When your brain looks older than expected: combined lifestyle risk and BrainAGE

Average consumption of different alcoholic beverages (beer as 0.2 l, red and white wine as one glass of 0.2 l, and spirits as 0.02 l) within the last 4 weeks was assessed via a self-report questionnaire (Schmermund et al. 2002). The proportion of pure alcohol within the specific beverage was then multiplied with the frequency of drinking. Next, all beverages per person were summed up, resulting in the total amount of pure alcohol consumption in grams per month (g/month). Alcohol consumption as assessed via self-report questionnaires has been shown to highly correlate with blood indices of alcohol consumption (Glovannucci et al. 1991), multiple weekly self-report diet records (Glovannucci et al. 1991) and transdermal alcohol-use assessment (Simons et al. 2015), thus providing adequate reliability and validity for most research purposes (Del Boca and Darkes 2003).

The degree of lifetime exposure to tobacco smoking was assessed as pack-years (Duriez et al. 2014; Franklin et al. 2014; Karama et al. 2015), calculated by multiplying the years of smoking with the self-reported number of smoked cigarettes per day.

Social integration was assessed using an adapted version of the social integration index (Berkman 2004). The present social integration index comprised the domains “Marital status” (married or cohabitating participants were scored a 2; single, never married, widowed, or divorced participants were scored a 0), “close ties” (sum score derived from the number of children, close relatives, and friends) and “membership in organizations” (sum score of the number of organizations participants were members in and participated in at least once a month). Organizations included were: sport clubs, regional clubs, hunting clubs, choirs, theater clubs, music clubs, occupational or labor unions, political clubs or parties, congregations, and self-help groups. The scores of all three domains were summed up into the social integration score.

To measure physical activity, we used the metabolic equivalent of task (MET; Ainsworth et al. (2000) measuring the energy expenditure of a given activity compared to rest. The compendium of physical activities (Ainsworth et al. 2000) provides a mean energy expenditure value per hour of each activity. Participants were asked to report up to four different sportive activities, carried out within the last month. Based on the MET values assigned to the activities listed in the compendium, MET values were assigned to each of the activities reported by the participants and multiplied by the duration in hours (per month). Finally, a sum score of all activities was calculated. In addition, body mass index (BMI) was measured, since it has been shown to affect the association between physical activity and brain volume (Ho et al. 2011).

The combined lifestyle risk score was constructed as used in a previous study (Bittner et al. 2019). That is, to standardize the measurements of the four lifestyle variables, we first transformed the raw score on each individual lifestyle variable into a z-score. Next, to obtain a risk score that indicated higher risk with higher values, we reversed signs of the protective behaviors (social integration and physical activity). To obtain a risk score where a value of zero would indicate a mathematical balance of negative and protective behaviors, an additional linear transformation of the z-transformed lifestyle behaviors was applied: The protective variables were linearly transformed into negative scores by subtracting the maximum value from each individual measurement, whereas risk behaviors were analogously transformed into positive scores by adding the minimum value to each individual measurement. Hence, all values for risk behaviors were positive. Finally, the linearly transformed values of all individual lifestyle variables were summed up into one combined lifestyle risk score.

Since we aimed at investigating the residual variability in brain aging, we considered chronological age to be a covariate as it was associated to the BrainAGE score (Fig. 2b). However, it is not conclusive, whether chronological age should be included within statistical analyses on BrainAGE, since BrainAGE is based on chronological age (Smith et al. 2019). We, therefore, conducted three different approaches to address this issue: approach 1 included chronological age as a predictor in the regression models of lifestyle on BrainAGE, approach 2 did not include chronological age within the models and for approach 3, chronological age was regressed out of BrainAGE, before using BrainAGE as a dependent variable.

The brains of women and men show differences with respect to their structural architecture, e.g., in the proportion of gray matter and the thickness of the cortex (Ritchie et al. 2018), as well as aging trajectories (Cosgrove et al. 2007; Franke et al. 2014; Gur and Gur 2017; Ruigrok et al. 2014; Wierenga et al. 2018), functional network organization sub-serving cognitive abilities (Jiang et al. 2020), as well as the morphology of the so-called social brain (Kiesow et al. 2020). Hence, we considered sex as a covariate in the statistical models, and separately examined sex differences in performance measures and lifestyle analyses.

