Introduction

The global demographic shift over the last few decades leading to increased proportion of older adults resulted in steadfastly raising numbers of people at risk of cognitive decline and dementia. This has long-lasting impact not only on the individual quality of life but also major economic impact on the societies worldwide [1]. While the ageing process is inevitable and comes with expectation of worsening in cognitive functioning, the literature examining cognitive ageing points to a striking heterogeneity in the rate of age-related deterioration in cognitive decline. It means that some older adults retain high levels of mental capacity, some undergo a gradual drop in cognitive functioning, and finally some experience a sharp cognitive decline (dementia) hindering their ability to undertake basic daily activities and affecting quality of life [2–6]. To keep up with the ageing population and to enable advances in medical/social care a better understanding of factors affecting these different cognitive outcomes is urgently needed [1,7].

Cognitive ageing is a highly complex process caused by a convergence of multiple neurobiological and neurophysiological changes influenced by environmental and genetic factors as well as gene-environment interactions [8–12]. Among environmental influences, sociodemographic and lifestyle factors seem to play a critical role [5,6]. It is well evidenced that the cognitive ageing is a result of structural and functional brain changes causing deterioration of cognitive performance. However, the interpretation of multifactorial influences on age-related changes in neural networks underlying the differential trajectories of cognitive ageing presents an ongoing challenge. There is a growing evidence that older adults, who have engaged across lifespan in cognitively and/or socially enriched environments exhibit greater resilience to cognitive decline and maintain better brain health and cognitive function in later year. The concept of cognitive reserve (CR) has been proposed to account for at least some the observed heterogeneity in healthy (non-dementia) cognitive ageing as well as susceptibility to dementia, pathological decline [5,6]. Within this framework, CR is indirectly measured by proxies of life experiences, in particular education but also occupation and leisure activities [5,13–16].

While the concept of CR offers appealing framework for understanding cumulative neural changes due to lifelong environmental influences, including specifically socio-behavioural factors, and their interplay with genetic variability, yet to date few theoretical and experimental attempts have been made in understanding the neurobiological origins of cognitive reserve in healthy ageing [5,17–21]. Here, we used one such theoretical framework (Robertson’s theory [18]) of CR to test the potential of machine learning methods to tease apart combinatorial effects of education (as CR proxy) and genetic influences on brain networks and cognitive outcomes (heterogeneity in cognitive function) in a large cross-sectional ageing cohort, the UK Biobank study. As UK Biobank has been created as a large-scale epidemiological resource with extensive sociodemographic questionnaires, physical measures, medical records, neuroimaging, and genetic data from middle and old age participants [22], it enables to tease apart multifactorial influences on cognitive ageing, using data science approaches on a scale not feasible before.

Robertson’s theory of CR proposes that life experiences acting on the noradrenergic system strengthen the right fronto-parietal networks to facilitate cognitive reserve and maintain cognition throughout lifetime [18,23]. This theory combines Stern’s observations on the effects of education in Alzheimer’s patients with animal research. Animal studies provide evidence that across the life-span neural networks are heavily subjected to regulatory influences of neuromodulators (dopamine, noradrenaline, serotonin, and acetylcholine). These neuromodulators are thought to maintain functional dynamics and interactions among large-scale brain networks and to optimize cognitive performance by signalling environmental inputs (for review see [24]). Strikingly, a growing body of evidence from human neuroimaging studies points to functional connectivity representing both within and between network interactions, as key to understanding CR and trajectories of cognitive decline [25–28].

In line with animal research Robertson’s theory stipulates that life experiences trigger continues engagement of several core cognitive process, all of which functionally rely on locus coeruleus—noradrenergic system and right fronto-parietal networks {18, 23]. A recent study linking cognitive performance and brain health in both non-demented older adults and Alzheimer’s patients to the volumetric estimates of locus coeruleus has provided compelling evidence that indeed the noradrenergic system might underpin the CR [19]. There is also evidence in support of right lateralised cognitive reserve network [17,20,21,29]. For example, our group has previously shown that life experiences mitigate age-related cognitive deficits by (i) preserving grey matter withing the right fronto-parietal regions [17] and (ii) offsetting age-related axonal dispersion within the right fronto-parietal white matter [21]. Interestingly, we previously demonstrated that both the composite measure of CR (derived based on education, occupation and cognitively stimulating leisure activities) and the education alone have an offsetting effect on age-related changes in attention function by preserving grey matter within the right fronto-parietal networks [17]. It should be noted that none of the previous studies examining neurobiological basis of Robertson’s theory of CR simultaneously tapped into the noradrenergic system and the right fronto-parietal networks, which current study addresses.

