Aging is a complex biological phenomenon, characterized by the gradual accumulation of molecular, cellular and tissue damage (Lopez-Otin et al., 2013). The body’s inability to repair itself precedes a decline in physiological functions, including sensory, motor, and cognitive functions that are important for maintaining independence and quality of life (Lopez-Otin et al., 2013; World Health Organization, 2015). The brain is susceptible to the effects of aging, undergoing many structural and functional changes over the lifespan. Most widely recognized is brain atrophy (the loss of tissue volume), which has been associated with a decline in cognitive function (Cabeza et al., 2018; Grajauskas et al., 2019). However, there is considerable diversity in the rate of aging, which is influenced by genetic and environmental factors.

Modelling ‘brain age’ has deepened our understanding of the interindividual differences in brain aging. This includes ‘brain age’, a measure of biological age that is predicted from neuroimaging at the individual level relative to a normative model. A person’s brain-predicted age difference (brain-PAD) represents the deviation of brain age from the normal aging trajectory. An older brain age relative to chronological age is considered a sign of ‘advanced’ brain aging (i.e., greater age-related brain atrophy), and has been linked with a lower cognition and a greater risk of dementia (Franke et al., 2010; Elliott et al., 2021). Conversely, a brain appearing younger than expected relative to one’s chronological age reflects ‘resilient’ brain aging, and has been predictive of better physical fitness and cognitive function (Kolbeinsson et al., 2020; Sanders et al., 2021). A number of brain age algorithms have been developed, including a model by Cole et al. (2018). This uses voxel-level measures of grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) volume to provide a single estimate of brain age measured across the whole brain. Using this model, we have found an association between advanced brain aging and poor cognitive processing (Wrigglesworth et al., 2022a), and identified sex differences in the change in brain aging over time (Wrigglesworth et al., 2022b).

As a surrogate measure of brain health, brain-PAD has provided useful insights into individual differences in biological aging trajectories, and their direct relationship with cognition, brain diseases, and other health outcomes (Lowe et al., 2016; Cole et al., 2020; Vidal-Pineiro et al., 2021). However, evidence suggests there may be additional heterogeneity hidden within populations of resilient or advanced brain agers (Eavani et al., 2018). For instance, Eavani et al. (2018) identified five different brain imaging phenotypes in cognitively unimpaired older adults with advanced brain aging, demonstrating differential neuroanatomical substrates, and thus potentially unique pathological pathways, underlying trajectories of unhealthy brain aging. Sub-categorizing people with a resilient or advanced brain-PAD may therefore be an important avenue of inquiry that could provide greater individual specificity in predicting distinct profiles of current and future health status. To address this hypothesis, this study aimed to examine heterogeneity within each category of brain-PAD relative to other health, behavioral or cognitive measures. We examined this question independently for each brain age group (advanced and resilient) to avoid assumptions regarding unique or overlapping outcomes within these qualitatively distinct cohorts. Through the use of data-driven approaches, we first identify healthy elderly people with advanced and resilient brain age from a large community-based cohort, then we identify unique clusters within these advanced and resilient brain age groups based on other health factors, and finally we determine whether these subprofiles are differentially associated with longitudinal change in cognitive function and frailty.

In a cohort of relatively healthy, cognitively unimpaired older adults, this study used a combination of brain age prediction, data-driven clustering and predictive modelling to determine whether there is heterogeneity within ‘resilient’ and ‘advanced’ brain age groups related to health, behavioral or cognitive measures, and how this heterogeneity relates to future longitudinal trajectories of cognition and frailty. We first demonstrate that there is health-related heterogeneity in the older population that is independent of brain aging. This was summarized by two qualitatively similar health-related clusters being identified in each brain age group, one largely characterized by physical qualities (i.e., low prevalence of obesity and better physical health; Class 1), and the other by poor cardiometabolic health and cognitive function (Class 2). Class membership was predictive of longitudinal change in cognitive function only for resilient brain agers.

To our knowledge, this is the first study to explore whether there are health-related profiles of brain aging that differentially contribute to changes in cognitive and functional performance. While brain age provides a simple and interpretable measure for understanding deviations from normative aging trajectories, prior evidence suggests there may be multiple imaging signatures that effectively characterize brain aging in older adults (Eavani et al., 2018). Data-driven approaches have also identified considerable diversity in neurological diseases, which holds great promise for precision medicine (Habes et al., 2020; Wen et al., 2022).

