nature medicine

Article https://doi.org/10.1038/s41591-023-02296-6

Heterogeneous aging across multiple organ

systems and prediction of chronic disease
and mortality

Received: 26 September 2022 Ye Ella Tian 1 , Vanessa Cropley1, Andrea B. Maier2,3,4,

Nicola T. Lautenschlager5,6, Michael Breakspear7,8 & Andrew Zalesky 1,9
Accepted: 9 March 2023

Published online: 6 April 2023

Biological aging of human organ systems reflects the interplay of age,
Check for updates chronic disease, lifestyle and genetic risk. Using longitudinal brain imaging
and physiological phenotypes from the UK Biobank, we establish normative
models of biological age for three brain and seven body systems. Here we
find that an organ’s biological age selectively influences the aging of other
organ systems, revealing a multiorgan aging network. We report organ age
profiles for 16 chronic diseases, where advanced biological aging extends
from the organ of primary disease to multiple systems. Advanced body
age associates with several lifestyle and environmental factors, leukocyte
telomere lengths and mortality risk, and predicts survival time (area under
the curve of 0.77) and premature death (area under the curve of 0.86). Our
work reveals the multisystem nature of human aging in health and chronic
disease. It may enable early identification of individuals at increased risk
of aging-related morbidity and inform new strategies to potentially limit
organ-specific aging in such individuals.

Age is the greatest common risk factor for chronic diseases1,2; how- diverse longitudinal populations15. A multiorgan characterization
ever, trajectories of age-related decline vary markedly between indi- of biological aging across major chronic diseases can facilitate new
viduals and differ across human organ systems3,4. Biological age is organ-specific therapeutic opportunities, yield disease-specific risk
thus recognized as a more informative marker of disease risk and calculators and elucidate factors that drive the divergence of an organ’s
mortality than chronological age5,6. As a result, cellular, molecular and biological age from chronological age. Elucidating such factors will
physiological aging biomarkers7,8 have been developed and studied inform strategies to potentially slow age-related decline, reduce the
across multiple species9–14. risk of chronic diseases and promote healthy longevity16–19.
To fulfill the clinical potential of this work, biological aging Biological aging of the human brain has been the focus of con-
clocks that are specific to particular organ systems, tissue types and siderable research20–22. Predictive models of brain age derived from
aging-related diseases are now required to be established in large and neuroimaging can infer apparent age based on brain structure and

1
Melbourne Neuropsychiatry Centre, Department of Psychiatry, Melbourne Medical School, The University of Melbourne, Melbourne, Victoria, Australia.
2
Healthy Longevity Translational Research Program, Yong Loo Lin School of Medicine, National University of Singapore, Singapore, Singapore. 3Centre
for Healthy Longevity, @AgeSingapore, National University Health System, Singapore, Singapore. 4Department of Human Movement Sciences,
@AgeAmsterdam, Amsterdam Movement Science, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands. 5Academic Unit for Psychiatry of Old Age,
Department of Psychiatry, Melbourne Medical School, The University of Melbourne, Melbourne, Victoria, Australia. 6NorthWestern Mental Health, Royal
Melbourne Hospital, Melbourne, Victoria, Australia. 7Discipline of Psychiatry, College of Health, Medicine and Wellbeing, The University of Newcastle,
Newcastle, New South Wales, Australia. 8School of Psychological Sciences, College of Engineering, Science and Environment, The University of
Newcastle, Newcastle, New South Wales, Australia. 9Department of Biomedical Engineering, Faculty of Engineering and Information Technology,
The University of Melbourne, Melbourne, Victoria, Australia. e-mail: ye.tian2@unimelb.edu.au; azalesky@unimelb.edu.au

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1221

Article https://doi.org/10.1038/s41591-023-02296-6

a b UKB participants
Aging clocks
Brain Medical history screening: Independent
Gray matter self-report, hospital inpatient and clinical data
White matter primary care records
Functional connectivity
Cognition Healthy individuals Individuals with
Immune diseases

Predicted age
Leukocytes
Erythrocytes Pulmonary
Thrombocytes FVC Training Test
C-reactive protein FEV1 20-fold
PEF Chronological age

Younger Older
Musculoskeletal
Cardiovascular Muscle strength
Pulse rate Bone mineral density
Systolic blood pressure Ankle spacing width
Diastolic blood pressure Height Aging profile Genes and environment
BMI
Waist circumference 16 chronic brain
Hip circumference and body diseases

Hepatic
Transferase enzymes Renal
Bilirubin Glomerular filtration
Albumin Electrolytes regulation Mortality prediction
5- and 10-year survival
Metabolic Premature death
Glucose
Lipids

Fig. 1 | Overview of study design. a, Organ systems for which normative gaps were determined for individuals with lifetime diagnoses of chronic diseases
models of biological age were established using organ-specific phenotypes. Key to investigate the relation of biological age with disease and mortality. Two
phenotypes are listed below each system. b, Predictive models of chronological independent datasets were used to validate brain age gap estimates in individuals
age were established using phenotypes from healthy adults and 20-fold cross- with dementia. Associations between genetic and environmental factors and
validation. Separate models were developed for each body and brain system and biological age were also investigated. UKB, UK Biobank. Image was created with
sex. Using models trained on healthy adults, personalized body and brain age BioRender.com.

function. The difference between chronological age and predicted pulmonary, musculoskeletal, immune, renal, hepatic and metabolic;
brain age, known as the brain age gap, provides a measure of biologi- Supplementary Table 1) and three brain (gray matter, white matter
cal age and can reveal insight into whether an individual’s brain seems and brain connectivity; Supplementary Table 2) systems. Cognitive
older or younger relative to same-aged peers21. While age gaps may performance formed an additional group (Supplementary Table 3).
first emerge in early life and accumulate across the lifespan, longi- Phenotypes for body systems were available for 143,423 individuals
tudinal increases in age gaps later in life relate more specifically to (age range 39–73 years, mean 56.7 ± 8.2, 79,980 males), and brain
aging-related decline. In principle, biological age can be estimated phenotypes were available for 36,901 individuals (age range 45–82
in vivo for organs and body systems other than the brain. Organ-specific years, mean 64.2 ± 7.5, 17,203 males). Cognitive phenotypes were
age gaps will enable concurrent investigation of biological aging across available for 32,317 individuals (age range 45–82 years, mean 65.1 ± 7.6,
multiple body and brain systems. To this end, we develop new assays 15,712 males).
to measure the biological age (age gap index) for seven body and three Healthy adults with no major medical conditions were selected
brain systems using imaging, physiological and blood-derived pheno- to train machine-learning models to predict individual chronologi-
types acquired cross-sectionally (body, n = 143,423; brain, n = 36,901) cal age (Methods). Separate predictive models of chronological age
and longitudinally (body, n = 1,220; brain, n = 1,294) in the UK Biobank were established for each body and brain system and sex. Subtracting
cohort. We aim to (1) map the influence of an organ’s biological age actual chronological age from predicted chronological age, referred
on the aging of other organ systems; (2) elucidate body and brain age to as the age gap, captures whether an individual’s organ system seems
profiles characteristic of 16 aging-related chronic diseases; (3) establish older (gap > 0) or younger (gap < 0) than population norms for the
whether organ-specific biological age associates with lifestyle factors individual’s chronological age and sex. Age gaps thus provide norma-
and leukocyte telomere length; and (4) predict the risk of mortality tive, organ-specific clocks of biological age (Fig. 1).
using body and brain age profiles. Our work reveals the heterogeneity
of biological aging across individuals and organs, and its relation to life- Normative aging models
style factors, risk of specific chronic diseases and mortality in midlife Chronological age could be predicted with modest to high accuracy for
and older adults. Quantifying the impact of major chronic diseases on body (female, r = 0.79, mean absolute error (MAE) = 3.71 years; male,
organ aging holds substantial promise for precision geriatric medicine r = 0.72, MAE = 4.46 years) and brain (female, r = 0.79, MAE = 3.52 years;
and related clinical translation. male, r = 0.80, MAE = 3.68 years) systems and cognition (female, r = 0.53,
MAE = 4.87 years; male, r = 0.54, MAE = 5.21 years; Fig. 2a). Prediction
Results accuracies varied between organ systems and sexes (Fig. 2b, Extended
Multimodal brain imaging, physiological and blood phenotypes Data Fig. 1 and Supplementary Table 4). Comparable accuracies
were grouped based on their relevance to the structure and func- for brain systems were achieved in additional datasets (female,
tion of specific organ systems; namely, seven body (cardiovascular, r = 0.82, MAE = 3.52; male, r = 0.85, MAE = 3.41; Extended Data Fig. 2a).

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1222

 Article https://doi.org/10.1038/s41591-023-02296-6

a Body Brain Cognition b

Model performance
1.0
Female
0.8

Correlation
Male

0.6
78 physiological measures 2,309 MRI-derived 29 cognitive
across 7 organ systems 0.4
brain measures measures
0.2
Female, n = 13,145 Female,!" n = 4,298 Female, n = 3,810 0
Predicted age (years)

90 90 90
#$ 8
80 80 80

MAE (years)
6
70 70 70
60 60 60 4
50 50 50
2
40 r = 0.79 40 r = 0.79 r = 0.53
40
MAE = 3.71 MAE = 3.52 MAE = 4.87
30 30 30 0
30 40 50 60 70 80 90 30 40 50 60 70 80 90 30 40 50 60 70 80 90

n
ic

e
Im c

og C
dy

M
M
l

lm l
na

ai
la

un

tio
i
Pu eta

ar

F
at

ol

W
G
Bo

Br
cu
on
Re

ep

ab

ni
Male, n = 15,444 Male, n = 3,624 Male, n = 3,357

el

as

et
H
sk
Predicted age (years)

90 90 90

ov

C
lo

di
cu
80 80 80

ar
us

C
M
70 70 70
60 60 60
c
50 50 50
Time of assessment
r = 0.72 r = 0.80 r = 0.54
40 40 40
MAE = 4.46 MAE = 3.68 MAE = 5.21
30 30 30 Body (t0) Body (t1) Brain (t0) Brain (t1)
30 40 50 60 70 80 90 30 40 50 60 70 80 90 30 40 50 60 70 80 90
Chronological age Chronological age Chronological age
(years) (years) (years)
2006–2010 2012–2013 2014–2020 2019–2021

d e Body age gap (t0, years) f Brain age gap (t0, years)
Body age gap Brain age gap
(t0, years) (t0, years)

Rate of change in body age gap Rate of change in brain age gap
(t1 – t0, months per year) (t1 – t0, months per year)
Muscle
Pulmon.
0.024 FC
0.31

Immune GM
0.0
41
Cardiac 0.051
0.29

Cardiac
0. 0.
03 28 31
WM 0.

93
7
0.073
0.29

0.
34
0.0 0. FC
Metab. 21 74 Muscle
0.0 0.031
0.36
–0.36

0.26

0.064 Brain
41
0.
Hepatic –0.
40 34
Body
0.0
4 7 0.0
65 WM 0.
0.22

–0
.29
0.023 GM
Renal
68 15
Pulmon. 0.0 0.
0.066 Metab.
Positive
Negative

Fig. 2 | Age prediction accuracy and multiorgan aging networks. a, Scatter- e, Influence of baseline organ age on longitudinal rate of change in organ age,
plots show associations between chronological and predicted age for prediction adjusting for overall body age gap, sex and chronological age at baseline.
models based on body (left), brain (mid) and cognitive (right) phenotypes. Lines An arrow from organ X to organ Y indicates that the age gap of X at baseline
of best fit are indicated with solid black lines. n, training sample size; r, Pearson significantly influences the rate of aging of Y (P < 0.05, FDR-corrected for
correlation coefficient. b, Bar plots show Pearson correlation coefficients 7 × 6 = 42 tests). f, Same as e but for brain systems (P < 0.05, FDR-corrected
(top) and MAE (bottom) quantifying age prediction accuracy (average of ten for 3 × 2 = 6 tests). Edge thickness and color reflect regression coefficients
repetitions of 20-fold cross-validation). c, Assessment timeline for measures (β) estimated for edges comprising the structural equation model. Including
of body and brain function. d, Influence of body age gaps (left) on brain age baseline and follow-up cognitive age gaps in the SEM did not reveal significant
gaps (right), adjusting for sex, chronological age and the time interval between influences of baseline tissue-specific brain age on the rate of cognitive aging.
assessments. Links are shown for significant influences inferred from SEM (false Metab., metabolism; pulmon., pulmonary; muscle, musculoskeletal; cardiac,
discovery rate (FDR)-corrected for eight body ages × four brain ages = 32 tests). cardiovascular; GM, gray matter; WM, white matter.

Applying the trained models to predict the chronological age of all (n = 1,294, 632 males; 2.0–2.7 years follow-up) systems. Chronological
participants resulted in personalized organ-specific age gaps. age was thus predicted at baseline (t0) and follow-up (t1), yielding two
Follow-up phenotype and imaging measurements were avail- age gaps for each organ per individual (Fig. 2c). This enabled estimation
able for body (n = 1,220, 837 males; 2.1–5.6 years follow-up) and brain of longitudinal rates of change in body and brain age.

