Prediction of brain age and cognitive age: Quantifying brain and cognitive maintenance in aging

Abstract The concept of brain maintenance refers to the preservation of brain integrity in older age, while cognitive reserve refers to the capacity to maintain cognition in the presence of neurodegeneration or aging‐related brain changes. While both mechanisms are thought to contribute to individual differences in cognitive function among older adults, there is currently no “gold standard” for measuring these constructs. Using machine‐learning methods, we estimated brain and cognitive age based on deviations from normative aging patterns in the Whitehall II MRI substudy cohort (N = 537, age range = 60.34–82.76), and tested the degree of correspondence between these constructs, as well as their associations with premorbid IQ, education, and lifestyle trajectories. In line with established literature highlighting IQ as a proxy for cognitive reserve, higher premorbid IQ was linked to lower cognitive age independent of brain age. No strong evidence was found for associations between brain or cognitive age and lifestyle trajectories from midlife to late life based on latent class growth analyses. However, post hoc analyses revealed a relationship between cumulative lifestyle measures and brain age independent of cognitive age. In conclusion, we present a novel approach to characterizing brain and cognitive maintenance in aging, which may be useful for future studies seeking to identify factors that contribute to brain preservation and cognitive reserve mechanisms in older age.


| Brain maintenance and cognitive reserve
Interindividual differences within older adult populations have led to a large number of studies focusing on risk and protective factors for cognitive decline in aging (Anatürk, Demnitz, Ebmeier, & Sexton, 2018;Bråthen, Lange, Fjell, & Walhovd, 2020;Nyberg, Fjell, & Walhovd, 2019;Sabia et al., 2019;Zsoldos et al., 2018), as well as factors that characterize successful aging or "SuperAgers" (Gefen et al., 2014;Harrison, Weintraub, Mesulam, & Rogalski, 2012;Rogalski et al., 2013;Yu et al., 2019). The maintenance of a "younger" brain, that is, the relative lack of aging-related changes including pathology, has been suggested as a main mechanism to preserve cognitive function into older age (Nyberg et al., 2012). For example, while decline in cortical thickness is commonly observed between midlife and late life, a unique group of older adults (i.e., "SuperAgers") do not exhibit this typical pattern of cortical atrophy. SuperAgers also appear to possess higher cortical thickness in some brain regions relative to younger individuals (Rogalski et al., 2013), and perform comparably to young adults on assessments of memory (Harrison et al., 2012;Sun et al., 2016). However, a number of clinical studies have reported weak associations between degree of brain pathology and relevant cognitive symptoms (Scarmeas & Stern, 2003), and as a consequence, reserve theories have gained prominence in the aging field (Driscoll et al., 2006;Snowdon, 2003;Stern, 2002;Stern, 2009).
The cognitive reserve theory seeks to explain why some individuals are able to maintain cognitive function in the presence of pathology or aging-related brain changes. It has been suggested that individuals with higher cognitive reserve process information more efficiently, enabling them to functionally adapt to brain aging and sustain greater pathology before cognitive impairments manifest (Stern, 2002). Education has been suggested to promote cognitive reserve by enhancing cognitive flexibility, and factors such as lifestyle behaviors may moderate the beneficial effect of education on cognition in older age (Cabeza et al., 2018). As there is no established method to directly measure cognitive reserve, the majority of reserve studies have used measures such as IQ or education as proxies of reserve (Valenzuela & Sachdev, 2006a;Valenzuela & Sachdev, 2006b), and focused on how cognitive function and brain metrics differ between individuals with high or low levels of this proxy measure (Cabeza et al., 2018).
In summary, preservation of brain structure and cognitive reserve mechanisms can both potentially contribute to a higher degree of maintained cognitive function in older age. Brain maintenance and cognitive reserve can thus be viewed as complementary constructs that may be malleable over the lifespan (Cabeza et al., 2018), potentially serving as targets for lifestyle interventions. However, despite their centrality to studies of aging (Stern, 2003), there remains no consistent approach to measuring and comparing these constructs. In this article, we refer to maintained brain structure and cognitive function relative to normative age trajectories as brain maintenance and cognitive maintenance. A strong association between the two would imply correspondence between degrees of brain and cognitive maintenance, while a weak association could reflect a lack of one-to-one relationships between these trajectories-potentially providing evidence for cognitive reserve mechanisms.