Growing evidence suggests that genetic variations, leading to either elevated or decreased levels of neuromodulators have impact on the functional dynamics of the neural networks, and underlie inter-individual variability in cognitive abilities across human lifetime [12,30]. Thus, along these lines Robertson’s theory of CR presents an interesting framework for understanding mechanism of cumulative socioeconomic and genetic influences on functional neural networks and whether such interplay contributes to CR. It enables to stipulate that genetic variability, enhancing neuromodulation across lifespan, and strengthening the right fronto-parietal networks in response to life experiences (e.g., education) could be advantageous in offsetting age-related cognitive decline. To test such proposal, we used here machine learning approaches, previously successfully applied by our group in large epidemiological studies [31–33] to model interactions between education and neurobiological features (measures of structural and functional fronto-parietal connectivity and genetic variants in noradrenergic signalling) on cognitive outcomes in the ageing UK Biobank cohort.

Materials and Methods

Modelling

As our objective was to model interactions between neurobiological features (both genetic and connectivity measures) and education (proxy of CR) on cognitive outcomes, we built regression models using general cognitive ability score (GCAS) as outcome. Using workflowsets and tidymodels, 5 different algorithms (generalized linear model (glm package), LASSO [51] (glmnet package), a single-hidden-layer neural network (nnet package), random forest (ranger package) and xgboost (xgboost package), were fitted to 14 different combinations of the features above extracted. More precisely, 5 different algorithms were run on data comprising: SNPs, FPC IDPs, AgeSex, SNPs x FPC IDPs, SNPs x CR (education), SNPs x AgeSex, FPC IDPs x CR (education), AgeSex x CR (education), SNPs x FPC IDPs x CR (education), SNPs x FPC IDPs x AgeSex, SNPs x AgeSex x CR (education), FPC IDPs x AgeSex x CR (education as proxy), SNPs x FPC IDPs x AgeSex x CR (education) information. In each algorithm and feature combination dataset (69 in total) (Fig 1), hyperparameters were tuned through grid.search and 100 cross-validation resamples fitted per model. To cut down on time, tune_race_anova [52] was used which eliminates tuning parameter combinations that are unlikely to be the best results. To do so, tune_race_anova uses a repeated measure ANOVA model just after an initial number of resamples have been evaluated. Performance was evaluated through Mean Absolute Error (MAE), where the sum of absolute errors between predicted GCAS and true GCAS is divided by the sample size.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Study Design.

We explored the interplay neurobiological features (genetic variants in noradrenergic signalling, structural and functional fronto-parietal connectivity) and education (proxy of CR) by testing regression models predicting general cognitive ability score (GCAS) as outcome. By splitting features into 4 datasets: GWAS derived SNPs, AgeXSex, CR with Education as proxy and imaging derived phenotypes representing structure and functional fronto-parietal connectivity (FPC IDPs), we generated 14 different dataset combinations. These 14 different feature combinations were then modelled with 5 different algorithms (Least Absolute Shrinkage and Selection Operator (LASSO), Generalized Linear Models, XGBoost, Random Forest and Neural Nets), generating 69 different models. Behaviour of these models was then evaluated through performance, using mean absolute error and feature importance. Given the disparity in modelling algorithms, different feature importance algorithms were also used (DALEX, model dependent variable importance assessment, permutation and SHAPLEY), and results were summarised as ranking scores. Top 10 features (sex and 9 neurobiological variables) selected as most important (top 2) per variable importance method, algorithm and dataset were then studied as mediators (9 neurobiological variables only) between education (CR) and GCAS and including age as covariate.

https://doi.org/10.1371/journal.pone.0350452.g001

Results

Modelling

To understand what data better captured cognitive outcome, different models were built predicting GCAS using both different feature combinations and modelling strategies (Fig 1). More specifically, generating all different feature combinations between brain imaging, genetic, demographics and education yielded 14 different datasets as described in Methods. All these datasets were then trained using 4 distinct machine learning algorithms combining both non-linear and linear methods (Random Forest, Neural Nets, LASSO and GLM), creating 69 different models. After splitting our dataset into training (5/6) and testing (1/6) data, models were trained, with hyperparameter tuning and resampling (100 fold-cross validation).