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1223

Article https://doi.org/10.1038/s41591-023-02296-6

Associated with older biological age

Average weekly alcohol consumption Body r = 0.1
Current tobacco smoking r = 0.2
Past tobacco smoking
r = 0.26
Pack years of smoking
Pack years of smoking (proportional) Cardiac
Daily smoker Associated with younger biological age
Townsend deprivation index
Index of multiple deprivation (England) Birth weight
Index of multiple deprivation (Wales) Average weekly alcohol consumption
Index of multiple deprivation (Scotland) Pulmon . Age started smoking in current smokers
Job involves mainly walking or standing Time since stopped smoking
Job involves heavy manual or physical work Age completed full time education
Job involves shift work College/university qualification
Unable to work because of sickness Muscle
In paid employment
Transport to work-public Usual walking pace
Frequency of strenuous sports in last 4 weeks Frequency of stair climbing in last 4 weeks
Time spent watching television Duration walking for pleasure
Salt added to food Summed days activity
Immune
Nap during day Above moderate/vigorous recommendation
Sleeplessness/insomnia Oily fish intake
Daytime dozing/sleeping (narcolepsy) Cheese intake
Falls in the last year Cereal intake
Long-standing illness, disability or infirmity
Renal Good sleep duration
Hearing (speech reception threshold) Ease of getting up in the morning
Visual acuity (logMAR) Age when periods started (menarche)
Mood swings Had menopause
Fed-up feelings Hepatic Number of live births
Nervous feelings Age at first live birth
Loneliness, isolation Age at menopause
Neuroticism score Greenspace %, buffer 1,000 m
Had menopause Metab. Greenspace %, buffer 300 m
Nitrogen dioxide air pollution Natural environment %, buffer 1,000 m
Nitrogen oxides air pollution Natural environment %, buffer 300 m
Particlar matter air pollution (PM10)
Particlar matter air pollution (PM2.5) Brain
PM2.5 absorbance
Cognitive age

Fig. 3 | Environmental/lifestyle associations with biological organ age. environmental/lifestyle factors, adjusting for chronological age and sex. Link
Icons represent organ systems for which biological age was estimated. Links are widths are proportional to absolute value of correlation coefficient magnitudes.
shown between environmental/lifestyle factors significantly associated with Environmental and lifestyle factors were assessed using 158 scales tapping early
organ-specific age gaps (P < 2.6 × 10−5, two-sided, t-test, Bonferroni-corrected). life experience, sociodemographics, lifestyle, psychosocial, local environmental
Links are suppressed for small effect sizes (|r| < 0.05). Left (right) list consists exposure, general health and cognitive age. Results for GM, WM and FC are shown
of factors positively (negatively) associated with body/brain age gaps. Partial in Supplementary Fig. 1.
correlation was used to test for associations between organ-specific age gaps and

Multiorgan aging networks musculoskeletal aging is a common sequela of aging across multiple
Given that organ systems dynamically interact via nervous, circulatory organ systems (Fig. 2e).
and lymphatic networks23, we hypothesized that an organ’s age would For the brain, advanced gray matter age leads to faster aging of
selectively influence the rate of aging of several connected organ sys- functional brain connectivity (each 1-year increase in gray matter age
tems. Using structural equation modeling (SEM) on organ age gaps, we at baseline leads to 28 d per year increase in the rate of aging of FC), but
found that advanced biological age of several body systems explains not the converse. A positive feedback loop is evident between FC and
advanced brain age (Fig. 2d). While these aging pathways are not nec- white matter (Fig. 2f). Patterns of interorgan synchrony in biological
essarily causal in the strict sense, they reveal directional relationships age are shown in Extended Data Fig. 3 and SEM estimates are provided
elucidated through an established process of casual structure discovery in Supplementary Table 5.
(Methods). For example, cardiovascular age demonstrates the strong-
est influence on brain age, where a 1-year increase in cardiovascular Genetic, environmental and lifestyle links with organ age
age explains a 0.074-year (27-d) increase in overall brain age and 19-d We next investigated genetic and environmental factors associated
and 27-d increases in functional connectivity (FC) and white matter with organ age. Partial correlations were used to test for associations
ages, respectively. between organ-specific age gaps and 158 environmental/lifestyle
We next tested whether an organ’s baseline biological age influ- measures, leukocyte telomere length and polygenic scores indexing
ences the rate of change in the biological age of other organs. A positive leukocyte telomere length (Methods). Several environmental and
influence would provide evidence consistent with faster aging24. Due lifestyle factors explain significant variation in the biological age of
to minimal overlap between individuals (n = 17) with both longitudinal multiple organs (P < 2.6 × 10−5, Bonferroni-corrected for 158 factors × 12
body and brain phenotypes, analyses were conducted separately for organ systems = 1,896 tests; Supplementary Table 6). For most body
body and brain systems. The influence of one organ’s age gap on the systems (Fig. 3), individuals appearing older than same-aged peers
rate of change in the age gap of each other organ was modeled using were more likely to have smoked tobacco, consumed more alcohol and
SEM (Methods), yielding multiorgan aging networks (body, Fig. 2e, experienced long-standing illness, had menopause early in life and lived
and brain, Fig. 2f). The networks reveal several putative aging path- in areas of greater socioeconomic inequality. In contrast, those who
ways. For example, advanced age of the pulmonary system leads to exercised (faster than usual walking pace), had a larger birth weight,
faster cardiovascular aging, which in turn results in faster aging of completed tertiary education and were older at first live birth were
the musculoskeletal and renal systems (Fig. 2e). The cardiovascular– more likely to appear younger. Some lifestyle factors exclusively associ-
renal–metabolic–musculoskeletal systems form positive feedback ate with organ-specific age gaps. For example, advanced pulmonary
loops, where faster aging is reinforced between organ systems. The system age associates with exposure to air pollution, but not natural/
musculoskeletal system is an in-degree hub, suggesting that faster green environments. Advanced brain age most strongly associates

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1224

Article https://doi.org/10.1038/s41591-023-02296-6

with smoking, alcohol consumption, long-standing illness and hearing the cardiovascular (schizophrenia), hepatic (diabetes, hypertensive
loss. Notably, advanced cognitive age not only significantly associates diseases and osteoporosis) and metabolic (osteoporosis and Parkin-
with advanced brain age, but also with advanced age of several body sonism) systems (Fig. 4b). Disease comorbidity does not explain het-
systems, including pulmonary and musculoskeletal systems. Asso- erogeneity in organ-specific age across brain versus non-brain disease
ciations with tissue-specific brain age are shown in Supplementary categories (Extended Data Fig. 6).
Fig. 1. Shorter leukocyte telomere lengths weakly associate with older We tested whether diagnostic markers confound the interpreta-
body (r = −0.033, P = 2.5 × 10−34), pulmonary (r = −0.023, P = 1.1 × 10−17), tion of disease-related aging effects. The exclusion of diagnostic mark-
immune (r = −0.037, P = 2.4 × 10−42) and renal (r = −0.02, P = 7.3 × 10−14) ers for diabetes (HbA1c) and CKD (cystatin C and creatinine) from the
age gaps. Similarly, polygenic scores indexing leukocyte telomere metabolic and renal aging models, respectively, did lead to decreases
length weakly associate with cardiovascular (r = 0.014, P = 2.7 × 10−7), in prediction accuracy of chronological age; however, significantly
pulmonary (r = 0.016, P = 1.5 × 10−9), immune (r = −0.015, P < 2.0 × 10−8) advanced age of the metabolic and renal systems in diabetes and CKD
and renal (r = −0.012, P = 1.7 × 10−5) age gaps. remained evident after these exclusions (Methods).
Given that some hallmarks of biological aging are also patho-
Biological organ age and chronic disease logical features of aging-related diseases8, we hypothesized that
To investigate the relationship between biological age and chronic phenotypic variation related to chronological age would covary with
disease risk, individuals with a lifetime diagnosis of a chronic disease disease-related phenotypic variation. Consistent with this hypothesis,
were grouped into 16 disease categories: Parkinsonism, multiple scle- estimated feature weights of the predictive models of chronological
rosis, stroke, dementia, depression, bipolar disorder, schizophrenia, age (one weight per phenotype; number of phenotypes, body, n = 78;
ischemic heart disease, hypertensive diseases, chronic obstructive brain, n = 2,309) significantly associate with disease-related pheno-
pulmonary disease (COPD), chronic kidney disease (CKD), diabetes, typic variation (body, female/male, r = 0.43/0.39; brain, female/male,
cirrhosis, osteoarthritis, osteoporosis and cancer. Additional datasets r = 0.47/0.52, P < 0.0001; Fig. 5a and stratified by disease in Supple-
independent of the UK Biobank were sourced to establish mild cognitive mentary Fig. 2).
impairment (MCI) and validation dementia cohorts. Disease categories
were selected based on lifelong contribution to brain-associated illness Biological organ age relates to disease progression
burden (depression, bipolar disorder and schizophrenia) or substan- For each disease category, we divided individuals into prodromal and
tial health burden in older adults, including disability and premature established disease groups, based on the date of first diagnosis (if
mortality25. Using the preceding normative models established for known) and the date of baseline assessment of body and brain function.
healthy individuals, biological age was estimated for each body and Individuals who did not experience disease onset or diagnosis before
brain system and disease category. the time of baseline assessment were considered as prodromal. Several
Body and brain systems of individuals with chronic disease are organ systems of the prodromal groups are significantly older than
significantly older on average than same-aged healthy peers (body, same-aged healthy peers (Extended Data Fig. 7a), although mean age
0.71–6.15 years older; brain, 0.68–4.64 years older). Individuals with gaps are larger for individuals with established diagnoses, compared to
CKD have the oldest body ages (mean age gap of 6.15 ± 9.32 years) of all prodromal individuals (Extended Data Fig. 7b). Hence, advanced body
16 disease categories, whereas Parkinsonism associates with advanced age predates disease diagnosis. Furthermore, between-group differ-
body age the least (mean age gap of 0.71 ± 5.04 years). Marked hetero- ences in mean age gaps (prodromal versus established diagnoses) for
geneity in organ-specific ages is evident between and within diseases. several body systems and diseases are significantly larger in groups with
Figure 4 shows organs with mean age gaps significantly different from established diagnoses, compared to prodromal groups (P < 2.6 × 10−4,
zero in each disease group (P < 2.6 × 10−4, Bonferroni-corrected for 16 Bonferroni-corrected for 12 organs × 16 diseases = 192 tests), although
diseases × 12 organs = 192 tests). Organs primarily affected by disease effect sizes are modest (Fig. 5b). These effects are greatest in CKD (renal,
pathology generally show the largest ages gaps on average (Extended 10.8 ± 1.2 years older), cirrhosis (musculoskeletal, 3.1 ± 0.8 years) and
Data Fig. 4) and show the largest effect sizes (Extended Data Fig. 5). diabetes (metabolic, 2.9 ± 0.1 years), where body systems are signifi-
For example, renal, pulmonary, metabolic and hepatic systems are the cantly older in established compared to prodromal groups.
oldest in CKD (8.03 ± 11.73 years, Cohen’s d = 0.92), COPD (6.19 ± 6.03 The rate of change in age gaps significantly associates with
years, Cohen’s d = 1.26), diabetes (5.17 ± 7.34 years, Cohen’s d = 0.91) the average age gap over the two assessment time points in indi-
and cirrhosis (4.29 ± 10.21 years, Cohen’s d = 0.57), respectively. Nota- viduals with chronic disease (body, β = 0.21, P = 0.01; brain, β = 0.44,
bly, for many disease categories, organs not typically implicated with P = 4.4 × 10−6; Fig. 5c). This suggests disease-related faster body and
disease-specific processes also show evidence of advanced biological brain aging, where each 1-year increase in the mean body (brain) age
age. For example, whereas advanced brain age is evident for most major gap associates with a 0.21 (0.44) months per year increase in the rate
brain disorders, such as multiple sclerosis (4.64 ± 5.39 years, Cohen’s of body (brain) aging. On the contrary, the rate of aging is constant in
d = 1.06), dementia (3.52 ± 5.19 years, Cohen’s d = 0.75) and Parkinson- healthy individuals (body, β = 0.07, P = 0.74; brain: β = −0.08, P = 0.66).
ism (2.26 ± 4.54 years, Cohen’s d = 0.58), individuals with non-brain Faster aging is evident for multiple organ systems in ischemic heart
disorders such as diabetes (2.14 ± 3.60 years, Cohen’s d = 0.65), CKD disease, hypertensive diseases, diabetes, osteoarthritis and can-
(1.66 ± 3.56 years, Cohen’s d = 0.50) and COPD (1.65 ± 3.86 years, cer (P < 0.05, FDR-corrected across six disease groups × 12 organ
Cohen’s d = 0.48) also show significantly advanced brain age (Fig. 4 systems = 72 tests). Depression shows no evidence of faster aging
and Extended Data Fig. 4) with moderate effect sizes (Extended Data (P > 0.05; Fig. 5d).
Fig. 5). CKD-related advanced body and renal ages were replicated in
a subgroup of individuals (n = 2,168, body, 5.85 ± 8.84 years; renal, Biological organ age predicts mortality risk
7.62 ± 11.08 years) who had not progressed to the end-stage renal dis- We sought to predict risk of mortality using body and brain age gaps.
ease. Dementia-related advanced brain age was replicated using two Mortality was determined using data linkages to national death reg-
additional cohorts (n = 284, 3.19 ± 6.13 years, t = 16.94, P < 2.23 × 10−308). istries in the UK. Cancer (29.6%), circulatory (26.7%) and respiratory
Brain aging was less pronounced in MCI (n = 780, 1.07 ± 4.25 years) (11.8%) diseases were the three main causes of death (Supplemen-
than dementia (t = 10.39, P = 3.56 × 10−25), but significantly greater tary Fig. 3). Survival after baseline assessment was ascertained up
than same-aged healthy peers (t = 10.39, P = 3.56 × 10−25, Extended to 13.41 years for body (n = 8,109, age of death 42–83 years, mean
Data Fig. 2). Although rare, some body systems are marginally younger 69.6 ± 7.3, 5,670 males) and 6.07 years for brain (n = 330, age of
than their chronological age for specific disease categories, including death 53–82 years, mean 70.7 ± 6.4, 203 males). Body (2.95 ± 6.56

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1225

Article https://doi.org/10.1038/s41591-023-02296-6

a
Body Cardiac Pulmon. Muscle Immune Renal Hepatic Metab. Brain GM WM FC

HC
Parkinsonism
Osteoarthritis
Osteoporosis
Cancer
Multiple sclerosis
Depression
Hypertensive diseases
Ischemic heart diseases
Bipolar
Stroke
Dementia
Schizophrenia
Diabetes
COPD
Cirrhosis
CKD

–1 4.5 8

b Mean age gap

Parkinsonism

Osteoarthritis

Osteoporosis

Cancer

Multiple sclerosis

Depression

Hypertensive diseases

Ischemic heart diseases

Bipolar

Stroke

Dementia

Schizophrenia

Diabetes

COPD

Cirrhosis

CKD

–1 0 1 2 3 4 5+
Years

Fig. 4 | Body and brain age in chronic disease. a, Distribution of body and mean that is not significantly different from the healthy group. The three axis
brain age gaps (columns) for 16 disease categories (rows), compared to healthy ticks on the horizontal axis from left to right for each distribution correspond
individuals (HC). Distributions are colored according to disease- and organ- to age gaps of −5, 0 and 5 years. b, Icons representing body systems and organs
specific mean age gaps. Colored distributions have a mean that significantly are positioned to indicate the mean age gap for each disease category. Icons are
differs from the healthy group (P < 2.6 × 10−4, two-sided, t-test, Bonferroni- not shown for organs with age gaps that do not significantly differ from zero.
corrected). Despite significant between-group differences, considerable overlap Organs with age gaps exceeding 5 years are truncated to 5 years for visualization
in distributions between disease categories and healthy individuals suggests that purposes. Disease categories are ordered from top to bottom according to
factors other than diagnostic status manifest notable heterogeneity of biological increasing body age gap.
age among individuals in the same category. Distributions colored gray have a

versus 0.57 ± 4.44 years, P < 0.0056, Bonferroni-corrected for nine between deceased and surviving individuals. Between-group differ-
body ages) and brain (1.72 ± 4.07 versus 0.48 ± 3.31 years, P < 0.0125, ences are also evident for specific body (Fig. 6a) and brain systems
Bonferroni-corrected for four brain ages) age gaps significantly differ (Supplementary Fig. 4a).