Definition of key concepts
Brain maintenance Preserved brain structure in older age relative to normative age trajectories

Cognitive maintenance
Preserved cognitive function in older age relative to normative age trajectories

Cognitive reserve
Preserved cognitive function in the presence of aging-related brain changes 1.2 | Prediction of brain age and cognitive age The application of machine learning to neuroimaging data has provided an avenue for estimating the apparent age of an individual's brain, and determining deviations from normative brain aging patterns (Cole, Marioni, Harris, & Deary, 2019). Studies in this area suggest that the difference between estimated "brain age" and chronological age (i.e., brain age gap [BAG]) varies between individuals, with positive BAG values (older brain age relative to chronological age) relating to poorer cognitive function (Boyle et al., 2020;Cole et al., 2018;Elliott et al., 2019). A recent multicohort study of 45,615 individuals further highlighted that BAG may be a sensitive marker of disease, with accelerated brain aging observed in a range of conditions including mild cognitive impairments, Alzheimer's disease, and depression . Mapping this metric onto the concepts introduced in Section 1.1, BAG estimated from structural MRI may reflect degree of brain maintenance (Steffener et al., 2016), with negative values suggesting preserved brain structure relative to what is expected based on normative trajectories (as illustrated in Figure 1).
Based on this rationale, machine learning algorithms can also be applied to cognitive data to provide an estimate of cognitive age, with the difference between cognitive and chronological age (i.e., cognitive age gap [CAG]) reflecting degree of cognitive maintenance (see Figure 1). The correlation between BAG and CAG would indicate the degree of correspondence between brain and cognitive maintenance or decline relative to normative aging patterns, as illustrated in Figure 1. Developing methods for characterizing brain and cognitive maintenance is timely, as there is a growing demand for lifestylebased interventions that may help prevent or delay age-and diseaserelated cognitive decline (Livingston et al., 2017).

| Lifestyle predictors of cognitive/neural integrity
The extent to which BAG and CAG relate to known predictors of healthy aging may shed light on the associations between lifestyle factors and maintenance/reserve mechanisms. For example, educational level may relate to cognitive maintenance independent of brain maintenance, potentially suggesting that education facilitates preserved cognition via cognitive reserve mechanisms. Health behaviors such as abstaining from smoking (Zhong, Wang, Zhang, Guo, & Zhao, 2015), limiting the amount of alcohol consumed (Rehm, Hasan, Black, Shield, & Schwarzinger, 2019), and achieving a sufficient amount of physical activity (PA; Guure, Ibrahim, Adam, & Said, 2017) could promote maintained cognitive function (Cadar et al., 2012;Clare et al., 2017;Sabia et al., 2009) and decreased dementia risk (Lourida et al., 2019) through brain preservation, as indicated by studies showing a link between these factors and estimates of BAG (Cole, 2020;Smith et al., 2020;Steffener et al., 2016).
However, a substantial proportion of studies on lifestyle factors and brain/cognitive aging are based on cross-sectional measurements (Anatürk et al., 2018;Scarmeas & Stern, 2003) or on self-reported history, which is highly susceptible to recall biases (Gow, Pattie, & Deary, 2017 suggested that there may be three major "clusters" of middle-aged and older adults, including (a) those who engaged in multiple healthy behaviors, (b) those who were socially and physically inactive but did not engage in risky behaviors, and (c) ex-smokers engaging in other risk behaviors (Liao et al., 2019). Other studies report between two and nine subgroups (reviewed in Noble, Paul, Turon, and Oldmeadow (2015)). The majority of this evidence, however, is based on cross-sectional survey data, and the ways in which brain preservation or cognitive reserve mechanisms may serve as potential pathways between lifestyle trajectories and late-life cognition are not well understood.
F I G U R E 1 Conceptual illustration of brain age (left) and cognitive age (right), where the distance between estimated brain age/cognitive age (colored dots) and the expected brain age/cognitive age (black line) represents brain age gap (BAG) and cognitive age gap (CAG)

