Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

Brain age predicts mortality

Brain age predicts mortality OPEN Molecular Psychiatry (2018) 23, 1385–1392 www.nature.com/mp ORIGINAL ARTICLE 1 2,3 2,4 2,4 2,4 2,4 2,3 2,3 2,5 JH Cole , SJ Ritchie , ME Bastin , MC Valdés Hernández , S Muñoz Maniega , N Royle , J Corley , A Pattie , SE Harris , 6 6,7 3 2,5,7 2 2,3 2,4 1 2,3 Q Zhang , NR Wray , P Redmond , RE Marioni , JM Starr ,SR Cox , JM Wardlaw , DJ Sharp and IJ Deary Age-associated disease and disability are placing a growing burden on society. However, ageing does not affect people uniformly. Hence, markers of the underlying biological ageing process are needed to help identify people at increased risk of age-associated physical and cognitive impairments and ultimately, death. Here, we present such a biomarker, ‘brain-predicted age’, derived using structural neuroimaging. Brain-predicted age was calculated using machine-learning analysis, trained on neuroimaging data from a large healthy reference sample (N = 2001), then tested in the Lothian Birth Cohort 1936 (N = 669), to determine relationships with age-associated functional measures and mortality. Having a brain-predicted age indicative of an older-appearing brain was associated with: weaker grip strength, poorer lung function, slower walking speed, lower fluid intelligence, higher allostatic load and increased mortality risk. Furthermore, while combining brain-predicted age with grey matter and cerebrospinal fluid volumes (themselves strong predictors) not did improve mortality risk prediction, the combination of brain-predicted age and DNA- methylation-predicted age did. This indicates that neuroimaging and epigenetics measures of ageing can provide complementary data regarding health outcomes. Our study introduces a clinically-relevant neuroimaging ageing biomarker and demonstrates that combining distinct measurements of biological ageing further helps to determine risk of age-related deterioration and death. Molecular Psychiatry (2018) 23, 1385–1392; doi:10.1038/mp.2017.62; published online 25 April 2017 INTRODUCTION increased brain atrophy for a given age, may well reflect latent neuropathological influences. A reliable and valid brain-based As the global population ages, the burden of disease is biomarker of ageing, that identifies individuals deviating from increasing. This has motivated research to understand the a healthy brain ageing trajectory, could have great utility in biological links between ageing and disease risk. There is efforts to combat age-associated neurodegeneration and its substantial heterogeneity in how the ageing process affects consequences. different individuals, indicating that people age at different rates, Neuroimaging is a powerful tool for deriving in vivo data on the due to both genetic and environmental influences. If the ageing brain, demonstrating both global and spatially-localised biological characteristics of these different rates of ageing can 10,11 relationships with normal ageing, and with age-associated be measured, then biomarkers of individual differences in the 12–15 cognitive decline. Recently, multivariate methods have been ageing process might help improve predictions of mortality and developed to define statistical models of healthy brain ageing. morbidity. Such biomarkers could potentially identify those at risk Using machine-learning analysis of neuroimaging data, chrono- of age-associated health problems years before symptoms appear, logical age can be accurately predicted in healthy individuals. and be used as outcome measures in trials of therapeutics aimed This provides a method of measuring whether a person’s brain at delaying the onset of age-related disease. Many different appears younger or older than their chronological age. Using this ageing biomarkers have been proposed, which tap into different model, deviations from healthy brain ageing have been iden- cellular and molecular aspects of ageing. For example, the so- 17 18 2,3 tified in Alzheimer’s disease, mild cognitive impairment, called ‘epigenetic clock’ uses measurements of DNA- schizophrenia and have been related to cognitive impairment methylation status at CpG sites across the genome, which can after traumatic brain injury. Furthermore, protective factors have be converted into an age prediction which correlates highly with been associated with a positive influence on brain ageing. For chronological age in healthy individuals. Other candidate ageing 4 5 example, years of education, physical exercise and practicing biomarkers include leucocyte telomere length, N-glycan profile meditation were recently linked to having younger-appearing and Ink4a/Arf locus expression. This diverse list of candidate 21,22 ageing biomarkers reflects the involvement of multiple biological brains. As these multivariate neuroimaging measures have been systems and the overall complexity of the ageing process in associated with age-related pathology and cognitive impairment, humans. Neurological aspects of ageing, such as cognitive decline and this raises the possibility that brain-based age predictions could dementia, are particularly deleterious to general health and well- be used as an ageing biomarker. A viable ageing biomarker must 8 9 23 being. Brain structure is well-known to alter throughout life, and relate to the risk of mortality and age-associated morbidity, deviations from this typical brain ageing trajectory, in terms of particularly if it is to have clinical utility. To establish what the 1 2 Computational, Cognitive and Clinical Neuroimaging Laboratory, Department of Medicine, Imperial College London, London, UK; Centre for Cognitive Ageing and Cognitive 3 4 Epidemiology, University of Edinburgh, Edinburgh, UK; Department of Psychology, University of Edinburgh, Edinburgh, UK; Brain Research Imaging Centre, Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, UK; Centre for Genomic and Experimental Medicine, MRC Institute of Genetics & Molecular Medicine, University of 6 7 Edinburgh, Edinburgh, UK; Institute for Molecular Bioscience, The University of Queensland, QLD, Australia and Queensland Brain Institute, The University of Queensland, QLD, Australia. Correspondence: Dr JH Cole, Medicine, Imperial College London, Computational, Cognitive and Clinical Neuroimaging Laboratory, Burlington Danes Building, Du Cane Road, London W12 0NN, UK. E-mail: [email protected] Received 4 October 2016; revised 18 January 2017; accepted 17 February 2017; published online 25 April 2017 Brain age predicts mortality JH Cole et al consequences of having a brain that appears older or younger Table 1. Lothian Birth Cohort 1936 characteristics than average for one’s chronological age, we estimated ‘brain- predicted age’ in a large, narrow age-range population cohort of All Male Female older adults (Lothian Birth Cohort 1936 (LBC1936), N = 669), using structural neuroimaging (T1-weighted magnetic resonance ima- N 669 352 317 Age 72.67 (0.73) 72.63 (0.71) 72.72 (0.74) ging (MRI)). We tested the association between brain-predicted Mini-mental state 29 (2) 29 (2) 29 (2) age difference (brain-PAD; calculated as brain-predicted age examination minus chronological age) and: mortality risk, disease prevalence, (median (IQR)) measures of physical and mental fitness (grip strength, walking Brain-predicted age 74.32 (8.72) 76.92 (8.64) 71.43 (7.88) speed, lung function and general fluid intelligence), and a Brain-PAD 1.65 (8.71) 4.29 (8.58) − 1.29 (7.87) composite measure of biological health (allostatic load). We g 0.03 (0.98) 0.01 (1.05) 0.06 (0.90) hypothesised that ‘older’ brain-PAD would be associated with Grip strength 28.79 (9.33) 35.38 (6.71) 21.45 (5.63) earlier mortality and more morbidity, poorer physical and FEV (l) 2.34 (0.68) 2.72 (0.62) 1.92 (0.44) cognitive fitness, and greater allostatic load. 6 metre walk time (s) 4.29 (1.21) 4.09 (1.11) 4.51 (1.27) Allostatic load − 0.03 (0.99) 0.09 (0.95) −0.15 (1.02) Further, it has been proposed that biological ageing occurs at Deceased (N) 73 43 30 different rates to different tissues or cells within the same person, the so-called ‘mosaic of ageing’. Hence, complementary Abbreviations: brain-PAD, brain-predicted age difference; FEV, forced information could be gained by combining ageing biomarkers expiratory volume in one second; g , fluid type general intelligence; IQR, derived from different sources. Conversely, if ageing occurs inter-quartile range. Mortality was ascertained between 5.4 and 7.9 years after neuroimaging assessment. Values reported are mean (s.d.) unless uniformly across the body, then diverse ageing biomarkers should otherwise specified. correlate highly. Here we explored these possibilities by examin- ing brain-predicted age in relation to molecular-genetic ageing biomarkers. We tested brain-PAD in combination with DNA- methylation-based age predictions using the ‘epigenetic-clock’ and leukocyte telomere length, both previously associated with and WM images were concatenated and converted into a similarity matrix 25,26 mortality, examining their influence on the relationship with of training subjects’ data, which to predict chronological age in a Gaussian age-related outcome measures. Finally, we considered how brain- Process regression. Model accuracy was then validated using ten-fold PAD related to more conventional imaging measures, previously cross-validation, comparing brain-predicted age with chronological age. shown to relate to ageing. The coefficients ‘learned’ from the full model (N = 2001) were then applied to the test data (LBC1936, N = 669) to make brain-based age predictions for these individuals. Brain-PAD (predicted age—chronological age) was then MATERIALS AND METHODS calculated and used for further statistical analysis. Full details of the participants, data acquisition and statistical methods used in the study are included in Supplementary Methods. Ageing fitness measures Five measures of ‘fitness’, or a healthy ageing phenotype, in older age Participants—Lothian Birth Cohort 1936 were considered: walking speed (time to walk 6 metres), right-hand grip strength (measured by a dynamometer), lung function (forced expiratory The LBC1936 is a longitudinal study of ageing based in the Edinburgh 27,28 and Lothians area of Scotland, UK. Most of the participants had volume in 1 s), cognitive function (fluid-type intelligence) and allostatic taken part in the Scottish Mental Survey of 1947, which involved a test of load. Allostatic load was derived from measures of: fibrinogen, general cognitive ability for almost all 11-year old children in the triglyceride, high-density lipoprotein, low-density lipoprotein, total choles- country at that time. At the first wave, 1,091 participants attended for terol, cholesterol high-density lipoprotein ratio, glycated haemoglobin, cognitive and medical testing (mean age 70 years, 548 = male, C-reactive protein, interleukin-6, body-mass index and blood pressure. All 543 = female). MRI testing began at the second wave, when 866 individuals measures used in the present analysis were collected at the same time as attended for cognitive, medical testing (mean age 73 years, 448 = male, the neuroimaging assessment. 418 = female), of whom 669 (352 = male, 317 = female) had MRI. This final cohort provided the data that were included in present analysis (Table 1). Mortality ascertainment The vast majority of participants were cognitively normal according to Mortality status was obtained via data linkage to the National Health mini-mental state examination, with 99.3% scoring ⩾ 24. Ethical approval Service Central Register, provided by the National Records of Scotland. for the LBC1936 was obtained from the Multi-Centre Research Ethics The LBC1936 research team are routinely informed of participant Committee for Scotland (MREC/01/0/56) and the Lothian Research Ethics deaths and cause of death approximately every 12 weeks. Most Committee (LREC/2003/2/29). Written informed consent was obtained recent ascertainment was at approximately age 79 years (range 78.7– from all subjects. 