Association of Air Pollution and Physical Activity With Brain Volumes

Abstract

Background and Objectives

In high-pollution areas, physical activity may have a paradoxical effect on brain health by increasing particulate deposition in the lungs. We examined whether physical activity modifies associations of air pollution (AP) with brain volumes in an epidemiologic framework.

Methods

The UK Biobank enrolled >500,000 adult participants from 2006 to 2010. Wrist accelerometers, multimodal MRI with T1 images and T2 fluid-attenuated inversion recovery data, and land use regression were used to estimate vigorous physical activity (VigPA), structural brain volumes, and AP, respectively, in subsets of the full sample. We evaluated associations among AP interquartile ranges, VigPA, and brain structure volumes and assessed interactions between AP and VigPA.

Results

Eight thousand six hundred participants were included, with an average age of 55.55 (SD 7.46) years. After correction for multiple testing, in overall models, VigPA was positively associated with gray matter volume (GMV) and negatively associated with white matter hyperintensity volume (WMHV), while NO₂, PM_{2.5absorbance}, and PM_2.5 were negatively associated with GMV. NO₂ and PM_{2.5absorbance} interacted with VigPA on WMHV (false discovery rate–corrected interaction p = 0.037). Associations between these air pollutants and WMHVs were stronger among participants with high VigPA. Similarly, VigPA was negatively associated with WMHV for those in areas of low NO₂ and PM_{2.5absorbance} but was null among those living in areas of high NO₂ and PM_{2.5absorbance.}

Discussion

Physical activity is associated with beneficial brain outcomes, while AP is associated with detrimental brain outcomes. VigPA may exacerbate associations of AP with white matter hyperintensity lesions, and AP may attenuate the beneficial associations of physical activity with these lesions.

Air pollution (AP) and physical activity (PA) have been differentially associated with a wide range of health effects,^1,2 although their interaction is poorly understood. Specifically, both have been associated with neurologic outcomes: PA has been associated with reduced risk of dementia and enhanced cognition and structural brain volumes,³ whereas AP is associated with higher risk of dementia,⁴ poorer cognition,⁵ and smaller brain volumes.^6-9

Increased respiratory rates during exercise may increase deposition of particulate matter (PM) in the lungs.¹⁰ Evidence suggests that the combined effects of AP and PA may be biologically significant, particularly for the brain. Aerobic exercise in moderately polluted environments reduces serum levels of brain-derived neurotrophic factor in humans and reduces brain-derived neurotrophic factor gene expression in rats.¹¹ In humans, the benefits of PA on cardiopulmonary outcomes are attenuated in areas of high AP,¹² and AP may have stronger associations with respiratory health among persons with high levels of PA.¹³ However, to our knowledge, no epidemiologic study has yet investigated the effects of this interaction on the brain.

Here, we investigate associations of PA and AP with structural brain volumes in the UK Biobank (UKB), the largest epidemiologic cohort with combined AP, neuroimaging, and objectively measured PA. We test the hypothesis that AP levels are negatively associated with structural gray matter volume (GMV) and white matter volume (WMV) and positively associated with white matter hyperintensity volumes (WMHVs). We also hypothesize that PA is associated with these volumes in the opposite direction and that there is an interaction of AP with PA on brain volumes such that AP has stronger associations with brain volumes among those with higher levels of PA.

Methods

Population and Substudies

The UKB enrolled >500,000 participants between 2006 and 2010 at 22 assessment centers throughout the United Kingdom.¹⁴ Briefly, individuals 40 to 69 years of age were invited to enroll if they lived within 40 km of an assessment center and were registered with the National Health Service.¹⁵ More than 500,000 consented to join this community-dwelling study cohort. At enrollment, participants answered questions about demographics and characteristics via touch-screen questionnaire and a computer-assisted interview. Subsets of the cohort were invited to enroll in substudies to gather additional data (eFigure 1, links.lww.com/WNL/B650), including an accelerometry substudy and multimodal MRI substudy. Invitations for each substudy were sent randomly to UKB participants, first to all participants with valid email addresses with continuing enrollment in the imaging, including postal invitations.^16,17 Individuals were included here if they were part of both the accelerometer and imaging substudies.