| Study aims
In the current study, we used machine learning models to estimate (a) brain age based on MRI-derived measures, and (b) cognitive age based on performance scores on several cognitive tests. For each participant, we calculated an estimate of BAG (predicted brain age minus chronological age, indicating degree of brain maintenance (Steffener et al., 2016)), and CAG (predicted cognitive age minus chronological age, i.e., degree of cognitive maintenance). We first correlated BAG with CAG to investigate the degree of correspondence between brain and cognitive maintenance. Second, we tested the associations of CAG and BAG with premorbid IQ and education, which are known as proxies of cognitive reserve (Cabeza et al., 2018). Finally, we 2 | MATERIALS AND METHODS

| Sample
Data from participants enrolled in the Whitehall II imaging substudy were examined. A detailed description of this cohort has been published previously (Filippini et al., 2014). In brief, these individuals were originally recruited in 1985 as part of a cohort of 10,308 civil servants based in London (Marmot & Brunner, 2005), and have since been regularly assessed (at 5 year intervals) on a range of lifestyle, biological, and cognitive variables. Between 2012 and 2016, a random sample of 800 individuals were invited to undergo an MRI brain scan and comprehensive cognitive battery, as part of the Whitehall II imaging substudy (Filippini et al., 2014).
The current sample was drawn from the Imaging Sub-study cohort, and included 537 participants who had provided lifestyle information for at least four previous study phases. Additional criteria for inclusion were complete MRI, cognitive and demographic data at the MRI phase, no artifacts or substantial motion in the MRI images, no structural abnormalities detected in the MRI scans (e.g., strokes or tumors) and no self-report neurological diagnoses (e.g., Parkinson's disease) or a current SCID diagnosis of depression, anxiety, psychosis, or cognitive disorder. A description of sample characteristics is provided in Table 1, and an overview of the study timeline is available in Figure 2.

| MRI data acquisition and processing
The participants underwent a 3 T MRI scan at the Centre for Func-

| MRI features
The GM feature preparation followed the procedures described in FLAIR images with Brain Intensity AbNormality Classification Algorithm, as described in Griffanti et al. (2016). To test for improvements in model performance using more detailed measures of WMH, we also included an extended model with separate measures for deep and periventricular WMH (Armstrong et al., 2020;Griffanti et al., 2018). The WM variables were residualized with respect to scanner, relative head motion, sex, and ethnic background using linear models.
A total of 1,118 GM features and 245 WM features were included in the brain age prediction model. For comparison purposes, a reduced model was run using one measure per modality as input variables (total GM volume, average FA, and total WMH volume).  (Nasreddine et al., 2005)), working memory (Digit Span: forward, backward, ascending sequence (Wechsler, 2008)), visual attention and task-switching (Trail Making Test; TMT A & B; (Reitan, 1958)), visuospatial memory (Rey-Osterrieth complex figure:

| Cognitive features
immediate and delayed recall, recognition (Osterrieth, 1944)), verbal memory (Hopkins Verbal Learning Test: immediate and delayed recall, recognition (Brandt, 1991)), semantic memory (Boston Naming Test For alcohol consumption, the measurement was based on selfreported total units consumed on a weekly basis (over the last year).
Individuals who self-reported as abstinent from alcohol (<1 unit/week) or engaged in light drinking (i.e., 1-7 units/week) were assigned a point of 1. As a prior analysis of the Whitehall sample (Topiwala et al., 2017) recently suggested that moderate levels of alcohol may be associated with adverse brain outcomes, individuals consuming more than 7 units per week received a score of 0 (classified as "unhealthy").
For smoker status, individuals were assigned a point of 1 if they had never smoked, whereas individuals who were current or previous smokers were given a score of 0 (as done in prior studies: (Sabia et al., 2009;Zhong et al., 2015)). At the MRI time point, smoker status was inferred from two items of the lifestyle questionnaire," Have you smoked cigarettes in the past four weeks?" and" How many cigarettes in a typical week?" For PA, an individual reporting at least 2.5 hr of moderate to vigorous intensity activities per week (Metabolic equivalent of task values > 3, assignments were coded as 1 based on previous findings (Sabia et al., 2017), with all other individuals scoring 0. The measurement and computation of these variables are described in greater detail in Sabia et al. (2017) To assess the significance of the general model performances, average R 2 , root mean square error (RMSE), and mean absolute error (MAE) were estimated for each model using cross-validation with 10 splits and 5 repetitions, and compared to null distributions calculated from 1,000 permutations. The 10-fold cross-validations produced an estimate of each measure for every individual in the sample.
To investigate the prediction accuracy, correlation analyses were run for predicted brain age versus chronological age and predicted cogni- where m1 = Model 1 (CAG); m2 = Model 2 (BAG); β = beta coefficients from the covariate adjusted linear regressions between variables of interest and BAG/CAG; σ = SEs of the beta coefficients; and ρ = age-adjusted correlation between BAG and CAG.