Final performance was assessed in the test set (Fig 2) and feature importance through 4 metrics (vip, permutation, shapley and DALEX) for all models is reported (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Results from All Regression Models.

Mean Absolute Error (MAE) of model combinations of AgeSex, fronto-parietal connectivity imaging derived phenotypes (FPC IDPs) and SNPs data for each of the 4 different machine learning algorithms are displayed (A. Neural Nets, B. LASSO, C. Random Forest, D. GLM and E. XGBoost), using the model with all datasets included as reference value (red line). In total, 69 models were analysed (with unique feature and algorithm combination). As a general trend, SNPs possess the least prediction ability, with the highest mean average error across all algorithms (higher loss of MAE with respect to baseline 0 (which is all datasets together). In contrast, including AgeSex x CR (education) guarantees a better performance across all datasets, and with IDPs x AgeSex x CR (education) generally yielding the best performance. All hyperparamters were tuned using tune race_anova with best performing hyperparameter combinations used and test set performance reported here. Error bars correspond to the 100 fold cross-validated performance results.

https://doi.org/10.1371/journal.pone.0350452.g002

A combination of imaging, demographic and education data was seen to be the best performing model in all algorithms, with these feature combinations yielding better performance than combining all features together (reference line in Fig 2). To better understand feature importance per model and algorithm, heatmaps were used to visualize results, with information on the models generated form imaging, demographic and education data (Fig 3). In this heatmap, the ranking of the top 9 features per feature importance methods applied to each model is reported, with Age and Education scoring first and second in all algorithms and methods but less consensus seen in the ranking of imaging data, with varying agreement between algorithms and feature importance methods. Results from the other models can be found in: https://github.com/InFlamUOB/CognitiveReserve

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Feature Importance Testing.

Heatmap of feature importance for the 4 out of the 5 different algorithms (LASSO and GLM as both linear models, are here only represented as LASSO results) and data comprising features from AgeSex, CR (education) and fronto-parietal connectivity imaging derived phenotypes (FPC IDPs). Most important features are ranked using 4 different feature importance algorithms (DALEX, model specific variable importance (VIP), permutation and Shapley values), with most important being 1 (yellow) and least 9 (purple). Only the 9 most important features for each models are here reported. All model-agnostic methods (permutation, DALEX and Shapley) had variable importance assessed though root mean-squared error (RMSE) and the absolute values of importance were assessed as ranking. In brackets, the normalized absolute importance value is reported as percentage.

https://doi.org/10.1371/journal.pone.0350452.g003

Mediation

Finally, wanting to better understand the relationship between education and cognitive outcome, we proceeded to explore the most frequent important features extracted from each variable importance method for each algorithm and dataset in a mediation model. Age was found as the most important variable across models and datasets when included (Fig 3). Thus, taking into account relevance of age to studied hypothesis and employed study sample (ageing UK Biobank cohort including middle age and older adults), we include age in our mediation analysis as an extra covariate and evaluate the top 9 frequent neurobiological features as mediators. More specifically, the direct effects of education (proxy of CR) on general cognitive ability score partialling out the effect of each of these features univariately was evaluated with the main effects reported in Table 1. Following that seen as best model (FPC IDPs, education and Age features), the variable seen to have a significant mediation effect was rFPN-dmDMN connectivity (Fig 4, Table 1). The results based on the imputed data are reported in the S6 Fig. rFPN-dmDMN connectivity was found as a significant mediator in both analyses, i.e., using only complete cases and imputed data, thus ascertaining the robustness of findings.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Mediation Analysis Results.

https://doi.org/10.1371/journal.pone.0350452.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Mediation Analysis Results. rFPN-dmDMN connectivity was found to be a significant mediator with education (proxy of CR) as predictor of General Cognitive Ability Score (GCAS) when controlling for Age (95% CI not crossing 0).

