ABSTRACT
We analyzed aging trajectories of complete blood counts (CBC) and their association with the incidence of chronic diseases and death in cohorts of aging individuals registered in the UK Biobank and National Health and Nutrition Examination Survey (NHANES) studies. Application of a pro-portional hazards model to the CBC data allowed us to identify the log-transformed hazard ratio as a natural biomarker of aging, which we have named the dynamic morbidity index (DMI). DMI increased with age in the UK Biobank and NHANES cohorts, was associated with frailty, and pre-dicted the prospective incidence of age-related diseases and death. To better understand the nature of DMI variations along individual aging trajectories, we acquired a sufficiently large longitudinal database of CBC measurements from a consumer diagnostics laboratory. We observed population DMI distribution broadening associated with a progressive loss of physiological resilience measured by the DMI inverse auto-correlation time. Extrapolation of this data suggested that DMI recovery time and variance would simultaneously diverge at a critical point of 120 – 150 years of age corresponding to a complete loss of resilience. We conclude that the criticality resulting in the end of life is an intrinsic biological property of an organism that is independent of stress factors and signifies a fundamental or absolute limit of human lifespan.
INTRODUCTION
Aging is manifested as a progressive functional decline leading to increasing prevalence [1, 2] and incidence of chronic age-related diseases (e.g., cancers, diabetes, cardiovascular diseases, etc. [3–5]) and disease-specific mortality [6]. Much of our current understanding of the relationship of aging with changes in physiological variables over an organism’s lifespan originates from large cross-sectional studies, including the National Health and Nutrition Examination Survey (NHANES; 40592 subjects, age range 18 – 85 y.o.) and UK Biobank (UKB; 471473 subjects, age range 39 – 73 y.o.) studies. Analysis of such large datasets has identified a number of “biological clocks”, including those reflecting age-related variations in blood markers [7], DNA methylation (DNAm) states [8, 9] or patterns of locomotor activity [10–12]. Typically, the physiological indices change from the levels observed in the young organism at approximately linear pace, much slower than would be expected from the ex-ponential Gompertz mortality law [13, 14]. Most factors that serve as biological clocks are easily measurable and, at least in principle, may be modified pharmacologically. An improved understanding of the relationship between the linear physiological state dynamics and the exponentially increasing morbidity, frailty and mortality observed during aging is needed to facilitate the rational design, development, and clinical validation of anti-aging interventions.
RESULTS
Identification of a new biomarker of aging based on blood cell counts
Complete blood count (CBC) measurements are included in standard blood tests and comprise the largest common subset of data available for both the NHANES and UKB study populations (see Table S1). In investigating whether CBC dynamics could be used to quantify aging, we found that co-clustering of age- and sex-adjusted CBC components revealed two dynamic subsystems associated with oxygenation and immune functions, represented by variations in red blood cell counts and total and mean corpuscular hemoglobin levels and with platelet and white blood cell counts, respectively (Fig. 1A).
Since rates of morbidity and mortality increase exponentially with age, a log-linear risk model is a good start-ing point for quantification of the aging process [12, 15]. We used the death register of the NHANES study (3792 death events observed in the follow-up by year 2015) and trained the Cox proportional hazards model [16] using the CBC measurements and sex variables (but not age) from 23807 study participants aged 40 y.o. and older. We therefore assumed that the risk of death depended on the organism state at the time of CBC measurement and the follow-up time only. The model yielded a log-hazards ratio (log-HR) estimate (a linear combination of log-transformed CBC variables) for every participant (see Table S2 for description of the model). After adjustment for sex and age, this predictor was demonstrated to be equally well associated with mortality in the NHANES study (HR = 1.43) and in the independent UKB study (HR = 1.39; Table S3), which was used as a validation dataset.
For the NHANES cohort, the CBC log-HR of the Cox mortality model gradually increased with age (Pearson’s correlation coefficient r = 0.29, p < 10−100, Fig. 1B). This parameter changed at a comparable rate in the independent UKB cohort; however, the correlation coefficient was lower in this cohort due to the more limited age range of the subjects (r = 0.13, p < 10−100, Fig. S1A).
To differentiate between the effects of chronic diseases and disease-free aging, we followed [17] and characterized the health status of each study participant based on their diagnosis with health conditions considered to be the most prevalent age-related conditions. The number of health conditions diagnosed for an individual was normalized to the total number of conditions included in the analysis to yield the “compound morbidity index” (CMI), a frailty index proxy with values ranging from zero to one. The list of health conditions common to the NHANES and UKB studies that were used for CMI determination.is given in Table S3 and Supplementary Information.
