Lifestyle and incident dementia: A COSMIC individual participant data meta‐analysis

The LIfestyle for BRAin Health (LIBRA) index yields a dementia risk score based on modifiable lifestyle factors and is validated in Western samples. We investigated whether the association between LIBRA scores and incident dementia is moderated by geographical location or sociodemographic characteristics.


Highlights
• A two-step individual participant data meta-analysis was conducted.
• This was done at a global scale using data from 21 ethno-regionally diverse cohorts.
• The association between a modifiable dementia risk score and dementia was examined.
• The association was modified by geographical region and age at baseline.
• Yet, modifiable dementia risk and protective factors appear relevant in all investigated groups and regions.

BACKGROUND
Due to the aging of the population and lack of curative treatments, dementia prevalence is expected to increase from 55 million cases worldwide in 2019 to 139 million cases by 2050, further heightening the already large burden of disease associated with this group of conditions. 1 However, evidence supporting the potential for dementia risk reduction by targeting modifiable risk and protective factors has been steadily accumulating. 2,3About 40% of all dementia cases worldwide have been estimated to be potentially attributable to 12 modifiable risk and protective factors. 4Many of these factors also overlap with the World Health Organization's recommendations to reduce dementia risk. 5wever, it is currently unclear whether dementia risk should be further personalized, or at least stratified, based on sociodemographic or other characteristics.Exploring risk stratification may lead to more accurate and inclusive risk estimates. 6For instance, the risk conferred by certain modifiable risk factors has been demonstrated to differ depending on the age of exposure.This has been the case for obesity, hypertension, and dyslipidemia which have been shown to be more strongly associated with dementia when their exposure occurs in midlife rather than in late life. 2,7,8However, little is known about whether other sociodemographic variables, such as sex, race and ethnicity, educational level, and socioeconomic position (SEP), or geographical location modify the association between modifiable risk and protective factors and incident dementia.Yet, this information may indicate where most preventive potential lies, which is important information for policy makers and individuals themselves.
3][14][15] There are reports of continent-level differences in the association between alcohol consumption and dementia risk 13 and a stronger association between diabetes and cognitive decline in samples from Asian countries compared to samples from Europe, North America, and Australia with predominantly White participants. 16Yet, in general the research on potential (ethno-) regional differences in associations between modifiable risk factors and future cognitive status is scarce.Therefore, risk personalization for all known modifiable risk and protective factors needs to be further investigated.In the case of dementia, composite modifiable risk scores such as the LIfestyle for BRAin Health (LIBRA) index 8,17 or the Australian National University Alzheimer's Disease Risk Index (ANU-ADRI 18 ) are also relevant to consider due to the syndrome's multifactor etiology. 19,20These composite risk scores combine the presence or absence of multiple factors into one numeric value that expresses dementia risk.They are especially valuable as they can aid in the identification of individuals at high risk and may facilitate the implementation of dementia risk reduction guidelines into practice. 19e LIBRA index includes only modifiable risk (ie, physical inactivity, current smoking, obesity, hypertension, dyslipidemia, diabetes, depression, coronary heart disease [CHD], and chronic kidney disease [CKD]) and protective (ie, low-to-moderate alcohol consumption, healthy diet, high cognitive activity) factors. 8,17The score has been well validated for predicting incident dementia in cohorts located in Western, high-income countries.

Contributing studies
In this individual-participant data meta-analysis, data from 21 cohorts, drawn from 17 countries across six continents, were used.All cohorts are part of the Cohort Studies of Memory in an International Consortium (COSMIC) collaboration. 21Details of the individual studies can be found in Table 1.
Exclusion criteria were the following: prevalent dementia, missing information on dementia status or length of follow-up, no available follow-up assessment, or insufficient data on modifiable risk and protective factors (ie, fewer than seven out of the 12 LIBRA factors, Figure 1).

