Physical inactivity, cardiometabolic disease, and risk of dementia: an individual-participant meta-analysis

Abstract Objective To examine whether physical inactivity is a risk factor for dementia, with attention to the role of cardiometabolic disease in this association and reverse causation bias that arises from changes in physical activity in the preclinical (prodromal) phase of dementia. Design Meta-analysis of 19 prospective observational cohort studies. Data sources The Individual-Participant-Data Meta-analysis in Working Populations Consortium, the Inter-University Consortium for Political and Social Research, and the UK Data Service, including a total of 19 of a potential 9741 studies. Review method The search strategy was designed to retrieve individual-participant data from prospective cohort studies. Exposure was physical inactivity; primary outcomes were incident all-cause dementia and Alzheimer’s disease; and the secondary outcome was incident cardiometabolic disease (that is, diabetes, coronary heart disease, and stroke). Summary estimates were obtained using random effects meta-analysis. Results Study population included 404 840 people (mean age 45.5 years, 57.7% women) who were initially free of dementia, had a measurement of physical inactivity at study entry, and were linked to electronic health records. In 6.0 million person-years at risk, we recorded 2044 incident cases of all-cause dementia. In studies with data on dementia subtype, the number of incident cases of Alzheimer’s disease was 1602 in 5.2 million person-years. When measured <10 years before dementia diagnosis (that is, the preclinical stage of dementia), physical inactivity was associated with increased incidence of all-cause dementia (hazard ratio 1.40, 95% confidence interval 1.23 to 1.71) and Alzheimer’s disease (1.36, 1.12 to 1.65). When reverse causation was minimised by assessing physical activity ≥10 years before dementia onset, no difference in dementia risk between physically active and inactive participants was observed (hazard ratios 1.01 (0.89 to 1.14) and 0.96 (0.85 to 1.08) for the two outcomes). Physical inactivity was consistently associated with increased risk of incident diabetes (hazard ratio 1.42, 1.25 to 1.61), coronary heart disease (1.24, 1.13 to 1.36), and stroke (1.16, 1.05 to 1.27). Among people in whom cardiometabolic disease preceded dementia, physical inactivity was non-significantly associated with dementia (hazard ratio for physical activity assessed >10 before dementia onset 1.30, 0.79 to 2.14). Conclusions In analyses that addressed bias due to reverse causation, physical inactivity was not associated with all-cause dementia or Alzheimer’s disease, although an indication of excess dementia risk was observed in a subgroup of physically inactive individuals who developed cardiometabolic disease.

Physical inactivity was defined as "No sport activities". For stroke ascertainment, only self-reports from annual follow-up surveys and mortality records were available. Dementia was defined using data from annual follow-up surveys requesting reported doctor-diagnosed demensis Alzheimer.

Health and Lifestyle Survey (HALS)
UK HALS is a nationwide sample survey of community dwelling adults in England, Scotland, and Wales. In 1984/1985, a total of 12,254 addresses were randomly chosen from Electoral Registers and one adult aged 18 years or over was selected from each household. A total of 9003 adults participated in the baseline examination. Ethical approval for the main HALS surveys was received from the BMA Ethical Committee before the launch of survey. Physical activity was measured with a list of 17 different sports (e.g., cycling, swimming, football) and 4 open-ended sports the participant could select freely. Physical inactivity was defined as not participating in any sport activities.
Participants were linked to mortality registers and deaths from dementia were defined using ICD

