Organ-specific proteomic aging clocks predict disease and longevity across diverse populations

Abstract

Aging and age-related diseases share convergent pathways at the proteome level. Here, using plasma proteomics and machine learning, we developed organismal and ten organ-specific aging clocks in the UK Biobank (n = 43,616) and validated their high accuracy in cohorts from China (n = 3,977) and the USA (n = 800; cross-cohort r = 0.98 and 0.93). Accelerated organ aging predicted disease onset, progression and mortality beyond clinical and genetic risk factors, with brain aging being most strongly linked to mortality. Organ aging reflected both genetic and environmental determinants: brain aging was associated with lifestyle, the GABBR1 and ECM1 genes, and brain structure. Distinct organ-specific pathogenic pathways were identified, with the brain and artery clocks linking synaptic loss, vascular dysfunction and glial activation to cognitive decline and dementia. The brain aging clock further stratified Alzheimer's disease risk across APOE haplotypes, and a super-youthful brain appears to confer resilience to APOE4. Together, proteomic organ aging clocks provide a biologically interpretable framework for tracking aging and disease risk across diverse populations.

Fig. 1. Organ-specific proteomic aging clocks and their associations with disease and mortality across diverse populations.

a, Performance of the organ aging models across the discovery cohort (UKB, n = 43,616) and external validation cohorts (CKB, n = 3,977; NHS, n = 800). Models were trained on organ-enriched proteins from the Olink Explore 3072 panel, which were identified by GTEx tissue expression data. Performance was assessed using Pearson correlations between predicted organ age and chronological age. The top 20 proteins included in each model are detailed in Extended Data Fig. 2. b, Cross-cohort consistency of the performance of proteomic organ aging clocks, assessed using Pearson correlation (left, UKB versus CKB; middle, UKB versus NHS; right, CKB versus NHS). c, Distribution of proteomic organ age gap across cohorts. Box bounds indicate the first quartile (Q1), median and Q3; whiskers extend to Q1 − 1.5 × interquartile range (IQR) and Q3 + 1.5 × IQR. d, Pairwise correlations among organismal and organ-specific age gaps in the UKB (mean r = 0.16), CKB (mean r = 0.19) and NHS (mean r = 0.10). e, Overlap in constituent proteins among the organismal aging clock and three representative organ-specific clocks (brain, artery and heart). f, Associations between organ-specific age gaps and the incidence of five NDs, five psychiatric disorders, seven other chronic physical diseases and all-cause mortality in the UKB (n = 43,616). Associations were externally validated in the CKB and NHS (Extended Data Fig. 3). HRs per 1-s.d. change in the organismal and ten organ-specific age gaps are shown for significant associations, estimated using separate Cox proportional hazards models for each outcome, with adjustments for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level and recruitment center. The number of incident cases is presented. Mean differences in organ age gaps at baseline between participants with and without corresponding ‘incident’ diseases are visualized. The right panel shows the relative contributions of organ age gaps to each outcome, calculated by scaling z-scores for significant organs so that they sum to 1. g, Association between organ age gaps and years since disease diagnosis in participants with prevalent diseases at the baseline proteomic assessment, assessed by Pearson correlation. h, Visualization of the brain age gap after prevalent diseases before baseline (reflecting disease progression) or before incident diseases (reflecting prodromal disease). Participants with incident diseases were matched by age (±2 years) and sex with five healthy controls without corresponding incident diseases during the follow-up. The associations of brain age gaps with CKD (n = 11,890), ACD (n = 5,760) and depression (n = 9,710) are shown as examples. Trajectories were fitted using Loess regression, with error bands indicating 95% confidence intervals (CIs). All regression models were adjusted for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level and recruitment center. All statistical tests are two-sided. The Benjamini–Hochberg FDR was used to correct for multiple comparisons in f and g. The asterisks denote FDR-adjusted P-value thresholds: *q < 0.05; **q < 0.01; ***q < 0.001. ProtAge, proteomic age; COPD, chronic obstructive pulmonary disease.

