age 50 and −3.846 mmHg per decade (P < 0.001) at age 80 in men (at 50, male sex–time interaction β = −0.2 mmHg per decade, P = 0.876; at 80, male sex–time interaction β = −2.7 mmHg per decade, P = 0.004). In other words, DBP decreased over time after age 50, and men experienced faster decline than women. Pulse pressure increased in women averaging 8.1 mmHg per decade (P < 0.001) at age 50, slowing to 7.6 mmHg per decade (P < 0.001) at age 80. In men, pulse pressure increased averaging 5.9 mmHg per decade (P < 0.001) at age 50, and increasing to 6.3 mmHg per decade (P < 0.001) at age 80 (at 50, male sex–time interaction β = −2.2 mmHg per decade, P = 0.132; at 80, sex–time interaction β = −1.4 mmHg per decade, P = 0.336). This suggests that pulse pressure increased over time linearly in both men and women. Carotid-femoral pulse-wave velocity (Fig. 6f), a parameter that conveys information on arterial stiffness, increased linearly over time, averaging 0.707 m s⁻¹ (P < 0.001) per decade. Creatinine clearance (corrected by body surface area) (Fig. 7a), an indicator of renal function, decreased by 14.2 mL min⁻¹/1.73 m² per decade (P < 0.001), the rate of decrease was progressively steeper with advancing age (β = 0.498 mL min⁻¹/1.73 m² per decade per year older, P < 0.001). Both total cholesterol and LDL cholesterol (Fig. 7b,c) increased earlier in life before decreasing in later life. Timing of the reversing trends was sex-dependent. For women, total cholesterol increased up to age 50 (β = 3.553 mg dL⁻¹ decade⁻¹, P = 0.230), and started declining at approximately age 58, while for men, total cholesterol started declining at 7.258 mg dL⁻¹ per decade up to age 50 (P for sex difference < 0.001). For women, LDL cholesterol increased longitudinally by 5.8 mg dL⁻¹ (P < 0.001) per decade up to age 50 and started declining at approximately 63 years old, while for men, LDL cholesterol started declining at 3.6 mg dL⁻¹ per decade up to age 50 (P for sex difference <0.001). HDL cholesterol (Fig. S3) decreased longitudinally at an average rate of 3.1 mg dL⁻¹ per decade (P = 0.013). Triglycerides (Fig. 7d) increased an average of 10.3 mg dL⁻¹ per decade at 50 (P = 0.009), but the rate of change decreased with advancing age, and started declining at approximately 71 years old for men and 79 years old for women.

Table 3.

Results of the mixed model analyses for variables from the homeostasis mechanisms domain^f