Health and Social Support (HeSSup), Finland
The Health and Social Support (HeSSup) study is a prospective cohort study of a stratified random sample of the Finnish population in the following four age groups: 20-24, 30-34, 40-44, and 50-54. The participants were identified from the Finnish population register and posted an invitation to participate, along with a baseline questionnaire, in 1998. A total of 23,842 had data on physical activity and dementia and were thus eligible for our meta-analyses. The Turku University Central Hospital Ethics Committee approved the study.
Physical inactivity was defined as "less than 0.5 hour of each (brisk walking, jogging, or running) per week." Participants were linked to drug reimbursement, hospitalisation and death registers. Dementia was defined using ICD-10, codes F00, F01, F02, F03, G30 and G31 (31.0, 31.1, 31.8). prospective cohort studies with data on physical activity and a follow-up for cause-specific deaths. In NHANES I, physical activity was measured with the question "Do you get much exercise in things you do for recreation?" with the response "little or no exercise" defined as physical inactivity (versus the responses "moderate exercise" and "much exercise"). In NHANES 2, the response options were the same but the question was "In things you do for recreation, for example, sports, hiking, dancing, and so forth, do you get much exercise, moderate exercise or little or no exercise?" with the response "little or no exercise" again defined as physical inactivity. In NHANES III the participants were asked how often they participated in different sports included in a list of 9 physical activities (e.g., walking, jogging, swimming) and 1 option of other activity. Physical inactivity was defined as not participating in any of these activities. In the continuous NHANES cohorts from 1999 to 2005, the participants were asked two questions about their moderate and vigorous physical activities: (1) "Over the past 30 days, did you do any vigorous activities for at least 10 minutes that caused heavy sweating, or large increases in breathing or heart rate?" and (2) "Over the past 30 days, did you do moderate activities for at least 10 minutes that cause only light sweating or a slight to moderate increase in breathing or heart rate?" Physical inactivity was defined as answering no to both of these questions. Stroke was defined using a broader definition of ICD-10 codes I60-I69. . The participants were asked how often they performed moderate and vigorous physical activities. Physical inactivity was defined as participating in no leisure-time aerobic activity that lasted at least 10 minutes. Stroke was defined using a broader definition of ICD-10 codes I60-I69. In NHIS 1990 and 1995, data on alcohol consumption were not available. NHIS 1995 also lacked information about smoking habits.
Reference: Dawson DA. Ethnic-differences in female overweight -data from the 1985 National-Health Interview Survey. Am J Public Health 1988; 78: 1326-29.

Still Working
Still Working is an ongoing prospective cohort study. In 1986, the employees (n = 12,173) at all Finnish centres of operation of Enso Gutzeit (a forestry products manufacturer) were invited to participate in a questionnaire survey on demographic, psychosocial and health-related factors. Physical activity was measured at study baseline in 1986 and 9058 provided data and were followed up for dementia. The study was approved by the ethics committee of the Finnish Institute of Occupational Health.
Physical inactivity was defined as "Sport activities less than a couple of times per month." Data on height and weight (BMI) were not available.

WOLF (Work, Lipids, and Fibrinogen) Norrland, Sweden
WOLF Norrland is a prospective cohort of 4,699 participants aged 19-65 working in companies in Jämtland and Västernorrland counties and with data on physical activity and dementia. At study baseline the participants underwent a clinical examination and completed a set of health questionnaires. The baseline assessment was undertaken at 13 occupational health service units in 1996-98. The Regional Research Ethics Board in Stockholm, and the ethics committee at Karolinska Institutet, Stockholm, Sweden approved the study.