Fig. 2. Association of proteomic organ aging clocks (versus established phenotypic aging clocks) with age-related traits, clinical markers, metabolites and lifestyle factors.

a, Pearson correlation of phenotypic age (KDMAge and PhenoAge) with chronological age. KDMAge and PhenoAge were trained using data from the National Health and Nutrition Examination Survey with an established algorithm and then mapped to the UKB data. b, Proteomic organ age gaps were only weakly correlated with established phenotypic age gaps. c, Association of proteomic organ age gaps and phenotypic age gaps with age-related traits, clinical markers, and cognitive and mental health measures. d, Associations of proteomic organ aging clocks with plasma metabolites measured using an NMR-based metabolomics platform. The clocks are broadly associated with an atherogenic metabolite profile. e, Associations of proteomic organ aging clocks with modifiable lifestyle factors (smoking, alcohol consumption, physical activity, TV watching/sedentary time, sleep duration, and intake of fruits and vegetables, oily fish, red meat and processed meat; n = 43,616). Associations with individual lifestyle factors (left) and with lifestyle risk categories based on the number of unhealthy factors (favorable: 0–2, intermediate: 3–5, unfavorable: 6–9) (right) are shown. Squares represent β coefficients, and error bars indicate the corresponding 95% CIs. Panels c–e display β coefficients from linear regression models (adjusted for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level and recruitment center in c and d; adjusted for age, sex, ethnicity, Townsend deprivation index and recruitment center in e). All statistical tests are two-sided. The Benjamini–Hochberg FDR was used to correct for multiple comparisons in c–e. The asterisks denote FDR-adjusted P-value thresholds: *q < 0.05; **q < 0.01; ***q < 0.001. Abbreviations are defined in Supplementary Table 13.

Fig. 3. Brain and peripheral organ aging in cognitive decline and NDs.

a, Associations of brain and peripheral organ (organismal, artery and heart) aging (age gaps) with baseline cognitive function in participants without NDs at baseline (n = 43,141). Associations were estimated by linear regression and presented as β coefficients. Higher reaction times and lower scores in visual memory, fluid intelligence and numerical memory indicate poorer cognitive function. b, Associations of brain and peripheral organ aging with the risk of transitioning from cognitively healthy to MCI (defined as a global cognitive score 1.5 s.d. below the education-adjusted baseline mean in healthy participants) over 8 years of follow-up (n = 39,684). Associations were estimated by logistic regression and presented as ORs. c, Associations of brain and peripheral organ aging with incident NDs and all-cause mortality in participants with baseline MCI over 13 years of follow-up (n = 3,337). d, Associations of brain and peripheral organ aging with incident NDs and all-cause mortality in healthy participants over 13 years of follow-up (n = 43,616). e, Associations between multiple markers (brain age gap, APOE ε4 heterozygotes, AD PRS, age, cognitive function, GDF15 and APOE protein) and the risk of incident AD over 13 years of follow-up (n = 43,616; 611 events for AD). Associations in c–e were estimated using Cox models and presented as HRs. Squares/circles represent effect sizes (β coefficients, ORs or HRs), and error bars indicate the corresponding 95% CIs in a–e. f, Cumulative incidence curves of AD across combined levels of brain age gap and AD PRS. Participants were grouped into five bins based on the combined standardized scores: bin −2 (<−1.5 s.d.), bin −1 (−1.5 to −0.5 s.d.), bin 0 (−0.5 to +0.5 s.d.), bin +1 (+0.5 to +1.5 s.d.) and bin +2 (>+1.5 s.d.). The displayed HR reflects the AD risk per 1-s.d. increase in the combined scores. g, Relative importance of individual proteins in predicting specific disease outcomes. For each disease, Cox models included the top 20 proteins in the brain aging clock, adjusting for covariates. In the left panel, the associations between each protein and incident disease are colored by z-score, with z-scores for associations with a P value of ≥0.05 set to 0. In the right panel, the relative importance of proteins significantly associated with each outcome is displayed. This was calculated as the proportion of each protein’s absolute z-scored β coefficient relative to the total sum of absolute z-scored β coefficients for all proteins significantly associated with the given disease. h, scRNA expression profiles of the top 20 brain aging proteins in the human brain. Mean normalized expression levels are shown across different cell types. Proteins enriched in the GO pathways GO:0099177 (regulation of trans-synaptic signaling) and GO:0061564 (axon development) are denoted in black. Proteins associated with two or more neuropsychiatric diseases are denoted in green. i, Associations of protein levels, bulk RNA expression and scRNA expression with age and AD for key proteins involved in brain aging and dementia risk across tissues (plasma and brain). Associations of plasma protein levels with age or AD were assessed in the UKB (n = 43,616) using linear regression (for age) and Cox models (for incident AD). Associations between gene expression in brain tissue and AD were evaluated using logistic regression based on data from ref. . Results were reported as β coefficients. All models were adjusted for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level and recruitment center in the UKB. All statistical tests are two-sided. The Benjamini–Hochberg FDR was used to correct for multiple comparisons. The asterisks denote FDR-adjusted P-value thresholds: *q < 0.05; **q < 0.01; ***q < 0.001. HC, healthy control.