79.7 years), which was between 5.4 and 7.9 years after neuroimaging assessment. Participants—brain-predicted age training cohort Further 2001 healthy individuals (age mean = 36.95 ± 18.12 years; age Molecular genetic biomarkers of ageing range = 18–90 years; males = 1016; females = 985) comprised the brain- Using whole blood samples, data for two candidate ageing biomarkers predicted age training cohort. These data were obtained via were generated. The ‘epigenetic clock’ was used to calculate predictions publicly-available repositories (Supplementary Table 5) and were screened of age based on DNA-methylation status at 450, 726 autosomal sites across according to local study protocols to ensure that they were 26,32 the genome, as per the previously reported ‘Horvath’ protocol. free of neurological and psychiatric disorders, had no history of head Leukocyte telomere length was measured using a protocol developed at trauma and other major medical conditions. Ethical approval for each University of Newcastle. initial study and subsequent data sharing was verified for each data repository. RESULTS Brain age prediction methods Chronological age can be predicted using neuroimaging The machine learning age-predictions methods using neuroimaging data A machine-learning model (Gaussian Processes), trained on the are outlined in Figure 1. Briefly, T1-weighted MRI scans were segmented brains of N = 2001 healthy adults, aged 18–90 years, can accurately into grey matter (GM) and white matter (WM) before being normalised in common space using non-linear spatial registration. Once normalised, GM predict chronological age using T1-weighted MRI scans Molecular Psychiatry (2018), 1385 – 1392 Brain variable y Model coefficients JH Cole et al Training Validation Accuracy T1-MRI with age label assessment GPR model Gaussian Process regression to predict age Defining brain-PAD Positive brain-PAD score = ‘older’ brain Chronological 50 Age 10-fold cross-validation Brain-PAD = 0, matched chronological and predicted age Testing Negative brain-PAD score = ‘younger’ brain Trained GPR Chronological Age model 72.9 Age prediction generated i i i i i i N = 2001 N = 669 ? ? ? ? ? Figure 1. Overview of study methods. Illustration of the methods used to generate brain-predict ages. 3D T1-weighted MRI scans were segmented into grey and WM before being normalised in common space using non-linear spatial registration. Normalised grey and WM images were concatenated and converted into vectors for each subject. These vectors were then projected into an NxN similarity matrix based on vector dot-products. (a) Once in similarity matrix form the training subjects’ data were used as predictors in a Gaussian Processes regression (GPR) with age as the outcome variable. (b) Model accuracy was assessed in a ten-fold cross-validation procedure, comparing brain- predicted age with original chronological age labels. (c) Model coefficients learned during training were then applied to the data from LBC1936 participants to make age predictions. (d) A metric to summarise the variation in predicted age was defined; the brain-predicted age difference (brain-PAD; predicted age—chronological age). LBC1936, Lothian Birth Cohort 1936; MRI, magnetic resonance imaging; WM, white matter. 20 40 60 80 50 60 70 80 90 100 Age (years) Age/Brain-predicted age (years) Figure 2. Brain-predicted age using structural neuroimaging in LBC1936. (a) Scatterplot showing the relationship between chronological age and brain-predicted age in the independent healthy cohort used as the training data (green diamonds) and the LBC1936 participants used as the test set (red circles). (b) Histogram showing the distributions of brain-predicted age (in blue) compared to the distribution in chronological age (in red). The substantially wider variability in brain-predicted age is evident. LBC1936, Lothian Birth Cohort 1936. (Figure 2a). Cross-validation results gave a correlation between Older adults show marked variation in structural brain ageing brain-predicted age and chronological age of r = 0.94, (Po0.001, The model coefficients ‘learned’ from the training dataset were corrected after 1000 permutations) and explained 88% of the applied to T1-weighted MRI scans acquired from the LBC1936 variance (R ). The mean absolute error of prediction was 5.02 years participants (Table 1) to generate a brain-predicted age. At the and the root mean square error was 6.31 years. This training stage time of scanning LBC1936 participants had a mean chronological validated our model of brain-predicted age, for use in predicting age of 72.67 (s.d. = 0.73) years and a mean brain-predicted age of age with neuroimaging data collected in other samples. 74.32 (s.d. = 8.72) years. The mean absolute error of age prediction Molecular Psychiatry (2018), 1385 – 1392 Brain variable x Brain-Predicted Age (years) Frequency Brain-Predicted Age Brain-Predicted Age Brain variable z Brain age predicts mortality JH Cole et al 1.00 1.0 Female/Alive Female/Deceased Male/Alive 0.69 Male/Deceased 0.8 0.95 0.66 0.59 0.90 0.6 0.50 0.85 0.4 Brain-PAD 0.80 0.2 −10 DNAm-PAD high Brain-PAD Brain-PAD + DNAm-PAD low Brain-PAD Telomere length 0.75 0.0 −20 72 74 76 78 80 1.0 0.8 0.6 0.4 0.2 0.0 Age (years) Specificity Sex/Deceased Figure 3. Association of mortality with brain-predicted age difference and DNA-methylation-predicted age difference. (a) Grouped scatterplot showing the relationship between brain-predicted age difference (brain-PAD) score (i.e., brain-predicted age—chronological age) and mortality (alive= blue, dead = red), sub-divided by sex (female = circle, male= triangle). Mortality status was determined ~ 6 years after MRI assessment. Horizontal black lines represent the median for each sub-group. (b) Kaplan–Meier plot of right-censored survival data since MRI assessment. The two lines represent a tertile split based on brain-PAD score, with highest 33.3% being classed as high brain-PAD (red line) indicating increased brain ageing and the lowest 33.3% (low brain-PAD, blue line) indicating reduced brain ageing. Crosses indicate censoring points (i.e. age at last survival ascertainment). Dotted lines represent the 95% confidence intervals. (c) Figure depicts the receiver operator characteristic (ROC) curves for four contrasting, nested, survival models. All models used mortality status as the response variables. The predictor variables were Brain-PAD (red line, model 1), DNAm-PAD (blue line, model 5), Brain-PAD+DNAm-PAD (green line, model 4), Telomere length+Brain-PAD+DNAm-PAD (grey line, model 3). The areas under the curve (AUC) are coloured-coded and appear next to each ROC curve. MRI, magnetic resonance imaging. in the LBC1936 participants was 7.08 years and the root mean other variables previously related to mortality in this sample was square error was 8.85 years. The variability in brain-predicted age considered in a fully-adjusted model, as per Marioni and was considerably greater than the variability in chronological age, colleagues. These were: Moray House Test IQ-type score at reflecting marked individual differences in brain structure in age 11, paternal social class (five-point scale), years of full-time participants aged approximately 73 (Figure 2b). As expected, education, APOE e4 carrier status, smoking status (never, brain-PAD scores did not correlate with chronological age ex-smoker, current smoker), and self-reported hypertension, (r = − 0.01, P = 0.79), indicating that deviations from healthy brain diabetes and cardiovascular disease. Brain-PAD remained signifi- ageing (that is, having an older- or younger-appearing brain) were cantly associated with mortality risk in this fully-adjusted model, not related to underlying chronological age. Females’ brain- with a slight attenuation of the effect size (HR = 1.051, 95% predicted ages were, on average, younger than their chronological CI = 1.020, 1.083, Po0.001; Supplementary Table S1). age (mean (s.d.) brain-PAD = − 1.29 (7.87) years), whereas males’ were older (mean (s.d.) brain-PAD = 4.29 (8.58) years). This sex Variability in apparent brain-ageing relates to physical and mental difference was statistically significant (Wilcoxon rank-sum test, fitness W = 35431, Po0.001), hence sex was included as a covariate in all Brain-PAD score was also significantly related to a number of further analyses. measures that reflect characteristics of physical and mental fitness in older age using linear regression (Supplementary Table 2). An Early mortality is associated with older-appearing brains older-appearing brain, as indicated by a higher brain-PAD score, Having a higher brain-PAD score (that is, a brain that appears was significantly associated with lower fluid cognitive perfor- older than one’s chronological age) was significantly associated mance (standardised beta = − 0.121, P = 0.007), weaker grip with mortality before the age of 80 (Po0.001); up to seven years strength (standardised beta = − 0.060, P = 0.020), poorer lung func- after neuroimaging assessment. Mean brain-PAD score for tion (standardised beta = − 0.072, P = 0.020) and slower walking deceased males (N = 43) and females (N = 30) was 8.13 (s.d. = speed (standardised beta = 0.133. P = 0.004). Higher brain-PAD 9.52) and 2.07 (s.d. = 9.27) years, respectively, compared to 3.76 score was also associated with higher allostatic load (standardised (s.d. = 8.32) and − 1.64 (s.d. = 7.65) years for surviving males and beta = 0.097, P = 0.020), a composite measure of biological and females (Figure 3). The relationship between mortality risk and physiological parameters, designed to reflect biological ‘wear-and- brain-PAD was tested using Cox proportional hazards regression tear’ accumulated over a lifetime of stress adaptation. Reported analysis, adjusting for age and sex. Survival was ascertained up to P-values were corrected for five tests using a 5% false discovery 7.9 years post-neuroimaging, and survival duration was right- rate. censored for surviving individuals based on days between neuroimaging assessment and mortality ascertainment. Each extra Brain-PAD is not related to the prevalence of morbidity year of brain-predicted age (that is, having a brain-PAD score of Next, we examined the relationship between brain-PAD and the +1) resulted in a 6.1% relative increase in the risk of death presence of self-reported cardiovascular disease, stroke, and between age 72 and 80 (hazard ratio (HR) = 1.061, 95% confidence diabetes. LBC1936 participants reported the following prevalence interval (CI) = 1.031, 1.091, Po0.001). The assumptions of propor- tional hazards were met by the model. An illustrative Kaplan– of disease: cardiovascular disease = 26.9% (N = 180), diabetes = Meier plot using the upper and lower tertiles of brain-PAD scores 10.2% (N = 68) and a history of stroke = 6.9% (N = 46). After in LBC1936 participants is shown in Figure 2b. The influence of adjusting for sex, there was no significant association between Molecular Psychiatry (2018), 1385 – 1392 Female/Alive Female/Deceased Male/Alive Male/Deceased Brain-PAD (years) Proportion survived Sensitivity Brain age predicts mortality JH Cole et al brain-PAD score and cardiovascular disease (P = 0.08), diabetes (635.2) base-pairs, while for males it was 3912.3 (783.3) base-pairs, (P = 0.14) or stroke (P = 0.85). which was significantly different (W = 45386, P = 0.001). There was no significant association between telomere length and brain-PAD (rho = 0.04, P = 0.31) or brain-predicted age (rho = 0.05, P = 0.23). Brain-PAD is not related to childhood IQ, life-course social factors Combining DNAm-PAD and telomere length with brain-PAD in or APOE e4 status a multivariate Cox regression, adjusted for age and sex, Brain-PAD was also not associated with potential life-course significantly predicted survival (N = 608, deceased N = 67, influences on ageing. Potential influences tested were: perfor- Po0.001). Within this model, brain-PAD (HR = 1.07, 95% CI = 1.04, mance on the Moray House Test at age 11 (P = 0.63), paternal 1.11, Po0.001) and DNAm-PAD (HR = 1.06, 95% CI = 1.02, 1.10, social class (P = 0.82), years of education (P = 0.45), neighbourhood Po0.001) were significant contributors to the prediction, while deprivation as indexed by the Scottish Index of Multiple telomere length was not (P = 0.97). A separate model using DNAm- Deprivation (P = 0.45), and the presence of an APOE e4 allele PAD alone also significantly predicted survival (HR = 1.06, 95% (P = 0.88). CI = 1.02, 1.09, Po0.001); however, this explained significantly less variance than a model using brain-PAD alone (AUC = 0.59 vs brain- Brain-PAD and conventional neuroimaging measures in relation to PAD alone AUC = 0.66, Po0.001). The combined model using survival brain-PAD and DNAm-PAD explained significantly more variance than either variable alone (combined model AUC = 0.69 vs brain- Brain-PAD was significantly correlated (positively or negatively) PAD alone AUC = 0.66 vs DNAm-PAD alone AUC = 0.59, Po0.001; with: GM, normal-appearing white matter, cerebrospinal fluid (CSF) see Figure 2c, Supplementary Table 4). This was also the case for and WM hyperintensity