	Estimate	SE	P-value
Interleukin – 6 (pg mL−1)a
Baseline age	0.016	0.006	0.003
Male	0.171	0.097	0.077
Time	0.937	0.201	<0.001
Baseline	−0.001	0.008	0.853
age × time
Male × time	0.165	0.196	0.398
CRP (mg L−1)b
Baseline age	0.006	0.003	0.062
Male	−0.391	0.095	<0.001
Time	0.134	0.154	0.384
Baseline	0.002	0.006	0.759
age × time
Male × time	0.148	0.133	0.266
Albumin (g dL−1)b
Baseline age	−0.006	0.00046214	<0.001
Male	0.053	0.013	<0.001
Time	−0.109	0.022	<0.001
Baseline	−0.000114	0.001	0.883
age × time
Male × time	−0.002	0.019	0.923
Haemoglobin (g dL−1)a
Baseline age	0.005	0.003	0.109
Male	1.325	0.055	<0.001
Time	−0.238	0.092	0.01
Baseline	−0.011	0.003	<0.001
age × time
Male × time	−0.441	0.068	<0.001
Red blood cell distribution width (%)c
Baseline age	0.016	0.002	<0.001
Male	−0.152	0.049	0.002
Time	0.496	0.081	<0.001
Baseline	0.004	0.002	0.092
age × time
Male × time	0.035	0.058	0.553
Absolute neutrophil count (cells μL−1)d
Baseline age	−1.628	3.221	0.613
Male	42.826	56.127	0.446
Time	407.025	96.237	<0.001
Baseline	6.963	3.082	0.024
age × time
Male × time	34.167	73.325	0.641
Fasting glucose (mg dL−1)c
Baseline age	0.197	0.026	<0.001
Male	3.312	0.465	<0.001
Time	5.510	0.819	<0.001
Baseline	−0.163	0.028	<0.001
age × time
Male × time	1.305	0.665	0.0499
Systolic blood pressure (mmHg)e
Baseline age	7.014	1.059	<0.001
Male	0.372	0.034	<0.001
Time	7.416	1.457	<0.001
Baseline	−2.493	1.857	0.179
age × time
Male × time	−0.032	0.068	0.633
Male × baseline	−0.217	0.048	<0.001
age
Male × baseline	−0.048	0.095	0.617
age × time
Diastolic blood pressure (mmHg)c
Baseline age	3.244	0.605	<0.001
Male	0.053	0.026	0.04
Time	−0.719	0.799	0.369
Baseline	−0.154	0.984	0.876
age × time
Male × time	−0.014	0.036	0.692
Male × baseline	−0.047	0.028	0.088
age
Male × baseline	−0.085	0.05	0.091
age × time
Pulse pressure (mmHg)c
Baseline age	3.756	0.866	<0.001
Male	0.316	0.038	<0.001
Time	8.093	1.19	<0.001
Baseline	−2.193	1.456	0.132
age × time
Male × time	−0.016	0.053	0.766
Male × baseline	−0.17	0.04	<0.001
age
Male × baseline	0.027	0.075	0.713
age × time
Carotid-Femoral pulse wave velocity (m s−1)a
Baseline age	0.05	0.005	<0.001
Male	0.564	0.085	<0.001
Time	0.707	0.16	<0.001
Baseline	0.006	0.006	0.342
age × time
Male × time	0.188	0.135	0.164
Creatinine clearance (mL min−1/1.73*m2)a
Baseline age	−0.653	0.071	<0.001
Male	4.052	1.237	0.001
Time	−14.207	2.637	<0.001
Baseline	−0.498	0.105	<0.001
age × time
Male × time	2.452	2.497	0.326
Total cholesterol (mg dL−1)a
Baseline age	0.321	0.096	0.001
Male	−18.403	1.710	<0.001
Time	3.553	2.958	0.230
Baseline	−0.470	0.101	<0.001
age × time
Male × time	−10.811	2.458	<0.001
Low-density lipoproteins (mg dL−1)b
Baseline age	−0.064	0.054	0.234
Male	−6.62	1.553	<0.001
Time	5.829	2.618	0.026
Baseline	−0.457	0.09	<0.001
age × time
Male × time	−9.387	2.191	<0.001
High-density lipoproteins (mg dL−1)b
Baseline age	0.084	0.027	0.002
Male	−13.29	0.782	<0.001
Time	−3.059	1.234	0.013
Baseline	0.035	0.033	0.298
age × time
Male × time	−0.899	0.805	0.264
Triglyceride (mg dL−1)a
Baseline age	0.401	0.141	0.004
Male	6.422	2.514	0.011
Time	10.338	3.976	0.009
Baseline	−0.359	0.118	0.002
age × time
Male × time	−2.839	2.831	0.316

^aAdjusted for method and time of measurement, baseline age²

^bAdjusted for method and time of measurement

^cAdjusted for time of measurement, baseline age²

^dAdjusted for race, method and time of measurement, baseline age²

^eAdjusted for time of measurement

^fBaseline age is centred at 50 years old, and time is scaled at unit of a decade.

Fig. 5.

Longitudinal fitted changes in selected phenotypes from the homeostatic mechanisms domain – Part I. The above plot showed the fitted longitudinal changes in IL-6 and CRP by decade.

Fig. 6.

Longitudinal fitted changes in selected phenotypes from the homeostatic mechanisms domain – Part II. The above plot showed the fitted longitudinal changes in albumin, haemoglobin, fasting glucose, systolic blood pressure, diastolic blood pressure and pulse-wave velocity by decade.