Fig. 4. Proteins involved in organismal and organ-specific aging clocks and their associations with age, dementia and vascular biology.

a, Associations between age and selected key proteins from organismal and organ-specific (brain, artery and heart) aging clocks that were also associated with dementia. The lines were fitted using Loess regression. Proteins in the organismal aging clock are preferentially colored according to their organ specificity when they are also included in organ-specific aging clocks (for example, ELN). Arterial proteins (for example, ELN and LTBP2), organismal proteins (for example, IGDCC4 and GDF15), as well as NEFL and GFAP showed earlier and steeper age-associated increases than other proteins. b, Summary of the associations between age and the proteins shown in a. Effect estimates from linear regression models (with age as the independent variable and protein levels as the dependent variable) and the corresponding significance levels are shown. c, Protein–protein interaction network identified through STRING analysis. Displayed are the interactions of the featured proteins from a and b, along with their interacting proteins that had a confidence score of ≥0.4. d, Enriched biological pathways among proteins involved in aging and dementia. Functional enrichment analysis was performed using GO terms, and enriched GO terms were identified using a hypergeometric test and corrected for multiple testing. e, Human scRNA expression of featured proteins in the brain and peripheral vasculature. The mean normalized expression and the proportion of cells expressing each gene are shown. These genes are predominantly expressed in endothelial cells, fibroblasts and SMCs in both the brain and peripheral vasculature. f, Levels of pericytes, SMCs, perivascular fibroblasts (P. Fibro) and arterial endothelial cells (Arterial) in patients with AD versus healthy controls (n = 17). Pericytes (P = 0.003), SMCs (P = 0.052) and arterial endothelial cells (P = 0.002) were decreased in AD, as assessed using the t-test. **P < 0.01. Box bounds indicate Q1, median and Q3; whiskers extend to Q1 − 1.5 × IQR and Q3 + 1.5 × IQR. g, Schematic model illustrating the contributions of synaptic and neuronal degradation, glial activation, vascular dysfunction and ECM alterations—as captured by the artery and brain aging clocks—to early cognitive impairments and NDs during biological aging. Panel g created with BioRender.com.

Fig. 5. Brain aging and dementia risk across APOE haplotypes.

a, Associations of APOE haplotypes (2/X, 3/3 (reference), 3/4 and 4/4) and the brain age gap with incident dementia (n = 29,634). Multivariable-adjusted HRs were estimated using Cox models. b, The association between brain aging and dementia was most pronounced among APOE4 homozygotes, as assessed using Cox models (P for interaction = 0.01; n = 29,634). Shown are the multivariable-adjusted HRs per 1-s.d. increase in the brain age gap, stratified by APOE haplotypes. Squares/circles represent HRs, and error bars indicate the corresponding 95% CIs in a and b. c, Cumulative incidence curves of dementia across joint categories of APOE haplotypes (3/3, 4/X) and brain ageotypes (super-youthful (↓), normal (–) and extremely aged (↑)). Compared to participants with APOE 3/3 and normal brain aging, APOE4 carriers with normal and extremely aged brains were at 3.6 and 11.0 times increased risk of dementia, respectively; those with APOE 3/3 and a super-youthful or extremely aged brain were at a 60% lower risk and three times increased risk, respectively. d, Association of age with the brain age gap and featured component proteins across APOE haplotypes. Among APOE4 homozygotes, a steep rise in the brain age gap and elevated levels of proteins implicated in AD pathology were observed between ages 55 and 65 years—approximately 5–10 years before the average age of dementia onset. e, Association of age with the artery age gap and featured component proteins across APOE haplotypes. No notable genotype-specific differences in age trajectories were observed (P for interaction = 0.16). Trajectories were fitted using Loess regression. The shading around the plotted lines in c–e indicates the 95% CI. All models were adjusted for age, sex, ethnicity, Townsend deprivation index and recruitment center.