Mediation diagram as extracted from mediate.diagram function from the psych package. Significant mediator (95% CI not crossing 0) when controlling for Age.

https://doi.org/10.1371/journal.pone.0350452.g004

Discussion

Using robust machine learning methods and mediation analysis applied to data from the UK Biobank Brain Imaging study [34], we tested here combinatorial effects of neurobiological features (genetic variants in noradrenergic signalling, measures of structural and functional fronto-parietal connectivity) and education (used here as proxy of CR) on cognitive outcomes in the UK Biobank ageing cohort. Initially, several different models were built with a large number of features. We then selected a small number of features, related to genetic variance, fronto-parietal connectivity alongside sex and age, for subsequent analysis using mediation testing. A variety of machine learning algorithms and feature importance methods were used given that there is no algorithm able to fit all data perfectly (“No Free Lunch Theorem” [58]). Therefore, testing more than one algorithm and more than one approach to interpret feature importance provided us with the possibility of acknowledging many more possible underlying relationships. In addition, we employed here feature importance testing as highly relevant for black box algorithms such as neural nets and random forest where behaviour of the model is more difficult to grasp. Based on these analyses we show that (1) interactions between education and neurobiological variables more fully explain general cognitive performance across multiple tests (general cognitive ability score) than either factor alone, and (2) among the examined neurobiological features measures of functional fronto-parietal connectivity are the strongest mediators of the association between education and cognitive outcomes. Our analyses have not provided any evidence of the mediating effects of genetic variability in noradrenergic signalling on interplay between education and cognitive outcomes.

Our main findings point to fronto-parietal resting state functional connectivity measures (rFPN-dmDMN, Fig 4) representing connections between the default mode network and the right fronto-parietal network as the significant mediator between education and cognitive outcomes in UK Biobank ageing population sample. This is in line with previous studies linking cognitive reserve to functional network dynamics [25–28]. Our findings specifically suggest that functional connectivity between several neural networks, including right fronto-parietal network, might be sensitive to the effects of education (proxy of CR) on cognition in middle and old age. Thus, alongside our previous studies [17,21], we provide here evidence in support of Robertson’s theory of CR [18,23]. Robertson’s theory stipulates that at the neural level the lifelong exposure to cognitive stimulation strengthens the right hemisphere fronto-parietal networks, which in turn offsets the symptoms of cognitive ageing. Our previous work [17,21] and that of others [29], indeed point to the right hemispheric fronto-parietal networks as neural substrates of CR. By contrast here we linked the phenomenon of CR to not only the right fronto-parietal network but also the default mode network encompassing additional bilateral fronto-parietal regions, as well as the connectivity between these two networks. The Robertson’s theory proposes that exposure to enriched environments acts via core cognitive processes, including alertness, sustained attention, and awareness which indeed are known to be subserved by the right fronto-parietal networks (e.g., [59–63]). However, the default mode networks implicated by our current study is also functionally linked to the cognitive processes described by Robertson’s theory. Furthermore, similarly to fronto-parietal network, the default mode network is known to be modulated by the locus coeruleus-noradrenergic (LC-NE) system [64–67], which is pivotal to Robertson’s theory.

It should be also noted that while the mentioned above studies, supporting Robertson’s theory, only examined the beneficial effects of CR on structural measures within fronto-parietal networks [17,21,29], the current study links the phenomenon of CR (measured here by proxy of education) to measures of functional connectivity. Several previous studies in non-demented older adults linked measures of functional connectivity to offsetting effects of CR either on global cognition or memory performance (e.g., [26,68–70]. However, most of these studies either explored connectivity measures derived from a task-based fMRI and/or the analysis were primarily restricted to measures of within-network connectivity, and additionally the previous findings were limited by relatively small sample sizes (less than 100 participants). By contrast we explored here between networks connectivity in a large cross-sectional UK Biobank cohort. But what is the most striking about our findings is that they point to the default mode network as the key player in between networks interactions underlying CR. As noted above this expands the Robertson’s proposal [18], which specifically stipulates that the hypothetical CR network consists of interconnected right lateral prefrontal lobe and right inferior parietal lobe regions. Our findings expand the proposed by Robertson CR network, to include bilaterally dorsal medial prefrontal cortex and posterior cingulate cortex.

In the context of ageing, the default mode network is perhaps the most studied of the resting state networks [71,72] with functional changes linked to cognitive decline and dementia [73–79]. In addition, when examining a small sample of Alzheimer’s patients, Bozzali and colleagues [80] previously found evidence that CR modulate connectivity within the default mode network. Finally, our results expand on newly published findings linking changes in connectivity between the default mode network and fronto-parietal networks to cognitive decline in elderly [81] as well as to the effects of beta-amyloid on cognitive status in Alzheimer’s patients [82] in line with classical account of CR [83,84] i.e., mitigating accumulated neuropathology.