Sociodemographic characteristics and geographical location
Age and sex data were available in all cohorts.Educational level was operationalized as total years of formal education.If unavailable, categories of educational attainment were converted to total years of formal education after consultation with local study coordinators.For a

Research in context
1. Systematic review: The authors reviewed the current state-of-evidence regarding modifiable risk factors for dementia using traditional sources (eg, PubMed).Risk personalization or stratification beyond age has rarely been explored whereas doing so could identify groups with the most preventive potential and is necessary for inclusive and accurate risk estimation.
2. Interpretation: Our results suggest that modifiable factors, compiled in a modifiable dementia risk score, are associated with dementia risk regardless of sex, age, educational level, socioeconomic position, or broad geographical region.However, this association was even stronger in younger individuals (≤75 years) and in Asian cohorts compared to European cohorts.non-Hispanic White, Hispanic White, non-Hispanic Black, and Hispanic Black).The only cohort in South America (Bambui) was from Brazil, a country known to have a very complex and hard to define racial and ethnic makeup. 432 modifiable risk and protective factors for dementia. 8,17Each risk and protective factor has an assigned weight based on the factor's relative risk for dementia from published meta-analyses. 8,17Based on the presence or absence of the factor, these weights (see Material S2) are summed to yield the total LIBRA score, ranging from −5.9 to +12.7.Higher scores indicate a less favorable combination ofrisk factors and a higher risk for dementia.[46][47][48] Here, the presence or absence of each factor was determined for every individual, based on the harmonization protocols described in Material S2.If possible, missing information at baseline was augmented with information at follow-up assessments.At least seven of the 12 factors had to be available for an individual to be included in the analyses (Figure 1).When not all factors were known, the LIBRA score was calculated with the available factors and rescaled to the full theoretical range (ie, when all 12 factors would be present) using the standard min-max normalization formula given as follows in Equation ( 1),

Modifiable risk and protective factors: LIBRA score
where a and b are the lower and upper limits, respectively, of the LIBRA score with the available factors.

Dementia incidence
Determination of incident dementia varied across studies.Twelve The harmonization protocol can be found in Material S2.

Statistical analyses
A two-step individual participant data meta-analysis approach was used.Potential associations between sociodemographic characteristics and LIBRA score at baseline were examined by calculating the cohort-specific differences in mean LIBRA score between the sociodemographic groups.Next, these mean differences were pooled using random-effects meta-analysis resulting in pooled mean differences and their 95% confidence intervals (CIs).
The LIBRA score was also compared between geographical locations.Here, mean LIBRA scores were pooled per continent and compared using a Cochran's Q test.The I 2 statistic was used to assess heterogeneity between studies.Meta-regression was used to confirm the results from the Cochran's Q test and estimate the mean difference in LIBRA score between continents.
Cox proportional hazard regression analysis was used to assess the association between the LIBRA score and dementia incidence in each cohort separately, resulting in hazard ratios (HRs) and their 95% CIs.
This was done for the continuous LIBRA score as well as for cohortspecific LIBRA score tertiles.In all analyses, age was used as the time and a late follow-up (>5 years after baseline).This allows us to distinguish between incident cases that were likely already in the preclinical stage at baseline, and cases that developed later.Each Cox regression analysis had to have at least five events (ie, incident dementia cases) per variable in the model (ie, LIBRA score + control variables). 53The proportional hazard assumption was assessed by testing of Schoenfeld residuals. 54One cohort (Gothenburg H70 Birth cohort study [the H70 study]) showed non-proportional hazards.As there is no straightforward remedy for this, without compromising the ability to compare and pool the results of analyses over all cohorts, this cohort was still included in the analyses.However, an additional sensitivity analysis without the H70 study was carried out.To examine the moderating effect of the sociodemographic variables described earlier, Cox regression analyses, stratified by these variables, were executed.Baseline age was stratified into individuals ≤75 versus older, as LIBRA was initially designed for capturing modifiable dementia risk in people aged 40 to 75 years and has been shown to perform less well after the age of 75. 7,44All cohort-specific HRs were pooled using random effects metaanalysis.Cochran's Q statistic for subgroup differences was used to test the potential moderating effect of the selected sociodemographic variables and geographical location on the association between LIBRA and dementia incidence with an alpha level of 0.10. 55The I 2 statistic was used to assess heterogeneity between studies.Additionally,

RESULTS
In total, 31,680 eligible individuals from 21 cohorts across six continents were included (mean age range at baseline: 52 to 85 [SD range across cohorts: 0 to 16], 58% female, Figure 1).A total of 2330 incident dementia cases were recorded during an overall median of 5.0 years of follow-up (range: 1.0 to 2.4 years).Of these, 1671