Standard Protocol Approvals, Registrations, and Patient Consents

All participants provided written informed consent. Ethics approval of the UKB study was given by the North West Multicentre Research Ethics Committee, the National Information Governance Board for Health & Social Care, and the Community Health Index Advisory Group. This study was approved by the University of Arizona Institutional Review Board, which determined that participant consent was not required for this analysis of entirely deidentified data and was waived.

Exposure and Outcome Assessment

Air Pollution

A pan-European land use regression (LUR) model following the European Study of Cohorts for Air Pollution Effects (ESCAPE) protocol was used to estimate annualized AP (nitrogen oxides [NO_X], nitrogen dioxide [NO₂], PM ≤2.5 µm [PM_2.5], PM 2.5–10 µm [PM_2.5-10], PM ≤10 µm [PM₁₀], and PM_{2.5absorbance} [a measure of black carbon]) for the 2010 calendar year at enrollment.^18,19 Descriptions are available on the UKB Data Showcase website. Briefly, LUR predicts monitored AP levels from land use characteristics. β Coefficients are then extracted from the predictive model and combined with values of land-use characteristics at other locations to predict AP levels in unmonitored areas. In this case, LUR models were used to estimate AP levels at participants' residences. PM and NO_X models were developed in separate LUR models.^18,19 PM variables were measured across 20 different European study areas, with 20 sites per area, from 2008 to 2011. Two sites were in the United Kingdom (Manchester and London/Oxford). Geospatial predictor variables (e.g., traffic density, road length, building canyons, population, green space) were used to model AP concentrations. LUR models were locally optimized at each site. NO₂ and NO_X LUR models were similarly constructed, although AP values were measured with Ogawa passive samplers at 36 study areas, with 40 or 80 sites per area. Three areas in the United Kingdom (Manchester, London/Oxford, and Bradford) were included, with measurements from 2008 to 2011. The r² for NO₂ and NO_X LUR models in the study region ranged from 0.83 to 0.91,¹⁹ and the r² for PM_2.5 and PM_{2.5absorbance} in included study regions ranged from 0.82 to 0.96.¹⁸ Models for PM measures that were >400 km from the center of London were below the ESCAPE threshold and are excluded by the UKB. Thus, 33,935 participants are missing PM measurements but do have NO₂ and NO_X measurements.

The UKB reports annual NO₂ estimates using European Union–wide AP maps, not ESCAPE methods, for 2005 to 2007. Correlations across the years 2005 to 2007 for NO₂ were high (>0.98, eFigure 2, links.lww.com/WNL/B650). Although 2010 values were measured with ESCAPE and the correlations capture differences in AP measurement methods in addition to temporal variability, ρ was still high when 2010 was predicted from earlier years (≥0.85) for NO₂. Thus, we assumed that the 2010 values are a measure of chronic AP.

AP measurements were treated continuously here and transformed so that a 1-unit increase in the AP measure was equivalent to an interquartile range (IQR) increase. To enhance interpretability, we used AP quartiles for strata-specific estimates of associations between vigorous PA (VigPA) and brain outcomes.

Physical Activity

Participants (N = 103,706) wore Axivity (Newcastle Upon Tyne, UK) AX3 triaxial accelerometers on their dominant wrists that recorded information on type, intensity, and duration of PA for a 1-week period between 2013 and 2015.¹⁶ Although early accelerometer-based studies used waist-worn devices, wrist-worn monitors are widely used in epidemiologic studies because participants can wear them continuously throughout the day and night and therefore have a higher level of participant compliance.¹⁶ Acceleration averages were calculated from accelerometer vector magnitude averaged over total wear time. After calibration, vector magnitudes were calculated from raw triaxial accelerometer signals, and 1 gravitational unit was removed from this signal (i.e., euclidean norm minus 1).¹⁶ Average vector magnitude was calculated over 5-second intervals, and values were imputed for time periods identified as nonwear,¹⁶ although total wear time averaged 6.8 of 7 total days. The UKB provides average vector magnitude for each individual and the distribution of time spent in different vector magnitude intensities.