In addition, higher education is associated with higher cognitive performance (Elias et al. 1997) and is generally considered a proxy for brain reserve, the ability to better tolerate age-related neuronal loss (Bartrés-Faz and Arenaza-Urquijo 2011; Christensen et al. 2008; Stern 2012). Furthermore, there may be associations between education, intelligence, and lifestyle behavior (Cullen et al. 2012; Fratiglioni et al. 2004), where less smoking has been found in more educated individuals (McKenna et al. 2003). Hence, we hypothesized general educational level as a possible confounding factor and added education as a covariate into the statistical model. General education was measured using the international standard classification of education (UNESCO 1997), a standard classification system with 10 levels, where higher levels indicate higher education. Further, cognitive, physical and mental health of our older participants was estimated. Physical and mental well-being was estimated using the SF36 self-report questionnaire (SF-36 Fragebogen zum Gesundheitszustand (Franke et al. 1998) assessing quality of life in relation to health. As body mass index (BMI) has been shown to influence the association between physical activity and brain structure (Ho et al. 2011) and memory performance (Floel et al. 2008), it was considered as a covariate in the analyses of the combined lifestyle risk score, as well as of physical activity. Cognitive performance was examined using the DemTect test (Kalbe et al. 2004), a screening tool for dementia symptoms including verbal and working memory, word fluency and intellectual flexibility examinations. The DemTect test was designed to provide information on the general cognitive status. The present study used these estimates, i.e., physical and mental well-being, as well as cognitive status, to provide more information on the general fitness of the here presented older adult sample.

The BrainAGE framework is based on a relevance vector machine (RVM) (Tipping 2001) that transforms training data into a high-dimensional space (Bennett and Campbell 2000) and translates features learned from a training sample for a specific outcome variable (i.e., age) onto an unkown test sample (Fig. 1). To train and test the BrainAGE framework with respect to prediction accuracy and reliability, we used a ‘leave 10 out’-procedure within the full cohort of 1000BRAINS, which spans 1,229 MR datasets with an age range from 18.5 to 87.0 years (M = 60.7 years, SD = 13.4 years; M_male = 60.7 years, SD_male = 14.0 years; M_female = 60.9 years, SD_female = 12.5 years). Even though our study focuses on the association between lifestyle behavior and brain aging in older adults, we chose to train the RVM on the full cohort of 1000BRAINS for two reasons: (i) from a neuropsychological perspective, brain aging is a highly complex phenomenon, which can - as a pattern - better be learned by the algorithm when having access to a larger age range, which in the full cohort spans the entire adult life span (18.5–87 years), i.e., the algorithm has access to phenotypically richer data. However, using the whole age range has methodological advantages as well (ii): having as many and as morphologically rich training data as possible within the learning phase results in higher (and therefore, more constant) degrees of freedom, higher predictive power, and higher accuracy. This has been shown even though training and test samples differ demographically [(Cole et al. 2015; Franke et al. 2010; Varikuti et al. 2018); for a recent review see Cole et al. 2019]. In the training stage, the input data (i.e., n - 10 of the whole sex-split sample) were used to train the BrainAGE framework using chronological age and the GM tissue probability maps (Franke et al. 2010). The algorithm was trained seperately for female and male participants, considering that different features may be relevant for age prediction in females and males, as has been shown for intelligence prediction (Jiang et al. 2020) and for social characteristics (Kiesow et al. 2020). Blind to their true chronological age, the patterns learned by the relevance vector regression were then transferred to the unknown 10 test participants that were left out during the training stage, such that estimated age was based only on their anatomical patterns within the GM maps. This procedure was repeated until estimated age was provided for the full cohort of 1000BRAINS. For each participant, true chronological age was then subtracted from estimated age:

Positive BrainAGE scores reflect accelerated aging, i.e., the estimated age is higher than the chronological age. In contrast, negative BrainAGE scores reflect decelerated aging, i.e., the estimated age is lower than the chronological age. Finally, a correction for a quadratic age trend, which is identifiable in Fig. 1a, was applied to the resulting BrainAGE values using spm_detrend (SPM12, The Wellcome Dept. of Imaging Neuroscience, London; www.​fil.​ion.​ucl.​ac.​uk/​spm).