Fig. 6. Comparison of predictive performance between proteomic aging clocks (brain and organismal) and clinical biomarkers for dementia and mortality.

a–f, Inclusion of proteomic aging clocks modestly improved risk prediction for all incident cases of ACD (a), AD (b) and VD (c), as well as for >10-year incident cases of ACD (d), AD (e) and VD (f). g, Receiver operating characteristic curve analyses for all-cause mortality. Receiver operating characteristic curve analyses based on logistic models were conducted to compare models incorporating basic demographic variables (age, sex and education) and aging clocks, with and without traditional clinical biomarkers. P values indicate the significance of differences in predictive performance with the model that includes the brain age gap, age, sex and education, estimated using 2,000 bootstrap iterations. h, Refined brain aging clocks using a reduced number of proteins selected using RFE. RFE was performed using SHAP values, with models iteratively fitted using fivefold cross-validation, reducing the number of proteins from 70 to 10. The shading around the plotted lines indicates the 95% CI. i, Predictive performance of the refined brain aging clock for dementia and mortality compared to that of the original model. Shown are concordance index (C-index) values across models.

Extended Data Fig. 1. Identification of organ-enriched genes and plasma proteins.

a, Tissue to organ mapping in GTEx. Tissues were mapped to corresponding organs according to the physiological function. Organ-level gene expression was established by identifying the maximum expression value among its tissue subtypes. Expression levels of APOD across tissues were illustrated as an example. b, Organ-level expression of APOD in GTEx. APOD shows a ≥ 4-fold higher expression in the brain than in other organs and is defined as brain-enriched. c, Organ-level fold-change distribution of Olink protein encoding genes in GTEx. d, Organ-level expression of 545 organ-enriched genes encoding Olink proteins in GTEx. Organ-enriched genes are denoted in white. e, Tissue-level expression of identified brain-enriched genes in HPA. f, Tissue-level expression of identified individual brain-enriched genes in HPA. g, Top 10 most FDR significant enrichment biological pathway enrichment for the identified brain-enriched proteins. Functional enrichment analysis was performed using GO terms and enriched GO terms were identified using a hypergeometric test and corrected for multiple testing. h, PC1 correlation pattern of organ-enriched proteins across cohorts. Overlap in numbers of protein signatures between the Olink-based organ aging clock and the SOMAscan-based clock by Oh et al. was shown in Supplementary Fig. 5. Source Data

Extended Data Fig. 2. Performance of organ aging models across cohorts and SHAP visualization of selected proteins.

a-k, Organ aging model performance in the discovery UKB cohort (i), the SHAP values of the top 20 selected proteins (ii), model performance in the external validation CKB cohort (iii), and model performance in the external validation NHS cohort (iv). The width of SHAP plot indicates the contribution of proteins to age and color gradient from blue (low) to red (high) indicates the direction of the protein effect on age. The positive direction on x axis indicates proteins are associated with older age, with the negative direction indicating younger age.

Extended Data Fig. 3. External validation of associations between proteomic organ aging clocks and diseases and mortality in the CKB and NHS.

a, Association of organ aging clocks with incident diseases and mortality in the full sample of the IHD case-cohort study in the CKB (n = 3,977; left panel) and in a subsample excluding enriched IHD patients (n = 2,029; right panel). Cox regression models were adjusted for age, sex, ethnicity, education, study region, smoking, and physical activity level. b, Association of organ aging clocks with incident diseases and mortality in the full sample of the colon cancer case-control study in the NHS (n = 774; left panel) and in a subsample of the control group (n = 387; right panel). Cox regression models were adjusted for age, ethnicity, neighborhood socioeconomic status, smoking, and physical activity level. c, Comparison of the association between organ aging clocks with incident diseases and death across cohorts. To account for potential bias from the case-control design, UKB (n = 43,616) estimates were compared with those from the control groups of the CKB (n = 2,029) and NHS (n = 387). Squares represent HRs, and error bars represent the corresponding 95% CIs. Detailed results are presented in Supplementary Tables 8, 9.