volume, whole brain cortical thickness, the fully-adjusted model covarying for potential influences on fractional anisotropy and mean diffusivity (Supplementary Figure 1). mortality risk. Prediction of ageing fitness measures was not When combining brain-PAD with these imaging measures to improved when combining brain-PAD and DNAm-PAD or brain- predict outcomes, brain-PAD contributed unique variance (deter- PAD and telomere length. mined using hierarchical variance partitioning) to each linear regression model (Supplementary Table 2). Although brain-PAD was not always the greatest contributor of unique variance to DISCUSSION outcome prediction, this analysis indicates that brain-PAD can add Here we showed that a neuroimaging-based marker of brain complementary information to models of age-related outcomes, ageing is associated with a greater risk of death and poorer over and above that gained from conventional neuroimaging physical and cognitive fitness in a large cohort of older adults. measures. Further, we assessed whether GM, normal-appearing Furthermore, we demonstrate that combining biological age white matter and CSF volume were associated with survival. Cox predictions generated from neuroimaging and DNA-methylation regression analyses, adjusted for age and sex, indicated that GM status data increases the accuracy of survival modelling. At ~ 73 and CSF volume (in ml) were associated with survival (GM: years of age, we found that people with brains that appeared HR = 0.991, 95% CI = 0.984, 0.998, P = 0.007; CSF: HR = 1.012, 95% older than their chronological age had, in addition to greater CI = 1.007, 1.017, Po0.001), where a having 1 ml lower GM or 1 ml mortality risk: weaker grip strength, poorer lung function, slower higher CSF volume was associated with a 1% increase in mortality walking speed, lower fluid general intelligence, and had been risk. Normal-appearing white matter volume was not associated exposed to greater allostatic load (a biological measure intended with mortality risk (P = 0.54). We then compared the predictive to summarise the cumulative effects of lifetime biological ‘wear value of linear combinations of brain-PAD with GM and CSF and tear’). The relationship between brain-PAD and survival was volume in Cox regression models (that is., brain-PAD+GM volume, independent of life-course influences on mortality including: brain-PAD+CSF volume, brain-PAD+GM volume+CSF volume). education, social class, childhood IQ, carrying an APOE e4 allele or Brain-PAD significantly related to survival in the paired combined the presence of age-associated illness. Furthermore, these factors models (Po0.05), indicating that it independently explained some were themselves not significantly associated with brain-PAD in variance relating to survival, when combined with GM volume or this sample. with CSF volume separately. However, when included alongside To the best of our knowledge, this is the first demonstration both GM volume and CSF volume, brain-PAD was no longer a that a neuroimaging-derived age prediction is associated with significant predictor of survival (P = 0.12), while the volumetric higher mortality risk. Such measures have been used in clinical measures remained significant (GM: z = − 3.78, Po0.001; CSF: samples, showing increased apparent brain age following trau- z = 4.56, Po0.001). For full details see Supplementary Table 3. matic brain injury and in individuals with mild cognitive, a key 18 21 risk factor for Alzheimer’s Disease. Higher levels of exercise Brain-PAD combined with DNA-methylation ‘age’ improves and meditation have been associated with lower brain age in survival modelling the healthy population, but the link with mortality is novel. This is Molecular genetic ageing biomarkers have also been proposed, crucial, as it supports the use of MRI as a screening tool to help hence we compared brain-PAD with DNA-methylation status and identify people at greater risk of general functional decline and leukocyte telomere length. DNA-methylation (DNAm) age was mortality during ageing. Brain-PAD has the potential to be predicted using Horvath’s ‘epigenetic clock’ method, in N = 620 estimated in large numbers of people, as MRI is collected routinely (female = 290, male = 330) participants, who had undergone in clinical settings. The success of projects like UK Biobank shows neuroimaging assessment. Mean DNAm-predicted age was 69.3 that acquiring MRI on a very large scale is feasible given the (s.d. = 6.2) years. Mean DNAm-predicted age difference (DNAm- appropriate infrastructure. PAD) was − 3.4 (s.d. = 6.1) years. Mean DNAm-PAD was similar for The combination of DNA-methylation-predicted age and males (−3.2, s.d. = 5.8) and females (−3.7, s.d. = 6.4), with no neuroimaging-predicted age is also novel. We found that, while statistically significant sex difference (W = 44803, P = 0.17). There brain-predicted age significantly out-performed DNA-methylation was no association between the DNAm-predicted age and brain- predicted age, there is greater added value gained when predicted age (rho = 0.001, P = 0.98) or between brain-PAD and combining these two approaches to predict survival. Previously, DNAm-PAD (rho = − 0.007, P = 0.85). Regarding telomere length, ‘DNA-methylation age’ has been related to mortality and 26,32 data were available in N = 653 participants (female = 309, male = ageing fitness, and in various clinical contexts including 35 36 37 344) with neuroimaging data. Telomere mean (s.d.) length was HIV, Down’s Syndrome and obesity. Interestingly, brain-PAD 3982.5 (711.7) base-pairs. Telomere length in females was 4045.5 and DNAm-PAD were not correlated, yet both related to survival Molecular Psychiatry (2018), 1385 – 1392 Brain age predicts mortality JH Cole et al 47–50 independently and improved survival prediction when analysed in with mortality. As proxies for systemic health (for example, combination, thus provided complementary information. This musculo-skeletal, respiratory, nervous, circulatory), they appear to demonstrates that contrasting approaches to estimating age relate to a common aspect of more general health of the whole biologically can be integrated to predict clinically-relevant out- body, likely due to the interactions between different human comes. Seemingly, epigenetic ageing in leukocytes and ageing biological systems. However, there also seems to be unique of brain structure are occurring independently, perhaps evidence variance in the relationship of these measures with mortality. This for a ‘mosaic’ of ageing, where biological ageing occurs at is supported by our finding that survival modelling accuracy was different rates in different systems or compartments within an improved when including multiple ageing fitness measures individual. This motivates further research that combines inde- alongside brain-PAD. Notably, brain-PAD remained the strongest pendent measures of biological ageing to develop a more global predictor in this combined model, which justifies further research ageing biomarker, which may further improve predictions of into the clinical applications of neuroimaging-based predictors of survival. mortality. Other neuroimaging measures have previously been associated Our study has some strengths and weaknesses, particularly with mortality in older adult population cohorts. These include relating to the cohorts under study. The sample size for both WM hyperintensities in adults aged 70–82 years, regional training and test sets is relatively large. One potential limitation is volume reductions at age 85 years and whole brain volume at the multiple sources of training data. Comprehensive demo- 40 41 42 78–85, 66–90 and 60–90 years. Visual assessment of infarcts, graphic data were not available on all these individuals. However, WM hyperintensities and atrophy also predicted mortality individuals in this sample were screened according to various 6 months after a stroke. This research supports the idea that criteria to ensure that were free of manifest neurological, the brain plays a central role in the ageing process and is sensitive psychiatric or major medical health issues. The LBC1936 is well- to the cumulative damage that accrues throughout life and characterised, allowing a broad exploration of relationships with increases mortality risk. That we can predict mortality before the brain-predicted ageing, particularly the follow-up to assess age of 80 using neuroimaging assessment at approximately age mortality. The limited age range of LBC1936 participants is a 73, fits with these previous reports. strength in that it eliminates the important confounding effect of Interestingly, when combining brain-PAD with GM and CSF chronological age, but it may limit generalisations to other age volume in a Cox regression, brain-PAD no longer significantly groups. However, this point in the life course is a timely juncture predicted survival. This indicates that the survival-related variance to assess brain ageing as individual differences have had time to in brain-PAD can be captured using more conventional volumetric accumulate though are unlikely to be widely confounded by measures. While brain-PAD did incrementally improve survival manifest neurodegenerative disease. The current analysis was prediction over individual volumetric measures, our results cross-sectional; therefore, we cannot determine whether the indicate that using a combination of GM and CSF volume is relationship between brain-predicted age and mortality risk varies potentially a strong biomarker of mortality. Nevertheless, these with age or where on the trajectory of atrophy an individual is. The volumetric measures appear less suitable as an ageing biomarker on-going nature of the LBC1936 study will allow future analysis of per se, as a linear model of GM, WM and CSF volume explained longitudinal data to determine whether trajectories of brain only 66% of variance in chronological age (mean absolute ageing are better indicators of future health outcomes than cross- error = 8.30 years, root mean square error = 10.53) in the training sectional measures. In addition, we only assessed all-cause dataset, compared with 88% using brain-predicted age. This mortality, which limits speculation about causal relationships demonstrates that in the context of developing an ageing between brain structural alterations and specific mortality causes, biomarker, there is benefit in using a machine-learning approach such as cardiovascular or neurological causes of death. Finally, the to analyse high-dimensional voxelwise T1-MRI data, compared to LBC1936 participants were not fully representative of the macroscopic volume measurements. Future steps to further population from which they were drawn. Compared to the full improve models of brain ageing and derived ageing biomarkers population who sat the cognitive test at age 11, LBC1936 could incorporate additional imaging modalities at the modelling participants had higher cognitive ability, and in later life were stage. This should capture further age-associated changes likely to be healthier than their peers in the general population. including WM hyperintensities using FLAIR-MRI, altered WM This may have been due to selection effects seen in most studies microstructure using diffusion-MRI and beta-amyloid deposition of ageing. That our sample might have missed individuals with using positron emission tomography. particularly poor health or high frailty. Hence we might have A key medical research goal is to identify reliable predictors of underestimated some of the effects reported here, as a small mortality, proxy measures of underlying pathological processes number of individuals with worse performance on measures of that increase mortality risk. For example, grip strength has been ageing fitness may not have been included. 