Following recent recommendations on age estimation models (Cole and Franke 2017) and the machine-learning literature (Bzdok and Ioannidis 2019; Yarkoni and Westfall 2017), we provide study sample-specific performance measures for the age estimation framework applied within this study. To test the accuracy of our model, we took the mathematical absolute value of each BrainAGE value and calculated the sample mean, which should then reflect the mean absolute error (MAE) of the brain age estimation. The smaller the MAE is, the better the performance of the estimation (Cole and Franke 2017). The idea behind is that each individual BrainAGE value can also be interpreted as the error of the age estimation model (i.e., the extent to which the estimation deviates from the true age). In addition, we examined Pearson correlations between estimated age, chronological age, and BrainAGE.

We calculated a between-subjects multivariate analysis of co-variance (MANCOVA) with sex as independent factor (male vs female), educational level as covariate and chronological age, estimated age, and BrainAGE as dependent variables. Sex differences in educational level were examined using a between-subject ANCOVA with independent factor sex (male vs female), covariate chronological age and education as dependent variable. In addition, we calculated Spearman correlations (method of choice for ordinally scaled variables, such as the ISCED) between education and BrainAGE. Further, we analyzed descriptive statistics of physical and mental well-being, as well as congitive status of the older subsample for sample description.

In our main analyses, we investigated the associations between lifestyle and BrainAGE using a twofold approach. First, we examined whether and how well BrainAGE would be predictable by our combined lifestyle risk score and second, we examined the prediction of BrainAGE by each individual lifestyle variable.

To investigate the linear effect of combining the four lifestyle variables into one score, we calculated multiple linear regressions (with IBM SPSS Statistics 26.0). The initial model (model 1) analyzing combined lifestyle risk used the combined lifestyle risk score and sex as explanatory variables. The second analysis introduced sex and the individual lifestyle behaviors (physical activity, social integration, alcohol consumption, and pack-years of cigarettes) as explanatory variables in model 1. In both analyses we used BrainAGE as dependent variable to be predicted. In a second model for each analysis, we applied a post hoc outlier exclusion (values > 3 SD) for each variable that showed a significant effect on BrainAGE. As additional covariates education and cognitive status were added within a third and a forth model. Please note, that each analysis step, i.e., model, was carried out for each of the approaches 1, 2 and 3. To assess whether the combined lifestyle risk score explaines more variance than the individual lifestyle variables, we compared the explained variance in R² of both approaches. Further we supplement our analyses with prediction metrics (mean absolute error, MAE) assessing how well the lifestyle models predict the individual expression of the BrainAGE score. MAE was calculated as the sum of the absolut value of the unstandardized residuals of each regression.

The datasets generated and/or analyzed during the current study will be made available from the corresponding author to other scientists on request in anonymized format and according to data protection policy in the ethics agreement.

In our first analysis of lifestyle, we investigated the association between combined lifestyle risk and BrainAGE. Table 4 shows descriptive statistics for all individual lifestyle variables as well as for the combined lifestyle risk score. Mean combined lifestyle risk was − 1.02 (SD = 2.1), reflecting a rather protective behavior within the selected older subsample of 1000BRAINS.

We now investigated how well combined lifestyle risk predicted BrainAGE in the multiple linear regressions implementing all three approaches to deal with chronological age [approach 1 including chronological age as predictor, 2 ignoring age and 3 examining the age-corrected residuals]. Please also see Table 5 for all regression statistics.