Fig. 7.

Longitudinal fitted changes in selected phenotypes from the homeostatic mechanisms domain – Part III. The above plot shows the fitted longitudinal changes in creatinine clearance, total cholesterol, LDL cholesterol and triglyceride by decade.

Neurodegeneration/neuroplasticity domain

Results concerning variables in the neurodegeneration/neuroplasticity domain are summarized in Table 4. Total brain and total white matter (Fig. 8a, b) declined linearly by 36.6 cm³ and 12.1 cm³ per decade, respectively (both P < 0.001). White matter declined faster among men than women by 3.5 cm³ per decade (P = 0.042). Grey matter (Fig. 8c) decreased at a rate of 25.9 cm³ per decade for women (P < 0.001) and decreased faster in men by 6.8 cm³ per decade (P = 0.016). Ventricular volume (Fig. 8d) increased at rate of 2.3 cm³ per decade among woman (P = 0.006) and increased faster in men than women by 3.1 cm³ per decade (P < 0.001). The decline in grey matter and the increase in ventricular volume were steeper at older ages (P < 0.001 for both). Fibular nerve conduction velocity (Fig. S4) declined linearly by an average of 2.06 m s⁻¹ per decade.

Table 4.

Results of the mixed model analyses for variables from the neurodegeneration/neuroplasticity domain^d

	Estimate	SE	P-value
Total brain volume (cm3)a,e
Baseline age	−2.82	0.107	<0.001
Male	1.549	3.395	0.648
Time	−36.56	4.04	<0.001
Baseline age × time	−0.222	0.166	0.183
Male × time	−5.222	3.519	0.138
White matter (cm3)b
Baseline age	−0.763	0.161	<0.001
Male	0.847	2.007	0.673
Time	−12.125	1.868	<0.001
Baseline age × time	−0.133	0.077	0.085
Male × time	−2.497	1.629	0.126
Grey matter (cm3)b
Baseline age	−1.884	0.177	<0.001
Male	1.099	2.206	0.618
Time	−25.887	3.229	<0.001
Baseline age × time	−0.473	0.133	<0.001
Male × time	−6.805	2.829	0.016
Ventricular volume (cm3)b
Baseline age	0.281	0.104	0.007
Male	2.684	1.36	0.049
Time	2.252	0.818	0.006
Baseline age × time	0.431	0.033	<0.001
Male × time	3.149	0.738	<0.001
Fibular nerve conduction velocity (m s−1)c
Baseline age	−0.138	0.014	<0.001
Male	0.533	0.342	0.119
Time	−2.057	0.379	<0.001
Baseline age × time	−0.01	0.016	0.559
Male × time	−0.692	0.374	0.064

^aAdjusted for baseline intracranial volume

^bAdjusted for baseline intracranial volume, baseline age²

^cAdjusted for measurement software, technician ID, baseline height and weight and baseline age²

^dBaseline age is centred at 50 years old, and time is scaled at unit of a decade.

^eNote that total brain volume includes not only grey matter, white matter, and ventricular volume, but also brain stem.

Fig. 8.

Longitudinal fitted changes in selected phenotypes from the neurodegeneration/neuroplasticity domain. The above plot shows the fitted longitudinal changes in total brain volume, total white matter, total grey matter and ventricular volume by decade.

Discussion

This manuscript used longitudinal data from the BLSA to delineate trajectories of phenotypic dimensions with ageing. Variables were selected within conceptual dimensions that mediate the effects of ageing and disease on functional status, namely changes in body composition, energetics, homeostatic mechanisms and neurodegeneration/ neuroplasticity. Our data represent the most comprehensive review of the phenotypic changes that occur with ageing since the original publication of the Normal Human Aging: Baltimore Longitudinal study of Aging in 1984 [67].

These findings suggest that the current commonly used approach of summarizing ageing through a linear combination of measures may be problematic. In fact, this approach assumes that these dimensions change linearly and proportionally with ageing, while it is evident from the data presented that most ageing manifestations accelerate – or even change direction – over time, so that differences between age 60 and 80 years are not equivalent to differences between age 30 and 50 years. The data reported here not only confirm the nonlinearity of phenotypic measures but also illuminate the possibility that ageing trajectories are often assessment specific.