We used measures of the fraction of time spent above a certain accelerometer magnitude threshold as markers of VigPA (≥425 milligravities [mg]) or moderate to vigorous PA (MVPA; ≥100 mg).²⁰ After evaluating variable distributions and balancing power for continuous variables against interpretability for interactions and clinical relevance, we used quartiles for average acceleration and objective 4-level ordinal measures of time spent in VigPA and time spent in MVPA according to accelerometry data. For VigPA, categories were 0, >0 to <15, 15 to <30, and ≥30 min/wk. For MVPA, our categories, were 0 to <1, 1 to <1.5, 1.5 to <2, and ≥2 h/d. To assess interactions between PA and AP, we selected a single PA measure (VigPA), which minimizes multiple testing while using a PA measure with high levels of respiration that is the focus of our study hypothesis. To convert fraction of time spent in vigorous activity to minutes per week, we multiplied the fraction of time by 10,080 min/wk.

Brain Imaging

As of 2019, multimodal MRI data for 21,407 individuals, collected in 2014 to 2015, were available for analysis. All UKB MRI scans for this analysis were acquired on a Siemens 3T Skyra system (Siemens Healthineers, Erlangen, Germany) with a standard 32-channel receive head coil from 1 imaging center. T1 images, which provide information about volumes of brain structures and tissues, were processed with the fully automated FSL pipeline that generated image-derived phenotypes, with 98% of participants having useable T1 images. T1 parameters included 3D magnetization-prepared rapid acquisition gradient echo, sagittal acquisition, 1-mm isotropic resolution, in-plane acceleration factor of 2, and inversion/repetition time of 880/2000 milliseconds. T2 fluid-attenuated inversion recovery data were processed for detection of focal high-signal regions in white matter (i.e., white matter hyperintensities) typically associated with white matter lesion load, and 97% of participants had useable T2 fluid-attenuated inversion recovery data. Parameters for the T2 protocol used 3-dimensional Sampling Perfection With Application Optimized Contrasts by Using Different Flip Angle Evolutions (SPACE), sagittal acquisition, 1.05 × 1.0 × 1.0–mm resolution, in-plane acceleration factor of 2, and inversion/repetition time of 1,800/5,000 milliseconds. All scans were prescan normalized (on-scanner bias-field corrected), and gradient distortion correction was applied during scan postprocessing. These methods and quality control procedures have been described previously.²¹

We evaluated associations with total GMV, total WMV, and WMHV. Brain variables were assessed for normality. All volumes were preadjusted for head size by multiplying raw volume by the head size scaling factor provided by the UKB. WMHVs were log transformed. All outcome variables were standardized to a mean of 0 and SD of 1. A positive β coefficient reflects associations with larger volumes (i.e., less brain atrophy) for WMV and GMV. For WMHVs, a larger β coefficient indicates associations with greater (worse) white matter lesion load. In sensitivity analyses, we excluded participants with unusually high WMHV values that were >2.5 SDs above the mean²² (n = 47).

Statistical Analysis

We conducted complete case analyses and restricted to participants with complete covariate, exposure, modifier, and outcome data (N = 8,600, eFigure 1, links.lww.com/WNL/B650). We excluded participants for whom accelerometry data were deemed unreliable by the UKB¹⁶ and participants who did not wear the accelerometer for >5 of the 7-day wear time for the study. We used linear regression to assess relationships among AP, PA, and brain structure and evaluated standard regression diagnostics. Overall associations were estimated for main effects of AP (NO_X, NO₂, PM_2.5, PM_{2.5absorbance,} PM_2.5-10, PM₁₀) and PA (VigPA, MVPA, overall accelerometry average) on brain structure with single-pollutant models. AP models were adjusted for the VigPA variable (because this was used in interactions and had strongest associations with brain outcomes), and the PA models were adjusted for NO₂ (because we had more observations with NO₂ than PM). We assessed 2-way interactions between each air pollutant and VigPA on each brain outcome. Strata-specific estimates for AP were obtained by changing the referent group of interaction terms in the model. For associations for which we observed interactions between AP and brain structure, we also evaluated associations between PA and brain structure by AP quartile. Last, we excluded outlier observations, defined as having both very high Cook distance (>3n) and leverage values (>1.5 p/n) from each model.^23,24

Covariates were selected a priori on the basis of hypothesized relationships with AP and brain structure. We evaluated covariates for potential collider and mediation biases in a directed acyclic graph.²⁵ All models were adjusted for race and ethnicity (white/other), sex, age at enrollment (continuous), urban or rural address, assessment center, and education (college or higher vs less). UKB participants also reported household income, although this was not available for almost 1,000 participants in our study. We thus excluded this variable from primary analyses but conducted sensitivity analyses with this covariate included (categories of <£18,000; £18,000–£30,999; £31,000–£51,999; £52,000–£100,000; >£100,000).