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1226

Article https://doi.org/10.1038/s41591-023-02296-6

a Aging versus disease b Established versus prodromal diagnosis

Females Males
20 20
Osteoporosis
Regression
coefficient

10 10
(aging)

Hypertensive diseases Body Cardiac Pulmon. Muscle

0 0
Ischemic heart disease
–10 r = 0.43 –10 r = 0.39
–5 –5
P = 7.21 × 10 P = 3.81 × 10 Stroke
–20 –20 Immune Renal Hepatic Metab.
–50 –25 0 25 50 –50 –25 0 25 50
Cancer
10 10
Osteoarthritis
5 5
Regression
coefficient

Diabetes
(aging)

0 0
Cirrhosis
–5 –5
r = 0.47 r = 0.52
P < 2.23 × 10
–308
P < 2.23 × 10
–308
CKD
–10 –10
–20 –10 0 10 20 –20 –10 0 10 20
–2 0 2 4 6 8 10 12
t-statistic t-statistic
(disease effects) (disease effects) Established – prodromal (years)

c d Sample size
(body : brain)
* ** * *
Ischemic heart disease 106 : 68
5 5
* *
Rate of change in gap

Hypertensive diseases 457 : 387

(months per year)

Diabetes 66 : 69
0 0
*
Osteoarthritis 245 : 210
Other* Other*
HC HC Depression 114 : 186
–5
–5 0 5
–5
–5 0 5 *
Gap (years) Gap (years) Cancer 179 : 127

5 10 15 20 25 30 35
Faster aging (days / year2)

Fig. 5 | Associations between aging and disease effects and progression. with bootstrapping (n = 1,000). c, Evidence of faster body (left) and brain (right)
a, Scatter-plots show associations between regression coefficients (feature aging in chronic disease. Lines of best fit show associations between the rate of
weights) of the age prediction models (y axis) and between-group differences change in age gap (y axis) and the average age gap across the two visits (x axis).
(disease versus health, t-statistics, two-sided) in body (n = 78, top) and brain Associations were tested separately in HCs (green) and individuals with chronic
(n = 2,309, bottom) features (x axis). The t-statistic quantifies disease-related diseases (other, orange). Asterisks indicate significant associations (body,
variation for each phenotype. Each data point represents a phenotype. Lines P = 0.01; brain, P = 4.4 × 10−6, two-sided, t-test, Bonferroni-corrected). d, Same
of best fit are indicated with solid black lines. r indicates Pearson correlation as c but stratified according to disease category and organ. Icons are positioned
coefficient. Supplementary Fig. 2 shows stratification by disease category. to indicate faster aging, quantified by the slope of the line of best fit between
b, Distributions show the difference in age gaps between individuals with rate of change in age gap and average age gap across the two visits. Associations
established diagnoses at the time of baseline assessment and prodromal were only tested for disease categories and/or organs consisting of more than 50
individuals who were first diagnosed after assessment, both normalized to age- individuals. Icons are shown for organs with a nominally significant association
matched peers. Distributions only shown for organs and disease categories with (P < 0.05, two-sided, t-test, uncorrected), with high icon opacity and asterisks
a significant difference in age gap (P < 3.9 × 10−4, two-sided, t-test, Bonferroni- indicating associations surviving FDR correction of 5%.
corrected). Distributions of the mean differences in age gap were estimated

Cox proportional hazards regression, where survival durations in the risk of mortality (area under the curve (AUC) = 0.75, loglikeli-
were right-censored for living individuals (body, n = 135,314; brain, hood = −9.19 ± 104; Fig. 6b). This model significantly outperformed a
n = 36,571), reveal that body age (Fig. 6b,c and Supplementary Table 9), baseline mortality model including only chronological age and sex
but not brain age (Supplementary Fig. 4b,c and Supplementary Table 10) (AUC = 0.72, loglikelihood = −9.29 ± 104, χ2 = 1.86 ± 103, P < 2.23 × 10−308).
is a significant predictor of mortality. In particular, adjusting for Several body systems (pulmonary, immune, renal and hepatic) remain
chronological age and sex, each 1 × s.d. increase in a person’s organ age significant mortality predictors, when controlling for existing disease
associates with a 7.3% (body, hazard ratio (HR) 1.073, 95% confidence diagnoses (Supplementary Fig. 5). A regression model including chron-
interval (CI) 1.037–1.136, P = 2.2 × 10−6), 3.6% (cardiovascular, HR 1.036, ological age, sex, all eight body age gaps, existing disease diagnoses,
95% CI 1.014–1.056, P = 2.7 × 10−3), 24.0% (pulmonary, HR 1.24, 95% CI general health (long-standing illness and disability) and key environ-
1.210–1.262, P = 2.6 × 10−89), 5.9% (immune, HR 1.059, 95% CI 1.045–1.078, mental/lifestyle factors such as smoking, exercise, tertiary education
P = 1.0 × 10−31), 15.1% (renal, HR 1.151, 95% CI 1.116–1.183, P = 3.0 × 10−24), and socioeconomic inequality yielded the most accurate (AUC = 0.771)
7.9% (hepatic, HR 1.079, 95% CI 1.051–1.104, P = 1.9 × 10−10) and 7.1% and best fitting model of mortality risk (loglikelihood = −8.97 ± 104,
(metabolic, HR 1.071, 95% CI 1.043–1.091, P = 3.2 × 10−9) relative increase χ2 = 4.98 ± 103, P < 2.23 × 10−308; Fig. 6c). After controlling for the

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1227

Article https://doi.org/10.1038/s41591-023-02296-6

a b 1.30 60
Deceased Surviving
1.25
10 50
Body age gap (years)

1.20 AUC = 0.75

HR (per s.d.)
1.15 40
5

z score
1.10 30
0
1.05
20
–5 1.00
0.95 10
–10
0.90 0
Body

Cardiovascular

Pulmonary

Musculoskeletal

Immune

Renal

Hepatic

Metabolic

Body
Cardiovascular
Pulmonary
Musculoskeletal
Immune
Renal
Hepatic
Metabolic
Age
Sex
Body
Cardiovascular
Pulmonary
Musculoskeletal
Immune
Renal
Hepatic
Metabolic
c 1.30 50
1.25
40
1.20 AUC = 0.77
HR (per 1 x s.d.)

30
1.15
z score

1.10 20
1.05
10
1.00
0
0.95
0.90 –10

Age
Sex
Ischemic heart disease
Hypertensive disease
Stroke
COPD
CKD
Diabetes
Cirrhosis
Osteoarthritis
Osteoporosis
Dementia
Parkinsonism
Multiple sclerosis
Schizophrenia
Depression
Bipolar
Cancer
Current tobacco smoking
Usual walking pace
Daily smoker
Long-standing illness
College/university qualification
Townsend deprivation
Body
Cardiovascular
Pulmonary
Musculoskeletal
Immune
Renal
Hepatic
Metabolic
Body
Cardiovascular
Pulmonary
Musculoskeletal
Immune
Renal
Hepatic
Metabolic

Fig. 6 | Body age and the risk of mortality. a, Body age gaps in deceased indicate 95% CI estimated with bootstrapping (n = 100). Center line indicates the
(n = 8,109) compared to surviving (n = 135,314) individuals. Asterisks indicate mean. Colored bars indicate organs with significant HRs (P < 0.005, two-sided,
significant between-group differences, controlling for chronological age and sex z-test, Bonferroni-corrected for ten dependent variables). c, Same as b, but
(P < 0.0056, two-sided, t-test, Bonferroni-corrected). b, Bar plots show mortality existing disease diagnoses, general health and key lifestyle factors are included
HRs per 1 × s.d. increase in organ-specific age (left) and corresponding z scores in the regression. Colored bars indicate organs with significant HRs (P < 0.0015,
(right). Chronological age and sex are included in the regression. Error bars two-sided, z-test, Bonferroni-corrected for 32 dependent variables).

above covariates, the composite body age gap outperformed all predictors for all three causes of mortality (Supplementary Table 13).
organ-specific age gaps in explaining mortality hazard, suggesting Finally, logistic models were developed to predict survival time (5 years,
that these factors are significant confounds. Replacing body age gaps AUC = 0.774 ± 0.006; 10 years, AUC = 0.770 ± 0.003) and premature
with body phenotypes associated with mortality, including systolic mortality (death before 70 years old, AUC = 0.86 ± 0.003; 75 years
blood pressure26, forced expiratory volume in 1 s (FEV1)27, handgrip old, AUC = 0.86 ± 0.003) using body age gaps (Extended Data Fig. 8).
strength28, C-reactive protein29, serum creatinine30, serum alanine and
aspartate aminotransferase31 and the total/high-density lipoprotein Discussion
cholesterol ratio32 did not improve the model accuracy (AUC = 0.770) By establishing normative models of aging-related decline for mul-
and fit (loglikelihood = −8.98 × 104; Supplementary Table 11). Similarly, tiple brain and body systems in the world’s largest population-based
replacing brain age gaps with several global brain measures, including biobank, we showed that aging is a complex, multisystem process,
whole brain volume of gray matter, cerebrospinal fluid, white matter, whereby the biological age of one organ system selectively influ-
white matter hyperintensity load and mean cortical thickness, mean ences the aging of multiple other systems via characteristic aging
fractional anisotropy and mean diffusivity did not improve the model pathways. While biological aging is an established concept12–14 and
(AUC = 0.722, loglikelihood = −3.14 × 103; Supplementary Table 12). earlier studies establish aging clocks for individual organs, includ-
Mortality risk associated with body (Supplementary Fig. 6) and brain ing the kidneys34, heart35, lungs36, skin and blood37, we derived the
(Supplementary Fig. 7) age remained largely unchanged after exclud- first whole-body multiorgan characterization of aging. Our organ
ing deaths subsequent to the date of coronavirus disease (COVID-19) clocks enabled elucidation of unique organ age profiles for 16 chronic
emergence in the UK33 (n = 1,033 for body age and n = 127 for brain age diseases and discovery of modifiable factors that can potentially lead
analyses), suggesting that COVID-19 did not significantly confound our to disease-specific longevity interventions targeted at specific body
mortality risk estimation models. Further analyses of mortality due to systems, ultimately extending lifespan.
specific disease cause, including cancer (AUC = 0.75; Supplementary Our work enhances the clinical utility of proxy measures of
Fig. 8a), circulatory diseases (AUC = 0.84; Supplementary Fig. 8b) and aging developed for older individuals, such as frailty indices38, as
respiratory diseases (AUC = 0.86; Supplementary Fig. 8c) reveal similar well as existing DNA methylation (epigenetic) clocks39,40. While epi-
results, where pulmonary, immune and renal age remain significant genetic clocks are clinically useful and provide important insights

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1228

Article https://doi.org/10.1038/s41591-023-02296-6

into aging biology across tissue types, it is now recognized that aging Chronological age and male sex were found to be the two strong-
varies markedly between organ systems and tissues, particularly in est mortality risk factors, consistent with previous literature43,44. After
disease41. Bespoke organ and disease-specific aging clocks are thus controlling for these two factors, organ ages remained strong mortality
needed to enhance the clinical utility of existing pan-tissue clocks, risk factors, particularly pulmonary age, followed by ages of the renal,
which do not readily differentiate between tissue components and hepatic, metabolic, immune and cardiovascular systems. Individuals
body systems15. Addressing this need, we showed that deviations who subsequently died had older-appearing brains compared to those
from expected aging-related decline can be detected in certain who survived, consistent with a previous study examining brain age
organs (but not all) years before disease diagnosis. These devia- and mortality45; however, advanced brain age did not predict increased
tions predict mortality, even after controlling for chronological age, risk of mortality. This may be due to the relatively low mortality rate in
disease burden and other risk factors. Our organ clocks could thus individuals with brain age estimates in the UK Biobank (mortality rate,
be used to identify individuals in midlife, before disease onset, who 330 of 36,901 = 0.0089) compared to the Lothian Birth Cohort 1936
may benefit from early interventions aimed at slowing the aging of used by Cole and colleagues (mortality rate, 73 of 669 = 0.11). Contin-
specific body systems and organs. ued follow-up of UK Biobank participants will likely yield more insight
Crucially, as with the frailty indices38, many of the biological mark- into the relationship between brain aging and mortality.
ers that inform our organ clocks are already widely assayed in primary Advanced pulmonary age was the strongest predictor of mortality
care (for example, full blood counts, renal and liver function, blood (HR 1.24), consistent with epidemiological observations of associations
pressure, lipids and glucose), are readily accessible at minimal cost between impaired lung function and increased risk of mortality27. While
(forced respiration, grip strength and waist circumference) or are reduced muscle strength, measured by handgrip strength, is commonly
accessible and cost-effective when benchmarked against the burden associated with increased risk of mortality46, older musculoskeletal
of chronic illness (brain magnetic resonance imaging (MRI) scans). age was not a significant risk factor of mortality when controlling for
Alongside the relatively modest computational burden of the model chronological age, sex and the age gaps of other organs. This is con-
algorithms (especially when pretrained), these considerations argue sistent with the configuration of the multiorgan aging network, where
for direct, cost-effective and feasible clinical implementation of organ the musculoskeletal system is a central hub, influenced by the extent
age in primary care. of aging of most other organ systems. The mortality risk explained
Our investigation into environmental and lifestyle factors can by musculoskeletal aging may thus be attributable to the biological
inform real-world personalized interventions targeted at specific age of other body systems. Advanced age of the pulmonary, immune,
body systems, through change of lifestyle, such as limiting tobacco renal and hepatic systems significantly raises a person’s mortality risk,
smoking and alcohol intake, exercise, education, sleep hygiene and beyond that explained by existing chronic diseases, chronological age
maternal nutrition, as well as efforts requiring national inputs such as and sex (Fig. 6). Whereas disease conditions that primarily affect these
reductions in socioeconomic inequality and air pollution, and improve- organ systems are common causes of death25, our results demonstrate
ments in residential greenspace and natural environment coverage. the uniqueness of biological age in explaining all-cause mortality,
Studying the impact of such interventions would provide causal evi- regardless of existing diseases.
dence for the conditional effects presently reported. The pulmonary, The divergence of an organ’s biological age from chronological age
metabolic and immune systems are promising organ-specific targets may emerge early in life and widen over the lifespan, increasing the risk
for interventions, given that these systems influence the rate of aging of chronic disease and mortality; however, the rate of aging reported
of multiple body systems (cardiovascular and musculoskeletal), via here more specifically reflects aging-related decline rather than early
interorgan aging pathways (Fig. 2e). Notably, the aging pathway linking life events. Interventions designed to delay the rate of organ aging may
pulmonary, cardiovascular and musculoskeletal systems recapitu- thus effectively delay disease onset, resulting in an extended healthy
lates the known epidemiological link between impaired lung function, lifespan. Further study is needed to determine whether interventions
weak muscle strength and elevated risk of adverse cardiovascular out- informed by the observational evidence reported here can reduce these
comes27,28. While aging-related brain gray matter loss is expected42, we risks and potentially slow organ aging in at-risk individuals. Further
found that advanced gray matter age substantially influences the rate work is also needed to determine the genetic influences on our organ
of aging-related decline in brain connectivity, but not the converse (Fig. clocks. We showed that leukocyte telomere lengths and genetic vari-
2f). Given that body phenotypes were measured several years before ants known to index leukocyte telomere length weakly associate with
brain phenotypes (4–14 years earlier; Fig. 2c), the estimated influence of several body ages. This complements a recent study13 showing the
advanced body age, particularly of the cardiovascular system on brain importance of the immune system (major histocompatibility complex
age (Fig. 2d), may thus reveal early signs of brain aging. on chromosome 6) and DNA repair pathways in aging.
While the organs that manifest primary disease processes seem the Our work addresses several recently identified challenges hinder-
oldest in individuals with the disease, we found that advanced organ ing the clinical translation of biological aging research15,19. We estab-
age is widespread, involving multiple body and brain systems. Brain lished bespoke clocks that measure the biological age of specific brain
systems of individuals with non-brain disorders, including diabetes, and body systems using markers that are routinely assayed in primary
chronic kidney, pulmonary and cardiovascular diseases seem sig- care, elucidated organ age profiles for prevalent chronic diseases
nificantly older than same-aged healthy peers, whereas body systems, and identified modifiable factors that can inform new strategies to
particularly pulmonary and renal systems, show signs of advanced potentially limit organ-specific aging. Our work maps a multiorgan
aging in individuals primarily diagnosed with major brain disorders, aging network for the human body.
including schizophrenia, dementia, bipolar disorder, depression, Several caveats pertain to our findings. First, biological aging is
multiple sclerosis and Parkinsonism. It is important to acknowledge multifaceted. As such, it is unlikely that a single index of organ aging
that some of our clocks are informed by diagnostic markers, which will be sufficient and conclusive. As with the continued refinement of
may potentially confound the interpretation of disease-related differ- epigenetic clocks over the last decade39,47–49, and ongoing deliberations
ences in biological age; however, confounding can be ruled out for the about how to define frailty in older people38,50, standardized measures
above examples where advanced age is evident for organs that are not of organ age remain to be developed. Second, body phenotypes were
informed by diagnostic markers relevant to the disease under consid- measured several years before brain phenotypes in the UK Biobank.
eration. Furthermore, excluding key diagnostic markers reduced model Due to the sequential and non-randomized participant assessment
performance but did not alter our conclusions about the association schedule, we were unable to assess the influence of brain aging on body
between disease and organ aging. systems. Future investigations, leveraging multiple cohort waves, may