A set of indicators exhibited linear longitudinal change (‘linear ageing phenotypes’), including resting metabolic rate, albumin, IL-6, pulse pressure, HDL cholesterol, red blood cell distribution width, nerve conduction velocity, total brain volume (after age 40) and white matter volume (after age 40), while another set of measures demonstrated longitudinal changes that progressively increased in magnitude with older age. These ‘accelerated ageing phenotypes’ include peak VO₂, cost–capacity ratio, FEV₁, FVC, haemoglobin, absolute neutrophil count, appendicular lean mass, creatinine clearance, grip strength, grey matter volume (after age 40) and ventricular volume (after age 40). It is possible that the age acceleration in at least some of these measures reflects a progressive decline in reserve capacity or resilience in the presence of disease processes. For example, most young people can afford to substantially increase their level of physical activity as desired because their cardiovascular system has the resilience to increase the amount of oxygen and substrate transported to muscle mitochondria, and the mitochondria can increase their oxidative capacity through training. However, as functional reserve and resiliency are reduced, increased effort for physical activity takes the individual closer to their true energetic limit and the effect of ageing becomes even more evident. In addition, the increased cost of movement coupled with lower energetic capacity further contributes to accelerated declines in performance. The trends in cost–capacity ratio show decline early in life, much before the age of 60, when age-related decline in mobility is usually detected. Similarly, the decline of reserve respiratory capacity and increased energetic cost of respiration may explain the accelerated decline of FEV₁ and FVC [51]. For the other variables that show this trend, namely, appendicular lean mass and brain volume in some regions (grey matter and ventricular volume), the reason for accelerated decline is less clear. Interestingly, the brain and lean body mass (mostly muscle or parenchymal tissues) have high energetic cost and the decline in energy production due to impaired micro-perfusion and intrinsic mitochondrial function may lead to progressive atrophy. Whether these changes reflect normal ageing or emerging pathology is unknown.

A third category of variables is those that are initially correlated with age in one direction but then switch direction later in life. These ‘bimodal ageing phenotypes’ include BMI, waist circumference, fat mass, fasting glucose, diastolic blood pressure, systolic blood pressure, total cholesterol, triglycerides and LDL cholesterol. The reasons for the observed trends are likely heterogeneous and include treatment effects such as medication use, lifestyle change (e.g. diet) and synergistic processes (e.g. weight loss begets SBP decline) [64, 68–72]. Triglycerides, LDL cholesterol and fat mass are strong predictors of cardiovascular and all-cause mortality in younger and middle-aged individuals but become risk factors for mortality after the age of 80, suggesting that such changes may be due to underlying pathology that is not well understood [73, 74]. For example, changes in fasting glucose and SBP are difficult to interpret and may be related to parallel changes in body composition, especially the decline in fat mass that occurs at older ages [75]. A possible hypothesis is that these changes are ‘permissive’ of very old age, and those who do not undergo these changes experience early mortality and thus do not contribute to late life estimates.

To the best of our knowledge, this is the first paper describing longitudinal trends in pre-defined ageing phenotypes across the lifespan. Though researchers have spent decades describing ageing phenotypes, it is rare that the longitudinal trends in several prespecified assessments are presented collectively using the same cohort, which may allow us to better understand the reciprocal temporal relationship of change and assess the stability of measurement among community dwelling older adults. For example, CRP has been proposed as an ageing phenotype [76–78]. Even though we did not observe an obvious overall trend of increasing CRP, we did observe statistically significant increasing CRP using the log-transformed CRP for the analysis (Table S7, Fig. S5). Plausibly, the increase in CRP may be limited to a portion of the population affected by pathology and therefore is not a true biomarker of ageing. Alternatively, the increase in CRP may be too slow to be observed on linear scale among healthy people and can only be detected via log-transformation, which magnifies the degree of the change at low CRP levels. Indeed, only 13.2% participants experienced increases in CRP over time, which is consistent with results from the English Longitudinal Study of Aging (14%) [79]. Another study using the data from Doetinchem Cohort Study with an average follow-up of 17.5 years also reported a non-significant longitudinal trend in CRP elevation among men [80]. In contrast, serum albumin, which has been included in other proposed metrics of ageing, appears to decline at a constant rate with ageing, and low albumin, independent of co-existing morbidity is associated with a several health outcomes, perhaps because low levels capture the cumulative effects of inflammation, poor nutrition and liver dysfunction [6, 54, 81–87].