Extended Data Fig. 4. Association of extreme ageotypes across multiple organs with diseases, multimorbidity, and mortality.

a, Overall association of extreme ageotypes (super ager and accelerated ager) with diseases and death (n = 43,616), assessed by Cox models. Accelerated ager and super ager was defined as ± 1.5 standard deviations beside the mean of z-scored age gap for at least one organ, respectively. In the UKB, 25% and 15% of participants had one or multiple extremely aged organs, respectively, while 27% and 20% had one or multiple extremely youthful organs. 14% had both extremely aged and youthful organs. b, Visualization of the organ-specific extreme ageotypes in relation to all-cause mortality, all-cause dementia, depression, and CKD (n = 43,616). Organ-specific models were ordered by the association of accelerated ageotypes with mortality for each outcome. c, Association of number of extreme ageotypes (super/extra) with mortality, dementia, depression, and chronic kidney disease. d, Association of organ age gap with multimorbidity. Multimorbidity is defined as having two or more incident neuropsychiatric diseases (left panel; 3,982 and 1,621 participants with one or two incident neuropsychiatric diseases), two or more other incident chronic diseases studied (middle panel; 6,658 and 2,116 participants with one or two incident physical diseases), or both incident neuropsychiatric and chronic diseases (right panel; 2,044 multimorbidity cases) during follow-up. Squares represent effect sizes (ORs, or HRs) and error bars the corresponding 95% CIs in b-d. All regression models were adjusted for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level, and recruitment center. The asterisks denote FDR-adjusted P value thresholds: *q < 0.05; **q < 0.01; ***q < 0.001. Error bars indicate 95% CIs.

Extended Data Fig. 5. Genetic determinants of brain aging clock with functional mapping and annotation.

a, Manhattan and QQ plots of GWAS on brain aging clock. Independent genome-wide significant SNPs were shown in Supplementary Table 15. b, Manhattan and QQ plots of MAGMA gene-based analysis from GWAS on brain aging clock. Input SNPs were mapped to 19839 protein coding genes. Significant genes are shown (P < 0.05/19839). c, MAGMA tissue enrichment analysis of brain ageing clock-related genes in 30 general tissue types from GTEx v8. d, Enrichment test of prioritized genes in differentially expressed gene (DEG) sets from GTEx v8 30 general tissue types. DEG at both sides are shown. No significant tissue enrichment was observed. e, Overlap and enrichment of input genes in GWAS catalog reported gene set (Supplementary Table 16).

Extended Data Fig. 6. Genetic determinants of organismal aging clock with functional mapping and annotation.

a, Manhattan and QQ plots of GWAS on organismal aging clock. Independent genome-wide significant SNPs were shown in Supplementary Table 17. b, Manhattan and QQ plots of MAGMA gene-based analysis from GWAS on organismal aging clock. Input SNPs were mapped to 19839 protein coding genes. Significant genes are shown (P < 0.05/19839). c, MAGMA tissue enrichment analysis of organismal ageing clock-related genes in 30 general tissue types from GTEx v8. No significant tissue enrichment was observed. d, Enrichment test of prioritized genes in differentially expressed gene (DEG) sets from GTEx v8 30 general tissue types. DEG at both sides are shown. Overlap and enrichment of input genes in GWAS catalog reported gene sets were shown in Supplementary Table 18.

Extended Data Fig. 7. Association of proteomic organ aging clocks with brain structures.

a, Association of proteomic aging clocks with grey matter volume of cortical regions. Number of significant associations with organ aging clocks were colored for each brain cortical region. b, Association of proteomic aging clock with cortical regions GMV. c, Association of proteomic aging clock with subcortical regions GMV, cerebellar regions GMV, and summary image-derived phenotypes (volume of brain, WMV, GMV, and tWMH). d, White matter tracts of interest. e, Association of brain and organismal aging clock with white matter tracts. Generally, brain and organismal aging clock were positively associated with MD and ISOVF, with brain aging clock negatively associated with FA, ICVF, and OD. Associations with brain structures in b, c, and e were estimated by linear regression, presented as beta coefficients. All regression models were adjusted for age, sex, ethnicity, Townsend deprivation index, physical activity level, and recruitment center. All statistical tests are two-sided. FDR using the Benjamini-Hochberg method was used to correct for multiple comparisons. The asterisks denote FDR-adjusted P value thresholds: *q < 0.05; **q < 0.01; ***q < 0.001. FA, fractional anisotropy; ICVF, intracellular volume fraction; ISOVF, isotropic volume fraction; MD, mean diffusivity; OD, orientation dispersion; L, right; R, left; AR, acoustic radiation; ATR, anterior thalamic radiation; CGC, cingulate gyrus part of cingulum; CGH, parahippocampal part of cingulum; CST, corticospinal tract; FMA, forceps major; FMI, forceps minor; IFO, inferior fronto-occipital fasciculus; ILF, inferior longitudinal fasciculus; MCP, middle cerebellar peduncle; ML, medial lemniscus; PTR, posterior thalamic radiation; SLF, superior longitudinal fasciculus; STR, superior thalamic radiation; UNC, uncinate fasciculus.