The association between frailty and estimated mortality risk is readily seen as the difference between the solid and dashed lines in Fig. 1B, which represent the mean log-HR values in the cohorts of healthy (“non-frail”, CMI < 0.1) and “most frail” (CMI > 0.6) NHANES participants, respectively. Given the association between allcause mortality risk and morbidity, we propose to refer to the log-HR of the mortality model as the “dynamic morbidity index” (DMI).
In groups stratified by increasing number of health condition diagnoses, the normalized distribution of DMI (after adjustment by mean DMI in age- and sex-matched cohorts of healthy subjects) exhibited a progressive shift to higher risk values (Fig. 1C for NHANES, Fig. S1B for UKB). For both NHANES and UKB, the largest shift was observed in the “most frail” (CMI > 0.6) population. The increasingly heavy tail at the high end of the DMI distribution in this group is characteristic of a mixture of at least two distinct states which occupy adjacent regions in the configuration space spanned by the DMI variable. Therefore, DMI displacement after adjustment for age and sex in cohorts of healthy subjects was expected to be determined by the fraction of “most frail” individuals in a cohort of any given age. This was confirmed to be true using the NHANES dataset (Fig. S2A; r = 0.83, p = 8.8 ×109). The fraction of surviving “most frail” subjects increased exponentially until the age corresponding to the end of healthspan was reached. The characteristic doubling rate constants for the “most frail” population fractions were 0.08 and 0.10 per year in the NHANES and the UKB cohorts, respectively, in comfortable agreement with the accepted Gompertz mortality doubling rate of 0.085 per year [18], see Fig. S2B.
Dynamic morbidity index and health risks
In the most healthy subjects, i.e. those with no diagnosed diseases at the time of assessment, the DMI predicted the future incidence of chronic age-related diseases observed during 10-year follow-up in the UKB study (Table S3) There was no relevant information available in NHANES. We tested this association using a series of Cox proportional hazard models trained to predict the age at the onset/diagnosis of specific diseases. We observed that the morbidity hazard ratios associated with the DMI relative to its mean in age- and sex-matched cohorts were statistically significant predictors for at least the most prevalent health conditions (those with more than 3000 occurrences in the UKB population). The effect size (HR ≈ 1.03 – 1.07) was the same regardless of whether a disease was diagnosed first in a given individual or followed any number of other diseases. Therefore, we conclude that the DMI is a characteristic of overall health status that is universally associated with the risks of developing the most prevalent diseases and, therefore, with the end of healthspan as indicated by the onset of the first morbidity (HR ≈ 1.05 for the “First morbidity” entry in Table S3).
In “non-frail” individuals with life-shortening lifestyles/behaviors, such as smoking, the DMI was also elevated, indicating a higher level of risks of future diseases and death (Fig. S2C). Notably, however, this effect appeared to be reversible: while the age- and sex-adjusted DMI means were higher in current smokers compared to non-smokers, they were indistinguishable between groups of individuals who never smoked and who quit smoking (c.f. [12, 19]).
Physiological state fluctuations and loss of resilience
To understand the nature of forces shaping the dynamics of the aging process at higher age/time-resolution, we acquired a large set of longitudinal CBC measurements from a clinical diagnostics laboratory. This dataset, which we refer to as GEROLONG, included 629 male and 1800 female subjects aged 35 – 90 with complete CBC analyses that were sampled 4 – 20 times within a period of up to 42 months.
As seen with the NHANES and UKB cohorts, DMI also increased with age in the longitudinal GEROLONG cohort. The average DMI value and its population variance at any given age were, however, considerably larger than those in the reference “non-frail” groups from the NHANES and UKB studies (see Fig. S1A). This difference likely reflects an enrollment bias: many of the GEROLONG blood samples were obtained from patients visiting clinic centers, presumably due to health issues. This could explain why the GEROLONG population appeared generally more frail in terms of DMI than the reference cohorts of the same age from other studies (Fig. S1A, compare the relative positions of the solid blue line and the two dashed lines representing the GERO-LONG cohort and the frail cohorts of the NHANES and UKB studies, respectively). There was no medical condition information available for the GEROLONG subjects. Hence, we used the mean DMI of the “most frail” NHANES and UKB participants (which coincided, approximately, with the mean DMI of all GEROLONG subjects) as the cutoff value to select “non-frail” GERO-LONG individuals.
Within the “non-frail” GEROLONG population, serial CBC measurements from individuals over a periods of time of up to three years revealed large stochastic fluctuations of the DMI around its mean values, which differed between individual study participants. The averaged DMI auto-correlation function is a basic property of a stochastic process (see e.g., [20]) and decayed exponentially as a function of the time delay between measurements within approximately a month (see Fig. 2B). As described in Supplementary Information, for a stationary process, the inverse auto-correlation time is an indicator of the relaxation (recovery) rate, characterizing the time scale involved in equilibration of a system’s state in response to external perturbations. We therefore propose using this quantity as a measure of an organism’s “resilience”, the capacity of an individual organism to resist and recover from the effects of physiological or pathological stresses [21, 22]).