| Associations with lifestyle trajectories
LCGA was employed to evaluate whether the sample could be subdivided into trajectory classes based on repeated measures of lifestyle index scores (Section 2.5). This is a widely used method that has previously been applied to identify trajectories of sleep (Zitser et al., 2020) and BMI (Vistisen et al., 2014) (Enders & Bandalos, 2001). All lifestyle index scores were standardized prior to the LCGA to minimize biases introduced by variation in the questionnaire between the prior phases and MRI scan.
The model with best fit was identified based on achieving the lowest Akaike information criterion (AIC), Bayesian information criterion (BIC), and sample-size adjusted BIC (aBIC), with an entropy value ≥0.8 and a p-value <.05 for the Vuong-Lo-Mendell-Rubin (VLMR) likelihood ratio and bootstrap likelihood ratio (BLRT) tests (Jung & Wickrama, 2008). A minimum of 5% of participants within each class was also required to minimize the likelihood that a class captured outliers only (Andruff, Carraro, Thompson, Gaudreau, & Louvet, 2009).
To test for relationships between lifestyle trajectories and estimates of brain and cognitive maintenance, BAG and CAG were used as dependent variables in separate linear regressions. Lifestyle trajectory classes were entered as independent variables and as dummy coding was used, the class with the least favorable lifestyle served as the reference category. In addition, a cumulative lifestyle score based on an average of lifestyle scores across all phases was examined as a potential predictor of BAG and CAG in separate regression analyses.
An average was selected over a summary score due to being less

| Multiple comparison corrections and sensitivity analyses
All hypothesis tests were two-sided. We report unadjusted p-value (p) and p-values corrected for multiple comparisons using false discovery rate (FDR) correction (Benjamini & Hochberg, 1995) (p corr ). As linear regression and z-tests for correlated samples were performed multiple times, we report the associations that showed a p-value <.05 but did not survive FDR-correction as trends, and associations that showed p corr <.05 as significant. The linear regression analyses and Z-tests for correlated samples were rerun after outliers on any of the variables of interest were excluded (i.e., individuals with values <Q1--3 × interquartile range [IQR] or values >Q3 + 3 × IQR).

| Brain age and cognitive age
Average R 2 , RMSE, and MAE for the brain age and cognitive age models are provided in Table 3. Supplementary Information (SI) Figure 1 shows the average R 2 , RMSE, and MAE for each of the models compared to null distributions. The number of permuted results from the null distribution that exceeded the mean from the cross-validation was 0 for both the brain age and cognitive age models (p = 1.00 × 10 −3