44,45 robustly associated with mortality, and is thought to be a The difference between neuroimaging-predicted age and proxy of the musculo-skeletal system. Similarly, brain-PAD may be chronological age is associated with survival in a large sample a general reflection of CNS health. Grip strength measures do not of older adults and relates to measures of cognitive and physical necessarily require a direct causal link with cardiovascular or all- fitness. Moreover, combining age-predictions from DNAm and cause mortality to be clinically useful; the same could apply to neuroimaging data increased the accuracy of survival modelling. brain-PAD. Moreover, the relevance of brain-predicted age for This study provides evidence that neuroimaging data can be used health is intuitively straightforward. Already, the UK National to construct a viable ageing biomarker, and potentially provides Health Service encourages people to complete a cardiovascular important prognostic information, particularly in combination with health question to determine their ‘heart age’ (www.nhs.uk/ complementary epigenetic ageing data. A global biomarker of Conditions/nhs-health-check/Pages/check-your-heart-age-tool. ageing has the potential to screen for asymptomatic individuals aspx). By analogy, ‘brain age’, or a global ‘biological age’, could be who are experiencing adverse ageing and thus are at increased used in public health settings to convey complex information to risk of future ill-health and could be used as a surrogate outcome patients in readily comprehendible terms. measure in clinical trials of neuroprotective treatments and anti- Brain-PAD related to all measures of ageing fitness. This ageing therapeutics. suggests that our measure of brain ageing relates to some more general facets of physiological ageing. Along with grip CONFLICT OF INTEREST strength, all these measures (lung function, gait speed, cognitive function and allostatic load) have been previously associated The authors declare no conflict of interest. Molecular Psychiatry (2018), 1385 – 1392 Brain age predicts mortality JH Cole et al ACKNOWLEDGMENTS 21 Steffener J, Habeck C, O'Shea D, Razlighi Q, Bherer L, Stern Y. Differences between chronological and brain age are related to education and self-reported physical We thank the Lothian Birth Cohort 1936 participants, radiographers at the Brain activity. Neurobiol Aging 2016; 40: 138–144. Research Imaging Centre (www.bric.ed.ac.uk/) and the researchers who collected and 22 Luders E, Cherbuin N, Gaser C. Estimating brain age using high-resolution pattern entered data used in this manuscript. This research and LBC1936 phenotype recognition: Younger brains in long-term meditation practitioners. Neuroimage collection were supported by Research Into Ageing and continues as part of The 2016; 134: 508–513. Disconnected Mind project (http://www.disconnectedmind.ed.ac.uk), funded by Age 23 Sprott RL. Biomarkers of aging and disease: introduction and definitions. UK. MRI acquisition and analyses were conducted at the Brain Research Imaging Exp Gerontol 2010; 45:2–4. Centre, Neuroimaging Sciences, Centre for Clinical Brain Sciences, University of 24 Cevenini E, Invidia L, Lescai F, Salvioli S, Tieri P, Castellani G et al. Human models Edinburgh, which is part of SINAPSE (Scottish Imaging Network—A Platform for of aging and longevity. Expert Opin Biol Ther 2008; 8: 1393–1405. Scientific Excellence) collaboration (www.sinapse.ac.uk/) funded by the Scottish 25 Cawthon RM, Smith KR, O'Brien E, Sivatchenko A, Kerber RA. Association between Funding Council and the Chief Scientist Office. This work was undertaken within the telomere length in blood and mortality in people aged 60 years or older. Lancet University of Edinburgh Centre for Cognitive Ageing and Cognitive Epidemiology 2003; 361: 393–395. (http://www.ccace.ed.ac.uk), part of the cross council Lifelong Health and Wellbeing 26 Marioni RE, Shah S, McRae AF, Chen BH, Colicino E, Harris SE et al. DNA methy- Initiative (MR/K026992/1), for which funding from the BBSRC and MRC is gratefully lation age of blood predicts all-cause mortality in later life. Genome Biol 2015; acknowledged. Simon Cox is supported by an MRC grant (MR/M013111/1). We would 16:25. like to thank Thomas von Zglinicki and Carmen Martin-Ruiz for their work on 27 Deary IJ, Gow AJ, Pattie A, Starr JM. Cohort profile: the Lothian Birth Cohorts of generating the telomere length data. Naomi Wray acknowledges funding from the 1921 and 1936. Int J Epidemiol 2012; 41:1576–1584. Australian National Health and Medical Research Council (613608, 1078901). 28 Deary IJ, Gow AJ, Taylor MD, Corley J, Brett C, Wilson V et al. The Lothian Birth Cohort 1936: A study to examine influences on cognitive ageing from age 11 to age 70 and beyond. BMC Geriatr 2007; 7:28. REFERENCES 29 Scottish Council for Research in Education. The Trend of Scottish intelligence: A 1 Vos T, Flaxman AD, Naghavi M, Lozano R, Michaud C, Ezzati M et al. Years lived Comparison of the 1947 and 1932 Surveys of the Intelligence of eleven-year-old with disability (YLDs) for 1160 sequelae of 289 diseases and injuries 1990–2010: a Pupils. University of London Press: London, 1949. systematic analysis for the Global Burden of Disease Study 2010. Lancet 2012; 30 Lara J, Godfrey A, Evans E, Heaven B, Brown LJE, Barron E et al. Towards mea- 380: 2163–2196. surement of the Healthy Ageing Phenotype in lifestyle-based intervention studies. 2 Bocklandt S, Lin W, Sehl ME, Sánchez FJ, Sinsheimer JS, Horvath S et al. Epigenetic Maturitas 2013; 76:189–199. predictor of age. PLoS ONE 2011; 6: e14821. 31 Booth T, Starr JM, Deary I. Modeling multisystem biological risk in later life: allostatic 3 Horvath S. DNA methylation age of human tissues and cell types. Genome Biol load in the lothian birth cohort study 1936. Am J Hum Biol 2013; 25:538–543. 2013; 14: R115. 32 Marioni RE, Shah S, McRae AF, Ritchie SJ, Muniz-Terrera G, Harris SE et al. The 4 Kruk PA, Rampino NJ, Bohr VA. DNA damage and repair in telomeres: relation epigenetic clock is correlated with physical and cognitive fitness in the Lothian to aging. Proc Natl Acad Sci USA 1995; 92:258–262. Birth Cohort 1936. Int J Epidemiol 2015; 44: 1388–1396. 5 Vanhooren V, Dewaele S, Libert C, Engelborghs S, De Deyn PP, Toussaint O et al. 33 Harris SE, Martin-Ruiz C, von Zglinicki T, Starr JM, Deary IJ. Telomere length and Serum N-glycan profile shift during human ageing. Exp Gerontol 2010; 45: aging biomarkers in 70-year-olds: the Lothian Birth Cohort 1936. Neurobiol Aging 738–743. 2012; 33: 1486.e1483–1488. 6 Krishnamurthy J, Torrice C, Ramsey MR, Kovalev GI, Al-Regaiey K, Su L et al. Ink4a/ 34 Miller KL, Alfaro-Almagro F, Bangerter NK, Thomas DL, Yacoub E, Xu J et al. Arf expression is a biomarker of aging. J Clin Invest 2004; 114: 1299–1307. Multimodal population brain imaging in the UK Biobank prospective epidemio- 7 Lopez-Otin C, Blasco MA, Partridge L, Serrano M, Kroemer G. The hallmarks logical study. Nat Neurosci 2016; 19: 1523–1536. of aging. Cell 2013; 153: 1194–1217. 35 Horvath S, Levine AJ. HIV-1 infection accelerates age according to the 8 Tucker-Drob EM. Neurocognitive functions and everyday functions change epigenetic clock. J Infect Dis 2015; 212: 1563–1573. together in old age. Neuropsychology 2011; 25: 368–377. 36 Horvath S, Garagnani P, Bacalini MG, Pirazzini C, Salvioli S, Gentilini D 9 Raz N, Rodrigue KM. Differential aging of the brain: patterns, cognitive correlates et al. Accelerated epigenetic aging in Down syndrome. Aging Cell 2015; 14: and modifiers. Neurosci Biobehav Rev 2006; 30:730–748. 491–495. 10 Sowell ER, Peterson BS, Thompson PM, Welcome SE, Henkenius AL, Toga AW. 37 Horvath S, Erhart W, Brosch M, Ammerpohl O, Von Schönfels W, Ahrens M et al. Mapping cortical change across the human life span. Nat Neurosci 2003; 6: Obesity accelerates epigenetic aging of human liver. Proc Natl Acad Sci USA 2014; 309–315. 111: 15538–15543. 11 Fjell AM, Westlye LT, Grydeland H, Amlien I, Espeseth T, Reinvang I et al. Critical 38 Sabayan B, van der Grond J, Westendorp RG, van Buchem MA, de Craen AJ. ages in the life course of the adult brain: nonlinear subcortical aging. Neurobiol Accelerated progression of white matter hyperintensities and subsequent Aging 2013; 34: 2239–2247. risk of mortality: a 12-year follow-up study. Neurobiol Aging 2015; 36: 2130–2135. 12 Ritchie SJ, Dickie DA, Cox SR, Valdes Hernandez MDC, Corley J, Royle NA et al. 39 Olesen PJ, Guo X, Gustafson D, Börjesson-Hanson A, Sacuíu S, Eckerström C et al. A Brain volumetric changes and cognitive ageing during the eighth decade of life. population-based study on the influence of brain atrophy on 20-year survival Hum Brain Mapp 2015; 36: 4910–4925. after age 85. Neurology 2011; 76:879–886. 13 Royle NA, Booth T, Valdés Hernández MC, Penke L, Murray C, Gow AJ et al. 40 Staff RT, Murray AD, Ahearn T, Salarirad S, Mowat D, Starr JM et al. Brain volume Estimated maximal and current brain volume predict cognitive ability in old age. and survival from age 78 to 85: The contribution of alzheimer-type magnetic Neurobiol Aging 2013; 34: 2726–2733. resonance imaging findings. J Am Geriatr Soc 2010; 58:688–695. 14 Deary IJ, Bastin ME, Pattie A, Clayden JD, Whalley LJ, Starr JM et al. White matter 41 Van Elderen SS, Zhang Q, Sigurdsson S, Haight TJ, Lopez O, Eiriksdottir G et al. integrity and cognition in childhood and old age. Neurology 2006; 66: 505–512. Brain volume as an integrated marker for the risk of death in a community-based 15 Shenkin SD, Bastin ME, MacGillivray TJ, Deary IJ, Starr JM, Rivers CS et al. Cognitive sample: age gene/environment susceptibility-reykjavik study. J Gerontol A Biol Sci correlates of cerebral white matter lesions and water diffusion tensor parameters Med Sci 2016; 71: 131–137. in community-dwelling older people. Cerebrovasc Dis 2005; 20: 310–318. 42 Ikram MA, Vernooij MW, Vrooman HA, Hofman A, Breteler MMB. Brain tissue 16 Franke K, Ziegler G, Klöppel S, Gaser C. Estimating the age of healthy subjects volumes and small vessel disease in relation to the risk of mortality. Neurobiol from T1-weighted MRI scans using kernel methods: exploring the influence of Aging 2009; 30:450–456. various parameters. Neuroimage 2010; 50: 883–892. 43 The IST-3 collaborative group. Association between brain imaging signs, early and 17 Franke K, Gaser C. Longitudinal changes in individual BrainAGE in healthy aging, late outcomes, and response to intravenous alteplase after acute ischaemic stroke mild cognitive impairment, and Alzheimer's Disease. GeroPsych 2012; 25: in the third International Stroke Trial (IST-3): secondary analysis of a randomised 235–245. controlled trial. Lancet Neurol 2015; 14:485–496. 18 Gaser C, Franke K, Klöppel S, Koutsouleris N, Sauer H. BrainAGE in mild cognitive 44 Leong DP, Teo KK, Rangarajan S, Lopez-Jaramillo P, Avezum A Jr, Orlandini A et al. impaired patients: predicting the conversion to Alzheimer's Disease. PLoS ONE Prognostic value of grip strength: findings from the Prospective Urban Rural 2013; 8: e67346. Epidemiology (PURE) study. Lancet 2015; 386:266–273. 19 Koutsouleris N, Davatzikos C, Borgwardt S, Gaser C, Bottlender R, Frodl T et al. 45 Sasaki H, Kasagi F, Yamada M, Fujita S. Grip strength predicts cause-specific Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker mortality in middle-aged and elderly persons. Am J Med 2007; 120: 337–342. of psychiatric disorders. Schizophr Bull 2013; 40: 1140–1153. 46 Sayer AA, Kirkwood TBL. Grip strength and mortality: a biomarker of ageing? 20 Cole JH, Leech R, Sharp DJ, for the Alzheimer's Disease Neuroimaging I. Prediction Lancet 2015; 386: 226–227. of brain age suggests accelerated atrophy after traumatic brain injury. Ann Neurol 47 Schunemann HJ, Dorn J, Grant BJ, Winkelstein W Jr., Trevisan M. Pulmo- 2015; 77: 571–581. nary function is a long-term predictor of mortality in the general popu- Molecular Psychiatry (2018), 1385 – 1392 Brain age predicts mortality JH Cole et al lation: 29-year follow-up of the Buffalo Health Study. Chest 2000; 118: This work is licensed under a Creative Commons Attribution 4.0 656–664. International License. The images or other third party material in this 48 Studenski S, Perera S, Patel K, Rosano C, Faulkner K, Inzitari M et al. Gait speed and article are included in the article’s Creative Commons license, unless indicated survival in older adults. J Am Med Assoc 2011; 305:50–58. otherwise in the credit line; if the material is not included under the Creative Commons 49 Swan GE, Carmelli D, Larue A. Performance on the digit symbol substitution test license, users will need to obtain permission from the license holder to reproduce the and 5-year mortality in the western collaborative group study. Am J Epidemiol material. To view a copy of this license, visit http://creativecommons.org/licenses/ 1995; 141:32–40. by/4.0/ 50 Seeman TE, McEwen BS, Rowe JW, Singer BH. Allostatic load as a marker of cumulative biological risk: MacArthur studies of successful aging. Proc Natl Acad © The Author(s) 2018 Sci USA 2001; 98: 4770–4775. Supplementary Information accompanies the paper on the Molecular Psychiatry website (http://www.nature.com/mp) Molecular Psychiatry (2018), 1385 – 1392 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Molecular Psychiatry Springer Journals

Brain age predicts mortality

Download PDF
8 pages

Loading next page...