We fitted the auto-correlation functions to an exponential function of the time delay and observed that recovery rates obtained from fitting to data in the subsequent age-cohorts decreased approximately linearly with age (Fig 2A). Extrapolation to older ages suggested that the equilibration rate vanishes and, hence, the recovery time becomes formally infinite, at an age of approximately 120 −150 y.o. (95% CI, 110 – 170 y.o.).
The variance of DMI increased with age in every dataset evaluated in this study. Since the dynamic range of random fluctuations would be inversely proportional to the recovery force (or the recovery rate, see [20]), we plotted the inverse variance of the DMI computed in sex- and age-matched cohorts of healthy persons (Fig. 2C). Again, extrapolation suggested that, if the tendency holds at older ages, the population variability would increase indefinitely at an age of approximately 120 – 150 y.o.
DISCUSSION
In this study, we investigated aging trajectories of human CBC values and their association with risks of chronic age-related diseases, mortality and life-shortening lifestyles. We produced a proportional hazards mortality model using a large NHANES dataset and defined its log-hazard ratio prediction as the dynamic morbidity index (DMI). This quantitative parameter displayed all of the expected properties of a biomarker of aging in several large independent datasets: DMI increased with age, was predictive of the prospective incidence of age-related diseases and death, and was associated with typical life-shortening lifestyles, such as smoking, and frailty. These findings support the idea that predictors from log-linear mortality or morbidity risk models can be effectively used to quantify the progress of aging and effects of lifestyles and diseases [12, 15, 23].
The simultaneous divergence of the organism state recovery times (critical slowing down) and the range of DMI variations (critical fluctuations) is characteristic of proximity of a critical point [20] at some advanced age over 100 y.o. Under these circumstances, the organism state dynamics are stochastic and dominated by the variation of the single dynamic variable associated with the criticality, the DMI (Fig. 2D). Schematically, far from the critical point (at younger ages), the organism state perturbations can be thought of as confined to the vicinity of a possible stable equilibrium state in a potential energy basin (A). Initially, the dynamic stability is provided by a sufficiently high potential energy barrier (B) separating this stability basin from the inevitably present dynamically unstable regions (C) in the space of physio-logical parameters. While in stability basin, an organism follows the traectory (D) of the equiliblium state, which is gradually displaced with aging even for the successfully aging individuals.
The DMI auto-correlation times (one-two months, see Fig. 2B) are much shorter than lifespan. The dramatic separation of time scales makes it very unlikely that the linear decline of the recovery force measured by the recovery rate in Fig. 2A can be explained by the dynamics of the organism state captured by the DMI variation alone. Therefore, we conclude that the progressive remodeling of the attraction basin geometry reflects adjustment of the DMI fluctuations to the slow independent process that is aging itself. In this view, the aging drift of the DMI mean in cohorts of healthy individuals (as in Fig. 1B) is the adaptive organism-level response to ever increasing stress produced by the aging process.
The dynamic range of the DMI fluctuations is inversely proportional to the recovery rate of the DMI fluctuations and hence the two parameters are dependent quantities. Each of them can be used equally well as a novel biomarker of aging. For example, recovery rate can serve as a biomarker of aging that is independent of DMI levels, but characterizes fluctuations of DMI on time scales of a few months or more and is associated with the progressive loss of physiological resilience. Such age-related remodeling of recovery rates has been previously observed in studies of various physiological and functional parameters in humans and other mammals. For example, in humans, a gradual increase in recovery time required after macular surgery was reported in sequential 10-year age cohorts [24] and age was shown to be a significant factor for twelve months recovery and the duration of hospitalization after hip fracture surgery [25, 26], coronary artery bypass [27], acute lateral ankle ligament sprain [28]. A mouse model suggested that the rate of healing of skin wounds can be a predictor of longevity [29].
In a reasonably smooth potential energy lanndscape forming the basin of attraction, the activation energy required for crossing the protective barrier (B) decreases along with the curvature at the same pace, that is, linearly with age. Whenever the protective barrier is crossed, dynamic stability is lost (see example trajec-tories 1 and 2 in Fig. 2D, which differ by the age of crossing) and deviations in the physiological parameters develop beyond control, leading to multiple morbidities, increasing frailty, and, eventually, death.
On a population level, activation into such a frail state is driven by stochastic forces and occurs approximately at the age corresponding to the end of healthspan, under-tood as “disease-free survival”. Since the probability of barrier crossing is an exponential function of the required activation energy (i.e., the barrier height) [20], the weak coupling between DMI fluctuations and aging is then the dynamic origin of the Gompertz mortality law. Since the remaining lifespan of an individual in the frail state is short, the proportion of frail subjects at any given age is proportional to the barrier crossing rate, which is an exponential function of age (see Fig. S2B).