Combined lifestyle risk significantly predicted BrainAGE in all three age approaches using sex as covariate: Higher combined lifestyle risk was predictive of higher BrainAGE, with a regression coefficient of β = 0.182 to β = 0.183, all p < 0.001. These results remained stable, even after outlier exclusion (all p < 0.001, Fig. 4; outlier marked with diamond shapes, Table 5), introducing education (all p < 0.001), as well as cognitive status (DemTect) and BMI (all p < 0.001) as covariates. All of these associations were significant at p_adjusted ≈ 0.003. Within the fully adjusted model, the combined lifestyle risk score predicted 2.3–2.5% of the variance in BrainAGE, with a MAE of about 3.93 years in all approaches.

In the next step, we examined which individual lifestyle variables were predictive of BrainAGE by including all four individual lifestyle variables into a multiple linear regression as explanatory variables, while correcting for sex. This was done for all three approaches. Overall, the model significantly predicted BrainAGE, F(6,615) = 5.14–5.17, all p =  < 0.001, R² = 0.040–0.048. Neither alcohol consumption (β = 0.049–0.053, all p > 0.197, R² = 0.002), nor social integration (β = − 0.054–0.055, all p > 0.169, R² = 0.002) significantly predicted BrainAGE (Suppl. Table 3). But a higher amount of smoking predicted higher BrainAGE, with β = 0.136–0.140, all p =  < 0.001. Smoking was still predictive of BrainAGE after outlier correction (β = 0.170–0.175, all p < 0.0001; Fig. 4; 10 outlier marked with diamond shapes), as well as adding educational level (β = 0.169–0.175, all p < 0.0001), and cognitive status (β = 0.173–0.181, all p < 0.0001) as covariates. In this fully adjusted model, pack-years of smoking predicted 2.9–3.2% of the variance in BrainAGE with a MAE of 3.88–3.90 years (Table 6) and remained significant as well after outlier exclusion (Fig. 5, marked with diamond shapes, β = − 0.147–0.149, all p < 0.001). Higher physical activity remained significantly associated with lower BrainAGE after correction for all three additional variables, i.e., BMI, education and cognitive status in all three approaches (β = − 0.144–0.146, all p < 0.001) above p_adjusted ≈ 0.003. It was predictive of 2.0–2.1% of variance in BrainAGE with a MAE of 3.90–3.93.

Finally, we tested for quadratic effects of the combined lifestyle risk score and individual lifestyle variables on BrainAGE. Lifestyle variables that were used as regressors in the curve fitting process were already corrected for sex, education, and, in case of the combined lifestyle risk score and physical activity, BMI. Therefore, the R² values within this estimation differ from those in the purely linear regressions. In the following section, the subscript “q” indicates that the regressors belong to the quadratic function (Fig. 6).

With n = 622 participants, an alpha level of 0.003 and number of predictors = 5 our power to detect an expected small (Cohen 2013) effect size of 0.05 (Karama, Bug and Head) was 0.99 for the regression of BrainAGE onto combined lifestyle risk within the whole subsample. With an expected effect size of 0.04 for pack-years of smoking our power was 0.97. With an expected effect size of 0.07 for physical activity, our power was 0.99.

One of the compelling results of the current study were the negative effects of smoking on the aged brain quantified in months, which were also not corroborated after correcting for education or cognitive status. Prior studies already hinted at an association between smoking and changes in GM. Regionally lower GM volume and density for smokers compared to non-smoking control participants have been reported in the prefrontal cortex and the cerebellum (Brody et al. 2004), the posterior cingulum, precuneus, right thalamus, and bilateral frontal cortex (Almeida et al. 2008), as well as the substantia nigra (Gallinat et al. 2006). Importantly, all of these studies had a fairly small sample size that either included younger adults only [n = 45, age range 22.4–38.3 years, Gallinat et al. (2006)], older adults only [n = 78, age range 71.6–78.9 years; Almeida et al. (2008)], or a large age range [n = 36, 21–65 years, Brody et al. (2004)]. In addition, results on the association of smoking with other brain metrics are quite heterogenous. Higher numbers of white matter hyperintensities (Longstreth et al. 2005), lower microstructural integrity (Gons et al. 2011), or infarcts (Howard et al. 1998) in smokers compared to non-smokers were reported. With population-based cohort imaging available, the sample sizes have substantially increased (Bamberg et al. 2015; Caspers et al. 2014; Miller et al. 2016; Van Essen et al. 2013), thus increasing generalizability of results to the general population. For example, Karama et al. (2015) showed that smoking was associated with widespread cortical thinning in a sample of 504 older adults particularly in prefrontal cortex, mostly omitting primary sensory areas. Still, none of the studies provided a quantification of the effect of smoking on the brain. We were able to predict an increase of 0.36 months of BrainAGE with each pack-year, with 2.9% of BrainAGE being predicted just by smoking (approach 3). Translating this result to our older adult study sample taking into account the average smoking behavior of 13.61 pack-years, an overall increase of 4.9 years of BrainAGE (13.61 × 0.36 month of BrainAGE) only by smoking can be stated.