To date, most models designed to capture phenotypic/biological ageing were developed by selecting dimensions cross-sectionally associated with age, and primarily validated using predictive models of multimorbidity and/or mortality as the reference outcomes, and only rarely using longitudinal change in phenotypes [25, 26, 30, 31, 33, 88–91]. Although these models have contributed substantially to our clinical understanding of the ageing process, our work shows that the validity, precision and robustness of these models can be further improved by considering longitudinal data and perform better especially later in the age spectrum when nonlinear trends are more likely to occur.

Assessing the pace of ageing beginning in mid- or even early-life is important for at least three reasons. First, it supports the clinical goal of identifying individuals who are experiencing accelerated ageing and might benefit from targeted prevention efforts sooner than later in the decline process. Second, a main reason to develop a phenotypic metric of ageing is to use it as a tool for discovery of underlying biological and physiological mechanisms that drive accelerated ageing. The last several decades of life, by far, the most informative period of life for such analysis, is where current metrics of ageing fail to recognize that certain phenotypes change their directionality with ageing, making it essential to capture early – as well as late – trajectories to increase understanding of these processes. Third, as our understanding of the biological mechanisms that drive the ageing process expand and interventions that may slow ageing are developed, valid outcome measures of their effectiveness will be needed, especially ones that are sensitive to changes of ageing and functional trajectories throughout life.

The current study shows that while a simple weighted linear combination of selected variables may produce a convenient metric, it may not yield a valid index of phenotypic age. A valid index of ‘phenotypic ageing’ should distinguish among ‘linear’, ‘accelerated’ and ‘bimodal’ ageing parameters and therefore must include nonlinear and interaction terms in its construction. This approach opens the door to a new model of geriatric medicine that identifies individuals at high risk of developing chronic conditions and facilitates implementation of interventions aimed at alleviating and/or preventing disease processes. For example, cost–capacity ratio (Fig. 4d) shows age-associated decline as early as the third decade of life and thus may serve as a candidate measure to identify individuals on an accelerated path to mobility loss [50]. Moreover, the identification of underlying biological mechanisms that correlate or predict the development of such accelerated ageing phenotypes in humans may allow for the identification of ‘accelerated ageing’ through a simple blood test.

The proposed variables and metrics can be used in clinical trials as either the target of an intervention or for the measurement of specific outcomes. Currently, the targets of several clinical trials includes reducing the pro-inflammatory state typically seen with ageing and evaluating response to treatment [92–94], yet results have been mixed. Indeed, inflammation is a strong predictor of multimorbidity and mortality, two of the most important consequences of the ageing process. Among the BLSA participants aged 60 and above, higher CRP plasma level is associated with 8% shorter time to death (Fig. S7) and 14% shorter time to developing two or more age-related morbidities (Fig. S8), which is consistent with previous findings from the InCHIANTI study [95–98]. Our findings support the importance of chronic inflammation as a phenotypic marker of ageing and provide a potential explanation for results from the CANTOS trial, a trial investigating the effect of a monoclonal antibody targeting at interleukin-1β, where the all-cause mortality did not differ between groups over a follow-up of 3.7 years, indicating longer term follow-up may be warranted. Finally, consistent with the Geroscience paradigm, biological mechanisms of ageing strongly contribute to the pathogenesis of most chronic diseases; therefore, defining accelerated ageing phenotypes provides a link between underlying biological mechanisms and clinical manifestations of functional decline, which may help reveal pathogenetic mechanisms not previously discovered [39, 99].