We corrected main effects and interaction effects for multiple testing using the Benjamini-Hochberg false discovery rate (FDR) method.²⁶ Because we a priori tested all main effects and interactions between AP and VigPA, we corrected interaction p values and main effects p values together. Because we used 4-level measures for PA in the main effects analyses, we corrected p trends from the tests of the ordinal variables. Thus, we corrected for 45 tests (27 main effects [6 AP and 3 PA on WM, GM, WMHV] and 18 interactions [6 AP by VigPA tests on WM, GM, and WMHV]). Because AP measures were highly correlated with each other and the brain volume measures were correlated with each other, such correction likely yields a conservative estimate.

In sensitivity analyses, we assessed several potential confounders, including season of PA, household income, time between AP assessment (December 31, 2010) and the start date of each participant's accelerometer study, and time between each participant's accelerometer and MRI study. We also assessed BMI at enrollment, current smoking, alcohol intake, and history of heart attack, high blood pressure, and diabetes. We also restricted analyses to participants whose accelerometry preceded their brain imaging, participants with WMHVs < 2.5 SDs above the mean, participants without a history of stroke, and participants without dementia before or during the study period. Last, we evaluated interactions between PA and age, between AP and age, and a 3-way interaction of PA, age, and AP on WMHV.

Data Availability

All data are available to any researcher on application to the UKB. Documentation and code will also be submitted to the UKB repository and will be made available from the corresponding author (M.A.F.) on request.

Results

Participants included 8,600 adults with complete AP, accelerometry, brain imaging, and covariate data. A subset of these had available data on WMHVs (n = 8,016). Participants were older (mean age 55 years), were mostly White, and lived in urban areas, with an average education below the college graduate level and moderate household incomes. Participants were slightly more likely to be women (54%) (Table 1). Participants included in this study also were more likely than participants in the overall UKB to have higher education levels and higher household incomes and to be of White ancestry (Table 1). AP levels were low to moderate, with overall averages below European targets for human health. These characteristics were significantly different compared to the entire study population (p < 0.05). No participants lived in areas where PM_2.5 or PM₁₀ yearly averages exceeded European Union targets for human health (25 µg/m³ for PM_2.5, 40 µg/m³ for PM₁₀), and only 2.48% of the included study population lived in areas where NO₂ exceeded standards for annual averages (40 µg/m³). In this study, the average time between the accelerometry substudy and the MRI substudy was 1.49 years.

Table 1.

Characteristics of the Study Population

In overall models, we report several associations among AP, PA, GMV, and WMHV (Table 2). The raw p values that corresponded to an FDR p value of <0.05 were <0.012. One-IQR increases in NO₂, PM_2.5, and PM_{2.5absorbance} were inversely associated with GMV (Table 2). Although effect sizes for these pollutants and WMV were similar, these p values did not meet the FDR threshold. PA was strongly associated with GMV and WMHV, with a dose-response effect across all categories of PA (Table 2). For instance, as VigPA increased, the magnitude of the associations with GMV similarly increased, from 0.101 to 0.149 to 0.199. Similarly, associations for VigPA with WMHVs strengthened across categories of VigPA (from −0.148 to −0.179 to −0.234). The FDR-corrected p trends for these associations were <0.001. We observed similar patterns for MVPA and average acceleration quartiles. We generally observed no consistent associations of PA with WMV.

Table 2.

Main Effects (β and 95% CI) of Air Pollution and Physical Activity on Brain Structures (N = 8,600)

In interaction models, we observed interactions between AP and VigPA on WMHVs (Table 3) but not GMV or WMV. Specifically, effect estimates for associations of NO₂ and PM_{2.5absorbance} with WMHVs increased as VigPA increased. The association for NO₂ and WMHVs was essentially null among those with 0 or 0 to <15 min/wk VigPA and elevated among those with ≥30 min/wk. Similarly, the association for PM_{2.5absorbance} and WMHVs was close to the null for those with <30 min/wk VigPA and elevated for those with ≥30 min/wk VigPA (Table 3). Although we observed similar strata-specific effect sizes for NO_X and PM_2.5 with WMHV, these interactions did not meet criteria for FDR-corrected significance. In examinations of strata-specific associations between VigPA and WMHVs, the inverse association of AP and VigPA on WMHVs was weakened in higher AP quartiles. Among those in the first quartile of NO₂ and PM_{2.5absorbance}, PA was strongly negatively associated with WMHVs (Figure 1). In higher AP quartiles, this association retreated toward the null, so that among those in the highest quartile of AP, VigPA was no longer negatively associated with WMHVs.