Our findings suggest that cluster formation based on cardiometabolic, physical and cognitive factors appear to be a stable and generalizable feature of the older population. This aligns with established literature (Oldridge et al., 2001; Waldstein and Katzel, 2004; Sullivan et al., 2007; Lai et al., 2020), and findings from the larger ASPREE study (19,114 participants from United States and Australia; Phyo et al., 2021a). In both brain age categories, “cluster 1” was characterized by lower prevalence of obesity and better perceived physical QoL (i.e., defined by physical functioning, role-physical, bodily pain and general health; Ware et al., 1996), while “cluster 2” had a higher prevalence of hypertension, poorer physical QoL, and a lower general cognitive status.

Our finding that living alone and higher cognitive performance are linked to low levels of obesity and better physical health was confined to people with advanced brain aging. This was despite the comparable proportion of these factors in the resilient and advanced brain age groups. Here, discrepancies may partly reflect a benefit of independent living on cognitive function through flexible neural processes, even with increasing brain atrophy (i.e., cognitive reserve; Evans et al., 2019; Stern et al., 2020; Villa et al., 2021). Living alone is a common occurrence for older adults due to the natural shift in social environments with aging (Evans et al., 2019; Bzdok and Dunbar, 2020; Villa et al., 2021). Although this imposes a greater risk for social isolation and loneliness (Finlay and Kobayashi, 2018), there is some evidence to suggest older people who live alone are more likely to engage in regular social activities (Evans et al., 2019). Social engagement requires complex cognitive processes that may help improve reserve (Stern et al., 2020). While this remains speculative, people with advanced brain aging may have maintained a higher cognitive performance through the influence of independent living on cognitive reserve, but this warrants further investigation (Evans et al., 2019; Stern et al., 2020; Villa et al., 2021).

Class membership was predictive of longitudinal change in cognitive function for resilient brain agers only and remained when assessed in the total cohort. In this subgroup of resilient brain agers, individuals with greater cognitive ability and physical functioning at baseline were more likely to show cognitive decline over 5 years. Conversely, poorer cognitive function and physical health were associated with improvement on the cognitive tests, on average, over the study period. This was a small effect that may reflect a ‘regression to the mean’ (i.e., resilient brain agers with a high or low cognitive score at baseline will present a score that is closer to the mean at follow-up; Barnett et al., 2005). However, the analogous subgroup of people with relatively poorer cognition and physical health in the advanced brain age group did not demonstrate a similar effect; these individuals declined in performance over time at the same rate as those in the high-cognition subgroup. Therefore, these findings could indicate a residual capacity, or “brain reserve” (Stern et al., 2020) in the resilient brain age group. For individuals already functioning at or near to ceiling for their age group, the most likely longitudinal trajectories are stability or decline, whereas for people operating below capacity at baseline, opportunities for performance improvement due to learning (e.g., repeated practice), motivation, or intervention are available. This same reserve is not available to people with an advanced brain age. Investigating the qualitative differences in longitudinal growth trajectories will overcome this study’s limitation of identifying latent classes cross-sectionally, and further our understanding of these findings. Interestingly, some epidemiological studies have also reported a decelerated cognitive decline, and a reduced risk of incident dementia, in older adults at increased risk of metabolic disease, including a high body fat mass and hypertension (Van den Berg et al., 2007; Forti et al., 2010; Watts et al., 2013). Although this finding may be specific to individuals with early neurodegenerative disease, as opposed to the healthy population (Watts et al., 2013), emerging evidence suggests a higher body fat mass may actually be beneficial for maintaining cognitive function in older adults, especially women (Luchsinger et al., 2013; Bohn et al., 2020). A meta-analysis of six prospective community-based studies also identified a reduced risk of dementia when hypertension is controlled through anti-hypertensive medications (Ding et al., 2020); more than half of the current NEURO cohort with hypertension were on treatment for high blood pressure. Further, individuals who have a stable cardiovascular risk profile over 5 years have a reduced likelihood of dementia compared with those with increasing risk over time (D’Agostino Sr et al., 2008; Farnsworth von Cederwald et al., 2022). While there remains insufficient evidence to provide a clear interpretation of how different lifestyle and health factors influence aging trajectories, future prospective analysis of these health clusters may improve our understanding of the interaction between brain aging and cognitive outcomes in older adults.