Statistical analysis
In the main analysis of all-cause dementia, Alzheimer's disease and each cardiometabolic disease, we used a 2-step individual-participant-data meta-analysis including study-specific analyses in the first step and pooling the study-specific estimates in the second. In each study, we performed Cox regression to generate hazard ratios and accompanying 95% confidence intervals (CI) for the association between physical inactivity (yes vs no) and the outcomes. Each participant was followed up from the date of physical inactivity assessment to the first record of dementia (or cardiometabolic disease of interest), death, or the end of follow-up.
To take into account that the associations are not necessarily similar across different settings, we present the summary hazard ratios from the random-effects models. Study-specific hazard ratios and their 95% confidence intervals (CIs) were combined using Knapp-Hartung estimator for between-study variance (these estimates are reported in text). 8 For comparison, the same meta-analyses were run using DerSimonian-Laird estimator for between-study variance (the default method in many software packages; these estimates are reported in appendix). 9 Two estimators were used because evidence from empirical and simulation studies suggests that the commonly used DerSimonian-Laird variance estimator can produce biased estimates particularly in meta-analyses based on small numbers of studies with moderate to substantial heterogeneity, 9 and Knapp-Hartung estimator can be less biased and more efficient. 8 We calculated I 2 and τ to estimate relative and absolute heterogeneity, respectively, among the study-specific estimates (in both indices, higher values denote greater heterogeneity). 10 We adjusted the hazard ratios for physical inactivity for age, sex, ethnicity, education/socioeconomic status (minimally-adjusted), and additionally BMI, smoking, and alcohol intake (multivariable-adjusted). We included in the analysis participants without missing data on the exposure, covariates in the minimally-adjusted model and outcome, but imputed missing covariates for BMI, smoking and alcohol intake, if the missingness was >10% (multivariate stochastic imputation with chained equations). In a preliminary analysis, we examined the association between physical inactivity and dementia ignoring potential non-proportionality; this approach corresponds to meta-analyses that are possible to conduct using only summary data from published studies.
We then examined whether the hazard ratio for physical inactivity was non-proportional over the follow-up using pooled individual-participant data from all cohort studies (these analyses were additionally adjusted for cohort). Two approaches were applied: Cox regression stratified by follow-up period (0 to <5 years, 5 to <10 years, 10 to <15 years, >15 years) and flexible parametric proportional-hazards for censored survival data on a log cumulative hazard scale. 11 In the latter analysis, we used the Akaike information criterion 12 to assist selection of the parametric model (the final model had two degrees of freedom for the restricted cubic spline function used for the baseline hazard rate and 1 degree of freedom for time-dependent effect of physical activity).
The analysis was then performed separately for incident dementia during the first 10 years of follow-up (when most dementia cases were expected to be at the preclinical or prodromal stage of dementia at the time of baseline physical inactivity measurement) and incident dementia from year 10 onwards for those who did not have dementia at year 10. As in the latter analysis the physical inactivity assessment was 10 years or more before recorded dementia, we assumed it was less likely affected by preclinical/prodromal stage of dementia. Similar analyses were performed for each cardiometabolic disease. The estimates for physical inactivity were adjusted for age (continuous variable), sex, ethnicity (white vs other) and education/socioeconomic status (high, intermediate, low). Multivariable-adjusted effect estimates were additionally adjusted for BMI (continuous variable), smoking status (current, ex-, never smoker), and alcohol intake (none, moderate, high).
Our analysis with follow-up starting from year 10 onwards is subject to regression dilution bias as people may change their physical activity during the first 10 years of follow-up. We assumed that the long-term level of physical activity has an impact on disease process. As the value of a single measurement of physical activity reflects both the usual level and random fluctuations unrelated to disease process, it will yield an underestimation of the true impact of physical inactivity on dementia. To address this potential source of bias, we corrected hazard ratios for regression dilution using Rosner method 13 and information from the stability of physical activity. The latter was estimated from repeated measurements of physical activity in two cohort studies with repeat physical inacitivty assessment (the Finnish Public Sector study and the Health and Social Support Study) and the pooled correlation coefficient was r=0.416 for 10-year stability.
For comparison to our analyses of physical inactivity and dementia, we examined the associations of physical inactivity with incident diabetes mellitus, coronary heart disease and stroke and treated them as positive controls. We assumed that if the positive control does not produce the expected result (i.e. an association of physical inactivity with incident type 2 diabetes, coronary heart disease and stroke is observed as these associations have previously been confirmed in metaanalyses of cohort studies and using randomized controlled trials on these or surrogate outcomes), 5-7 our measurement of physical inactivity or analytic procedure might not be correct and thus also findings of dementia might not be valid. Expected associations between physical inactivity and these cardiometabolic diseases, in turn, support the validity of our analytic approach.
To assess dose-response pattern, we repeated the main analyses using a 3-level physical activity measure as the exposure. This measure was available from the Finnish Public Sector study (FPS), the Health and Social Support study (HeSSup), WOLF, GAZEL, and the Still Working study.