 
/lp/springer-journals/brain-age-predicts-mortality-DJyFOnfrVR

References (54)

Publisher
Springer Journals
Copyright
Copyright © 2018 by The Author(s)
Subject
Medicine & Public Health; Medicine/Public Health, general; Psychiatry; Neurosciences; Behavioral Sciences; Pharmacotherapy; Biological Psychology
ISSN
1359-4184
eISSN
1476-5578
DOI
10.1038/mp.2017.62
Publisher site
See Article on Publisher Site

Abstract

OPEN Molecular Psychiatry (2018) 23, 1385–1392 www.nature.com/mp ORIGINAL ARTICLE 1 2,3 2,4 2,4 2,4 2,4 2,3 2,3 2,5 JH Cole , SJ Ritchie , ME Bastin , MC Valdés Hernández , S Muñoz Maniega , N Royle , J Corley , A Pattie , SE Harris , 6 6,7 3 2,5,7 2 2,3 2,4 1 2,3 Q Zhang , NR Wray , P Redmond , RE Marioni , JM Starr ,SR Cox , JM Wardlaw , DJ Sharp and IJ Deary Age-associated disease and disability are placing a growing burden on society. However, ageing does not affect people uniformly. Hence, markers of the underlying biological ageing process are needed to help identify people at increased risk of age-associated physical and cognitive impairments and ultimately, death. Here, we present such a biomarker, ‘brain-predicted age’, derived using structural neuroimaging. Brain-predicted age was calculated using machine-learning analysis, trained on neuroimaging data from a large healthy reference sample (N = 2001), then tested in the Lothian Birth Cohort 1936 (N = 669), to determine relationships with age-associated functional measures and mortality. Having a brain-predicted age indicative of an older-appearing brain was associated with: weaker grip strength, poorer lung function, slower walking speed, lower fluid intelligence, higher allostatic load and increased mortality risk. Furthermore, while combining brain-predicted age with grey matter and cerebrospinal fluid volumes (themselves strong predictors) not did improve mortality risk prediction, the combination of brain-predicted age and DNA- methylation-predicted age did. This indicates that neuroimaging and epigenetics measures of ageing can provide complementary data regarding health outcomes. Our study introduces a clinically-relevant neuroimaging ageing biomarker and demonstrates that combining distinct measurements of biological ageing further helps to determine risk of age-related deterioration and death. Molecular Psychiatry (2018) 23, 1385–1392; doi:10.1038/mp.2017.62; published online 25 April 2017 INTRODUCTION increased brain atrophy for a given age, may well reflect latent neuropathological influences. A reliable and valid brain-based As the global population ages, the burden of disease is biomarker of ageing, that identifies individuals deviating from increasing. This has motivated research to understand the a healthy brain ageing trajectory, could have great utility in biological links between ageing and disease risk. There is efforts to combat age-associated neurodegeneration and its substantial heterogeneity in how the ageing process affects consequences. different individuals, indicating that people age at different rates, Neuroimaging is a powerful tool for deriving in vivo data on the due to both genetic and environmental influences. If the ageing brain, demonstrating both global and spatially-localised biological characteristics of these different rates of ageing can 10,11 relationships with normal ageing, and with age-associated be measured, then biomarkers of individual differences in the 12–15 cognitive decline. Recently, multivariate methods have been ageing process might help improve predictions of mortality and developed to define statistical models of healthy brain ageing. morbidity. Such biomarkers could potentially identify those at risk Using machine-learning analysis of neuroimaging data, chrono- of age-associated health problems years before symptoms appear, logical age can be accurately predicted in healthy individuals. and be used as outcome measures in trials of therapeutics aimed This provides a method of measuring whether a person’s brain at delaying the onset of age-related disease. Many different appears younger or older than their chronological age. Using this ageing biomarkers have been proposed, which tap into different model, deviations from healthy brain ageing have been iden- cellular and molecular aspects of ageing. For example, the so- 17 18 2,3 tified in Alzheimer’s disease, mild cognitive impairment, called ‘epigenetic clock’ uses measurements of DNA- schizophrenia and have been related to cognitive impairment methylation status at CpG sites across the genome, which can after traumatic brain injury. Furthermore, protective factors have be converted into an age prediction which correlates highly with been associated with a positive influence on brain ageing. For chronological age in healthy individuals. Other candidate ageing 4 5 example, years of education, physical exercise and practicing biomarkers include leucocyte telomere length, N-glycan profile meditation were recently linked to having younger-appearing and Ink4a/Arf locus expression. This diverse list of candidate 21,22 ageing biomarkers reflects the involvement of multiple biological brains. As these multivariate neuroimaging measures have been systems and the overall complexity of the ageing process in associated with age-related pathology and cognitive impairment, humans. Neurological aspects of ageing, such as cognitive decline and this raises the possibility that brain-based age predictions could dementia, are particularly deleterious to general health and well- be used as an ageing biomarker. A viable ageing biomarker must 8 9 23 being. Brain structure is well-known to alter throughout life, and relate to the risk of mortality and age-associated morbidity, deviations from this typical brain ageing trajectory, in terms of particularly if it is to have clinical utility. To establish what the 1 2 Computational, Cognitive and Clinical Neuroimaging Laboratory, Department of Medicine, Imperial College London, London, UK; Centre for Cognitive Ageing and Cognitive 3 4 Epidemiology, University of Edinburgh, Edinburgh, UK; Department of Psychology, University of Edinburgh, Edinburgh, UK; Brain Research Imaging Centre, Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, UK; Centre for Genomic and Experimental Medicine, MRC Institute of Genetics & Molecular Medicine, University of 6 7 Edinburgh, Edinburgh, UK; Institute for Molecular Bioscience, The University of Queensland, QLD, Australia and Queensland Brain Institute, The University of Queensland, QLD, Australia. Correspondence: Dr JH Cole, Medicine, Imperial College London, Computational, Cognitive and Clinical Neuroimaging Laboratory, Burlington Danes Building, Du Cane Road, London W12 0NN, UK. E-mail: [email protected] Received 4 October 2016; revised 18 January 2017; accepted 17 February 2017; published online 25 April 2017 Brain age predicts mortality JH Cole et al consequences of having a brain that appears older or younger Table 1. Lothian Birth Cohort 1936 characteristics than average for one’s chronological age, we estimated ‘brain- predicted age’ in a large, narrow age-range population cohort of All Male Female older adults (Lothian Birth Cohort 1936 (LBC1936), N = 669), using structural neuroimaging (T1-weighted magnetic resonance ima- N 669 352 317 Age 72.67 (0.73) 72.63 (0.71) 72.72 (0.74) ging (MRI)). We tested the association between brain-predicted Mini-mental state 29 (2) 29 (2) 29 (2) age difference (brain-PAD; calculated as brain-predicted age examination minus chronological age) and: mortality risk, disease prevalence, (median (IQR)) measures of physical and mental fitness (grip strength, walking Brain-predicted age 74.32 (8.72) 76.92 (8.64) 71.43 (7.88) speed, lung function and general fluid intelligence), and a Brain-PAD 1.65 (8.71) 4.29 (8.58) − 1.29 (7.87) composite measure of biological health (allostatic load). We g 0.03 (0.98) 0.01 (1.05) 0.06 (0.90) hypothesised that ‘older’ brain-PAD would be associated with Grip strength 28.79 (9.33) 35.38 (6.71) 21.45 (5.63) earlier mortality and more morbidity, poorer physical and FEV (l) 2.34 (0.68) 2.72 (0.62) 1.92 (0.44) cognitive fitness, and greater allostatic load. 6 metre walk time (s) 4.29 (1.21) 4.09 (1.11) 4.51 (1.27) Allostatic load − 0.03 (0.99) 0.09 (0.95) −0.15 (1.02) Further, it has been proposed that biological ageing occurs at Deceased (N) 73 43 30 different rates to different tissues or cells within the same person, the so-called ‘mosaic of ageing’. Hence, complementary Abbreviations: brain-PAD, brain-predicted age difference; FEV, forced information could be gained by combining ageing biomarkers expiratory volume in one second; g , fluid type general intelligence; IQR, derived from different sources. Conversely, if ageing occurs inter-quartile range. Mortality was ascertained between 5.4 and 7.9 years after neuroimaging assessment. Values reported are mean (s.d.) unless uniformly across the body, then diverse ageing biomarkers should otherwise specified. correlate highly. Here we explored these possibilities by examin- ing brain-predicted age in relation to molecular-genetic ageing biomarkers. We tested brain-PAD in combination with DNA- methylation-based age predictions using the ‘epigenetic-clock’ and leukocyte telomere length, both previously associated with and WM images were concatenated and converted into a similarity matrix 25,26 mortality, examining their influence on the relationship with of training subjects’ data, which to predict chronological age in a Gaussian age-related outcome measures. Finally, we considered how brain- Process regression. Model accuracy was then validated using ten-fold PAD related to more conventional imaging measures, previously cross-validation, comparing brain-predicted age with chronological age. shown to relate to ageing. The coefficients ‘learned’ from the full model (N = 2001) were then applied to the test data (LBC1936, N = 669) to make brain-based age predictions for these individuals. Brain-PAD (predicted age—chronological age) was then MATERIALS AND METHODS calculated and used for further statistical analysis. Full details of the participants, data acquisition and statistical methods used in the study are included in Supplementary Methods. Ageing fitness measures Five measures of ‘fitness’, or a healthy ageing phenotype, in older age Participants—Lothian Birth Cohort 1936 were considered: walking speed (time to walk 6 metres), right-hand grip strength (measured by a dynamometer), lung function (forced expiratory The LBC1936 is a longitudinal study of ageing based in the Edinburgh 27,28 and Lothians area of Scotland, UK. Most of the participants had volume in 1 s), cognitive function (fluid-type intelligence) and allostatic taken part in the Scottish Mental Survey of 1947, which involved a test of load. Allostatic load was derived from measures of: fibrinogen, general cognitive ability for almost all 11-year old children in the triglyceride, high-density lipoprotein, low-density lipoprotein, total choles- country at that time. At the first wave, 1,091 participants attended for terol, cholesterol high-density lipoprotein ratio, glycated haemoglobin, cognitive and medical testing (mean age 70 years, 548 = male, C-reactive protein, interleukin-6, body-mass index and blood pressure. All 543 = female). MRI testing began at the second wave, when 866 individuals measures used in the present analysis were collected at the same time as attended for cognitive, medical testing (mean age 73 years, 448 = male, the neuroimaging assessment. 418 = female), of whom 669 (352 = male, 317 = female) had MRI. This final cohort provided the data that were included in present analysis (Table 1). Mortality ascertainment The vast majority of participants were cognitively normal according to Mortality status was obtained via data linkage to the National Health mini-mental state examination, with 99.3% scoring ⩾ 24. Ethical approval Service Central Register, provided by the National Records of Scotland. for the LBC1936 was obtained from the Multi-Centre Research Ethics The LBC1936 research team are routinely informed of participant Committee for Scotland (MREC/01/0/56) and the Lothian Research Ethics deaths and cause of death approximately every 12 weeks. Most Committee (LREC/2003/2/29). Written informed consent was obtained recent ascertainment was at approximately age 79 years (range 78.7– from all subjects. 