LIBRA index across sociodemographic strata and geographical location
Potential baseline differences in mean LIBRA scores between sociodemographic strata and geographical location were assessed.Mean (95% CI) LIBRA scores were 0.5 (0.27 to 0.72) points higher (worse) in older individuals (≤75 vs >75 years old).LIBRA scores were also higher in individuals with fewer years of formal education or a lower SEP.For education, compared to 6 to 11 years, those with <6 years were 0.52 points higher and those with ≥12 years were 0.58 points lower (all p < 0.05).For SEP, compared to those with an intermediate SEP, those with a lower SEP were 0.39 points higher and those with a high SEP were 0.39 points lower (all p < 0.05).Mean LIBRA scores did not differ significantly between males and females.Forest plots of individual study data can be found in Material S4.LIBRA scores were also compared based on the geographical location of the cohort (Figure 2).This showed large heterogeneity both within and between continental regions.LIBRA scores were on average 2.1 (95% CI = −3.8 to −0.3) points lower in Asian cohorts compared to European cohorts (p = 0.020).

LIBRA and dementia incidence
On a continuous scale, a one-point increase in the LIBRA score was significantly associated with a higher risk for dementia (HR = 1.06, 95% CI = 1.04 to 1.08) in main Model 2 (Figure 3 and Table 3).Per one SD increase in LIBRA (over all participants in all cohorts combined: SD = 3.26), this translates to HR = 1.21 (95% CI = 1.14 to 1.29).Heterogeneity was limited (I 2 = 31.1%).The hazard for dementia was 33% F I G U R E 2 Mean LIBRA score by cohort and continent.Expansions of cohort name abbreviations are presented in Table 1.CI, confidence interval; LIBRA, LIfestyle for BRAin Health.

F I G U R E 3
Hazard ratio for incident dementia per one-point increase in LIBRA score.Model 2 (main model), controlled for age (time scale), sex, and years of formal education.Expansions of cohort name abbreviations are presented in Table 1.CI, confidence interval; HR, hazard ratio; LIBRA, LIfestyle for BRAin Health.
higher in the highest LIBRA tertile compared to the lowest (Model 2, Table 3).
When examining early and late follow-up, the LIBRA index was associated with dementia incidence, both in individuals who developed dementia within 5 years (HR = 1.06, 95% CI = 1.03 to 1.09, Model 2) and in individuals who developed dementia after more than 5 years (also HR = 1.06, 95% CI = 1.03 to 1.09, Model 2).

Potential moderating effect of sociodemographic characteristics and geographical location on the association between LIBRA and incident dementia
Potential interactions between LIBRA, sociodemographic characteristics and geographical location were explored using meta-analysis and Cochran's Q statistic for subgroup differences (Figure 4 and Material S5), as well as meta-regression (Material S6).
The association between LIBRA and dementia incidence did not differ between males and females, regardless of timing during the follow-up.Further, years of formal education and SEP did not interact with LIBRA.When comparing the association between LIBRA and incident dementia among younger (≤75 years old) and older (>75 years old) individuals across the entire follow-up time, evidence for an interaction was observed.With higher LIBRA scores, the risk increase was larger in younger individuals (1.08, 95% CI = 1.05 to 1.11) than in older individuals (1.04, 95% CI = 1.02 to 1.07, p interaction = 0.035).
Meta-regression confirmed that median age at baseline was indeed a significant moderator.However, this interaction was only apparent in the early follow-up period (Cochran's Q test, Model 2 younger: 1.12, 95% CI = 1.09 to 1.15, vs older: 1.03, 95% CI = 1.01 to 1.06, p interaction < 0.001).When considering late follow-up cases  (occurring after at least 5 years), baseline age did not alter the association between LIBRA and dementia incidence (Cochran's Q test, Model 2 younger: 1.04, 95% CI = 1.00 to 1.08, vs older: 1.06, 95% CI = 1.02 to 1.10, p interaction = 0.545).Meta-regression analysis confirmed these results: median baseline age was a significant moderator in the early follow-up period but not in the late follow-up period.Lastly, across the entire follow-up period, the association between LIBRA and incident dementia was found to be stronger in Asian cohorts (1.10, 95% CI = 1.07 to 1.14) compared to European cohorts (1.04, 95% CI = 1.02 to 1.07, p interaction = 0.069).However, when examining this difference for early and late follow-up separately, there was no significant effect modification by geographical region (early follow-up Asian cohorts: 1.09, 95% CI = 1.06 to 1.13), vs early follow-up European cohorts: 1.05, 95% CI = 1.00 to 1.10, p interaction = 0.599; late follow-up Asian cohorts: 1.14, 95% CI = 0.98 to 1.31, vs late follow-up European cohorts: 1.06, 95% CI = 1.01 to 1.10, p interaction = 0.610).These findings were confirmed in the meta-regression analysis.Full results on the stratified meta-analysis can be found in Material S5 and detailed results of the meta-regression can be found in Material S6.A sensitivity analysis omitting the H70 study cohort because of non-proportional hazards was also run.The results were similar and can be seen in Material S7.