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1229

Article https://doi.org/10.1038/s41591-023-02296-6

reveal bidirectional brain-body influences. Third, most participants 20. Kaufmann, T. et al. Common brain disorders are associated with
enrolled in the UK Biobank are from a white ethnic background. Inclu- heritable patterns of apparent aging of the brain. Nat. Neurosci.
sion of participants from a diversity of ethnicities, demographic and 22, 1617–1623 (2019).
socioeconomic backgrounds will be required to assess the generaliz- 21. Cole, J. H. & Franke, K. Predicting age using neuroimaging:
ability of our current finding. Finally, some imaging modalities (for innovative brain ageing biomarkers. Trends Neurosci. 40,
example, carotid imaging) were only acquired in select individuals, 681–690 (2017).
limiting the data available for some organ clocks. Clinical translation 22. Bashyam, V. M. et al. MRI signatures of brain age and disease over
could proceed by adding these to clinical assays, and/or removing the lifespan based on a deep brain network and 14468 individuals
others based upon the tradeoff between their added predictive value worldwide. Brain 143, 2312–2324 (2020).
versus their cost. 23. Bartsch, R. P., Liu, K. K., Bashan, A. & Ivanov, P. Network
physiology: how organ systems dynamically interact. PLoS ONE
Online content 10, e0142143 (2015).
Any methods, additional references, Nature Portfolio reporting sum- 24. Vidal-Piñeiro, D. et al. Individual variations in ‘brain age’ relate to
maries, source data, extended data, supplementary information, early-life factors more than to longitudinal brain change. eLife 10,
acknowledgements, peer review information; details of author contri- e69995 (2021).
butions and competing interests; and statements of data and code avail- 25. Vos, T. et al. Global burden of 369 diseases and injuries in 204
ability are available at https://doi.org/10.1038/s41591-023-02296-6. countries and territories, 1990–2019: a systematic analysis for
the Global Burden of Disease Study 2019. Lancet 396, 1204–1222
References (2020).
1. Niccoli, T. & Partridge, L. Ageing as a risk factor for disease. Curr. 26. Port, S., Demer, L., Jennrich, R., Walter, D. & Garfinkel, A. Systolic
Biol. 22, R741–R752 (2012). blood pressure and mortality. Lancet 355, 175–180 (2000).
2. Hou, Y. et al. Ageing as a risk factor for neurodegenerative 27. Duong, M. et al. Mortality and cardiovascular and respiratory
disease. Nat. Rev. Neurol. 15, 565–581 (2019). morbidity in individuals with impaired FEV1(PURE): an
3. Elliott, M. L. et al. Disparities in the pace of biological aging international, community-based cohort study. Lancet Glob. Health
among midlife adults of the same chronological age have 7, e613–e623 (2019).
implications for future frailty risk and policy. Nat. Aging 1, 28. Celis-Morales, C. A. et al. Associations of grip strength with
295–308 (2021). cardiovascular, respiratory, and cancer outcomes and all cause
4. Tuttle, C. S. L. et al. Cellular senescence and chronological age mortality: prospective cohort study of half a million UK Biobank
in various human tissues: a systematic review and meta-analysis. participants. BMJ 361, k1651 (2018).
Aging Cell 19, e13083 (2020). 29. Zacho, J., Tybjaerg-Hansen, A. & Nordestgaard, B. G. C-reactive
5. Khan, S. S., Singer, B. D. & Vaughan, D. E. Molecular and protein and all-cause mortality—the Copenhagen City Heart
physiological manifestations and measurement of aging in Study. Eur. Heart J. 31, 1624–1632 (2010).
humans. Aging Cell 16, 624–633 (2017). 30. Newsome, B. B. et al. Long-term risk of mortality and end-stage
6. Ferrucci, L. et al. Measuring biological aging in humans: a quest. renal disease among the elderly after small increases in serum
Aging Cell 19, e13080 (2020). creatinine level during hospitalization for acute myocardial
7. Xia, X., Wang, Y., Yu, Z., Chen, J. & Han, J.-D. J. Assessing the rate infarction. Arch. Intern. Med. 168, 609–616 (2008).
of aging to monitor aging itself. Ageing Res. Rev. 69, 101350 31. Lee, T. H., Kim, W. R., Benson, J. T., Therneau, T. M. &
(2021). Melton, L. J. 3rd Serum aminotransferase activity and
8. López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, mortality risk in a United States community. Hepatology 47,
G. The hallmarks of aging. Cell 153, 1194–1217 (2013). 880–887 (2008).
9. Lu, Y. et al. Reprogramming to recover youthful epigenetic 32. Lewington, S. et al. Blood cholesterol and vascular mortality by
information and restore vision. Nature 588, 124–129 (2020). age, sex, and blood pressure: a meta-analysis of individual data
10. Palliyaguru, D. L. et al. Study of longitudinal aging in mice: from 61 prospective studies with 55,000 vascular deaths. Lancet
presentation of experimental techniques. J. Gerontol. A 76, 370, 1829–1839 (2007).
552–560 (2020). 33. Mahase, E. Covid-19: UK records first death, as world’s cases
11. Schaum, N. et al. Ageing hallmarks exhibit organ-specific exceed 100000. BMJ 368, m943 (2020).
temporal signatures. Nature 583, 596–602 (2020). 34. Belur Nagaraj, S., Kieneker, L. M. & Pena, M. J. Kidney Age Index
12. Ahadi, S. et al. Personal aging markers and ageotypes (KAI): a novel age-related biomarker to estimate kidney function
revealed by deep longitudinal profiling. Nat. Med. 26, in patients with diabetic kidney disease using machine learning.
83–90 (2020). Comput. Methods Programs Biomed. 211, 106434 (2021).
13. Nie, C. et al. Distinct biological ages of organs and systems 35. Wells, S., Kerr, A., Eadie, S., Wiltshire, C. & Jackson, R. ‘Your Heart
identified from a multi-omics study. Cell Rep. 38, 110459 (2022). Forecast’: a new approach for describing and communicating
14. Belsky, D. W. et al. Quantification of biological aging in young cardiovascular risk? Heart 96, 708–713 (2010).
adults. Proc. Natl Acad. Sci. USA 112, E4104–E4110 (2015). 36. Morris, J. F. & Temple, W. Spirometric “lung age” estimation for
15. Bell, C. G. et al. DNA methylation aging clocks: challenges and motivating smoking cessation. Prev. Med. 14, 655–662 (1985).
recommendations. Genome Biol. 20, 249 (2019). 37. Horvath, S. et al. Epigenetic clock for skin and blood cells applied
16. Kaeberlein, M., Rabinovitch, P. S. & Martin, G. M. Healthy aging: to Hutchinson Gilford Progeria Syndrome and ex vivo studies.
the ultimate preventative medicine. Science 350, 1191–1193 Aging 10, 1758–1775 (2018).
(2015). 38. Hoogendijk, E. O. et al. Frailty: implications for clinical practice
17. Kennedy, B. K. et al. Geroscience: linking aging to chronic and public health. Lancet 394, 1365–1375 (2019).
disease. Cell 159, 709–713 (2014). 39. Horvath, S. DNA methylation age of human tissues and cell types.
18. Kaeberlein, M. Longevity and aging. F1000Prime Rep. 5, 5 (2013). Genome Biol. 14, 3156 (2013).
19. Cole, J. H., Marioni, R. E., Harris, S. E. & Deary, I. J. Brain age 40. Hannum, G. et al. Genome-wide methylation profiles reveal
and other bodily ‘ages’: implications for neuropsychiatry. Mol. quantitative views of human aging rates. Mol. Cell 49, 359–367
Psychiatry 24, 266–281 (2019). (2013).

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1230

Article https://doi.org/10.1038/s41591-023-02296-6

41. Slieker, R. C., Relton, C. L., Gaunt, T. R., Slagboom, P. E. & 48. Levine, M. E. et al. An epigenetic biomarker of aging for lifespan
Heijmans, B. T. Age-related DNA methylation changes are and healthspan. Aging 10, 573–591 (2018).
tissue-specific with ELOVL2 promoter methylation as exception. 49. Belsky, D. W. et al. DunedinPACE, a DNA methylation biomarker of
Epigenetics Chromatin 11, 25 (2018). the pace of aging. eLife 11, e73420 (2022).
42. Resnick, S. M., Pham, D. L., Kraut, M. A., Zonderman, A. B. & 50. Markle-Reid, M. & Browne, G. Conceptualizations
Davatzikos, C. Longitudinal magnetic resonance imaging studies of frailty in relation to older adults. J. Adv. Nurs. 44,
of older adults: a shrinking brain. J. Neurosci. 23, 3295–3301 58–68 (2003).
(2003).
43. Austad, S.N. in Handbook of the Biology of Aging 7th edn (eds Publisher’s note Springer Nature remains neutral with regard to
Masoro, E.J. & Austad, S.N.) Ch. 23 (Academic Press, 2011). jurisdictional claims in published maps and institutional affiliations.
44. Makeham, W. M. On the law of mortality and the construction of
annuity tables. J. Inst. Actuar. 8, 301–310 (1860). Springer Nature or its licensor (e.g. a society or other partner) holds
45. Cole, J. H. et al. Brain age predicts mortality. Mol. Psychiatry 23, exclusive rights to this article under a publishing agreement with
1385–1392 (2018). the author(s) or other rightsholder(s); author self-archiving of the
46. Leong, D. P. et al. Prognostic value of grip strength: findings from accepted manuscript version of this article is solely governed by the
the Prospective Urban Rural Epidemiology (PURE) study. Lancet terms of such publishing agreement and applicable law.
386, 266–273 (2015).
47. Lu, A. T. et al. DNA methylation GrimAge strongly predicts lifespan © The Author(s), under exclusive licence to Springer Nature America,
and healthspan. Aging 11, 303–327 (2019). Inc. 2023