79.7 years), which was between 5.4 and 7.9 years after neuroimaging assessment. Participants—brain-predicted age training cohort Further 2001 healthy individuals (age mean = 36.95 ± 18.12 years; age Molecular genetic biomarkers of ageing range = 18–90 years; males = 1016; females = 985) comprised the brain- Using whole blood samples, data for two candidate ageing biomarkers predicted age training cohort. These data were obtained via were generated. The ‘epigenetic clock’ was used to calculate predictions publicly-available repositories (Supplementary Table 5) and were screened of age based on DNA-methylation status at 450, 726 autosomal sites across according to local study protocols to ensure that they were 26,32 the genome, as per the previously reported ‘Horvath’ protocol. free of neurological and psychiatric disorders, had no history of head Leukocyte telomere length was measured using a protocol developed at trauma and other major medical conditions. Ethical approval for each University of Newcastle. initial study and subsequent data sharing was verified for each data repository. RESULTS Brain age prediction methods Chronological age can be predicted using neuroimaging The machine learning age-predictions methods using neuroimaging data A machine-learning model (Gaussian Processes), trained on the are outlined in Figure 1. Briefly, T1-weighted MRI scans were segmented brains of N = 2001 healthy adults, aged 18–90 years, can accurately into grey matter (GM) and white matter (WM) before being normalised in common space using non-linear spatial registration. Once normalised, GM predict chronological age using T1-weighted MRI scans Molecular Psychiatry (2018), 1385 – 1392 Brain variable y Model coefficients JH Cole et al Training Validation Accuracy T1-MRI with age label assessment GPR model Gaussian Process regression to predict age Defining brain-PAD Positive brain-PAD score = ‘older’ brain Chronological 50 Age 10-fold cross-validation Brain-PAD = 0, matched chronological and predicted age Testing Negative brain-PAD score = ‘younger’ brain Trained GPR Chronological Age model 72.9 Age prediction generated i i i i i i N = 2001 N = 669 ? ? ? ? ? Figure 1. Overview of study methods. Illustration of the methods used to generate brain-predict ages. 3D T1-weighted MRI scans were segmented into grey and WM before being normalised in common space using non-linear spatial registration. Normalised grey and WM images were concatenated and converted into vectors for each subject. These vectors were then projected into an NxN similarity matrix based on vector dot-products. (a) Once in similarity matrix form the training subjects’ data were used as predictors in a Gaussian Processes regression (GPR) with age as the outcome variable. (b) Model accuracy was assessed in a ten-fold cross-validation procedure, comparing brain- predicted age with original chronological age labels. (c) Model coefficients learned during training were then applied to the data from LBC1936 participants to make age predictions. (d) A metric to summarise the variation in predicted age was defined; the brain-predicted age difference (brain-PAD; predicted age—chronological age). LBC1936, Lothian Birth Cohort 1936; MRI, magnetic resonance imaging; WM, white matter. 20 40 60 80 50 60 70 80 90 100 Age (years) Age/Brain-predicted age (years) Figure 2. Brain-predicted age using structural neuroimaging in LBC1936. (a) Scatterplot showing the relationship between chronological age and brain-predicted age in the independent healthy cohort used as the training data (green diamonds) and the LBC1936 participants used as the test set (red circles). (b) Histogram showing the distributions of brain-predicted age (in blue) compared to the distribution in chronological age (in red). The substantially wider variability in brain-predicted age is evident. LBC1936, Lothian Birth Cohort 1936. (Figure 2a). Cross-validation results gave a correlation between Older adults show marked variation in structural brain ageing brain-predicted age and chronological age of r = 0.94, (Po0.001, The model coefficients ‘learned’ from the training dataset were corrected after 1000 permutations) and explained 88% of the applied to T1-weighted MRI scans acquired from the LBC1936 variance (R ). The mean absolute error of prediction was 5.02 years participants (Table 1) to generate a brain-predicted age. At the and the root mean square error was 6.31 years. This training stage time of scanning LBC1936 participants had a mean chronological validated our model of brain-predicted age, for use in predicting age of 72.67 (s.d. = 0.73) years and a mean brain-predicted age of age with neuroimaging data collected in other samples. 74.32 (s.d. = 8.72) years. The mean absolute error of age prediction Molecular Psychiatry (2018), 1385 – 1392 Brain variable x Brain-Predicted Age (years) Frequency Brain-Predicted Age Brain-Predicted Age Brain variable z Brain age predicts mortality JH Cole et al 1.00 1.0 Female/Alive Female/Deceased Male/Alive 0.69 Male/Deceased 0.8 0.95 0.66 0.59 0.90 0.6 0.50 0.85 0.4 Brain-PAD 0.80 0.2 −10 DNAm-PAD high Brain-PAD Brain-PAD + DNAm-PAD low Brain-PAD Telomere length 0.75 0.0 −20 72 74 76 78 80 1.0 0.8 0.6 0.4 0.2 0.0 Age (years) Specificity Sex/Deceased Figure 3. Association of mortality with brain-predicted age difference and DNA-methylation-predicted age difference. (a) Grouped scatterplot showing the relationship between brain-predicted age difference (brain-PAD) score (i.e., brain-predicted age—chronological age) and mortality (alive= blue, dead = red), sub-divided by sex (female = circle, male= triangle). Mortality status was determined ~ 6 years after MRI assessment. Horizontal black lines represent the median for each sub-group. (b) Kaplan–Meier plot of right-censored survival data since MRI assessment. The two lines represent a tertile split based on brain-PAD score, with highest 33.3% being classed as high brain-PAD (red line) indicating increased brain ageing and the lowest 33.3% (low brain-PAD, blue line) indicating reduced brain ageing. Crosses indicate censoring points (i.e. age at last survival ascertainment). Dotted lines represent the 95% confidence intervals. (c) Figure depicts the receiver operator characteristic (ROC) curves for four contrasting, nested, survival models. All models used mortality status as the response variables. The predictor variables were Brain-PAD (red line, model 1), DNAm-PAD (blue line, model 5), Brain-PAD+DNAm-PAD (green line, model 4), Telomere length+Brain-PAD+DNAm-PAD (grey line, model 3). The areas under the curve (AUC) are coloured-coded and appear next to each ROC curve. MRI, magnetic resonance imaging. in the LBC1936 participants was 7.08 years and the root mean other variables previously related to mortality in this sample was square error was 8.85 years. The variability in brain-predicted age considered in a fully-adjusted model, as per Marioni and was considerably greater than the variability in chronological age, colleagues. These were: Moray House Test IQ-type score at reflecting marked individual differences in brain structure in age 11, paternal social class (five-point scale), years of full-time participants aged approximately 73 (Figure 2b). As expected, education, APOE e4 carrier status, smoking status (never, brain-PAD scores did not correlate with chronological age ex-smoker, current smoker), and self-reported hypertension, (r = − 0.01, P = 0.79), indicating that deviations from healthy brain diabetes and cardiovascular disease. Brain-PAD remained signifi- ageing (that is, having an older- or younger-appearing brain) were cantly associated with mortality risk in this fully-adjusted model, not related to underlying chronological age. Females’ brain- with a slight attenuation of the effect size (HR = 1.051, 95% predicted ages were, on average, younger than their chronological CI = 1.020, 1.083, Po0.001; Supplementary Table S1). age (mean (s.d.) brain-PAD = − 1.29 (7.87) years), whereas males’ were older (mean (s.d.) brain-PAD = 4.29 (8.58) years). This sex Variability in apparent brain-ageing relates to physical and mental difference was statistically significant (Wilcoxon rank-sum test, fitness W = 35431, Po0.001), hence sex was included as a covariate in all Brain-PAD score was also significantly related to a number of further analyses. measures that reflect characteristics of physical and mental fitness in older age using linear regression (Supplementary Table 2). An Early mortality is associated with older-appearing brains older-appearing brain, as indicated by a higher brain-PAD score, Having a higher brain-PAD score (that is, a brain that appears was significantly associated with lower fluid cognitive perfor- older than one’s chronological age) was significantly associated mance (standardised beta = − 0.121, P = 0.007), weaker grip with mortality before the age of 80 (Po0.001); up to seven years strength (standardised beta = − 0.060, P = 0.020), poorer lung func- after neuroimaging assessment. Mean brain-PAD score for tion (standardised beta = − 0.072, P = 0.020) and slower walking deceased males (N = 43) and females (N = 30) was 8.13 (s.d. = speed (standardised beta = 0.133. P = 0.004). Higher brain-PAD 9.52) and 2.07 (s.d. = 9.27) years, respectively, compared to 3.76 score was also associated with higher allostatic load (standardised (s.d. = 8.32) and − 1.64 (s.d. = 7.65) years for surviving males and beta = 0.097, P = 0.020), a composite measure of biological and females (Figure 3). The relationship between mortality risk and physiological parameters, designed to reflect biological ‘wear-and- brain-PAD was tested using Cox proportional hazards regression tear’ accumulated over a lifetime of stress adaptation. Reported analysis, adjusting for age and sex. Survival was ascertained up to P-values were corrected for five tests using a 5% false discovery 7.9 years post-neuroimaging, and survival duration was right- rate. censored for surviving individuals based on days between neuroimaging assessment and mortality ascertainment. Each extra Brain-PAD is not related to the prevalence of morbidity year of brain-predicted age (that is, having a brain-PAD score of Next, we examined the relationship between brain-PAD and the +1) resulted in a 6.1% relative increase in the risk of death presence of self-reported cardiovascular disease, stroke, and between age 72 and 80 (hazard ratio (HR) = 1.061, 95% confidence diabetes. LBC1936 participants reported the following prevalence interval (CI) = 1.031, 1.091, Po0.001). The assumptions of propor- tional hazards were met by the model. An illustrative Kaplan– of disease: cardiovascular disease = 26.9% (N = 180), diabetes = Meier plot using the upper and lower tertiles of brain-PAD scores 10.2% (N = 68) and a history of stroke = 6.9% (N = 46). After in LBC1936 participants is shown in Figure 2b. The influence of adjusting for sex, there was no significant association between Molecular Psychiatry (2018), 1385 – 1392 Female/Alive Female/Deceased Male/Alive Male/Deceased Brain-PAD (years) Proportion survived Sensitivity Brain age predicts mortality JH Cole et al brain-PAD score and cardiovascular disease (P = 0.08), diabetes (635.2) base-pairs, while for males it was 3912.3 (783.3) base-pairs, (P = 0.14) or stroke (P = 0.85). which was significantly different (W = 45386, P = 0.001). There was no significant association between telomere length and brain-PAD (rho = 0.04, P = 0.31) or brain-predicted age (rho = 0.05, P = 0.23). Brain-PAD is not related to childhood IQ, life-course social factors Combining DNAm-PAD and telomere length with brain-PAD in or APOE e4 status a multivariate Cox regression, adjusted for age and sex, Brain-PAD was also not associated with potential life-course significantly predicted survival (N = 608, deceased N = 67, influences on ageing. Potential influences tested were: perfor- Po0.001). Within this model, brain-PAD (HR = 1.07, 95% CI = 1.04, mance on the Moray House Test at age 11 (P = 0.63), paternal 1.11, Po0.001) and DNAm-PAD (HR = 1.06, 95% CI = 1.02, 1.10, social class (P = 0.82), years of education (P = 0.45), neighbourhood Po0.001) were significant contributors to the prediction, while deprivation as indexed by the Scottish Index of Multiple telomere length was not (P = 0.97). A separate model using DNAm- Deprivation (P = 0.45), and the presence of an APOE e4 allele PAD alone also significantly predicted survival (HR = 1.06, 95% (P = 0.88). CI = 1.02, 1.09, Po0.001); however, this explained significantly less variance than a model using brain-PAD alone (AUC = 0.59 vs brain- Brain-PAD and conventional neuroimaging measures in relation to PAD alone AUC = 0.66, Po0.001). The combined model using survival brain-PAD and DNAm-PAD explained significantly more variance than either variable alone (combined model AUC = 0.69 vs brain- Brain-PAD was significantly correlated (positively or negatively) PAD alone AUC = 0.66 vs DNAm-PAD alone