This study has a number of strengths. Firstly, we derived estimates of brain age from MRI following a rigorous quality assessment and using a publicly sourced model that has been previously validated using an independent test cohort (Cole et al., 2018), and which we have previously shown is not biased by chronological age in our population (Wrigglesworth et al., 2022a). We also identified subclasses of people within the resilient and advanced brain age groups using a person-centered approach that focuses on identifying homogenous groups of people rather than homogeneous groups of variables (Muthén and Muthén, 2012). Further, our application of LCA allowed for quantification of uncertainty of class membership through posterior probabilities, and permitted the usage of mixed data types (Andreopoulos et al., 2009; Sinha et al., 2021). A robust approach to assessing model fit and classification accuracy was adopted through the combined use of multiple inferential (i.e., LR chi-square goodness of fit) and descriptive tools (i.e., AIC/BIC information criterion). Finally, phenotypic changes in frailty and cognitive function were investigated prospectively using data from a large community-based older cohort, focusing on a stage of life characterized by great heterogeneity of aging trajectories.

There are several limitations that must also be considered. First, the moderate correlation between brain age and chronological age for our study cohort was weaker than achieved for the held-out test (n = 857, r = 0.973) and validation cohorts (n = 611, r = 0.947; Cole et al., 2018). Here, a less than accurate fit of the model may relate to the narrow age-range of our study sample, which contrasts the broad age-range of the cohorts used to train and validate the model (18 to 92 and 90 years respectively). This is further supported by the larger mean error of prediction when adjusted for age-range (0.28 vs. 0.05 and 0.07 years for held-out and test datasets, respectively).

We identified latent classes cross-sectionally, rather than exploring the qualitative differences in longitudinal growth trajectories (Nylund et al., 2007; Eavani et al., 2018). General limitations of LCA include handling missing data, sensitivity to extreme outliers, and reliance on relatively large sample sizes for reliability (Mccutcheon, 1987; Nylund et al., 2007; Collins and Lanza, 2009; Berlin et al., 2014). We mitigated the influence of these limitations by evaluating multiple fit criteria and selecting variables with a relative frequency greater than 10% of the sample (Sinha et al., 2021). Our study sample included primarily white people who were generally healthy, clinically cognitively unimpaired [i.e., no dementia diagnosis or 3MS (Teng and Chui, 1987) score ≥ 78], and had completed 12 or more years of formal education, potentially limiting generalizability to the wider population (Masyn, 2013). Longitudinal study attrition may have included individuals with greater health deficits and poorer cognitive function. Lastly, given the modest number of participants within each class, we may not have been sufficiently powered to detect subtle interindividual differences in the longitudinal change in frailty index and cognitive function over the three and 5 year follow-up period, respectively (Sinha et al., 2021).

In conclusion, this study is the first to demonstrate heterogeneity related to health, behavior, cognitive function, quality of life within the construct of brain age prediction in an older community-dwelling cohort. Given the large spectrum of interacting factors in a real-world setting, it is critical to our understanding of brain age to consider the influence of biological and functional measures in aggregate. Identification of differences in health-related characteristics based on brain aging may help accurately predict long term outcomes, prevention and treatment strategies.

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by Monash University Human Research and Ethics Committee. The participants provided their written informed consent to participate in the ASPREE and NEURO studies, and study procedures were conducted in accordance with institutional guidelines.

JW and IH contributed to the conception and design of the study. JW performed the statistical analysis and wrote the first draft of the manuscript. JR helped with the analysis and interpretation of the results. All authors contributed to the article and approved the submitted version.

ASPREE received funding from the National Institute on Aging and the National Cancer Institute at the National Institutes of Health (U01AG029824), Monash University, the Victorian Cancer Agency, and the Australian National Health and Medical Research Council (NHMRC, grant numbers 334047 and 1127060). NEURO received funding from NHMRC (1086188) and support from Monash Biomedical Imaging. JW is the recipient of a Research Training Program stipend, awarded by Monash University and the Australian government. IH is supported by a NHMRC Fellowship (APP1106533). JR is funded by a NHMRC Dementia Research Leader Fellowship (APP1135727). Funders neither directed the conduction of this study, nor the decision to publish these findings.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.