Nature Medicine | Volume 29 | May 2023 | 1221–1231 1231

Article https://doi.org/10.1038/s41591-023-02296-6

Methods Follow-up data were available in 1,220 individuals (age range 44–75
Participants years, mean 61.6 ± 7.7 at the second visit, 837 males) at 2.1–5.6 years
Individuals (n = 502,504; 229,122 males) participating in the UK follow-up. Phenotypes were grouped based on relevance to the struc-
Biobank51 were analyzed for the primary study (project ID 60698). ture and function of each organ systems, forming seven phenotype
They were aged 37–73 years at the time of recruitment (2006–2010) groups. A predictive model of chronological age was established using
and underwent extensive questionaries, physical assessments, blood all phenotypes comprising a given organ-specific phenotype group
and urine sample assays and genome-wide genotyping at 22 assess- (see below). Additionally, a whole-body predictive model was estab-
ment centers throughout the UK. A subset of individuals (n = 20,345; lished using all body phenotypes, irrespective of organ grouping. Key
9,938 males) was followed up during 2012–2013 for repeated physi- measures used to assess individual organ function are as follows (also
cal and physiological assessments. Multimodal brain imaging52 was see Fig. 1a):
acquired during the third visit (2014–2020) at three mirrored imaging • Cardiovascular system: pulse rate, systolic blood pressure and
centers located at Manchester, Reading and Newcastle, respectively, diastolic blood pressure.
in 49,002 individuals (23,710 males). Follow-up brain imaging was • Pulmonary system: FVC, FEV1, PEF and FEV1:FVC ratio.
conducted from 2019 onwards in 1,503 individuals (754 males), provid- • Musculoskeletal system: handgrip strength, standing height,
ing a longitudinal sample enabling estimation of the rate of change in weight, BMI, waist and hip circumference, waist:hip circumfer-
biological age. An assessment timeline is shown in Fig. 2c. Each step in ence ratio, heal bone mineral density, ankle spacing width,
the assessment and processing of biological samples was handled and blood biochemical markers such as phosphatase, calcium,
monitored centrally to minimize biases across recruitment centers. phosphate and vitamin D.
We found that biological age (age gaps) showed negligible site-related • Immune system: C-reactive protein and blood hematology tests
(Supplementary Fig. 9), ethnicity-related (Supplementary Fig. 10) of leukocytes, erythrocytes, thrombocytes and hemoglobin.
and longitudinal subsampling-related (Supplementary Fig. 11) varia- • Renal system: biomarkers associated with glomerular filtration
tion. The UK Biobank has approval from the North West Multi-centre and electrolyte regulation, including creatinine (enzymatic),
Research Ethics Committee to obtain and disseminate data and samples potassium and sodium in urine, albumin, urea, urate, creatinine,
from the participants (http://www.ukbiobank.ac.uk/ethics/). Written cystatin C, phosphate and total protein in blood.
informed consent was obtained from all participants. Details of partici- • Hepatic system: alanine aminotransferase, aspartate ami-
pants consisting of the independent validation cohorts are described notransferase, γ-glutamyl transferase, direct and total bilirubin,
below (see Replication of brain age prediction in additional datasets). albumin, alkaline phosphatase and total protein in blood.
• Metabolic system: blood biomarkers associated with lipids and
Body age phenotypes glucose metabolism, including apolipoprotein A, apolipopro-
Physical and physiological measures known to index the function, tein B, cholesterol, glucose, glycated hemoglobin, high-density
structure and/or general health of the cardiovascular, pulmonary, lipoprotein cholesterol, direct low-density lipoprotein choles-
musculoskeletal, immune, renal, hepatic and metabolic systems were terol, lipoprotein A and triglycerides.
selected, resulting in 101 organ-specific phenotypes. Physical meas-
ures included standing height, weight, body mass index (BMI), hip Several blood biomarkers, including insulin-like growth factor
and waist circumferences, handgrip strength, ultrasound heel bone 1, testosterone and sex hormone-binding globulin were not assigned
densitometry, spirometry and cardiorespiratory fitness. Physiologi- to any of the seven systems and were only used for overall body age
cal assessments included blood pressure, pulse rate, arterial stiffness, gap estimation. Post-hoc analysis was performed to investigate
blood hematology and blood and urine biochemical assays. Further the potential confounding effect of antihypertensive medications
steps included: (angiotensin-converting enzyme inhibitors, angiotensin receptor
1. Averaging measures if tested for left and/or right side of the blockers, beta-blockers, calcium channel blockers and thiazide diu-
body, such as handgrip strength, heel bone mineral density and retic agents) on the estimation of cardiovascular age. This grouping of
ankle spacing width. medication categories is consistent with a previous UK Biobank study53.
2. Averaging measures if tested more than once at the same visit, Adjusting for chronological age and sex, we found no significant differ-
such as diastolic and systolic blood pressure and pulse rate ence in the estimated cardiovascular age between individuals who regu-
were each measured twice. larly take antihypertensive medications (mean age gap = 0.32 ± 3.58
3. Selecting the best performance among multiple repeated tests years) and individuals who do not take any antihypertensive medica-
at the same visit, such as spirometry test for lung function, tions (mean age gap = 0.46 ± 3.94 years, t = 1.04, P = 0.29), suggesting
including forced vital capacity (FVC), FEV1 and peak expira- that antihypertensive medications are not significant confounds.
tory flow (PEF). Each participant was asked to conduct up to
three blows (lasting for at least 6 s) within a period of approxi- Brain age phenotypes
mately 6 min. The quality of each blow result was automatically Structural and functional brain phenotypes (n = 2,309; Supplemen-
detected by the device and only the best performing blow of tary Table 2) derived from three neuroimaging modalities, including
acceptable quality was selected. The FEV1:FVC ratio was also T1-weighted MRI, diffusion MRI (dMRI) and resting-state functional MRI
computed and used for body age estimation. (fMRI) were sourced from the UK Biobank52,54. The image processing
4. Excluding measures with missing responses in more than 30% pipeline, artifact removal, cross-modality and cross-individual image
of individuals. As such, measures of cardiorespiratory fitness alignment, quality control and phenotype calculation are described in
(missing proportion, 85%), arterial stiffness (69%), urine micro- detail in the central UK Biobank brain imaging documentation (https://
albumin (76%), blood estradiol (79%) and rheumatoid factor biobank.ctsu.ox.ac.uk/showcase/showcase/docs/brain_mri.pdf) and
(93%) were excluded. by Alfaro-Almagro and colleagues54. Participants with missing entries
for any of the 2,309 phenotypes were discarded, resulting in a final
This resulted in 78 body phenotypes for chronological age pre- sample consisting of 36,901 individuals (age range 45–82 years, mean
diction (Supplementary Table 1). Participants with missing responses 64.2 ± 7.5 at the first imaging visit, 17,203 males) for brain age analyses.
for any of the 78 phenotypes were then excluded, resulting in a final Repeated brain imaging phenotypes were available in 1,294 individuals
sample consisting of 143,423 individuals (age range 39–73 years, mean (age range 50–83 years, mean 65.2 ± 7.2 at the second imaging visit, 632
56.7 ± 8.2 at the baseline assessment of body function, 79,980 males). males) at 2.0–2.7 years follow-up.

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Predictive models of chronological age were established using of the interquartile range of the predicted variable (chronological age).
imaging-derived phenotypes (IDPs) pertaining to gray matter struc- The SVM was solved using sequential minimal optimization, using a gap
ture, white matter microstructure and brain FC. Additionally, a tolerance of 0.001. More specifically, linear SVM regression involved
whole-brain predictive model was established using all brain pheno- fitting the linear function
types (n = 2,309), irrespective of brain tissue class. IDPs used to assess
f (x) = xβ + b
individual brain systems are as follows:
• Gray matter: regional gray matter volume, cortical thickness and
surface area, as derived from T1-weighted MRI (number of IDPs, for each organ system, where x is the matrix of organ-specific pheno-
578). types (subjects × phenotypes), β is the fitted model coefficients and
• White matter: dMRI-derived microstructural measures of white b is the model offset. To estimate β and b for each organ system, the
matter tracts including mean fractional anisotropy and mean following objective function was minimized:
diffusivity (number of IDPs, 92) N
• Brain FC: connectivity strengths between 55 functional brain H (β) = 0.5β ′ β + C ∑ (sn + s∗n )
networks derived from resting-state fMRI (IDPs, 1,485 connec- n=1

tion pairs).
subject to the constraints |agen − (xn β + b)| ≤ ϵ; agen − (xn β + b) ≤ ϵ + sn;
Other IDPs such as regional gray/white matter intensity contrast (xn β + b) − agen ≤ ϵ + s∗n and sn, s∗n ≥ 0 for all individuals n = 1,…N in the
from T1-weighted MRI and volumes of ventricles were only included training dataset. In this formulation, sn and s∗n are the slack variables for
in the whole-brain predictive model. each individual, ϵ is the model residual and C is the box constraint
constant (C = 1 in this work). Consistent with recent work55, using a
Cognitive phenotypes nonlinear kernel function (Gaussian or polynomial) did not improve
Cognitive tests assessing reasoning, memory, attention, processing model performance. The predicted chronological age of the individual
speed and executive function were conducted on the same day of with index n comprising the test dataset was given by
brain imaging in 45,930 individuals (22,307 males). A predictive model
ˆ
a gen = xn β + b
of chronological age was established using 29 distinct measures of
cognitive performance (Supplementary Table 3). Dummy variables
were generated for categorical responses. Several cognitive tests, Model performance was quantified using the Pearson correlation
including trail making, matrix pattern completion, tower rearranging coefficient (r) and MAE between predicted and chronological age in the
and symbol digit substitution test were only added to the assessment test sets. The 20-fold cross-validation procedure was repeated for ten
battery from 2016 onwards, resulting in an incomplete assessment for trials, randomizing the assignment of individuals to folds for each trial.
some participants. These participants were omitted, yielding a final Owing to the large sample size, variation across different train–test
sample consisting of 32,317 individuals (age range 45–82 years, mean data splits was negligible (rs.d.<0.002, MAEs.d. < 0.005). Optimization
65.1 ± 7.6, 15,712 males) for cognitive age analyses. of hyperparameters (box constraint, kernel function and ε) did not
substantially improve performance of the predictive models trained on
Normative aging models all body or brain phenotypes and thus, hyperparameter optimization
Support vector machines (SVMs) were trained to predict an individual’s was not conducted for the organ-specific models.
chronological age using body (n = 28,589, age range 40–70 years, To determine chance-level prediction accuracy intervals, chron-
mean 52.7 ± 7.8, 15,444 males), brain (n = 7,922, age range 46–82 years, ological age was randomly permuted among individuals and each
mean 61.8 ± 7.3, 3624 males) and cognitive (n = 7,167, age range 47–82 organ-specific predictive model was re-trained using the permuted
years, mean 62.6 ± 7.3, 3,357 males) phenotypes in healthy individu- data. This was repeated for 5,000 permutations to generate an empiri-
als, defined as no self-reported and healthcare-documented lifetime cal null distribution for MAE, under the null hypothesis of an absence of
chronic medical conditions (see section on Health outcomes and clini- predictive utility of body and brain phenotypes on chronological age
cal characterization). Compared to linear regression, SVM regression prediction. The observed MAE for all predictive models was less than
can provide improved robustness to outliers and overfitting. It auto- the fifth percentile of the MAE null distribution, enabling rejection of
matically learns the relative value of each phenotype toward predict- the null hypothesis. The FDR was controlled at 5% across all predictive
ing age and fits a hyperplane to the phenotype data. Using 20-fold models (n = 26) using the Benjamini–Hochberg procedure56.
cross-validation, predictive models were developed for each body Prediction accuracies varied considerably between organ systems
(n = 8) and brain (n = 4) system as well as for cognitive performance (Fig. 2a). While prediction accuracy often improved with the number
(Fig. 2b). Separate models were trained for males and females. Each of features available, this was not always the case. For example, the
model accepted an individual’s organ-specific phenotypes and yielded least accurate organ-specific model (immune system) consisted of
an estimate of chronological age based on these phenotype inputs. the largest number of features (n = 33), whereas the best performing
Individuals were thus characterized by 12 organ-specific predictions model consisted of 11 phenotypes pertaining to renal function. Models
of chronological age. developed for the pulmonary (n = 4) and cardiovascular system (n = 3)
For each 20-fold cross-validation iteration, a linear SVM was also outperformed the immune system. Prediction accuracy variation
trained to predict chronological age using individuals consisting of 19 between organ models could be due to (1) insufficient or inaccurate
folds (training set). The fitted regression coefficients (feature weights) phenotype ensembles to fully characterize an organ’s age-related
were then applied iteratively to the held-out set of individuals (test set), decline; or (2) complex trajectories of age-related decline that are
resulting in a predicted chronological age for each healthy individual. nonlinearly related to chronological age. Regarding the latter con-
In this way, the prediction model was never trained using the same sideration, deep neural networks and nonlinear learners could have
individuals for which it was applied to predict age, minimizing the improved the prediction accuracies reported here, as suggested in
risk of overfitting. All measures except for categorical variables were recent brain age prediction studies55,57–60. Regarding the former, we
standardized by weighted column mean and s.d., computed within note that most physiological measures (for example, blood biochem-
the training set before each iteration of model training. For all mod- istry and urine assays) used in this study are validated with rigorous
els, the SVM box constraint and kernel scale were set to unity, and the quality control procedures (https://biobank.ctsu.ox.ac.uk/crystal/
half-width of the epsilon-insensitive band was set to a tenth of the s.d. ukb/docs/biomarker_issues.pdf) and are commonly used in clinical

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

settings as diagnostic tools to assess organ-specific function and gen- adjust the age gaps for individuals comprising the test set. Legacy
eral health. For example, elevated serum liver enzyme levels often studies66–70 typically quantify biological age using a linear combination
reflect hepatocyte damage or cholestasis61; however, key phenotypes of chronological age and selected physiological phenotypes13,71–73 and
for some body systems were unavailable. For example, inflammatory are sometimes referred to as ‘phenotypic age’12,14,48,74. In contrast, for
cytokines would have enabled a more holistic characterization of the age gap index used here, chronological age is the prediction target
immune function62, whereas cardiac and carotid imaging would have (independent variable). An advantage of the age gap index is that it is
enabled detailed assessment of heart function and atherosclerotic an inherently personalized measure, cross-validated, independent
plaque morphology63. Of note, cardiac and carotid imaging were not of chronological age and thereby directly indexes deviations from
primarily used to estimate cardiovascular age in our study as they were population norms. Likewise, compared to the commonly used frailty
available from 2014 onwards (third visit) in the UK Biobank. The lack of index38, which characterizes an overall functional decline in older
temporal correspondence between cardiac and carotid imaging data people (usually >65 years) by counting the number of health deficits
and other body phenotypes (blood biochemistry and urine assays) present75, an advantage of the age gap index is that it is an organ-specific
precluded a concurrent investigation of biological age across multiple aging measure applicable across the lifespan.
body systems. In supplementary analyses, we established a revised Longitudinal assessments enabled estimation of the rate of change
cardiovascular normative model that includes heart MRI and carotid in age gaps, providing organ-specific estimates of the rate of aging3,14.
ultrasound data and compared the accuracy of the revised model to Of note, all normative models were trained using phenotypes measured
our original cardiovascular model. at baseline and subsequently applied to predict chronological age in
We found modest improvement in the age prediction using the the follow-up data. Let gapt0 (years) be the organ age gap estimated
revised model (12% and 15% reduction in MAE in females and males, at baseline, gapt1 (years) be the age gap at follow-up and T be the time
respectively, Supplementary Fig. 12). This suggests that our original interval (in years) between baseline and follow-up assessment. The
model using blood pressure indices and pulse rate provides a reason- rate of aging was estimated as ∆ = 12 × (gapt1 − gapt0) / T, expressed
able estimation of cardiovascular age. Recent work13 measures cardio- in units of months per year. To test for faster organ aging, the rate of
vascular age using a combination of blood pressure, blood markers change, ∆ was regressed against the average age gap (gapt1 + gapt0) / 2,
(such as glucose and lipids) and physical fitness (such as vital capac- and a significant positive association between these quantities across
ity). In contrast, we use blood-derived glucose and lipids and vital individuals provided evidence consistent with faster aging24. The slope
capacity measures to instead inform aging models of the metabolic of the regression line provided an estimate of the putative acceleration
and pulmonary system, respectively. This exemplifies the need for rate (months / year2). Note that this is effectively a population-level
future standardization of organ age measures and we suggest that estimate and it does not necessarily imply faster or accelerated aging
cardiovascular age, as measured by Nie and colleagues13, is based on a for any individual participant.
broader characterization of the cardiovascular system and combines
some features of our metabolic and pulmonary age models. Replication of brain age prediction in additional datasets
In supplementary analyses, we also established revised metabolic To assess our normative aging model in an older age cohort, we per-
and renal normative models that exclude diagnostic markers for dia- formed supplementary analyses combining brain MRI data from the
betes (HbA1c) and CKD (cystatin C and creatinine), respectively. We Australian Imaging, Biomarkers and Lifestyle (AIBL) Flagship Study of
found that prediction accuracy worsens (without HbA1c: females, Ageing (n = 650) (https://aibl.csiro.au/) and the Alzheimer’s Disease
r = 0.45, MAE = 5.48; males, r = 0.19, MAE = 6.59) compared to our origi- NeuroImaging Initiative (ADNI, n = 1,677) (http://adni.loni.usc.edu/).
nal model (with HbA1c: females, r = 0.50, MAE = 5.30; males, r = 0.23, The two cohorts comprise individuals diagnosed with MCI and demen-
MAE = 6.51); however, the metabolic system still shows significantly tia as well as healthy individuals, thus facilitating external validation for
advanced biological age in diabetes (age gap of 0.18 ± 4.64 years, our normative brain aging model and the relationship between brain
t = 6.02, P = 1.79 × 10−9) after excluding this diagnostic marker. Simi- age and neurodegenerative diseases.
larly, a model excluding cystatin C and creatinine led to less accurate AIBL study methodology has been reported previously76. The
prediction of chronological age (females, r = 0.43, MAE = 5.61; males, AIBL study was approved by the institutional ethics committees of
r = 0.36; MAE = 6.15) compared to our original model (females, r = 0.53, Austin Health, St Vincent’s Health, Hollywood Private Hospital and
MAE = 5.22; males, r = 0.45; MAE = 5.83). Nevertheless, the renal system Edith Cowan University and all volunteers gave written informed con-
still shows significantly advanced age in CKD (age gap of 3.14 ± 6.32 sent before participating in the study. ADNI was launched in 2003
years, t = 35.52, P < 2.23 × 10−308) after excluding these two diagnostic as a public–private partnership, led by Principal Investigator M.W.
markers. Age prediction models trained using healthy individuals were Weiner. The primary goal of ADNI has been to test whether serial MRI,
applied to predict the chronological age of individuals diagnosed with positron emission tomography, other biological markers and clinical
one or more diseases (see Health outcomes and clinical characteriza- and neuropsychological assessment can be combined to measure
tion). For this purpose, all models were re-trained on the full sample the progression of MCI and early Alzheimer’s disease. For up-to-date
of healthy individuals. information, see www.adni-info.org. As per ADNI protocols, all pro-
cedures performed in ADNI studies involving human participants
Age gap index were in accordance with the ethical standards of the institutional and/
Subtracting actual chronological age from predicted chronological or national research committee and with the 1964 Helsinki Declara-
age, referred to as the age gap, provides a normalized measure of the tion and its later amendments or comparable ethical standards. More
extent to which an individual’s organ system appeared older (gap > 0) details can be found at adni.loni.usc.edu.
or younger (gap < 0) than same-aged peers of the same sex. Age gaps Details of brain image acquisition can be found elsewhere76,77.
were estimated for each organ, yielding a multiorgan assay of biologi- T1-weighted MRI brain images acquired at baseline assessments were
cal age for each individual. Chronological age was regressed from all used in this study. Consistent with the brain image processing pipeline
estimated age gaps to adjust for regression-toward-the-mean bias64,65 used for the primary cohort (UK Biobank), MRI brain images were
and the residuals of this regression defined adjusted age gaps. All age processed using FreeSurfer v.6 (ref. 78), resulting in 578 regionally
gaps in this study were adjusted as such. Regressing the square of specific MRI-derived phenotypes representing regional gray matter
chronological age in addition to chronological age had minimal impact volume, cortical thickness and surface area. The Destrieux atlas79
on the adjusted age gaps. Regression coefficients for performing age was used for cortical parcellation. Additional segmentations of hip-
gap adjustment were fitted using the training set and then used to pocampal subfields80, amygdala81 and thalamic82 nuclei and brainstem