AUC = 0.59, Po0.001; with: GM, normal-appearing white matter, cerebrospinal fluid (CSF) see Figure 2c, Supplementary Table 4). This was also the case for and WM hyperintensity volume, whole brain cortical thickness, the fully-adjusted model covarying for potential influences on fractional anisotropy and mean diffusivity (Supplementary Figure 1). mortality risk. Prediction of ageing fitness measures was not When combining brain-PAD with these imaging measures to improved when combining brain-PAD and DNAm-PAD or brain- predict outcomes, brain-PAD contributed unique variance (deter- PAD and telomere length. mined using hierarchical variance partitioning) to each linear regression model (Supplementary Table 2). Although brain-PAD was not always the greatest contributor of unique variance to DISCUSSION outcome prediction, this analysis indicates that brain-PAD can add Here we showed that a neuroimaging-based marker of brain complementary information to models of age-related outcomes, ageing is associated with a greater risk of death and poorer over and above that gained from conventional neuroimaging physical and cognitive fitness in a large cohort of older adults. measures. Further, we assessed whether GM, normal-appearing Furthermore, we demonstrate that combining biological age white matter and CSF volume were associated with survival. Cox predictions generated from neuroimaging and DNA-methylation regression analyses, adjusted for age and sex, indicated that GM status data increases the accuracy of survival modelling. At ~ 73 and CSF volume (in ml) were associated with survival (GM: years of age, we found that people with brains that appeared HR = 0.991, 95% CI = 0.984, 0.998, P = 0.007; CSF: HR = 1.012, 95% older than their chronological age had, in addition to greater CI = 1.007, 1.017, Po0.001), where a having 1 ml lower GM or 1 ml mortality risk: weaker grip strength, poorer lung function, slower higher CSF volume was associated with a 1% increase in mortality walking speed, lower fluid general intelligence, and had been risk. Normal-appearing white matter volume was not associated exposed to greater allostatic load (a biological measure intended with mortality risk (P = 0.54). We then compared the predictive to summarise the cumulative effects of lifetime biological ‘wear value of linear combinations of brain-PAD with GM and CSF and tear’). The relationship between brain-PAD and survival was volume in Cox regression models (that is., brain-PAD+GM volume, independent of life-course influences on mortality including: brain-PAD+CSF volume, brain-PAD+GM volume+CSF volume). education, social class, childhood IQ, carrying an APOE e4 allele or Brain-PAD significantly related to survival in the paired combined the presence of age-associated illness. Furthermore, these factors models (Po0.05), indicating that it independently explained some were themselves not significantly associated with brain-PAD in variance relating to survival, when combined with GM volume or this sample. with CSF volume separately. However, when included alongside To the best of our knowledge, this is the first demonstration both GM volume and CSF volume, brain-PAD was no longer a that a neuroimaging-derived age prediction is associated with significant predictor of survival (P = 0.12), while the volumetric higher mortality risk. Such measures have been used in clinical measures remained significant (GM: z = − 3.78, Po0.001; CSF: samples, showing increased apparent brain age following trau- z = 4.56, Po0.001). For full details see Supplementary Table 3. matic brain injury and in individuals with mild cognitive, a key 18 21 risk factor for Alzheimer’s Disease. Higher levels of exercise Brain-PAD combined with DNA-methylation ‘age’ improves and meditation have been associated with lower brain age in survival modelling the healthy population, but the link with mortality is novel. This is Molecular genetic ageing biomarkers have also been proposed, crucial, as it supports the use of MRI as a screening tool to help hence we compared brain-PAD with DNA-methylation status and identify people at greater risk of general functional decline and leukocyte telomere length. DNA-methylation (DNAm) age was mortality during ageing. Brain-PAD has the potential to be predicted using Horvath’s ‘epigenetic clock’ method, in N = 620 estimated in large numbers of people, as MRI is collected routinely (female = 290, male = 330) participants, who had undergone in clinical settings. The success of projects like UK Biobank shows neuroimaging assessment. Mean DNAm-predicted age was 69.3 that acquiring MRI on a very large scale is feasible given the (s.d. = 6.2) years. Mean DNAm-predicted age difference (DNAm- appropriate infrastructure. PAD) was − 3.4 (s.d. = 6.1) years. Mean DNAm-PAD was similar for The combination of DNA-methylation-predicted age and males (−3.2, s.d. = 5.8) and females (−3.7, s.d. = 6.4), with no neuroimaging-predicted age is also novel. We found that, while statistically significant sex difference (W = 44803, P = 0.17). There brain-predicted age significantly out-performed DNA-methylation was no association between the DNAm-predicted age and brain- predicted age, there is greater added value gained when predicted age (rho = 0.001, P = 0.98) or between brain-PAD and combining these two approaches to predict survival. Previously, DNAm-PAD (rho = − 0.007, P = 0.85). Regarding telomere length, ‘DNA-methylation age’ has been related to mortality and 26,32 data were available in N = 653 participants (female = 309, male = ageing fitness, and in various clinical contexts including 35 36 37 344) with neuroimaging data. Telomere mean (s.d.) length was HIV, Down’s Syndrome and obesity. Interestingly, brain-PAD 3982.5 (711.7) base-pairs. Telomere length in females was 4045.5 and DNAm-PAD were not correlated, yet both related to survival Molecular Psychiatry (2018), 1385 – 1392 Brain age predicts mortality JH Cole et al 47–50 independently and improved survival prediction when analysed in with mortality. As proxies for systemic health (for example, combination, thus provided complementary information. This musculo-skeletal, respiratory, nervous, circulatory), they appear to demonstrates that contrasting approaches to estimating age relate to a common aspect of more general health of the whole biologically can be integrated to predict clinically-relevant out- body, likely due to the interactions between different human comes. Seemingly, epigenetic ageing in leukocytes and ageing biological systems. However, there also seems to be unique of brain structure are occurring independently, perhaps evidence variance in the relationship of these measures with mortality. This for a ‘mosaic’ of ageing, where biological ageing occurs at is supported by our finding that survival modelling accuracy was different rates in different systems or compartments within an improved when including multiple ageing fitness measures individual. This motivates further research that combines inde- alongside brain-PAD. Notably, brain-PAD remained the strongest pendent measures of biological ageing to develop a more global predictor in this combined model, which justifies further research ageing biomarker, which may further improve predictions of into the clinical applications of neuroimaging-based predictors of survival. mortality. Other neuroimaging measures have previously been associated Our study has some strengths and weaknesses, particularly with mortality in older adult population cohorts. These include relating to the cohorts under study. The sample size for both WM hyperintensities in adults aged 70–82 years, regional training and test sets is relatively large. One potential limitation is volume reductions at age 85 years and whole brain volume at the multiple sources of training data. Comprehensive demo- 40 41 42 78–85, 66–90 and 60–90 years. Visual assessment of infarcts, graphic data were not available on all these individuals. However, WM hyperintensities and atrophy also predicted mortality individuals in this sample were screened according to various 6 months after a stroke. This research supports the idea that criteria to ensure that were free of manifest neurological, the brain plays a central role in the ageing process and is sensitive psychiatric or major medical health issues. The LBC1936 is well- to the cumulative damage that accrues throughout life and characterised, allowing a broad exploration of relationships with increases mortality risk. That we can predict mortality before the brain-predicted ageing, particularly the follow-up to assess age of 80 using neuroimaging assessment at approximately age mortality. The limited age range of LBC1936 participants is a 73, fits with these previous reports. strength in that it eliminates the important confounding effect of Interestingly, when combining brain-PAD with GM and CSF chronological age, but it may limit generalisations to other age volume in a Cox regression, brain-PAD no longer significantly groups. However, this point in the life course is a timely juncture predicted survival. This indicates that the survival-related variance to assess brain ageing as individual differences have had time to in brain-PAD can be captured using more conventional volumetric accumulate though are unlikely to be widely confounded by measures. While brain-PAD did incrementally improve survival manifest neurodegenerative disease. The current analysis was prediction over individual volumetric measures, our results cross-sectional; therefore, we cannot determine whether the indicate that using a combination of GM and CSF volume is relationship between brain-predicted age and mortality risk varies potentially a strong biomarker of mortality. Nevertheless, these with age or where on the trajectory of atrophy an individual is. The volumetric measures appear less suitable as an ageing biomarker on-going nature of the LBC1936 study will allow future analysis of per se, as a linear model of GM, WM and CSF volume explained longitudinal data to determine whether trajectories of brain only 66% of variance in chronological age (mean absolute ageing are better indicators of future health outcomes than cross- error = 8.30 years, root mean square error = 10.53) in the training sectional measures. In addition, we only assessed all-cause dataset, compared with 88% using brain-predicted age. This mortality, which limits speculation about causal relationships demonstrates that in the context of developing an ageing between brain structural alterations and specific mortality causes, biomarker, there is benefit in using a machine-learning approach such as cardiovascular or neurological causes of death. Finally, the to analyse high-dimensional voxelwise T1-MRI data, compared to LBC1936 participants were not fully representative of the macroscopic volume measurements. Future steps to further population from which they were drawn. Compared to the full improve models of brain ageing and derived ageing biomarkers population who sat the cognitive test at age 11, LBC1936 could incorporate additional imaging modalities at the modelling participants had higher cognitive ability, and in later life were stage. This should capture further age-associated changes likely to be healthier than their peers in the general population. including WM hyperintensities using FLAIR-MRI, altered WM This may have been due to selection effects seen in most studies microstructure using diffusion-MRI and beta-amyloid deposition of ageing. That our sample might have missed individuals with using positron emission tomography. particularly poor health or high frailty. Hence we might have A key medical research goal is to identify reliable predictors of underestimated some of the effects reported here, as a small mortality, proxy measures of underlying pathological processes number of individuals with worse performance on measures of that increase mortality risk. For example, grip strength has been ageing fitness may not have been included. 