To examine robustness of our findings, we performed pre-selected subgroup analyses by sex, age (threshold 60 years), study-specific physical inactivity prevalence (threshold 40%) and outcome ascertainment method (electronic records from morbidity registers, mortality registers, and both). Due to smaller sample sizes in these subgroups, the analyses were based on pooled data across all cohorts rather than meta-analysis of study-specific estimates and were adjusted for study in addition to other covariates.
To address potential survival bias we conducted a Fine and Gray competing risk analysis with dementia and death as outcomes. 14 This analysis was confined to the 5 studies with dementia identified using morbidity and mortality data. The late onset of disease may introduce survival bias. As the onset of dementia is at an older age than those of diabetes and coronary heart disease, we repeated the analysis of physical inactivity, diabetes, coronary heart disease and stroke in participants who were alive at age 65 and free of the health outcome of interest at that age. In this group, mean age at recorded incident diabetes, coronary heart disease or stroke was comparable to the mean age of recorded dementia in the entire cohort, thus reducing any differences in survival bias between analyses of dementia and the other health outcomes. The purpose of this sensitivity analysis was to compare the association between physical inactivity and cardiometabolic diseases to that between physical inactivity and dementia when the age of disease onset is the same for cardiometabolic diseases and dementia.
To assess the association of physical inactivity with dementia in relation to cardiometabolic disease (i.e. having one or more of diabetes, coronary heart disease and stroke), we formed two dementia endpoints for participants with no cardiometabolic disease at baseline and no dementia at year 10: (1) incident cardiometabolic disease followed by incident dementia (defined as cases, all others as non-cases) and (2) incident dementia without preceding cardiometabolic disease (defined as cases, all others as non-cases).
Censoring was at the date of dementia, death or end of follow-up. We tested whether physical inactivity was differently associated with these outcomes using the following formula: χ 2 (1 degree of freedom) = (b1-b2) 2 / (SE1 2 + SE2 2 ), where bi is parameter estimate for event endpoint i SEi is standard error for event endpoint i. 15 In these analyses, pooled data were used. Morbidity and mortality data for these disease trajectories were available from 5 studies: the Finnish Public Sector study (FPS), the Health and Social Support study (HeSSup), WOLF, GAZEL, and the Still Working study.
We used SAS (version 9.4) to analyse physical inactivity-health outcome associations separately in study-specific data. R (version 3.3.1) was used in meta-analyses to combine studyspecific estimates.
In relation to the follow-up from year 10 onwards for incident dementia, hazard ratios favoured risk factor status for physical inactivity in 12 studies and a protective effect in 7 studies. In this latter follow-up for Alzheimer's disease, 6 hazard ratios favoured risk status and 7 a protective effect. In contrast, despite heterogeneity in study-specific estimates, the hazard ratios for physical inactivity favoured risk factor status in 16-17 of the 18 studies on diabetes in both follow-ups. This was the case for 18-19 of the 19 studies on incident coronary heart disease and 15-18 of the 19 studies on incident stroke.
The mean age at ascertainment of dementia was high (80.6 years), compared to the mean age for diabetes, coronary heart disease and stroke (66.8, 75.1 and 73.4 years), raising the possibility of more severe survival bias as an explanation for weaker associations of physical inactivity in relation to dementia compared to cardiometabolic disease. This possibility was not supported by Fine and Gray competing risk analyses for dementia (eFigure 7), or a sensitivity analysis of incident cardiometabolic disease in a group of participants without cardiometabolic disease at age 65. In the latter analysis, the excess risk of diabetes, coronary heart disease and stroke associated with physical inactivity was evident (hazard ratios 1.27, 95% CI 1.16-1.39; 1.14, 95% CI 1.07-1.21; and 1.07 95% CI 0.97-1.17, respectively) in spite of the mean age at ascertainment of the health outcome being comparable to that for dementia (77.9, 81.9 and 80.5, respectively). eFigure 8 shows that adding previous long-term follow-up studies [16][17][18] to our meta-analysis of the association between physical inactivity and dementia when follow-up for dementia in our studies was started at year 10 did not change conclusions from our main analysis. eTable 1. Cohort characteristics eFigure 3. Hazard ratio for the association of physical inactivity with risk of Alzheimer's disease in the first 10 years of follow-up in participant without dementia at baseline and from year 10 onwards in those without dementia at year 10 (age-, sex-, ethnicity and SES/education-adjusted random-effects meta-analysis) eFigure 5. Hazard ratio for the association of physical inactivity with risk of coronary heart disease in the first 10 years of follow-up in participant without the disease at baseline and from year 10 onwards in those without the disease at year 10 (age-, sex-, ethnicity and SES/education-adjusted random-effects meta-analysis) eFigure 6. Hazard ratio for the association of physical inactivity with risk of stroke in the first 10 years of follow-up in participant without stroke at baseline and from year 10 onwards in those without stroke at year 10 (age-, sex-, ethnicity and SES/education-adjusted random-effects meta-analysis) eFigure 8. Hazard ratio for the association of physical inactivity with risk of dementia and Alzheimer's disease from year 10 onwards in participants without dementia at year 10 in the present study and those for the association during a long follow-up in three previously published studies (random-effect meta-analysis)