There was a high variance in BrainAGE in individuals, who never smoked (139 female, 114 male, Fig. 4). Within this group, BrainAGE scores were very high, as well as very low, aggregating to a mean BrainAGE of almost zero. This finding is comparable to the considerable variance in cortical thickness in the rarely smoking participants observed by Karama et al. (2015). In the current study, the more the participants smoked, the stronger was the relationship between higher pack-years and higher BrainAGE, suggesting that this effect was mostly driven by high lifetime smoking (Fig. 4). This was also revealed when comparing never, moderate and severe smokers, where the brains of severe smokers (BrainAGE = 1.61) appeared significantly older than those of never (BrainAGE = − 0.05) and moderate (BrainAGE = − 0.03) smokers. It is particularly important to note that this observation cannot be translated simply to the assumption that rare smoking has no effect on the brain. Rather, rare smoking may manifest in other metrics for healthy brain aging, even if alterations in brain structure would not be present: For example, in our previous study, we found no association between smoking and cortical surface measures in older adults. Instead, more smoking was associated with higher resting-state functional connectivity (RSFC), which may be a compensation mechanism for accelerated brain aging (Bittner et al. 2019). In addition, activity differences in task-based fMRI (Lawrence et al. 2002; Tanabe et al. 2011), as well as receptor differences between smokers and non-smokers (Feduccia et al. 2012; Mukhin et al. 2008) were described. Studies on RSFC in relation to long-term effects of smoking and not acute effects of nicotine are rather rare. It may thus be of particular interest to investigate general differences in brain function, e.g., in RSFC, associated with light smoking (Janes et al. 2012; Pariyadath et al. 2014; Zhou et al. 2017), even though light smoking seems not to be heavily associated to differences in brain structure.

The underlying mechanisms driving the association between smoking and changes in brain structure are still unclear. Smoking could potentially act via atherosclerotic processes, which may impact the aging brain and thus accelerate brain aging (Freund et al. 1993; Mucha et al. 2006; Prescott et al. 1998; Pujades-Rodriguez et al. 2015). Possibly, the measured increase in BrainAGE might also be attributable to the direct toxic effects of tobacco smoke onto the cerebro-vascular system, which includes oxidative stress within the cells and results in apoptosis (Swan and Lessov-Schlaggar 2007). However, since BrainAGE takes the whole GM volume of an individual into account, drawing inferences about any molecular mechanisms or disentangle regional differences that drive the association between stronger smoking and accelerated brain aging as reflected by higher BrainAGE remains for future studies. Still, our results provide evidence, that smoking may be one of the detrimental factors contributing to an unfavorable soil, which then together with additional risk factors, promotes alterations within the brain or accelerated already existing processes of structural decline.

The protective effect of physical activity on GM volume has been discussed to be regionally specific [see review by Erickson et al. (2014)]. Our results support that higher physical activity is predictive of lower BrainAGE, thus younger looking brains. Physical activity, therefore, does not only seem to affect specific brain regions, but also the multidimensional pattern of brain aging itself. Most previous studies comprised intervention trainings, where training was systematic, regular, and highly controlled (Erickson et al. 2014). The present study adds to this by demonstrating that higher physical activity is associated with decelerated brain aging (lower BrainAGE scores) in a large sample of “normal” aging older adults using a comprehensive, epidemiologically motivated measurement of physical activity, i.e., the metabolic equivalent (Ainsworth et al. 2000; Bus et al. 2011; Floel et al. 2010; Milanovic et al. 2013; Pierce et al. 2007; Ruscheweyh et al. 2011; Wagner et al. 2012). This measurement is drawn from self-reports that summarize all sorts of sports older adults engage in and is likely to reflect the average daily physical activity, in contrast to highly controlled intervention settings. The present association between self-reported physical activity and BrainAGE is, therefore, not as strong as reported effects of fitness training on, e.g., the hippocampus (Erickson et al. 2011), but is likely reflecting a natural, and therefore, more generalizable relationship.