DISCUSSION
In this individual participant data meta-analysis, modifiable dementia risk profiles from diverse ethno-regional groups were examined in 21 prospective cohort studies from 17 countries.An unfavorable modifiable risk profile was associated with an increased risk for dementia, and this association was stronger with younger baseline ages (≤75 years), specifically for dementia cases occurring in the first 5 years of follow-up.Across the entire follow-up time, the association between LIBRA index and incident dementia appeared stronger in Asian compared to European cohorts.However, this interaction disappeared when considering early and late follow-up cases separately.Importantly, no interactions between modifiable risk profiles and other sociodemographic variables (ie, sex, years of formal education, SEP) were observed.
First, this study examined the association between modifiable risk factors and dementia incidence at a global scale.So far, the link between modifiable risk factors and dementia risk has been predominantly based on studies from high income Western countries, while estimates from low-and middle-income countries (LMICs) were lack- ing.Yet, the largest increase in dementia prevalence is projected for LMICs. 1,57The current study included cohorts from six LMICs (Bambui from Brazil, CLAS, and SAS from China, LRGS TUA from Malaysia, EPI-DEMCA from the Republic of Congo, and ISA from Nigeria; see Table 1 for expansions of cohort name abbreviations).Our findings suggest that modifiable risk factors are relevant for dementia risk reduction in different parts of the world and might have important implications for global efforts to reduce dementia risk. 1,58ross the entire follow-up period, the association between the LIBRA score and incident dementia was stronger in Asian cohorts compared to European cohorts.6][57] However, this stronger association between LIBRA and dementia risk within the Asian cohorts was not observed when considering early and late follow-up cases separately.This may be because of insufficient power to detect in an interaction, especially within late follow-up.In early follow-up, however, sample size of Asian cohorts was still large.Upon further investigation, a suppressor effect between follow-up time and geographical region was noticed, in which controlling for follow-up resulted in a larger estimated effect of geographical region, as well as a significant effect of follow-up time (with longer follow-up time being associated with a larger HR).Without controlling for region, no significant effect of follow-up time was noticed as Asian cohorts tended to have a shorter follow-up time compared to European and American cohorts.We therefore contemplate that the unobserved interaction between LIBRA and geographical region in early follow-up is mostly due to the fact that region and follow-up time suppress each other when they are not both included in the model.When splitting survival time, the effect of follow-up time is modelled less well.Other work looking at ethnic and/or regional differences in susceptibility is limited.One 10/66 Dementia Research Group population-based study compared multiple dementia risk prediction models (including modifiable risk factors) for their predictive validity in LMICs (China and several Latin American countries). 59All assessed prediction models were developed based on Western, and predominantly White, populations.They found that ANU-ADRI and the Brief Dementia Screening Indicator (BDSI 60 ) replicated well in the studied LMICs, but the Cardiovascular Risk Factors, Aging and Dementia (CAIDE 61 ) risk score did not.It remains unclear whether there are ethno-regional differences in susceptibility to modifiable risk factors for dementia.More research is needed.
No interactions were observed between the LIBRA score and sex, years of formal education, or SEP.A healthy lifestyle or favorable risk profile thus appears to be associated with a lower risk for dementia to a similar degree in all assessed groups.It is important to note that individuals with more years of formal education or a higher SEP tended to have a more favorable risk factor profile.[15] Age interacted with the LIBRA score, which is in line with earlier findings, demonstrating that LIBRA is (more) predictive of incident dementia in individuals in mid-to-early-late life compared to individuals in late-late life.For example, the Cambridge City over-75s cohort study found no association between the LIBRA score-nor its individual risk factors-and dementia risk in individuals over 85 years old, 7 whereas an association between LIBRA and incident dementia is consistently reported in younger populations. 10,17,46Similar results were observed in another study using data from the European populationbased DESCRIPA study. 44From a life course perspective, one may indeed expect more benefit from adhering to a healthy lifestyle at a younger age, before the pathological process of decades of exposure to risk factors has caused extensive damage. 62Interestingly, this interaction was specifically observed for early dementia cases that occurred within the first 5 years after baseline.In other words, when a dementia diagnosis occurs within 5 years after the risk profile assessment (when the likelihood for reverse causality is much larger), the risk profile is more strongly associated with dementia risk in the younger (≤75 years) individuals.Potentially, the older the age at baseline, the more time might be needed for effects of the risk factor profile on dementia risk to become noticeable, as exposure to risk factors before that time has played a larger role.However, this result may also be related to a larger proportion of the sample being Asian in early follow-up compared to late follow-up.Asian cohorts tended to be younger and were associated with a larger HR compared to European cohorts.Overall, this demonstrates again that it is important to consider the timing of these diagnoses.
The current results hint at universal prevention initiatives.However, this study only considered a small part of the many steps there are between the establishment of a modifiable risk factor (or clustering of multiple risk and protective factors) and the actual evidence-based implementation of effective strategies aimed at altering the exposure to them (for example, this is also dependent on risk factor prevalence or cost-effectiveness).