While on one hand the limitation of the current study is that we only examined between network connections, on the other hand several theories of cognitive ageing explicitly address changes in functional correlations between functional networks associated with either preservation or decline in cognitive performance in ageing. Our results pointing to between the fronto-parietal network and the default mode network connectivity are thus compelling, in terms of mechanistic significance and insights into cognitive ageing. Specifically, it has been suggested that processes such as compensation and dedifferentiation both might be accompanied by stronger correlations among functionally unrelated networks or weaker anti-correlations among competing networks [85–87]. In this context the default mode network is of particular interest as the existing evidence points to anti-correlations between the default mode network and various other networks including right fronto-parietal networks at rest and during task performance in young healthy participants [88,89]. The functional coupling between the two networks has been shown to be essential for executive function and cognitive control, in particular attention, working memory and goal directed behaviours [90–92]. Strikingly, it has also been demonstrated that in elderly participants the default mode network is significantly less deactivated during task performance and that this diminished suppression might underly poorer cognitive performance [75,93,94]. Finally, there is evidence that functional connectivity between the fronto-parietal network and the default mode network changes across lifetime with implications for cognitive decline [81,93]. Thus, future research should address both cross-sectionally and longitudinally whether the offsetting effects of exposure to CR proxies on cognitive functioning in older adults translate into changes in co-activations and anti-correlations between the default mode networks and task-related fronto-parietal networks while participants are performing cognitive tasks using functional MRI.

To our knowledge, only one previous study examined underpinnings of the effects of CR on cognitive outcomes in the UK Biobank cohort [95]. However, the scope, research objectives and methods employed by Jin et al [95] and here are very different. Jin and colleagues specifically conducted a series of studies systematically exploring the effects of different CR proxies (education, leisure activities, fluid intelligence, social interactions, and physical activity) on the relationship between structural brain measures (global volume, regional volumetric measures and cortical thickness within brain areas know to deteriorate in dementia patients) and cognitive outcomes indexed by performance on tests within 4 separate cognitive domains. The main aim of their research was to explore the reported inconsistencies in the literature concerning the moderating effects of various CR proxies on the links between brain structure and cognitive abilities in an ageing population. This was motivated by the premise that CR explains the discrepancy between accumulated brain damage and observed cognitive performance in accordance with classical theory of cognitive reserve as proposed by Stern and based on his early work in Alzheimer’s patients [83,84]. By contrast our analyses were motivated by the Robertson’s proposal [18,23] that CR, here represented by education strengthen the right fronto-parietal networks to offset age-related cognitive decline. Thus, the only link between Jin et al [95] and our study, conducted using data from UK Biobank cohort, is that both provide support for the notion of beneficial effects of CR on cognitive ageing.