Table 3.

Modification of Associations of AP IQRs With WMHVs, by VigPA (N = 8,016^a)

Figure 1. Associations of VigPA With WMHV, by Air Pollution Quartiles.

Estimates show the associations between vigorous physical activity (VigPA; binary for >30 min/wk, which corresponds to the fourth category) and white matter hyperintensity volume (WMHV), by quartile of air pollution exposure, for NO₂ and particulate matter (PM)_{2.5 absorbance} (N = 8,016). Models control for reported education, race and ethnicity, age at recruitment, assessment center, and urban or rural status. Bars represent 95% confidence intervals (CIs). *Significance at the false discovery rate–corrected p < 0.05 level.

In sensitivity analyses, adjusting for season of PA measurement, income, time between AP and MRI, and time between MRI and accelerometry measures did not affect interpretation or change β coefficients by >15% (eTables 1 and 2, links.lww.com/WNL/B650). We similarly did not observe notable changes when controlling for BMI, smoking, and alcohol use (eTables 1b and 2b) or when restricting to participants with accelerometry measured before MRI (eTable 2a) and when excluding participants with WMHVs >2.5 SDs above the mean (n = 47, eTable 2a). Interpretations were unchanged when we excluded participants with a history of dementia or stroke and controlled for diabetes, heart attacks, and high blood pressure (eTables 1b and 2b). At FDR-corrected significance, we did not observe any interaction of PA with age on any of the outcomes; any 3-way interaction among VigPA, age, and AP; and only very limited evidence of interactions between pollutants and age. We did observe an interaction between PM_2.5-10 and age for GMV and WMHV (interaction p = 0.01 for both), such that PM_2.5-10 was negatively associated with GMV and positively associated with WMHV among participants <50 years of age but not older participants (eTable 3).

Discussion

This is the first study to examine interactions between PA and AP on structural brain volumes and one of a few studies to report overall associations between AP and brain structure. In addition, this study was performed on the largest database of brain structural imaging and accelerometer-measured PA currently available. Thus, our results represent a well-powered assessment of whether AP and PA modify each other's association with brain volumes in the middle-aged to older adult brain.

We consistently found strong positive associations between higher PA and GMV and negative associations with WMHV in overall analyses, which is consistent with prior literature.²⁷ We also report that AP was inversely associated with brain outcomes; NO₂, PM_2.5, and PM_{2.5absorbance} in particular were inversely associated with GMV. Although we also report similar effect sizes and directionality for other air pollutants and WMV, these p values did not withstand FDR correction. Supporting evidence for these associations in the literature is sparse. Two studies in the Women's Health Initiative reported associations of PM_2.5 with WMV and GMV among women,^6,7 which is consistent with both our FDR-corrected association for PM_2.5 with GMV and the elevated (although not statistically significant) association we report for PM_2.5 with WMV. In one report from the Framingham Offspring study,⁸ authors found associations between PM_2.5 and total cerebral brain volume and covert brain infarcts, while others reported associations between PM_2.5 and PM₁₀ and deep GMV.⁹ In addition, recent work has shown no association among WMHV and PM_2.5, NO_x, and NO₂ and no association between WMHV and distance to roads,²⁸ while others reported associations of PM₁₀ and NO₂ with worsening white matter grade.⁴ Wilker et al.⁸ paradoxically observed that greater distance to road (less AP) was associated with more WMHV. While their findings are unexpected, Wilker et al. note that these associations with distance to road did not hold for a binary measure of extensive WMHV. They describe extensive WMHV as being more clinically relevant in their sample because extensive WMHV, but not linear WMHV, had previously been associated with cognitive function in their study. Other studies found associations between black carbon (e.g., PM_{2.5absorbance}) or traffic-related AP and worse cognitive performance,⁵ and worse cognitive performance has been associated with greater WMHVs.²⁹