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

substructures83 were performed using FreeSurfer v.7. The quality of the regression coefficients, where J denotes the total number of organ
T1 images was automatically assessed using the Euler number, an index systems. If an edge was detected by FGES but the edge’s regression
generated by FreeSurfer that measures the topological complexity of coefficient did not survive FDR correction, the edge was removed.
a reconstructed cortical surface84. Following previous recommenda- Hence, drawing an edge required both statistical and causal evidence. In
tions85, images with a Euler number less than −217 were associated with Fig. 2e,f, a positive regression coefficient provided evidence consistent
poor quality and thus discarded (AIBL, n = 111; ADNI, n = 104). Images with faster aging of organ Y, relative to organ X (a unit increase in the
with any MRI-derived phenotypes residing more than 6 × s.d. from the age of X predicted an increase of β in the rate of aging of Y). In Fig. 2e,f,
median were also discarded (AIBL, n = 7; ADNI, n = 17). the baseline age gap node and rate of aging node were merged into a
Given the older age range of the AIBL and ADNI cohorts compared single node for each organ to provide a succinct network representa-
to the UK Biobank, the normative aging model established in the UK tion for visualization purposes. The Tetrad software package v.6.8.1
Biobank cohort could not be directly applied to the two external data- (https://github.com/cmu-phil/tetrad) was used to perform the FGES
sets. We therefore re-trained the brain age prediction model for gray heuristic, bootstrapping and parameter estimation. Default parameter
matter phenotypes in a combined group of healthy individuals across settings for FGES were used (Chickering rule; BIC penalty discount of
the three datasets. Before model training, data harmonization was 0.5; T-depth of −1).
performed using ComBat (https://github.com/Jfortin1/ComBatHar-
monization)86,87 to control for variation in brain phenotypes due to Genetic, environmental and lifestyle factors
differences in scanners and datasets. Of note, images acquired from Telomere length shortening and human aging are linked89 and thus
MRI scanners with fewer than ten scanned individuals were further we tested for associations between organ-specific age gaps and the
discarded (n = 186, ADNI) to ensure reliable harmonization, resulting in relative leukocyte telomere length (adjusted for technical param-
532 AIBL (age range 55–96 years, mean 72.4 ± 6.4, 218 males, four scan- eters90) measured at baseline assessment, as well as genetic variants
ners) and 1,370 ADNI (age range 50–95 years, mean 72.7 ± 7.5, 656 males, known to index leukocyte telomere length. Following an established
76 scanners) individuals for further analyses. Age, sex and diagnostic method91, a polygenic score for leukocyte telomere length was com-
status were included as biological covariates in the harmonization. puted for each individual based on nine single-nucleotide poly-
Consistent with our main findings in the UK Biobank cohort, we morphisms associated with leukocyte telomere length92–94. Larger
found that chronological age could be predicted with high accuracy polygenic scores associate with longer leukocyte telomere length
using gray matter phenotypes (female, r = 0.82, MAE = 3.52; male, and vice versa. Age gaps for body and brain systems were also tested
r = 0.85, MAE = 3.41; Extended Data Fig. 2a). Gray matter feature weights for associations with numerous environmental and lifestyle factors.
were highly consistent between the original and the re-trained model We selected 158 variables that tapped individual differences in early
(female, r = 0.86, P < 2.23 × 10−308; male, r = 0.87, P < 2.23 × 10−308; life experience (for example, birth weight, breastfed, adoption,
Extended Data Fig. 2b). The re-trained brain age prediction model was maternal smoking and traumatic events), sociodemographics (for
applied to estimate brain gray matter age for individuals diagnosed MCI example, education, neighborhood measure of deprivation, job sta-
and dementia. We found that brain age seems significantly older in indi- tus and parenting), lifestyle (for example, smoking, alcohol intake,
viduals diagnosed with MCI (n = 780, mean age gap of 1.07 ± 4.25 years, diet, exercise, sleep and e-device use), psychosocial (for example,
P = 3.56 × 10−25) and dementia (n = 284, mean age gap of 3.19 ± 6.13 years, social support and mood status), local environmental exposures
P < 2.23 × 10−308) than same-aged healthy peers (Extended Data Fig. 2c). (for example, air and noise pollution, greenspace and coastal prox-
imity), general health (for example, menstrual cycle, menopause,
Structural equation modeling long-standing illness and disability, hearing, vision and falls) and
SEM was used to infer the influence of each organ’s baseline age gap cognitive ability. Individual variation in cognitive ability was meas-
on the follow-up age gap (Fig. 2d) or rate of aging, ∆ (Fig. 2e,f) of other ured using the cognitive age gap, inferred from the above-described
organ systems. The fast-greedy equivalence search (FGES) heuristic cognitive age prediction model. Dummy variables were generated
for continuous variables was performed to search for causal Bayesian for categorical responses. Several variables were curated to enable
networks and determine the highest scoring model. FGES is a Bayesian more intuitive interpretation than the native UK Biobank coding.
heuristic that starts with an empty graph and adds edges to improve The curation procedure includes:
the score function (Bayesian Information Criterion), until no more • Reponses indicating less than one unit in time, distance,
edges can be added. It then performs a backward search that removes frequency and quantity were originally coded as −10 and were
edges until no edge removal increases the score function88. The search recoded to 0 for all relevant variables, including food intake
was constrained to edge modeling influences consistent with the flow frequencies, time spent on watching television, using computer
of time. To this end, age gaps measured at the same assessment were and driving, distance between home and job workplace, etc.
forbidden from influencing each other, age gaps measured at follow-up • Individuals who did not provide a valid answer, originally coded
were forbidden from influencing age gaps measured at baseline and as −3 (prefer not to answer) or −1 (do not know) were labeled as
rates of aging were forbidden from influencing baseline age gaps. missing responses.
The FGES heuristic was repeated for 500 bootstrapped samples and • An average weekly alcohol consumption (in UK standard units)
edges present in 50% of the samples formed a final consensus network was computed by combining information on each person’s
structure. Regression was used to estimate residual variances and coef- response to questionnaire on weekly and monthly intake of
ficients (β) for the edges consisting of the final consensus network. a variety of beverage type, including red wine, white wine/
Sex, age at baseline and whole-body (Fig. 2e) and whole-brain (Fig. 2f) champagne, beer/cider, spirits and fortified wine, consistent
age gaps were regressed from all organ-specific age gaps and FGES was with previous literature95,96. Specifically, weekly alcohol intake
performed on the resulting residuals. The time interval between base- data were collected from individuals who indicated that they
line and follow-up measurements was also regressed from baseline and drink more often than once or twice a week, whereas monthly
follow-up age gaps, if appropriate (Fig. 2d). Influences inferred from alcohol intake was collected from individuals who drink alcohol
FGES were represented using multiorgan networks, where each organ one to three times a month or on special occasions. The alcohol
was denoted with a distinct network node. A directed edge was drawn consumption of individuals who indicated that they never drink
from organ X to organ Y only if the age gap of X at baseline significantly was set to zeros.
influenced the follow-up age gap of Y (Fig. 2d) or rate of aging of Y • Reponses to current tobacco smoking: 1, yes, on most or all days;
(Fig. 2e,f), following FDR correction at 0.05 across the set of J( J − 1) / 2 2, only occasionally; and 0, no, were recoded to: 2, yes, on most

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

or all days; 1, only occasionally; and 0, no, so that higher scores lookup table (‘all_lkps_maps_v2.xlsx’) provided by the UK Biobank
denoted higher frequency of current tobacco smoking. (https://biobank.ndph.ox.ac.uk/showcase/showcase/auxdata/prima-
• Responses to past tobacco smoking were originally coded as: 1, rycare_codings.zip). For Read v.2, the mapping was only performed
smoked on most or all days; 2, smoked occasionally; 3, just tried when the read code matched to a single ICD-9 or ICD-10 code. For Read
once or twice; and 4, I have never smoked. Responses were thus CTv.3, the mapping was only performed for read code flagged as exact
reversed so that higher scores denoted higher frequency of past one-to-one mapping (‘E’) or target concept more general (‘G’) that had
tobacco smoking. been completely refined (‘C’). Whereas self-report provided past and
• Individuals who smoked tobacco on most or all days in the past current medical conditions, healthcare records enabled a lifetime
or current were labeled as ‘daily smokers’, notwithstanding assessment of a participant’s health outcomes.
varied definitions of smokers97. Based on self-report and healthcare records, we defined a healthy
• Time since stopped smoking was computed for past tobacco aging group and 16 clinical groups consisting of individuals with a life-
smokers by subtracting ‘age stopped smoking’ from their time diagnosis of Parkinsonism, multiple sclerosis, stroke, dementia,
chronological age at the assessment. depression, bipolar disorder, schizophrenia, ischemic heart disease,
• Age started smoking in either past or current smokers was hypertensive diseases, COPD, CKD, diabetes, cirrhosis, osteoarthritis,
derived. osteoporosis and cancer. Each disease category was defined broadly
• An overall fruit and vegetable consumption (per day) was com- with all causes and subtypes included. For example, the COPD group
puted by summing fresh fruit, dried fruit, salad, cooked and raw included self-reported COPD, emphysema/chronic bronchitis and
vegetables intake98. emphysema; healthcare recorded ICD-9 coded emphysema (code
• Individuals who slept 7–8 h per night were labeled as had ‘good 492) and chronic airway obstruction not elsewhere classified (496)
sleep duration’99. and ICD-10 coded emphysema (code J43), MacLeod syndrome ( J430),
• Reponses to facial aging were recoded to 1, younger than you panlobular emphysema ( J431), centrilobular emphysema ( J432), other
are; 2, about your age; and 3, older than you are, such that higher emphysema ( J438), emphysema unspecified ( J439), other COPD ( J44),
scores indicated older-appearing faces. COPD with acute lower respiratory infection ( J440), COPD with acute
• Women who had no regular length of menstrual cycle were exacerbation unspecified ( J441), other specified COPD ( J448) and
labeled as had irregular menstrual cycle. COPD unspecified ( J449). Supplementary Table 7 lists diagnostic
• Women who were not sure if they have had menopause because codes related to each of the 16 disease categories. For each individual,
of hysterectomy or other reasons were labeled as missing the recorded date of diagnosis was compared across self-report and
responses. healthcare sources for each disease category, to determine whether
• Hearing, as measured by the speech reception threshold, was the illness onset/diagnosis preceded or occurred after the baseline
averaged for left and right ears. assessment of body and brain function; however, the earliest state
• Visual acuity, as measured by the logarithm of the minimum of disease onset for some individuals may have not been captured
angle of resolution (logMAR), was averaged for left and right because the data from general practitioners only covered approxi-
eyes. mately 45% of the UK Biobank cohort; in contrast to the more than
87% coverage of hospital inpatient records. To enable comparisons,
Supplementary Table 6 provides a full list of selected variables. the healthy aging group included individuals with no self-reported
The original UK Biobank field IDs of variables were provided where and/or healthcare-documented lifetime chronic medical conditions.
applicable. Proportions of healthy individuals included in the body (28,589 of
Partial correlation was used to test for associations between 143,423 = 0.199) and brain (7,909 of 36,901 = 0.214) analyses subsam-
organ-specific age gaps and genetic, environmental and lifestyle fac- ples are slightly greater than the proportion in the full sample (91,808
tors, adjusting for chronological age and sex. Due to the very large sam- of 502,504 = 0.183; body, χ2 = 203.43, P < 2.2 × 10−16, brain, χ2 = 228.01,
ple sizes, statistical significance was Bonferroni-corrected for genetic P < 2.2 × 10−16). This suggests that individuals for whom organ age was
and environmental/lifestyle factors separately at P < 0.004 (12 organ estimated are not necessarily representative of the full cohort. Demo-
systems) and P < 2.6 × 10−5 (158 factors × 12 organ systems = 1,896 tests), graphic details of individuals comprising each defined clinical group
respectively. A minimum effect size threshold of |r | >0.05 was enforced are provided in Supplementary Table 8.
to suppress weak associations for visualization purpose (Fig. 3). Of Individuals diagnosed with more than one disease category
note, Bonferroni correction was used to control the family-wise error throughout their lifetime were assigned to multiple disease groups.
when sample sizes were very large (>10,000). The FDR was controlled As shown in Extended Data Fig. 6, most individuals included in
at 5% using the Benjamini–Hochberg procedure described elsewhere56. either body or brain aging analyses (n = 169,109; 90,918 males) were
linked with a single diagnostic category (n = 52,113, 54.1%) and the
Health outcomes and clinical characterization proportion of individuals comorbid with 2–8 conditions was 28.3%
Diagnoses and medical conditions of participants were obtained (n = 27,215), 11.6% (n = 11,171), 4.1% (n = 3,991), 1.4% (n = 1,308), 0.4%
through self-report (verbal interview at assessment centers, UK (n = 394), 0.1% (n = 103) and 0.026% (n = 25), respectively. The largest
Biobank Field IDs 20001 and 20002) and healthcare records (for exam- number of comorbidities was nine, in four individuals (0.0042%).
ple, hospital inpatient and primary care) from the UK National Health Extended Data Fig. 6b,c shows a comorbidity network, representing a
Services (NHS). Hospital inpatient records were summarized by distinct population-level lifetime co-occurrence of the 16 disease categories.
International Classification of Diseases and Related Health Problems The extent of comorbidity was quantified by correlating (Pearson
(ICD)-9 and/or ICD-10 coded primary and/or secondary diagnoses correlation) the presence of categorical diagnoses (1, yes or 0, no)
for participants whose health outcomes resulted in a hospital admis- across individuals for each sex. Permutation testing (n = 10,000) was
sion. Summary inpatient diagnoses (field IDs 41270 and 41271) in the used to estimate P values, and significant correlations (P < 4.2 × 10−4)
July 2020 release were used in this study. Primary care data (field ID were Bonferroni-corrected for (16 × 15) / 2 = 120 disease pairs. Results
42040) were sourced at record-level on 26 November 2020. Of note, were broadly consistent between females (Extended Data Fig. 6b) and
primary care data in relation to clinical events were recorded by health males (Extended Data Fig. 6c) and that the 16 disease categories were
professionals working at general practices using Read Codes v.2 (Read parsed into two large comorbid groups, corresponding to major brain
v.2) and Read Codes Clinical Terms v.3 (Read CTv.3). Diagnoses coded and body disorders. Notably, brain disorders were also comorbid with
in Read were mapped to corresponding ICD codes according to the diseases primarily implicated in body organ systems, with stronger