Strengths and limitations
Strengths of this study include its overall sample size and sufficiently high number of incident dementia cases for finding associations.Due to the relatively large amount of individual participant data, we were able to use more refined and elaborate analysis methods.Nonetheless, it is important to consider that during the process of data harmonization, information can get lost which may result in bias.The global context of this study is uncommon and leads to good external validity.However, data from South America, Africa, Oceania, and certain Asian regions (especially the Middle East and Central Asia) were still limited.Therefore, our results on potential effect modification by geographical region need to be replicated.Furthermore, potential differences based on geographical location, as well as race and ethnicity, should ideally be examined.Due to the nature of the data here, with a very high level of overlap between these two variables, doing so was not reasonable.
Instead, we had to focus on one variable and geographical location was chosen because it was readily available for all cohorts and meant a larger sample size.Race and ethnicity were not adequately assessed in several of the cohorts that we aimed to include.Further, we were not able to control for apolipoprotein E (APOE) ε4 allele carriership and information on treatments (eg, for managing dyslipidemia or hypertension) was not always available.Another limitation lies in the fact that we looked at the presence of risk factors at just one timepoint during the lives of these individuals.Also, the studied outcome was incident dementia diagnosis, which is not equivalent to the onset of underlying disease, and may also depend on the healthcare system, testing procedures, and so on.Additionally, for some cohorts many individuals were excluded as they did not meet the preset inclusion criteria.They often differed from the included individuals (see Material S3).In general, younger, more educated, and healthier individuals were included.
Included individuals were more likely to have dyslipidemia compared to excluded individuals.Loss to follow-up or dropout due to worse health (ie, multiple comorbidities) is a recurrent observation in longitudinal studies but cannot be remediated easily.Indeed, most people were excluded because they only had cross-sectional data.This could have resulted in an underestimation of observed associations.
Taken together, these results suggest that a "brain-healthy" lifestyle is associated with a lower risk for dementia years later, regardless of geographical location or sociodemographic characteristics.However, exposure to these modifiable risk and protective factors varied greatly across the ethno-regionally diverse cohorts.Further research towards risk stratification and personalization is needed and should focus on a life course approach, examining the association between a brainhealthy lifestyle and dementia risk at different timepoints during the entire life.