Variation across studies may be due to differences in sample demographics, AP measurement approaches, study sites, enrollment dates, and urban/rural differences. Differences may also be due to varying composition of PM across sites, which may have differing sources and regulations. In addition, AP levels in UK regions here were generally low. Thus, we may observe stronger associations in regions with higher pollution. To place our findings in context, effect sizes of AP IQRs shown here were modest compared to the effect of 1 year of age. A 1-year increase in age had a β of ≈−0.07 for GMV and 0.06 for WMHV. NO₂ had a β of −0.042 on GMV and 0.029 on WMHV in main effect models, which are roughly equivalent to half the effect of 1 year of age. In PA subgroup analyses, the largest effect size in the high VigPA group was 0.071 for PM_{2.5absorbance}, which approximates the effect of 1 year on WMHV. In addition, the main effects of VigPA were much larger; >30 minutes of VigPA had β values of 0.199 for GMV and −0.234 for WMHV. Thus, the putative effect of regular VigPA on brain structure is approximately equivalent to being 3 years younger.

We found that NO₂ and PM_{2.5absorbance} are positively associated with WMHVs among those with high levels of VigPA and that the possible beneficial effect of PA on WMHVs appears to be neutralized among participants with higher pollution exposures. Recent epidemiologic studies from Europe reported no PA/AP interaction for myocardial infarction,³⁰ mortality,³¹ or arterial stiffness.³² However, a randomized trial in London showed that short-term AP exposures neutralized beneficial effects of walking on cardiopulmonary outcomes,¹² which is consistent with our findings.

WMHVs are considered to have vascular origins. Elevated blood pressure is positively associated with WMHVs, and visceral fat, smoking, and inflammatory markers have been associated with WMHVs.³³ Larger WMHVs are also associated with other progressive pathologic changes that include demyelination, astrogliosis, axonal disruption,³⁴ Alzheimer disease and dementia, cognitive decline, and increased risk of stroke and mortality.^29,34 AP has consistently been associated with similar vascular outcomes.³⁵ Hypothesized mechanisms for AP on cardiovascular or neurologic outcomes broadly include inflammation and oxidative stress^36,37 and DNA damage, neuroinflammation, and accumulation of β-amyloid.¹¹ In addition, both WMHVs and AP have been associated with microglial activation.^33,38 These shared associations between WMHVs and AP indicate that a common biological response may be related to both WMHVs and AP, although this may vary for gases (NO₂/NO_X), fine particulates (PM_2.5), and larger particulates (PM₁₀). Specifically, NO₂ exposure in rats has been shown to increase risk of cerebrovascular disease by inducing excitotoxicity and increasing expression of proteins that mediate long-term potentiation and influence synaptic plasticity.³⁶ PM_2.5, but not NO₂, has been shown to reduce flow-mediated vasodilation.^37,39 Evaluating different air pollutants may help identify specific mechanisms linking AP and PA to brain health.

Our study had several strengths and limitations. Strengths include the large dataset, use of accelerometers to objectively assess PA, and validated AP measures with excellent model performance. Limitations center predominantly around temporality of measurements. We used 2010 AP values because this was the only year for which all components were available and the latest year for which AP data were available. However, correlations across years are likely high; measurements of NO₂ are reported for earlier years, and correlations across years for which NO₂ was estimated with the same methods are quite high (ρ > 0.98; eFigure 2, links.lww.com/WNL/B650). In addition, several studies in Europe and in North America have demonstrated that LUR models are temporally stable over short and long time frames of 7 to 12 years.^40-43 An additional limitation is that the accelerometry substudy took place in 2013 to 2015. Our hypothesis regarding the biological interpretation of interactions between PA and AP relies in part on the assumption that the 2 measures represent simultaneous exposures. Although longitudinal data are limited, there is evidence that changes in objectively measured PA over short time periods are modest, and PA at baseline is highly predictive of PA measured at a later time point in middle-aged adults.^44,45 In addition, time between substudies was not a confounder in sensitivity analyses, and AP measures in our study were highly correlated over years. However, some of our participants may have moved between enrollment and substudies. It is possible that some participants with better brain function were more mobile. Conversely, people with deteriorating brain function may tend to move to care homes or move in with relatives. Prior studies that have estimated this bias report that ignoring mobility generally biases results toward the null.⁴⁶ Thus, our results may be an underestimate of true associations. There may also be exposure misclassification for participants who have high-intensity occupations with high levels of AP (e.g., loading dock workers who work around diesel-fueled vehicles). This, too, would bias estimates toward the null because their overall AP exposures would be underestimated.