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

effect sizes observed in males. For example, depression was signifi- survival and premature death beyond established predictors, includ-
cantly associated with osteoarthritis, COPD and hypertensive diseases ing chronological age, sex, existing disease diagnoses and key lifestyle
and dementia was associated with CKD and stroke. factors. The six models were as below:
• Model 1, chronological age and sex.
Mortality risk prediction • Model 2, chronological age, sex and eight body system ages.
Mortality data released on 4 March 2021 were used in this study. Indi- • Model 3, chronological age, sex and existing diagnoses of the
vidual mortality status (date of death) was determined using data 16 disease categories.
linkages to national death registries in the UK, including NHS Digital • Model 4, chronological age, sex, eight body age gaps and exist-
(England and Wales) and NHS Central Register (Scotland). Data linkage ing diagnoses of the 16 disease categories.
procedures and steps in data cleaning and validation are described • Model 5, chronological age, sex, eight body age gaps, existing
in detail in the central UK Biobank death linkage documentation: diagnoses of the 16 disease categories, general health and key
(https://biobank.ndph.ox.ac.uk/showcase/showcase/docs/Death- lifestyle factors included in the Cox regression analysis.
Linkage.pdf). • Model 6, chronological age, sex, existing diagnoses of the 16 dis-
Mortality was confirmed in 8,109 (age of death 42–83 years, ease categories, general health and key lifestyle factors included
mean 69.6 ± 7.3, 5,670 males) and 330 (age of death 53–82 years, mean in the Cox regression analysis.
70.7 ± 6.4, 203 males) individuals after baseline assessment of body
and brain function, respectively. We first compared the mean age gap Chronological age and body age gaps were standardized by mean
between deceased and living individuals for each organ system using a and s.d. before model training. The AUC of the receiver operating char-
two-sample t-test. Cox proportional hazards regression was then used acteristic curve was used to quantify prediction accuracy. Confidence
to estimate the risk of mortality associated with organ-specific age intervals were estimated with bootstrapping (100 samples).
gaps. Last, we developed a logistic model using tenfold cross-validation
to predict an individual’s 5- and 10-year survival and premature mortal- Reporting summary
ity based on organ-specific age gaps. The risk of mortality associated Further information on research design is available in the Nature Port-
with body and brain age were estimated separately because of the time folio Reporting Summary linked to this article.
difference in assessments (Fig. 2c).
The Cox proportional hazards model was applied under the Data availability
assumption that mortality HRs in relation to organ age gaps do not Data were obtained from the UK Biobank, the ADNI and the AIBL Flag-
change over time for any individual. Therefore, the estimated HR rep- ship Study of Ageing. Participant age gaps for all body and brain sys-
resented the relative risk of death for each unit increase in age gap, tems estimated in this study will be returned to the UK Biobank to
compared to the baseline hazard, which was defined as the mean age strengthen the resource and facilitate access to other researchers
gap across individuals. To enable comparisons, each organ age gap for future research. Researchers can register to access all data used
was first standardized by mean and s.d. Two Cox regression mod- in this study via the UK Biobank Access Management System (https://
els were then formulated, where one model estimated the mortality bbams.ndph.ox.ac.uk/ams/) and the ADNI database (https://adni.
HRs per 1 × s.d. increase in organ-specific age gap, adjusting for sex loni.usc.edu/).
and chronological age (standardized) and the second model further
adjusted for existing diagnoses, general health (long-standing illness) Code availability
and key lifestyle factors, including smoking, exercise, socioeconomic MATLAB (R2021a, MathWorks) code for conducting the core analyses is
inequality (deprivation) and tertiary education. Key lifestyle factors available on GitHub (https://github.com/yetianmed/BioAge). SEM was
were selected based on (1) significant associations with body/brain age performed using the Tetrad software package v.6.8.1 (https://github.
gap (ranked within the top 20 out of 158 measures); and (2) no missing com/cmu-phil/tetrad). The organ images shown in Fig. 1 were created
responses among deceased individuals. Survival was ascertained up to with BioRender.com. Other figures were created using visualization
13.41 (mean 7.47 ± 3.2) and 6.07 (mean 2.47 ± 1.49) years after baseline routines in MATLAB.
assessment of body and brain function, respectively. Living individuals
were right-censored, where survival duration was calculated as days References
between the date of body or brain assessment and the date of mortality 51. Sudlow, C. et al. UK Biobank: an open access resource for
ascertainment (4 March 2021). identifying the causes of a wide range of complex diseases of
Finally, using tenfold cross-validation, logistic models were middle and old age. PLoS Med. 12, e1001779 (2015).
trained to predict the probability of an individual’s survival time and 52. Miller, K. L. et al. Multimodal population brain imaging in the UK
premature mortality based on body age gaps. A predictive model was Biobank prospective epidemiological study. Nat. Neurosci. 19,
not developed for brain age gaps, given the lack of evidence for a link 1523–1536 (2016).
between advanced brain age and mortality risk, as estimated from Cox 53. Wu, Y. et al. Genome-wide association study of medication-use
regression (Supplementary Fig. 4b,c). For survival time (T) prediction, a and associated disease in the UK Biobank. Nat. Commun. 10,
nominal logistic regression was fitted to classify whether an individual 1891 (2019).
was deceased within 5 years (T < 5) or had lived more than 5 years (T ≥ 5, 54. Alfaro-Almagro, F. et al. Image processing and quality control
deceased or living) after assessment of body function. The same model for the first 10,000 brain imaging datasets from UK Biobank.
was fitted for 10-year survival time. All living individuals had lived more NeuroImage 166, 400–424 (2018).
than 10 years (minimal survival time of 10.43 years) after body function 55. Schulz, M.-A. et al. Different scaling of linear models and deep
assessment. Similarly, logistic models were developed to predict an learning in UK Biobank brain images versus machine-learning
individual’s premature death. Premature death is typically defined as datasets. Nat. Commun. 11, 4238 (2020).
death occurring before the mean age of death in a certain population, 56. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate:
which is approximately 75 years in the UK100. Death before age 70 years a practical and powerful approach to multiple testing. J. R. Stat.
was also considered based on the mean age of death of the current Soc. B Methodol. 57, 289–300 (1995).
UK Biobank cohort (n = 34,003, mean age of death 69.6 ± 7.4 years). 57. Jonsson, B. A. et al. Brain age prediction using deep learning
Prediction models were estimated based on five groups of predictors uncovers associated sequence variants. Nat. Commun. 10,
to assess the extent to which body age gaps improved prediction of 5409 (2019).

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

58. Dinsdale, N. K. et al. Learning patterns of the ageing brain in MRI 80. Iglesias, J. E. et al. A computational atlas of the hippocampal
using deep convolutional networks. NeuroImage 224, 117401 formation using ex vivo, ultra-high resolution MRI: application to
(2021). adaptive segmentation of in vivo MRI. NeuroImage 115, 117–137
59. Cole, J. H. et al. Predicting brain age with deep learning from (2015).
raw imaging data results in a reliable and heritable biomarker. 81. Saygin, Z. M. et al. High-resolution magnetic resonance imaging
NeuroImage 163, 115–124 (2017). reveals nuclei of the human amygdala: manual segmentation to
60. Peng, H., Gong, W., Beckmann, C. F., Vedaldi, A. & Smith, S. M. automatic atlas. NeuroImage 155, 370–382 (2017).
Accurate brain age prediction with lightweight deep neural 82. Iglesias, J. E. et al. A probabilistic atlas of the human thalamic
networks. Med. Image Anal. 68, 101871 (2021). nuclei combining ex vivo MRI and histology. Neuroimage 183,
61. Giannini, E. G., Testa, R. & Savarino, V. Liver enzyme alteration: a 314–326 (2018).
guide for clinicians. CMAJ 172, 367–379 (2005). 83. Iglesias, J. E. et al. Bayesian segmentation of brainstem structures
62. Dembic, Z. in The Cytokines of the Immune System (ed. Dembic, in MRI. NeuroImage 113, 184–195 (2015).
Z.) Ch. 4, 99–122 (Academic Press, 2015). 84. Dale, A. M., Fischl, B. & Sereno, M. I. Cortical surface-based
63. Johri, A. M. et al. Recommendations for the assessment of analysis. I. Segmentation and surface reconstruction. NeuroImage
carotid arterial plaque by ultrasound for the characterization of 9, 179–194 (1999).
atherosclerosis and evaluation of cardiovascular risk: from the 85. Rosen, A. F. G. et al. Quantitative assessment of structural image
American Society of Echocardiography. J. Am. Soc. Echocardiogr. quality. NeuroImage 169, 407–418 (2018).
33, 917–933 (2020). 86. Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in
64. Le, T. T. et al. A nonlinear simulation framework supports adjusting microarray expression data using empirical Bayes methods.
for age when analyzing brainAGE. Front. Aging Neurosci. 10, 317 Biostatistics 8, 118–127 (2006).
(2018). 87. Fortin, J.-P. et al. Harmonization of cortical thickness
65. Smith, S. M., Vidaurre, D., Alfaro-Almagro, F., Nichols, T. E. & measurements across scanners and sites. NeuroImage 167,
Miller, K. L. Estimation of brain age delta from brain imaging. 104–120 (2018).
NeuroImage 200, 528–539 (2019). 88. Ramsey, J., Glymour, M., Sanchez-Romero, R. & Glymour, C. A
66. Klemera, P. & Doubal, S. A new approach to the concept and million variables and more: the fast greedy equivalence search
computation of biological age. Mech. Ageing Dev. 127, 240–248 algorithm for learning high-dimensional graphical causal models,
(2006). with an application to functional magnetic resonance images. Int.
67. Krøll, J. & Saxtrup, O. On the use of regression analysis for the J. Data Sci. Anal. 3, 121–129 (2017).
estimation of human biological age. Biogerontology 1, 363–368 89. Blackburn, E. H., Epel, E. S. & Lin, J. Human telomere biology: A
(2000). contributory and interactive factor in aging, disease risks, and
68. Nakamura, E. A study on the basic nature of human biological protection. Science 350, 1193–1198 (2015).
aging processes based upon a hierarchical factor solution of 90. Codd, V. et al. Measurement and initial characterization of
the age-related physiological variables. Mech. Ageing Dev. 60, leukocyte telomere length in 474,074 participants in UK Biobank.
153–170 (1991). Nat. Aging 2, 170–179 (2022).
69. Nakamura, E., Miyao, K. & Ozeki, T. Assessment of biological age 91. Demanelis, K., Tong, L. & Pierce, B. L. Genetically increased
by principal component analysis. Mech. Ageing Dev. 46, 1–18 telomere length and aging-related traits in the UK Biobank. J.
(1988). Gerontol. A 76, 15–22 (2021).
70. Voitenko, V. P. & Tokar, A. V. The assessment of biological age 92. Codd, V. et al. Identification of seven loci affecting mean telomere
and sex differences of human aging. Exp. Aging Res. 9, 239–244 length and their association with disease. Nat. Genet. 45, 427
(1983). (2013).
71. Chan, M. S. et al. A biomarker-based biological age in UK Biobank: 93. Mangino, M. et al. Genome-wide meta-analysis points to CTC1 and
composition and prediction of mortality and hospital admissions. ZNF676 as genes regulating telomere homeostasis in humans.
J. Gerontol. A 76, 1295–1302 (2021). Hum. Mol. Genet. 21, 5385–5394 (2012).
72. Kuo, C.-L. et al. Biological aging predicts vulnerability to COVID-19 94. Mangino, M. et al. DCAF4, a novel gene associated with leucocyte
severity in UK Biobank participants. J. Gerontol. A 76, e133–e141 telomere length. J. Med. Genet. 52, 157–162 (2015).
(2021). 95. Clarke, T. K. et al. Genome-wide association study of alcohol
73. Zhong, X. et al. Estimating biological age in the Singapore consumption and genetic overlap with other health-related traits
Longitudinal Aging Study. J. Gerontol. A 75, 1913–1920 (2019). in UK Biobank (N=112117). Mol. Psychiatry 22, 1376–1384 (2017).
74. Liu, Z. et al. A new aging measure captures morbidity and 96. Wood, A. M. et al. Risk thresholds for alcohol consumption:
mortality risk across diverse subpopulations from NHANES IV: a combined analysis of individual-participant data for 599912
cohort study. PLoS Med. 15, e1002718 (2018). current drinkers in 83 prospective studies. Lancet 391, 1513–1523
75. Rockwood, K. & Mitnitski, A. Frailty in relation to the accumulation (2018).
of deficits. J. Gerontol. A 62, 722–727 (2007). 97. Husten, C. G. How should we define light or intermittent smoking?
76. Ellis, K. A. et al. The Australian Imaging, Biomarkers and Lifestyle Does it matter? Nicotine Tob. Res. 11, 111–121 (2009).
(AIBL) study of aging: methodology and baseline characteristics 98. Cassidy, S., Chau, J. Y., Catt, M., Bauman, A. & Trenell, M. I.
of 1112 individuals recruited for a longitudinal study of Alzheimer’s Cross-sectional study of diet, physical activity, television viewing
disease. Int. Psychogeriatr. 21, 672–687 (2009). and sleep duration in 233 110 adults from the UK Biobank; the
77. Petersen, R. C. et al. Alzheimer’s Disease Neuroimaging Initiative behavioural phenotype of cardiovascular disease and type 2
(ADNI): clinical characterization. Neurology 74, 201–209 (2010). diabetes. BMJ Open 6, e010038 (2016).
78. Fischl, B. et al. Whole brain segmentation: automated labeling 99. Shan, Z. et al. Sleep duration and risk of type 2 diabetes: a
of neuroanatomical structures in the human brain. Neuron 33, meta-analysis of prospective studies. Diabetes Care 38, 529–537
341–355 (2002). (2015).
79. Destrieux, C., Fischl, B., Dale, A. & Halgren, E. Automatic 100. Lewer, D. et al. Premature mortality attributable to socioeconomic
parcellation of human cortical gyri and sulci using standard inequality in England between 2003 and 2018: an observational
anatomical nomenclature. NeuroImage 53, 1–15 (2010). study. Lancet Public Health. 5, e33–e41 (2020).