Another important contribution of this work is the identification of sex differences in these ageing phenotypes which were observed for total lean mass, appendicular lean mass, fat mass, mid-thigh area, resting metabolic rate, peak VO₂, cost–capacity ratio, FEV₁, FVC, blood pressures, total cholesterol, LDL, triglyceride, grey matter volume and ventricular volume. These findings might inspire researchers to investigate the sex difference of these phenotypes more extensively. For example, as the sex difference in cardiovascular phenotypes have been reported [58, 63], we also illustrated these differences by fitting models with and without three-way interactions, which may sometimes give us a clearer picture of the role of sex in the trajectories of these phenotypes (Fig. S6 and Table S8). There are several potential explanations for these differences, including menopausal changes in hormones (e.g. oestrogen, testosterone, endothelin-1) [100–102], telomere length and attrition [103, 104], exposure to oxidative stress [101, 105, 106], genetics (e.g. Y-chromosome, X-chromosome-related mechanism, nuclear structure) [107–109] and epigenetic factors [108, 109]. Though sex differences were not extensively examined (e.g. including the three-way interaction between baseline age, time and sex) in this review, our work points to substantial differences in the rate of change in many phenotypes between men and women and highlight areas for future research. Specifically, it is unclear whether a valid index of ‘phenotypic ageing’ account for sex differences or separate ‘phenotypic ageing’ indexes should be developed for men and women.

Substantially, more work in this field needs to be done. For example, although we delineated longitudinal changes of different phenotypic markers with ageing, whether these changes are mutually independent, evolve in clusters or are fully harmonic is unknown. Do individuals with accelerated ageing experience accelerated phenotypes across all dimensions or only a specific set of dimensions? In other words, is there a shared mechanism of ageing, or several shared mechanisms of ageing that play out differently across groups of individuals? Machine learning may be needed to incorporate both cross-sectional measurement and longitudinal change to improve the predictive and concurrent criterion validity.

The uniqueness of the BLSA in terms of richness of phenotypes and repeated measures allows us to delineate a comprehensive set of age trajectories across several dimensions. However, it would be difficult to replicate these findings in a different population and should be a consideration in future studies. Another limitation of this study is that the whole set of measures is only available for participants who were seen in the clinic and not for home visits, which may have led to underestimation of changes near the end-of-life. Third, the strict enrolment criteria in this study may have selected a particularly healthy population, although this potential source of bias is offset by the fact that most BLSA participants remain in the study for many years. In spite of these limitations, this study provides the most comprehensive view of phenotypic ageing currently available in the literature. It can serve as a reference for researchers who study ageing as well as a foundation for developing a solid metric of the pace of ageing that may have innumerable clinical applications.

In conclusion, this study examined the longitudinal change in several indicators of ageing phenotypes across four predetermined domains. Findings confirm the critical importance of incorporating longitudinal assessment of several measures representing central phenotypical domains as an approach to identify and differentiate rate of ageing.

Supplementary Material

Supplemental Information

Figure S1. Longitudinal fitted change in mid-thigh mass by decade.

Figure S2. Longitudinal fitted changes in red blood cell width distribution and absolute neutrophil count by decade.

Figure S3. Longitudinal fitted change in HDL cholesterol by decade.

Figure S4. Longitudinal fitted change in nerve conduction velocity by decade.

Figure S5. Longitudinal fitted change in log CRP.

Figure S6. Longitudinal fitted change in cardiovascular parameters.

Figure S7. Cumulative incidence of death by CRP level.

Figure S8. Cumulative incidence of age-related morbidities by CRP level.

Table S1. Brief comparison of the measurements collected in BLSA and the measurements used in the relevant published paper.

Table S2. The inclusion and exclusion criteria implemented in BLSA.

Table S3. Baseline characteristics of BLSA participants in this analysis.

Table S4. Sex distribution and number of subjects for variables used in this report.

Table S5. Length of follow up and number of repeated observations for variables used in this report.

Table S6. Results of the mixed model analyses for additional longitudinal change in mid-thigh area.

Table S7. Results of the mixed model analyses for log-transformed CRP.

Table S8. Results of the mixed model analyses for cardiovascular phenotypes with and without the three-way interaction (baseline age X male sex X time).