44,45 robustly associated with mortality, and is thought to be a The difference between neuroimaging-predicted age and proxy of the musculo-skeletal system. Similarly, brain-PAD may be chronological age is associated with survival in a large sample a general reflection of CNS health. Grip strength measures do not of older adults and relates to measures of cognitive and physical necessarily require a direct causal link with cardiovascular or all- fitness. Moreover, combining age-predictions from DNAm and cause mortality to be clinically useful; the same could apply to neuroimaging data increased the accuracy of survival modelling. brain-PAD. Moreover, the relevance of brain-predicted age for This study provides evidence that neuroimaging data can be used health is intuitively straightforward. Already, the UK National to construct a viable ageing biomarker, and potentially provides Health Service encourages people to complete a cardiovascular important prognostic information, particularly in combination with health question to determine their ‘heart age’ (www.nhs.uk/ complementary epigenetic ageing data. A global biomarker of Conditions/nhs-health-check/Pages/check-your-heart-age-tool. ageing has the potential to screen for asymptomatic individuals aspx). By analogy, ‘brain age’, or a global ‘biological age’, could be who are experiencing adverse ageing and thus are at increased used in public health settings to convey complex information to risk of future ill-health and could be used as a surrogate outcome patients in readily comprehendible terms. measure in clinical trials of neuroprotective treatments and anti- Brain-PAD related to all measures of ageing fitness. This ageing therapeutics. suggests that our measure of brain ageing relates to some more general facets of physiological ageing. Along with grip CONFLICT OF INTEREST strength, all these measures (lung function, gait speed, cognitive function and allostatic load) have been previously associated The authors declare no conflict of interest. Molecular Psychiatry (2018), 1385 – 1392 Brain age predicts mortality JH Cole et al ACKNOWLEDGMENTS 21 Steffener J, Habeck C, O'Shea D, Razlighi Q, Bherer L, Stern Y. Differences between chronological and brain age are related to education and self-reported physical We thank the Lothian Birth Cohort 1936 participants, radiographers at the Brain activity. Neurobiol Aging 2016; 40: 138–144. Research Imaging Centre (www.bric.ed.ac.uk/) and the researchers who collected and 22 Luders E, Cherbuin N, Gaser C. Estimating brain age using high-resolution pattern entered data used in this manuscript. This research and LBC1936 phenotype recognition: Younger brains in long-term meditation practitioners. Neuroimage collection were supported by Research Into Ageing and continues as part of The 2016; 134: 508–513. Disconnected Mind project (http://www.disconnectedmind.ed.ac.uk), funded by Age 23 Sprott RL. Biomarkers of aging and disease: introduction and definitions. UK. MRI acquisition and analyses were conducted at the Brain Research Imaging Exp Gerontol 2010; 45:2–4. Centre, Neuroimaging Sciences, Centre for Clinical Brain Sciences, University of 24 Cevenini E, Invidia L, Lescai F, Salvioli S, Tieri P, Castellani G et al. Human models Edinburgh, which is part of SINAPSE (Scottish Imaging Network—A Platform for of aging and longevity. Expert Opin Biol Ther 2008; 8: 1393–1405. Scientific Excellence) collaboration (www.sinapse.ac.uk/) funded by the Scottish 25 Cawthon RM, Smith KR, O'Brien E, Sivatchenko A, Kerber RA. Association between Funding Council and the Chief Scientist Office. This work was undertaken within the telomere length in blood and mortality in people aged 60 years or older. Lancet University of Edinburgh Centre for Cognitive Ageing and Cognitive Epidemiology 2003; 361: 393–395. (http://www.ccace.ed.ac.uk), part of the cross council Lifelong Health and Wellbeing 26 Marioni RE, Shah S, McRae AF, Chen BH, Colicino E, Harris SE et al. DNA methy- Initiative (MR/K026992/1), for which funding from the BBSRC and MRC is gratefully lation age of blood predicts all-cause mortality in later life. Genome Biol 2015; acknowledged. Simon Cox is supported by an MRC grant (MR/M013111/1). We would 16:25. like to thank Thomas von Zglinicki and Carmen Martin-Ruiz for their work on 27 Deary IJ, Gow AJ, Pattie A, Starr JM. Cohort profile: the Lothian Birth Cohorts of generating the telomere length data. Naomi Wray acknowledges funding from the 1921 and 1936. Int J Epidemiol 2012; 41:1576–1584. Australian National Health and Medical Research Council (613608, 1078901). 28 Deary IJ, Gow AJ, Taylor MD, Corley J, Brett C, Wilson V et al. The Lothian Birth Cohort 1936: A study to examine influences on cognitive ageing from age 11 to age 70 and beyond. BMC Geriatr 2007; 7:28. REFERENCES 29 Scottish Council for Research in Education. The Trend of Scottish intelligence: A 1 Vos T, Flaxman AD, Naghavi M, Lozano R, Michaud C, Ezzati M et al. Years lived Comparison of the 1947 and 1932 Surveys of the Intelligence of eleven-year-old with disability (YLDs) for 1160 sequelae of 289 diseases and injuries 1990–2010: a Pupils. University of London Press: London, 1949. systematic analysis for the Global Burden of Disease Study 2010. Lancet 2012; 30 Lara J, Godfrey A, Evans E, Heaven B, Brown LJE, Barron E et al. Towards mea- 380: 2163–2196. surement of the Healthy Ageing Phenotype in lifestyle-based intervention studies. 2 Bocklandt S, Lin W, Sehl ME, Sánchez FJ, Sinsheimer JS, Horvath S et al. Epigenetic Maturitas 2013; 76:189–199. predictor of age. PLoS ONE 2011; 6: e14821. 31 Booth T, Starr JM, Deary I. Modeling multisystem biological risk in later life: allostatic 3 Horvath S. DNA methylation age of human tissues and cell types. Genome Biol load in the lothian birth cohort study 1936. Am J Hum Biol 2013; 25:538–543. 2013; 14: R115. 32 Marioni RE, Shah S, McRae AF, Ritchie SJ, Muniz-Terrera G, Harris SE et al. The 4 Kruk PA, Rampino NJ, Bohr VA. DNA damage and repair in telomeres: relation epigenetic clock is correlated with physical and cognitive fitness in the Lothian to aging. Proc Natl Acad Sci USA 1995; 92:258–262. Birth Cohort 1936. Int J Epidemiol 2015; 44: 1388–1396. 5 Vanhooren V, Dewaele S, Libert C, Engelborghs S, De Deyn PP, Toussaint O et al. 33 Harris SE, Martin-Ruiz C, von Zglinicki T, Starr JM, Deary IJ. Telomere length and Serum N-glycan profile shift during human ageing. Exp Gerontol 2010; 45: aging biomarkers in 70-year-olds: the Lothian Birth Cohort 1936. Neurobiol Aging 738–743. 2012; 33: 1486.e1483–1488. 6 Krishnamurthy J, Torrice C, Ramsey MR, Kovalev GI, Al-Regaiey K, Su L et al. Ink4a/ 34 Miller KL, Alfaro-Almagro F, Bangerter NK, Thomas DL, Yacoub E, Xu J et al. Arf expression is a biomarker of aging. J Clin Invest 2004; 114: 1299–1307. Multimodal population brain imaging in the UK Biobank prospective epidemio- 7 Lopez-Otin C, Blasco MA, Partridge L, Serrano M, Kroemer G. The hallmarks logical study. Nat Neurosci 2016; 19: 1523–1536. of aging. Cell 2013; 153: 1194–1217. 35 Horvath S, Levine AJ. HIV-1 infection accelerates age according to the 8 Tucker-Drob EM. Neurocognitive functions and everyday functions change epigenetic clock. J Infect Dis 2015; 212: 1563–1573. together in old age. Neuropsychology 2011; 25: 368–377. 36 Horvath S, Garagnani P, Bacalini MG, Pirazzini C, Salvioli S, Gentilini D 9 Raz N, Rodrigue KM. Differential aging of the brain: patterns, cognitive correlates et al. Accelerated epigenetic aging in Down syndrome. Aging Cell 2015; 14: and modifiers. Neurosci Biobehav Rev 2006; 30:730–748. 491–495. 10 Sowell ER, Peterson BS, Thompson PM, Welcome SE, Henkenius AL, Toga AW. 37 Horvath S, Erhart W, Brosch M, Ammerpohl O, Von Schönfels W, Ahrens M et al. Mapping cortical change across the human life span. Nat Neurosci 2003; 6: Obesity accelerates epigenetic aging of human liver. Proc Natl Acad Sci USA 2014; 309–315. 111: 15538–15543. 11 Fjell AM, Westlye LT, Grydeland H, Amlien I, Espeseth T, Reinvang I et al. Critical 38 Sabayan B, van der Grond J, Westendorp RG, van Buchem MA, de Craen AJ. ages in the life course of the adult brain: nonlinear subcortical aging. Neurobiol Accelerated progression of white matter hyperintensities and subsequent Aging 2013; 34: 2239–2247. risk of mortality: a 12-year follow-up study. Neurobiol Aging 2015; 36: 2130–2135. 12 Ritchie SJ, Dickie DA, Cox SR, Valdes Hernandez MDC, Corley J, Royle NA et al. 39 Olesen PJ, Guo X, Gustafson D, Börjesson-Hanson A, Sacuíu S, Eckerström C et al. A Brain volumetric changes and cognitive ageing during the eighth decade of life. population-based study on the influence of brain atrophy on 20-year survival Hum Brain Mapp 2015; 36: 4910–4925. after age 85. Neurology 2011; 76:879–886. 13 Royle NA, Booth T, Valdés Hernández MC, Penke L, Murray C, Gow AJ et al. 40 Staff RT, Murray AD, Ahearn T, Salarirad S, Mowat D, Starr JM et al. Brain volume Estimated maximal and current brain volume predict cognitive ability in old age. and survival from age 78 to 85: The contribution of alzheimer-type magnetic Neurobiol Aging 2013; 34: 2726–2733. resonance imaging findings. J Am Geriatr Soc 2010; 58:688–695. 14 Deary IJ, Bastin ME, Pattie A, Clayden JD, Whalley LJ, Starr JM et al. White matter 41 Van Elderen SS, Zhang Q, Sigurdsson S, Haight TJ, Lopez O, Eiriksdottir G et al. integrity and cognition in childhood and old age. Neurology 2006; 66: 505–512. Brain volume as an integrated marker for the risk of death in a community-based 15 Shenkin SD, Bastin ME, MacGillivray TJ, Deary IJ, Starr JM, Rivers CS et al. Cognitive sample: age gene/environment susceptibility-reykjavik study. J Gerontol A Biol Sci correlates of cerebral white matter lesions and water diffusion tensor parameters Med Sci 2016; 71: 131–137. in community-dwelling older people. Cerebrovasc Dis 2005; 20: 310–318. 42 Ikram MA, Vernooij MW, Vrooman HA, Hofman A, Breteler MMB. Brain tissue 16 Franke K, Ziegler G, Klöppel S, Gaser C. Estimating the age of healthy subjects volumes and small vessel disease in relation to the risk of mortality. Neurobiol from T1-weighted MRI scans using kernel methods: exploring the influence of Aging 2009; 30:450–456. various parameters. Neuroimage 2010; 50: 883–892. 43 The IST-3 collaborative group. Association between brain imaging signs, early and 17 Franke K, Gaser C. Longitudinal changes in individual BrainAGE in healthy aging, late outcomes, and response to intravenous alteplase after acute ischaemic stroke mild cognitive impairment, and Alzheimer's Disease. GeroPsych 2012; 25: in the third International Stroke Trial (IST-3): secondary analysis of a randomised 235–245. controlled trial. Lancet Neurol 2015; 14:485–496. 18 Gaser C, Franke K, Klöppel S, Koutsouleris N, Sauer H. BrainAGE in mild cognitive 44 Leong DP, Teo KK, Rangarajan S, Lopez-Jaramillo P, Avezum A Jr, Orlandini A et al. impaired patients: predicting the conversion to Alzheimer's Disease. PLoS ONE Prognostic value of grip strength: findings from the Prospective Urban Rural 2013; 8: e67346. Epidemiology (PURE) study. Lancet 2015; 386:266–273. 19 Koutsouleris N, Davatzikos C, Borgwardt S, Gaser C, Bottlender R, Frodl T et al. 45 Sasaki H, Kasagi F, Yamada M, Fujita S. Grip strength predicts cause-specific Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker mortality in middle-aged and elderly persons. Am J Med 2007; 120: 337–342. of psychiatric disorders. Schizophr Bull 2013; 40: 1140–1153. 46 Sayer AA, Kirkwood TBL. Grip strength and mortality: a biomarker of ageing? 20 Cole JH, Leech R, Sharp DJ, for the Alzheimer's Disease Neuroimaging I. Prediction Lancet 2015; 386: 226–227. of brain age suggests accelerated atrophy after traumatic brain injury. Ann Neurol 47 Schunemann HJ, Dorn J, Grant BJ, Winkelstein W Jr., Trevisan M. Pulmo- 2015; 77: 571–581. nary function is a long-term predictor of mortality in the general popu- Molecular Psychiatry (2018), 1385 – 1392 Brain age predicts mortality JH Cole et al lation: 29-year follow-up of the Buffalo Health Study. Chest 2000; 118: This work is licensed under a Creative Commons Attribution 4.0 656–664. International License. The images or other third party material in this 48 Studenski S, Perera S, Patel K, Rosano C, Faulkner K, Inzitari M et al. Gait speed and article are included in the article’s Creative Commons license, unless indicated survival in older adults. J Am Med Assoc 2011; 305:50–58. otherwise in the credit line; if the material is not included under the Creative Commons 49 Swan GE, Carmelli D, Larue A. Performance on the digit symbol substitution test license, users will need to obtain permission from the license holder to reproduce the and 5-year mortality in the western collaborative group study. Am J Epidemiol material. To view a copy of this license, visit http://creativecommons.org/licenses/ 1995; 141:32–40. by/4.0/ 50 Seeman TE, McEwen BS, Rowe JW, Singer BH. Allostatic load as a marker of cumulative biological risk: MacArthur studies of successful aging. Proc Natl Acad © The Author(s) 2018 Sci USA 2001; 98: 4770–4775. Supplementary Information accompanies the paper on the Molecular Psychiatry website (http://www.nature.com/mp) Molecular Psychiatry (2018), 1385 – 1392

Journal

Molecular Psychiatry – Springer Journals

Published: Apr 25, 2017

There are no references for this article.