| Associations with education and premorbid IQ
The CAG/BAG associations with education and premorbid IQ are provided in Table 5. Both higher premorbid IQ and years of education related to a younger cognitive age relative to chronological age (β = T A B L E 3 Average R 2 , RMSE, and MAE ± SDs for each of the models, and correlations (r) between predicted brain/cognitive age and chronological age. 95% confidence intervals are indicated in square brackets. RMSE and MAE are reported in years T A B L E 5 Associations between CAG/BAG estimates and education, premorbid IQ, and cumulative healthy lifestyle scores. 95% confidence intervals are indicated in square brackets. Model 1 was adjusted for chronological age only. Associations that were significant (before multiple comparison corrections) were submitted to linear regression with additional covariates (i.e., Model 2). Model 2 was adjusted for age, sex, education, ethnicity, BMI, and mutual adjustments between BAG and CAG. Note, where education and premorbid IQ were the independent variables of interest, mutual adjustments were also included in Model 2 (with BMI excluded from the list of covariates). Confidence intervals are indicated in square brackets. P-values are provided before and after FDR-correction. premorbid IQ or education, as shown in Table 5. Z-tests for correlated samples (using Equation (1)) showed that premorbid IQ related more strongly to CAG relative to BAG (Z = −3.13, p < .001, p corr = .472) as shown in Table 6.

| Lifestyle trajectories
The results of the LCGA are provided in SI Table S6 and Figure 3. were identified for CAG, as shown in Table 5.

| Associations with cumulative lifestyle scores
The CAG/BAG associations with cumulative lifestyle scores are provided in Table 5. Cumulative lifestyle scores associated negatively with BAG (β = −.06, 95% CI = [−0.12, −0.01], p = .01). However, this association did not survive corrections for multiple comparisons, as shown in Table 5 Table 6.

| Sensitivity analyses
After the exclusion of extreme outliers (n of remaining sample = 532),

| Post-hoc analyses
Given the significant association for cumulative lifestyle indices (prior to multiple comparison corrections; p = .029, p corr = .064), we investigated whether the use of binarized indices to compute our lifestyle score may have attenuated our ability to detect the associations of interest. Specifically, a cumulative lifestyle measure based on continuous versions of the variables (where available) was examined in a series of post hoc analyses. As shown in SI Table S7, a negative association between cumulative lifestyle scores and BAG was detected in F I G U R E 3 Trajectories of the three-class solution that was identified as the best fit, which are described in Section 3.3. Shaded areas reflect 95% confidence intervals both Model 1 (β = −.24, 95% CI = [−0.35, −0.13], p = <.001, p corr = <.001) and Model 2 (β = −.23, 95% CI = [−0.34-0.13], p = <.001, p corr = <.001). No significant associations were found for CAG (SI Table S7, p's > .05). Cumulative lifestyle scores were more strongly associated with BAG than CAG (Z = 2.813, p = .005, p corr = .015), SI Table S8.
Additionally, cumulative alcohol consumption was more strongly associated with BAG than CAG (Z = −2.456, p = .014, p corr = .021), whereas for smoker status over time, there was no difference in the strength of associations (Z = −1.566, p = .117, p corr = .117, SI Table S8).