Our analysis also did not consider the duration or intensity of individual activities, and future work should examine individual physical activity bout characteristics to refine our understanding of PA-related AP exposure. We also examined only VigPA in interactions, and these interactions may differ for different intensities. In addition, some of the proposed biological mechanisms rely on the assumption that PA is taking place in an area where AP is representative of the LUR-modeled AP levels at their residence. In support of this, previously published data show that indoor, outdoor, and personal AP levels are well correlated with each other, such that ambient AP measures are a good proxy for personal exposures,^47,48 with 1 study specifically in England.⁴⁸ However, future work is needed to better detail individual exposure to AP according to the type and location of PA. Finally, aspects of our sample and geographic location may limit generalizability. Our study population was predominantly of European descent. Because cardiovascular risk factors vary by race and ethnicity, it may be important to evaluate these associations in populations with different cardiovascular risk. The UKB population is healthier than the general UK population, which may indicate some selection bias,¹⁵ underscoring the need to replicate our findings. However, prior studies have argued that although the UKB is not suitable for estimating generalizable disease prevalence and incidence rates, findings are externally valid for providing scientific inferences of measures of associations between exposures and health outcomes.^15,49 Our study sample additionally had slightly different demographics from those at enrollment, which may indicate the potential for our estimates to be skewed by selection bias and may reduce generalizability of our findings. Although we controlled for these demographic characteristics, there may be residual bias, again underscoring the need to replicate these findings. In addition, levels of pollution are slightly lower in the imaging subsample compared to the full UKB sample. Although differences are not large, they underscore that interactions between PA and AP are found even at moderate levels of AP. Last, overall AP levels in the United Kingdom are lower than in large cities in some other countries.⁵⁰ We suspect that the relationship among VigPA, AP, and brain health will be similar in other countries and that higher levels of pollution will strengthen these relationships; however, future work should examine more polluted areas.

Despite potential limitations, this study provides evidence that even in areas of moderate pollution, high-intensity PA may exacerbate associations between AP and worse brain health. Furthermore, although AP may neutralize beneficial effects of PA on certain brain structures, we did not observe any association between PA and worse brain outcomes. Our findings are consistent with previous studies that have not found main effects of AP on WMHV. However, it is possible that effects may be stronger in environments with higher pollution. Future work in areas with higher levels of pollution may find these associations strengthened and potentially that they occur at lower exercise intensities. While speculative, our results suggest that vascular mechanisms may play a role in these associations, consistent with previous work showing that even short-term AP exposure during exercise may have adverse cardiovascular impacts.¹²

If these findings are replicated further, policy could be structured to minimize exposure to AP during exercise. Because most AP sources are traffic related, promoting running or bicycling along paths far from heavily trafficked roads may reduce AP-related risks. In addition, risk assessment measures for AP should consider those with high PA as a subpopulation of possible concern.

Glossary

AP

air pollution

ESCAPE

European Study of Cohorts for Air Pollution Effects

FDR

false discovery rate

GMV

gray matter volume

IQR

interquartile range

LUR

land use regression

mg

milligravities

MVPA

moderate to vigorous PA

NO_X

nitrogen oxides

PA

physical activity

PM

particulate matter

VigPA

vigorous PA

WMHV

white matter hyperintensity volume

WMV

white matter volume

Appendix. Authors

Study Funding

Funding provided by P30AG019610, R01AG049464, R56AG067200, R00ES028743, R01HL136528, McKnight Brain Research Foundation, and the State of Arizona and Arizona Department of Health Services.

Disclosure

No authors have corporate sponsorship to disclose. M.A. Furlong is supported by NIH and the US Centers for Disease Control and Prevention. G.E. Alexander is supported by NIH, the McKnight Brain Research Foundation, and the State of Arizona and Arizona Department of Health Services. Y.C. Klimentidis and D.A. Raichlen are supported by NIH. Go to Neurology.org/N for full disclosures.