3 .
Future directions: These findings need to be replicated.More data and research are needed particularly regarding geographical region.Cohort studies conducted in multiple regions would be ideal.Future work should also examine potential interactions between individual modifiable risk factors (rather than an index) and sociodemographic or environmental characteristics.stratified analysis, years of formal education were categorized into low (<6 years of formal education), intermediate (6 to 11 years of formal education) and high (≥12 years of formal education).SEP was categorized as low, middle, or high within each cohort, based on several alternative measures, predominantly income and occupation, to maximize harmonization potential.See Material S2 for the sociodemographic characteristics' harmonization protocols.Geographical location was defined as the continent where the cohort was sampled.Within nearly every cohort the sample was predominantly homogeneous in terms of broad racial and ethnic group.Race and ethnicity, and geographical location, overlapped almost entirely: non-Hispanic White in Europe and Australia, Asian in Asia, and non-Hispanic Black in Africa.The cohorts in North America included one predominantly non-Hispanic White (MYHAT), one fully Hispanic (SALSA, no further specification on White or Black but mostly of Mexican origin), and one mixed (EAS: studies used Diagnostic and Statistical Manual of Mental Disorders (DSM) III or IV criteria.49,50A few cohorts employed other assessment methods, such as the Clinical Dementia Rating (CDR51 ), the Mini-Mental State Examination (MMSE52 ), or a self-reported F I G U R E 1 Flow chart of sample selection.LIBRA, LIfestyle for BRAin Health (index).diagnosis.It was not possible to distinguish between different etiologies, as brain imaging, biomarkers, or autopsy data were not available.
scale and dementia was treated as the failure event.Survival time was defined as the age at study entry until the age at the date of dementia diagnosis (if reported, otherwise calculated as the mid-point between waves) or study exit (ie, date of last interview or date of death, whichever came first).Three different models were run, controlling for the following variables: Model 1 = crude model (age only); Model 2 = Model 1 + sex and years of formal education (main model); and Model 3 = Model 2 + SEP.Model 3 was defined as a separate model because data on SEP were not available in a significant portion of the sample.Also because of this, Model 2 is regarded as the main model.As dementia's pathology develops over many years before diagnosis, additional analyses were carried out in which survival time within each cohort was split up into an early follow-up (≤5 years after baseline) meta-regression was used to examine potential sources of heterogeneity (including proportion of females, median age, median follow-up time, number of available LIBRA factors in a cohort, continent, and gross-domestic product per capita of the country where the cohort was based 56 ).All tests were carried out two-sided with an alpha level of 0.05 unless indicated otherwise.All analyses were conducted with Stata 17.0 (StataCorp, College Station, Texas, United States).

F I G U R E 4
). Model 2 (main model): Model 1 + sex and years of formal education.Model 3: Model 2 + socioeconomic position.Abbreviations: CI, confidence interval; HR, hazard ratio; LIBRA, LIfestyle for BRAin Health.a per one-point increase.Hazard ratio for incident dementia per one-point increase in LIBRA score per sociodemographic group.Model 2 (main model), controlled for age (time scale), sex, and years of formal education.CI, confidence interval; HR, hazard ratio; LIBRA, LIfestyle for BRAin Health; SEP, socioeconomic position.

AFFILIATIONS 1 Alzheimer
Centrum Limburg, Department of Psychiatry and Neuropsychology, Mental Health and Neuroscience (MHeNs) Research Institute, Maastricht University, Maastricht, The Netherlands

TA B L E 1 Contributing studies, in alphabetical order. Study Abbreviation Location (continent) Year started n included (%) of total sample
Cohort-specific population characteristics.
TA B L E 2Note: LIBRA theoretical range: −5.9 to +12.7, missing data: CLAS: sex: n = 2, years of formal education: n = 6.Expansions of cohort name abbreviations are presented in Table1.Abbreviations: IQR, interquartile range; LIBRA, LIfestyle for BRAin Health.dementiacaseswere diagnosed ≤5 years after baseline, and 659 cases were diagnosed later (>5 years after baseline).Detailed, cohortspecific population characteristics can be found in Table2.The cohortspecific distribution of individual LIBRA factors can be consulted in

TA B L E 3
Association between LIBRA score and dementia incidence.