The end of healthspan can therefore be viewed as a form of a nucleation transition [20], corresponding in our case to the spontaneous formation of states corresponding to chronic diseases out of the metastable phase corresponding to healthy organisms. The DMI is then the order parameter associated with the organism-level stress responses at younger ages and plays the role of the “reaction coordinate” of the transition to the frail state later in life. All chronic diseases contributing to frailty and death in our model originate from the dynamic instability associated with single protective barrier crossings. This is, of course, a simplification and yet the assumption could naturally explain why mortality and the incidence of major age-related diseases increase exponentially with age at approximately the same rate [3].
The DMI is an organism level variable and, as such, is not a property of any specific functional subsystem or organism compartment. Indeed, the associations of individual CBC components with the DMI extend over the functional clustering shown in Fig. 1A. In the vicinity of a critical point, fluctuations associated with the critical mode become amplified and hence the DMI should be identifiable with the signal components explaining most of the variance in virtually every biological signal (see, e.g., [7, 30] for reviews including performance assessment of variance based biological age models (such as Principal Component Analysis)-based markers of aging). For example, in our recent study [12], we observed that the first principal component score in the configuration space spanning the physical activity acceleration/deceleration statistics was strongly associated with mortality, incidence of first morbidity, and health status.
According to the presented model, external stresses (such as smoking) or diseases produce perturbations that modify the shape of the effective potential leading to the shift of the equilibrium DMI position. For example, the mean DMI values in cohorts of individuals who never smoked or who quit smoking are indistinguishable from each other, yet significantly different from (lower than) the mean DMI in the cohort of smokers (Fig. S2C). Thus, the effect of the external stress factor is reflected by a change in the DMI and is reversed as soon as the factor is removed. These findings agree with earlier observa-tions suggesting that the effects of smoking on remaining lifespan and on the risks of developing diseases are mostly reversible once smoking is ceased well before the onset of chronic diseases [12, 19].
The reversible character of the DMI deviations in healthy subjects suggests that the relationship between the DMI and survival shows signs of “antagonistic pleiotropy”. On one hand, the elevation of physiological variables associated with the DMI indicates reversible activation of the most generic protective stress responses at younger ages, when the organism state is dynamically stable. Moderately elevated DMI levels are therefore a marker of generic stress that can measured by molecular markers (e.g., C-reactive protein) and affects general physical and mental health status [23]. On the other hand, the excessive DMI observed in older organisms can be thought of as an aberrant activation of stress-responses beyond the dynamic stability range. This is a characteristic of frailty, multiple chronic diseases and death.
We propose that therapies targeting frailty-associated phenotypes (e.g., inflammation) would, therefore, produce distinctly different effects in disease-free versus frail populations. In healthy subjects, who reside in the region of the stability basin (B) (see Fig. 2D), a treatment-induced reduction of DMI would quickly saturate over the characteristic auto-correlation time and lead to a moderate decrease in long-term risk of morbidity and death without a change in resilience. Technically, this would translate into an increase in healthspan, although the reduction of health risks would be transient and disappear after cessation of the treatment. In frail individuals, how-ever, the intervention could reduce frailty, thus increasing lifespan beyond healthspan.
The emergence of chronic diseases and frailty out of increasingly unstable fluctuations of the organism state provides the necessary dynamic argument to support the derivation of the Gompertz mortality law in the Strehler-Mildvan theory of aging [31]. In [32, 33], the authors suggested that the exponential growth of disease bur-den observed in the National Population Health Survey of Canadians over 20 y.o. could be explained by an age-related decrease in organism recovery in the face of a constant rate of exposure to environmental stresses. Our study provides evidence suggesting that vanishing resilience cannot be avoided even in the most successfully aging individuals and, therefore, could explain the very high mortality seen in cohorts of super-centennials characterized by the so-called compression of morbidity (late onset of age-related diseases [34]). Formally, such a state of “zero-resilience” at the critical point corresponds to the absolute zero on the vitality scale in the Strehler-Mildvan theory of aging, thus representing a natural limit on human lifespan.
The proximity to the critical point indicates that the apparent human lifespan limit is not likely to be improved by therapies aimed against specific chronic diseases or frailty syndrome. Thus, no strong life extension is possible by preventing or curing diseases without interception of the aging process, the root cause of the underlying loss of resilience. We do not foresee any laws of nature prohibiting such an intervention. Therefore, further development of the aging model presented in this work has the potential to eventually lead to experimental demonstration of a meaningful life-extending therapy.
MATERIALS AND METHODS
Full details for the materials and methods used in this study, including information of the CBC parameters, Cox proportional hazards model, health conditions, and analysis of physiological state fluctuations are provided in the Supplementary Information.