One of the objectives of the current study was to explore a premise that an interplay between accumulative effects of CR (represented by education) and genetic variance in noradrenergic signalling on fronto-parietal neural networks might offset age-related cognitive decline. While previous studies indeed linked genetic variations in neurotransmitter signalling to heterogeneity in cognitive ageing [12,30,96,97], we have not found any evidence of the mediating effects of genetic variability in noradrenergic signalling on interplay between education and cognitive outcomes. It should be noted that the previous evidence comes from hypothesis driven genetic association studies using single-nucleotide polymorphisms (SNPs). By contrast we employed here GWAS based approach to find genetic variants associated with education, cognitive outcomes and fronto-parietal connectivity (structural and functional). While the GWAS analyses were substantially restricted to the preselected gene set, 154 genes totalling 464,939 SNPs (fraction of the number of SNPs typically employed in GWAS, i.e., approximately 10 million SNPs) the negative findings might not be that surprising considering a relatively small sample comparing to other GWAS studies examining genetic underpinnings of education and cognition (e.g., [98,99]). Thus, the size of study sample constitutes a limitation when exploring genetic influences. The negative findings might also reflect potential shortcomings of the taken here genetic approach to testing Robertson’s theory. Many epidemiological studies rather than using GWAS employ methods such as polygenic risk scores (PRS) or Mendelian randomisation (MR) to study genetic influences and causal associations with specific traits. It would be particularly interesting to follow up with MR approach combined with longitudinal data. However, based on the current study, use of PRS would be unlikely beneficial. Our framework contains tests, LASSO and logistic regression, that are conceptually analogous to PRS. In addition, while PRS assumes a strictly linear aggregation of SNP effects, our framework extends that by evaluating nonlinear models (random forest, XGBoost, neural networks), allowing for more flexible representations of genetic effects. This implies that a conventional PRS would not yield better predictive utility. Another potential limitation is use of genetic variance in noradrenergic signalling as variable depicting noradrenergic modulation rather than examining the locus coeruleus (LC) function as more direct measure of noradrenergic activity. It is plausible that the noradrenergic genetic variance does not sufficiently capture cumulative influences of cognitively enriched environments as per Robertson’s theory of CR. However, it should be noted that the beneficial effects of noradrenergic signalling and the LC function in the context of both cognitive decline and progression of Alzheimer’s disease have been recently questioned by many (for comprehensive review see [100]. Furthermore, it is plausible that education measure used here as sole proxy of CR does not reflects the complexity of lifelong cognitively stimulating experiences influencing noradrenergic function. Finally, our analyses have not controlled for sociodemographic and health conditions, potentially either hindering reported findings or their generalisation [101]. Thus, further work should consider a broader approach to factors contributing to CR such as lifelong learning (degree type, part-time and continuing education), occupational, leisure and social activities alongside broader socioeconomic and demographic factors across lifespan. Although in defence of employed here CR proxy, it could be argued that education lessens the decline in cognitive functioning as we age, not only via enhancing neural resources early in life (childhood till early adulthood) with effects lasting later in life, but also at least partially influences professional and lifestyle choices throughout life, i.e., other proxy measures of CR. On contrary, some argue that education reflects only individual differences in cognitive abilities that persist from childhood to old age (for review see [5,102]). The assumption is that education attenuates the decline in cognitive functioning as we age, likely via enhancing neural resources early in life (childhood till early adulthood) with effects lasting later in life.

One key consideration when interpreting our results is the way cognitive functioning was assessed. The UK Biobank cognitive battery has been designed to be “a broad but shallow” tool to quickly assess a range of cognitive abilities. Thus, we have opted to use GCAS as a measure better representing cognitive heterogeneity (shared variance between various administered tests) in the studied population compared to individual tests. This is a suggested approach when considering shared genetic/environmental/sociodemographic influences [36,41] with a similar variance percentages (~40%) reported in each. Nevertheless, the use of domain specific tests should be considered in future research.

Finally, the scope and size of UK Biobank data resources are unprecedented, however similarly to other large scale epidemiological datasets, UK Biobank database includes large proportion of incomplete cases. Here, out of an initial sample of 48,561 participants, only 12,076 complete cases were found (i.e., including all brain imaging data, cognitive, genetic, demographic and education variables). The missing cases cause potential limitations due to loss of statistical power (e.g., in GWAS as discussed above) and by contributing to risk of bias affecting findings. To address that we have first examined the pattern and distribution of missing data and after determining that missing values were missing at random, we repeated all analyses using imputed data. Importantly, we again found the measures of fronto-parietal connectivity to be the strongest mediators of the effect of education on cognitive outcomes.

In summary, the data presented here not only add to our understanding of the neurobiological underpinnings of CR but also the role of between network connectivity and the default mode network in cognitive ageing. By employing comprehensive machine learning methods, we have demonstrated that between networks functional fronto-parietal connectivity is a strong mediator of the association between education used her as proxy of cognitive reserve and cognitive outcomes in ageing population. Our findings provide additional support for Robertson’s theory of CR [18,23] as well as the recent finding in Alzheimer’s patients linking between networks fronto-parietal connectivity to offsetting effects against amyloid burden [82]. To strengthen evidence supporting Robertson’s theory by validating causal pathway, future work should include longitudinal analyses with data from currently ongoing UK Biobank follow-up study. As socioeconomic status, ethnicity and multiple health conditions across the lifespan have been shown to have significant impact on cognitive ageing trajectories [101,103], it is crucial that any future work also considers these factors in a more inclusive and broader context of CR and its mechanistic underpinnings.

Supporting information

Acknowledgments

This research has been conducted using the UK Biobank Resource under Application Numbers 29447.