Extended Data Fig. 8. Organismal and brain aging in mental well-being and psychiatric diseases.

a, Associations of brain and peripheral organ (kidney, intestine, and pancreas) aging (age gaps) with PHQ-4 score and depressive and anxiety symptoms at baseline in participants without neurodegenerative and psychiatric diseases (n = 37,764). Associations were estimated by linear regression, presented as beta coefficients. b, Associations of brain and peripheral organ aging with future mental health conditions at follow-up in participants without neurodegenerative and psychiatric diseases (n = 11,504). c, Associations of brain and peripheral organ aging with risk of incident psychiatric diseases and all-cause mortality in participants with psychological distress (symptoms of depression and anxiety) at baseline over 13 years follow-up (n = 3,958). d, Associations of brain and peripheral organ aging with incident NDs and all-cause mortality in healthy participants over 13 years follow-up (n = 43,616). e, Association between multiple markers (brain age gap, PRS for depression, and age) and risk of incident depression over 13 years follow-up (n = 43,616). Squares represent effect sizes (beta coefficients, ORs, or HRs) and error bars the corresponding 95% CIs in a-e. f, Cumulative incidence curves of depression across combined levels of brain age gap and PRS for depression. Participants were grouped into five bins based on the combined standardized scores: bin -2 (< -1.5 SD), bin -1 (-1.5 to -0.5 SD), bin 0 (-0.5 to +0.5 SD), bin +1 ( + 0.5 to +1.5 SD), and bin +2 (> +1.5 SD). Displayed HR reflects the depression risk per 1 SD increase in the combined scores. g, Change of top proteins in organ-specific (brain, kidney, intestine, and pancreas) aging clock that were also related to depression/anxiety with chronological age. Trajectories with age were fitted by Loess regression. Proteins in organismal model are preferentially colored as organ-specific if they are also included in organ-specific aging clocks. h, Effect of age on proteins in g. i, Protein-protein interaction network identified through STRING analysis. Displayed are interactions of the featured proteins from g and h, and their interacting proteins with a confidence score ≥ 0.4. j, Enriched biological pathways for featured proteins involved in aging and depression/anxiety. All relevant models were adjusted for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level, and recruitment center. Error bars indicate 95% CIs.

Extended Data Fig. 9. Performance of refined brain and organismal aging clocks with reduced number of proteins and their association with diseases and mortality.

a, RFE using SHAP values for brain and organismal aging clocks. Models were fitted iteratively using 5-fold cross-validation starting from full protein panels (230 for organismal aging and 70 for brain aging) down to a single protein. At each step, the protein with the smallest absolute mean SHAP value across folds was removed. The R² for explained variance in chronological age was reported as the average across the five folds. Refined brain clocks with 10 proteins and organismal clocks with 20 proteins were ultimately identified. SHAP values for the refined protein panel are shown. b, Comparison of the predictive performance for chronological age between the full and refined versions of the brain/organismal aging clocks across cohorts. The orange bar denotes the percentage of feature reduction. c-e, Comparison of the association of the full versus refined versions of aging clocks with diseases and mortality in the UKB (n = 43,616), CKB (n = 3,977), and NHS (n = 774). Squares represent HRs and error bars the corresponding 95% CIs. Cox regression models were adjusted for age, sex, ethnicity, Townsend deprivation index, smoking, physical activity level, and recruitment center in UKB; for age, sex, ethnicity, education, study region, smoking, and physical activity in CKB; and for age, ethnicity, neighborhood socioeconomic status, smoking, and physical activity in NHS.