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Acknowledgements ADNI data are disseminated by the Laboratory for Neuro Imaging at
This research has been conducted using data from UK Biobank the University of Southern California. We thank S.M. Smith (University
(https://www.ukbiobank.ac.uk/), a major biomedical database. We are of Oxford) and D. Vidal-Piñeiro (University of Oslo) for their feedback
grateful to UK Biobank for making the data available and to all study and discussion on data analyses. Y.E.T. was supported by the Mary
participants, who generously donated their time to make this resource Lugton Postdoctoral Fellowship. A.Z. was supported by a National
possible. Some of the data used in the preparation of this article Health and Medical Research Council (NHMRC) grant (APP1142801
were obtained from the AIBL Flagship Study of Ageing, funded by and APP118153), V.C. was supported by an NHMRC grant (APP1177370)
the Commonwealth Scientific and Industrial Research Organisation, and M.B. was supported by an NHMRC grant (APP1152623 and
which was made available at the ADNI database (https://adni.loni.usc. APP2008612).
edu/). The AIBL researchers contributed data but did not participate in
analysis or writing of this report. AIBL researchers are listed at: https:// Author contributions
aibl.csiro.au/. Some of the data used in preparation of this article Conceptualization was the responsibility of Y.E.T., A.Z., V.C. and
were obtained from the ADNI database (https://adni.loni.usc.edu/). As M.B. Methodology was the responsibility of Y.E.T., A.Z. and M.B.
such, the investigators within the ADNI contributed to the design and Investigation was carried out by Y.E.T. and A.Z. Visualization was
implementation of ADNI and/or provided data but did not participate carried out by Y.E.T. and A.Z. Project administration was performed by
in analysis or writing of this report. A complete listing of ADNI Y.E.T. Supervision was the responsibility of A.Z. and V.C. Writing of the
investigators can be found at: http://adni.loni.usc.edu/wp-content/ original draft was carried out by Y.E.T. and A.Z. and review and editing
uploads/how_to_apply/ADNI_Acknowledgement_List.pdf. Some of was conducted by Y.E.T., A.Z., V.C., M.B., A.B.M. and N.T.L.
the data collection and sharing for this project was funded by ADNI
(National Institutes of Health grant U01 AG024904) and Department Competing interests
of Defense ADNI (award number W81XWH-12-2-0012). ADNI is funded The authors declare no competing interests.
by the National Institute on Aging, the National Institute of Biomedical
Imaging and Bioengineering and through generous contributions Additional information
from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Extended data is available for this paper at
Drug Discovery Foundation; Araclon Biotech; BioClinica; Biogen; https://doi.org/10.1038/s41591-023-02296-6.
Bristol-Myers Squibb Company; CereSpir; Cogstate; Eisai; Elan
Pharmaceuticals; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Supplementary information The online version
Roche and its affiliated company Genentech; Fujirebio; GE Healthcare; contains supplementary material available at
IXICO; Janssen Alzheimer Immunotherapy Research & Development; https://doi.org/10.1038/s41591-023-02296-6.
Johnson & Johnson Pharmaceutical Research & Development;
Lumosity; Lundbeck; Merck & Co.; Meso Scale Diagnostics; NeuroRx Correspondence and requests for materials should be addressed to
Research; Neurotrack Technologies; Novartis Pharmaceuticals Ye Ella Tian or Andrew Zalesky.
Corporation; Pfizer; Piramal Imaging; Servier; Takeda Pharmaceutical
Company; and Transition Therapeutics. The Canadian Institutes of Peer review information Nature Medicine thanks Janine Bijsterbosch
Health Research is providing funds to support ADNI clinical sites in and the other, anonymous, reviewer(s) for their contribution to the
Canada. Private sector contributions are facilitated by the Foundation peer review of this work. Primary Handling editor: Jerome Staal,
for the National Institutes of Health (https://www.fnih.org/). The in collaboration with the Nature Medicine team
grantee organization is the Northern California Institute for Research
and Education, and the study is coordinated by the Alzheimer’s Reprints and permissions information is available at
Therapeutic Research Institute at the University of Southern California. www.nature.com/reprints.

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 1 | Age prediction accuracy. Scatter plots show associations between chronological and predicted age for prediction models based on body and
brain phenotypes as well as phenotypes pertaining to each individual organ system. Lines of best fit indicated with solid black lines. r: Pearson correlation coefficients;
MAE: mean absolute error.

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 2 | Replication of predictive models for brain gray individuals diagnosed with mild cognitive impairment (MCI, n = 780, mean age
matter age. (a) Scatter plots show associations between chronological age and gap=1.07 ± 4.25 years) and dementia (n = 284 mean age gap=3.19 ± 6.13 years),
predicted age for prediction model based on brain gray matter phenotypes in a compared to healthy individuals (HC). The mean age gap significantly differs
combined group of healthy individuals from the Australian Imaging, Biomarkers across the three groups (F-statistic=157.49, p = 4.71 × 10−68, two-sided). Asterisks
and Lifestyle Flagship Study of Ageing (AIBL, n = 396, 154 males), the Alzheimer’s indicate significant between-group differences, adjusting for chronological
Disease NeuroImaging Initiative (ADNI, n = 467, 192 males) and the UK Biobank age and sex (MCI vs HC, t = 10.39, p = 3.56 × 10−25; dementia vs HC, t = 16.94,
(n = 7,922, 3,624 males). Lines of best fit indicated with solid black lines. n: p < 2.23 × 10−308; MCI vs dementia: t = 10.76, p = 1.11 × 10−25). The bottom and top
training sample size; r: Pearson correlation coefficients; MAE: mean absolute edges of the boxes indicate the 25th and 75th percentiles of the distribution,
error. (b) Scatter plots show associations between gray matter feature weights respectively. The central line indicates the median. The whiskers extend to
estimated from the original age prediction model (primary) and the re-trained the most extreme data points that are not considered outliers (1.5-times the
model using the replication cohort. Lines of best fit indicated with solid black interquartile range).
lines. r: Pearson correlation coefficients. (c) Gray matter age (that is, age gap) in

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 3 | Synchrony among organ-specific age gaps. (a) two groups based on interorgan synchrony in age gaps (Group I: renal, hepatic,
Synchrony in biological ages between each pair of body systems at baseline musculoskeletal; Group II: pulmonary, cardiovascular, metabolic, immune). (c)
assessment was estimated using partial correlation, adjusting for sex and & (d) Same as (A) & (B) but the synchrony in age gaps is shown for different brain
chronological age. Correlation coefficients of significant pairs of correlations systems at baseline and follow-up assessment respectively. Biological age is
(p < 0.002, two-sided, t-test, Bonferroni-corrected for 21 pairs) are indicated most strongly synchronized between white and gray matter, whereas functional
in the matrix (left) and also visualized as a graph (right). In the graph, each connectivity is only weakly synchronized with other brain systems (Bonferroni-
node represents one of the 7 body organs and the edges between them indicate corrected for 3 correlations, p < 0.017, two-sided). GM, gray matter; WM, white
correlations. Edge thicknesses are proportional to correlation coefficients. Edges matter; FC, functional connectivity. Ward’s linkage clustering was used to
are suppressed for small effect sizes (|r|<0.05) (b) Same as (A) but shows the determine the reordering and the cluster tree shown.
correlations at follow-up assessment. Body systems can be differentiated into

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 4 | Relationship between chronic disease and organ- representation. The font size was normalized according to the mean age gap
specific biological age. (a) A clock face represents the extent of body aging for across the 16 disease groups within each organ system. Diseases for which
16 disease categories. Body age is older (younger) in a clockwise (anticlockwise) organs appear older than chronological age (gap>0) are colored black, whereas
direction, with a body age gap of zero at the 12 o’clock position. Bar plot diseases for which organs appear younger (gap<0) are colored blue. COPD,
shows the mean body age gap in each disease, sorted from the smallest to the chronic obstructive pulmonary disease; CKD, chronic kidney disease. Organ
largest value. (b) Same as panel (A) but shows the mean brain age gap across image was created with BioRender.com.
disease. Dashed arm indicates replication dementia cohort. (c) Word-cloud

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 5 | Effect size of body and brain age in chronic disease. indicate the effect size for each disease category. Icons are not shown for organs
Effect sizes of differences in organ-specific age gaps between each disease with mean age gaps that do not significantly differ from zero (p < 2.6 × 10−4, two-
category and healthy comparison group were quantified using the Cohen’s d. The sided, t-test, Bonferroni corrected, Fig. 4a). Disease categories are ordered from
Cohen’s d value was multiplied by the sign of the mean between-group difference top to bottom according to increased mean body age gaps as shown in Fig. 4b.
in age gap. Icons representing body systems and organs are positioned to

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 6 | Disease comorbidity. (a) Bar plots show the number estimate p-values and significant correlations were Bonferroni corrected
of lifetime comorbid diagnoses for individuals who completed assessment for (16 × 15)/2 = 120 disease pairs (p < 4.2 × 10−4, one-sided). Non-significant
of body (left) or, brain (middle) function and all individuals (right). (b) correlations were suppressed from the correlation matrix (left) and
Comorbidity network for females. The Pearson correlation coefficient the network graph (right). Edge thickness is modulated by correlation
was used to quantify the extent of lifetime comorbidity between each coefficients. (c) Same as (d) but for males. Also see Methods.
pair of disease categories. Permutation testing (n = 10,000) was used to

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 7 | See next page for caption.

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 7 | Biological organ age in chronic disease at different Bonferroni corrected for 16 disease categories × 8 body systems = 128 tests).
illness stages. (a) Distribution of body age gaps (columns) for 16 disease Distributions colored gray have a mean that is not significantly different from
categories (rows) in individuals at prodromal stage, compared to healthy the healthy group. (b) Same as (A) but in individuals with established diagnosis.
individuals (HC, first row). Distributions are colored according to disease- Prodromal groups for brain imaging data were insufficient to investigate the
and organ-specific mean age gaps. Colored distributions have a mean that impact of disease progression on brain age.
significantly differs from the healthy group (p < 3.9 × 10−4, two-sided, t-test,

Nature Medicine
Article https://doi.org/10.1038/s41591-023-02296-6

Extended Data Fig. 8 | Survival time and premature death prediction. p = 9.44 × 10−47/p = 1.56 × 10−54, two-sided, t-test). Nevertheless, Model 5, which
A logistic regression model was trained (10-fold cross-validation) to predict includes all predictors, achieves the most accurate predictions of survival time
an individual’s 5- and 10-year survival (left) and premature death (defined as (5-year: AUC = 0.774 ± 0.006; 10-year: AUC = 0.770 ± 0.003) and premature
death before 70 or 75 years old, right). Boxplots show prediction accuracy, as death (70 years old: AUC = 0.86 ± 0.003; 75 years old: AUC = 0.86 ± 0.003).
quantified with area under curve (AUC). A hierarchy of six logistic models was Omitting body age gaps (Model 6) leads to significantly reduced accuracy
established to determine the extent to which biological age improves prediction (5-year/10-year: p = 4.43 × 10−29/p = 4.89 × 10−40; 70 years old/75 years old:
of survival time and premature death above and beyond established predictors p = 1.86 × 10−42/p = 4.31 × 10−46, two-sided, t-test) for predictions of survival time
(that is, chronological age, sex, diagnoses, lifestyle factors). For prediction of (5-year: AUC = 0.76 ± 0.005; 10-year: AUC = 0.76 ± 0.003) and premature death
both survival time and premature death, the model including body age gaps (70 years old: AUC = 0.85 ± 0.003; 75 years old: AUC = 0.85 ± 0.003). Confidence
(Model 2) significantly outperforms the model including only chronological intervals for AUC estimated with bootstrapping (100 samples). The bottom and
age and sex (Model 1, 5-year/10-year: p = 4.55 × 10−133/p = 1.15 × 10−141; 70 years top edges of the boxes Indicate the 25th and 75th percentiles of the distribution,
old/75 years old: p = 3.40 × 10−77/p = 1.19 × 10−87, two-sided, t-test). Similarly, respectively. The central line indicates the median. The whiskers extend to
the model including body age gaps (Model 4) significantly outperforms the most extreme data points that are not considered outliers (1.5-times the
the model with only chronological age, sex and existing diagnoses (Model interquartile range).
3, 5-year/10-year: p = 3.52 × 10−58/p = 2.02 × 10−88; 70 years old/75 years old:

Nature Medicine