| DISCUSSION
The present study is the first to demonstrate that by combining machine learning, MRI, and cognitive data, it is possible to quantify the concepts of brain maintenance (i.e., BAG) and cognitive maintenance (i.e., CAG) in parallel. Overall, we found no significant association between BAG and CAG. While this potentially supports the idea that there is limited correspondence between brain and cognitive maintenance in older age (Stern, 2002), the accuracy of the cognitive age prediction was relatively low, which may have limited sensitivity to detecting BAG and CAG relationships. However, the lack of association was also replicated when BAG was tested against measures of global and domain-specific cognitive function. Moreover, premorbid IQ was related to CAG independent of BAG, in line with established literature highlighting IQ as a potential proxy for cognitive reserve (Anthony & Lin, 2018). Our results are, however, in contradiction with other studies showing consistent links between BAG and cognitive performance (e.g., Boyle et al., 2020;Cole et al., 2018;Elliott et al., 2019). This discrepancy could potentially be attributed to disparities between the cognitive tests administered, or differences in sample characteristics.
4.1 | Brain maintenance versus cognitive maintenance: Potential utility of machine learning While brain age prediction is frequently used as a metric of brain preservation in aging (Franke & Gaser, 2019), there is currently no established method for capturing preserved cognitive function relative to degree of brain decline in cross-sectional studies. To date, the mea- Measures of cognitive reserve have traditionally consisted of either a stand-alone measure of education or premorbid IQ, or a combination of these indices (alongside occupational complexity and engagement in leisure activities (Valenzuela & Sachdev, 2006a;Valenzuela & Sachdev, 2006b)). The most commonly used index of reserve is education, with the rationale being that education may contribute to reserve by promoting cognitive flexibility (Stern, 2002) or accumulation of neural resources (Cabeza et al., 2018). Despite evidence to suggest that education is consistently linked to a lower risk of dementia (Anstey, Ee, Eramudugolla, Jagger, & Peters, 2019), and may moderate the relationship between neuropathology and cognitive performance (Bennett et al., 2003;Rentz et al., 2010), the use of education as a "proxy" of cognitive reserve is conceptually problematic: although cognitive activity may promote cognitive reserve, such mechanisms are unlikely to represent the only pathway between education and late-life cognition. Genetics and early life factors such as birth weight, childhood IQ, and parental education have been shown to influence brain and cognition continuously across the lifespan (Finkel, Reynolds, McArdle, & Pedersen, 2005;Lyons et al., 2009;Walhovd et al., 2016), and associations between education and cognitive function may be further mediated by socioeconomic status, risk of disease (e.g., heart disease, stroke and diabetes (Jones et al., 2011;Reed et al., 2010)), lifestyle factors, as well as individual potential for neural compensation and plasticity (Cabeza et al., 2018;Fjell et al., 2014). Educational attainment may therefore be better regarded as one of several predictors of cognitive maintenance in aging.
Another important consideration is that cognitive reserve refers to the deviation between observed cognitive function and what is expected based on the degree of age-or disease-related brain changes (Stern, 2002). Therefore, any measure of cognitive reserve needs to take into account the disparity between observed cognitive performance and underlying brain structure. In line with this, some efforts to improve measurements of this construct have defined cognitive reserve as the residual variance in episodic memory, after adjusting for demographic variables and structural MRI abnormalities (Reed et al., 2010). This approach to characterizing cognitive reserve as model "residuals" appears promising (Reed et al., 2010;Zahodne et al., 2013). For example, the residual term has been linked to risk of conversion from mild cognitive impairment to dementia, and has been shown to modify the relationship between brain atrophy and cognitive decline (Reed et al., 2010).
However, this body of evidence has focused exclusively on a single cognitive domain (memory), and considered only a handful of MRI markers of brain integrity, including global and hippocampal GM and WMHs. While the approach in the current study also utilizes model residuals, the use of machine learning provides the advantage of integrating a substantially greater number of MRI variables, as well as a large body of tests measuring performance across multiple cognitive domains. Furthermore, the age prediction models add a dimension to cross-sectional studies by capturing deviations from normative aging trajectories.
By generating separate measures of brain and cognitive maintenance, we are able to shed light on cognitive reserve mechanisms by (a) examining the relationship between brain and cognitive maintenance, (b) identifying discrepancies between their associations with variables of interest, and (c) mutually adjusting for these variables in the same model. In this way, our approach may contribute to addressing crucial questions in the aging field, including whether lifestyle factors predominately contribute to brain maintenance or cognitive reserve in late life.
4.2 | Favorable lifestyles are linked to higher brain maintenance The extent to which brain maintenance and cognitive reserve form pathways between favorable lifestyles and cognitive maintenance in aging is currently unclear. In our main analyses, only a weak association was found between more favorable lifestyles between midlife and late life and brain maintenance, and the association did not statis- our findings suggest that when brain and cognitive maintenance is of focus, lifestyle indices should integrate health behaviors in their continuous forms, where available, as this may improve sensitivity to the small effects often reported for lifestyle variables (Corley, Cox, & Deary, 2018). Using longitudinal assessments of BAG and CAG may also provide further information about how lifestyle affects brain and cognitive trajectories over time, as only cross-sectional estimates of these constructs were available in the current study.
Examining how genetic variability (Lee, 2013) mediates the associations of interest is also an important "next step" for future studies, to better capture interactions between genes and environment. For example, Lourida et al. (2019)) found that favorable lifestyles were linked to reduced dementia risk, specifically in individuals with a genetic risk of dementia. IQ is also known as a heritable trait (Plomin & Deary, 2015), and a possible genetic correlation between this trait and cognitive reserve could indicate that the phenotypic relationship observed between these factors reflects common underlying genetic architecture (Lee et al., 2018).
Future studies may therefore consider integrating polygenic risk scores and genetic pleiotropy to further capture the variance in CAG and BAG.