Several mechanisms how higher physical activity or fitness levels may act protectively on the aging brain have been discussed, such as the upregulation of neurotrophic factors, including brain-derived-neurotrophic factor [BDNF, (de Melo Coelho et al. 2013; Neeper et al. 1996; Piepmeier and Etnier 2015) Ruscheweyh et al. 2011] and granulocyte-colony stimulating factor (G-CSF; Flöel et al. 2010), which significantly impact synaptic efficacy, neuronal connectivity, and use-dependent plasticity. Here, use-dependent plasticity may play a crucial role in the sense that physical activity as one kind of training would lead to better preservation of those brain structures needed to perform the activity engaged in (Bittner et al. 2019; Colcombe et al. 2003; Vaynman and Gomez-Pinilla 2005; Vaynman et al. 2004). Several studies have shown that training-induced preservation or even adaptation of brain regions is possible in adults (Churchill et al. 2002; Draganski et al. 2004; Erickson et al. 2011; Kramer and Erickson 2007) or older adults in particular (Boyke et al. 2008). Nevertheless, other studies have discussed, that the most important factor for preservation of brain structure is effortful learning (Shors 2014). During physical activity, both processes are possible, but could not be controlled in the epidemiological setting of the present study. However, physical activity is widely known for its benefitting effects on the general health of the human organism (World Health Organization 2019). Possibly enhancing the resilience of the organism, as well as higher neuronal integrity (Engeroff et al. 2019), it may, therefore, contribute to a more protective soil, where pathological processes may have lesser effect (McDonough and Allen 2019). Future studies examining specific mechanisms are needed, though.

We did not find an association between BrainAGE and social integration or alcohol consumption. Regional differences in brain structure associated to social integration, as well as alcohol consumption could be present, but might not have been identifiable with the specific approach of the current study. Several explanations might hold for these observations. To date, the number of studies investigating social integration in relation to structural brain decline in older adults is relatively small (for a recent review, see Anaturk et al. 2018). In addition, most studies report effects for composite measurements of cognitive and social activities, which do not clearly differentiate between social and cognitive components (Gow et al. 2012; Hafsteinsdottir et al. 2012; Vaughan et al. 2014). Further, composite measures for social activities when investigated in combination with additional lifestyle behaviors (Bittner et al. 2019) have been assessed. Therefore, future studies could shed light on effects of social activities with low cognitive versus high cognitive demands to answer the question whether the cognitive or the social component of social integration contributes to brain reserve, the amount to which age-related GM loss can be tolerated without showing deficiencies (Stern, 2012). In addition, we used a quantitative measurement of social integration. Studies have shown that older adults engage in relationships with a focus on quality rather than quantity (Carstensen et al. 1999). Hence, future studies would be needed to address the association between quality of relationships and brain aging. Further, differences in brain structure related to social integration may be regional or subtle (James et al. 2012), which also seems to be the case for alcohol consumption (Topiwala et al. 2017). Even though accelerated brain aging has been shown in patients with alcoholism (Pfefferbaum et al. 1995), differences related to alcohol consumption in the normal population may not be as strong or only identifiable if several risk behaviors co-occur (Bittner et al. 2019). Further, alcohol consumption may affect other brain parameters earlier such as WM lesions (den Heijer et al. 2004) or RSFC (Vergara et al. 2017). In addition, effects of alcohol consumption may also be non-linear (Mukamal et al. 2001), which we could not identify in the present study, but might be interesting for future studies to further investigate into.