| Strengths and limitations
The multimodal MRI and detailed cognitive battery administered to most participants allowed us to generate a large number of MRI and cognitive features to be used in the estimation of brain/cognitive age.
While it is possible that measurement error introduced by cognitive tests may account for the reduced prediction accuracy of our cognitive model when compared to the MRI-based predictions, re-running these analyses with latent factors did not substantially modify the pre- applied to a test set with a wider range of cognitive scores could potentially improve the sensitivity to variation in cognitive age estimates. Given that a train/test split of the current sample (n = 537) would lower the sample size, this approach would reduce the model performance  as well as the statistical power to detect associations with the predictors of interest. Hence, prospective studies utilizing larger cohorts with a wider age range would also be more suitable for investigating how modifying the composition of training sets influences model performance.
Due to the relatively low accuracy of the cognitive age prediction in the current study, the weak association between the BAG and CAG estimates should be interpreted with caution. However, BAG did not correlate with measures of global or domain-specific cognitive function (SI Table S5), indicating that the lower accuracy of the cognitive age prediction could not fully explain the lack of association between the BAG and CAG estimates.
As compared to more traditional measures such as total GM volume and general scores of cognitive function, age prediction models distill a rich variety of brain characteristics and cognitive performance scores into single estimates per individual, which can be compared to normative age trajectories. However, summarizing measures across brain regions and cognitive domains cannot provide information about specialized cognitive networks potentially linked to regional brain characteristics. Combined with models that estimate regional brain aging patterns (de Lange, Anatürk, et al., 2020;Eavani et al., 2018;Kaufmann et al., 2019;Smith et al., 2020), a development of processspecific cognitive models may provide a more detailed and accurate estimate of the relationship between brain and cognitive maintenance. Furthermore, dynamic brain processes that are not captured by structural measures may play an important role in reserve mechanisms (Franzmeier et al., 2017;Sala-Llonch et al., 2012). Future studies including models based on functional MRI data may thus help to further elucidate the mechanisms involved in cognitive maintenance in aging.
One main strength of this study is the longitudinal and repeat data available on lifestyle indices (over a mean of 17 years), which was provided by a large and highly characterized sample. However, only smoker status, alcohol intake, and PA were included in the lifestyle indices, and a several other factors, including diet (Lourida et al., 2019) and sleep patterns (Tanaka & Shirakawa, 2004), may form part of a healthy lifestyle. A more comprehensive characterization of lifestyle trajectories could thus potentially increase their explanatory value.
Another consideration is that only 17.5% of the sample were female, reflecting sex distributions in the British civil service at the time of initial recruitment (1980s), which limits the generalizability of the reported results. Consistent with previous studies investigating the WHII imaging sub-sample (Demnitz et al., 2017;Sexton et al., 2017), a "healthy volunteer effect" was further observed (SI Table S9), such that included individuals were significantly younger and had higher MoCA scores than excluded individuals. Participants in our study further appeared to be more educated, more physically active, less overweight, and less likely to smoke than the general population, based on nationally representative data sources (SI Table S10). However, no statistical comparisons were conducted due to substantial differences in the methods used to collect this information. As these sample characteristics are likely to limit the generalizability of our results, replication in more representative samples are required. Lastly, no causal conclusions can be drawn due to the observational nature of the study.

| CONCLUSION
In conclusion, this study describes a novel method for characterizing brain and cognitive maintenance relative to normative trajectories in aging. While prospective studies are required to validate these metrics in other cohorts with different sample characteristics, the presented approach provides a method for quantifying and comparing brain and cognitive age estimates, which may be valuable for future studies seeking to investigate factors underlying maintenance and reserve mechanisms in older age.