Association between Fine Particulate Matter Exposure and Cerebrospinal Fluid Biomarkers of Alzheimer’s Disease among a Cognitively Healthy Population-Based Cohort

Background: Epidemiological evidence suggests air pollution adversely affects cognition and increases the risk of Alzheimer’s disease (AD), but little is known about the biological effects of fine particulate matter (PM2.5, particulate matter with aerodynamic diameter ≤2.5μm) on early predictors of future disease risk. Objectives: We investigated the association between 1-, 3-, and 5-y exposure to ambient and traffic-related PM2.5 and cerebrospinal fluid (CSF) biomarkers of AD. Methods: We conducted a cross-sectional analysis using data from 1,113 cognitively healthy adults (45–75 y of age) from the Emory Healthy Brain Study in Georgia in the United States. CSF biomarker concentrations of Aβ42, tTau, and pTau, were collected at enrollment (2016–2020) and analyzed with the Roche Elecsys system. Annual ambient and traffic-related residential PM2.5 concentrations were estimated at a 1-km and 250-m resolution, respectively, and computed for each participant’s geocoded address, using three exposure time periods based on specimen collection date. Associations between PM2.5 and CSF biomarker concentrations, considering continuous and dichotomous (dichotomized at clinical cutoffs) outcomes, were estimated with multiple linear/logistic regression, respectively, controlling for potential confounders (age, gender, race, ethnicity, body mass index, and neighborhood socioeconomic status). Results: Interquartile range (IQR; IQR=0.845) increases in 1-y [β:−0.101; 95% confidence interval (CI): −0.18, −0.02] and 3-y (β:−0.078; 95% CI: −0.15, −0.00) ambient PM2.5 exposures were negatively associated with Aβ42 CSF concentrations. Associations between ambient PM2.5 and Aβ42 were similar for 5-y estimates (β:−0.076; 95% CI: −0.160, 0.005). Dichotomized CSF variables revealed similar associations between ambient PM2.5 and Aβ42. Associations with traffic-related PM2.5 were similar but not significant. Associations between PM2.5 exposures and tTau, pTau tTau/Aβ42, or pTau/Aβ42 levels were mainly null. Conclusion: In our study, consistent trends were found between 1-y PM2.5 exposure and decreased CSF Aβ42, which suggests an accumulation of amyloid plaques in the brain and an increased risk of developing AD. https://doi.org/10.1289/EHP13503


67
As life expectancy rises and the U.S. population pyramid continues to age, we are seeing an 68 increase in chronic non-communicable age-related conditions, including Alzheimer's disease and 69 related dementias (AD/ADRD). With the multifactorial nature of AD pathogenesis, building evidence 70 around preventable environmental exposures may ultimately reduce inequities and improve health 71 outcomes. In this regard, accumulating epidemiological evidence demonstrates an association 72 between exposure to air pollution and the prevalence of AD/ADRD. Most of the evidence to date has 73 focused on links between fine particulate matter (PM2.5), a mixture of fine particles in the air, and 74 cognitive function, incident cognitive impairment, or dementia. 1-4 Incident dementia is often studied 75 using diagnostic codes on insurance billing claims and medical records, facilitating studies with large 76 sample sizes; however, billing data are known to miss true dementia cases 1, 5 and there are no 77 diagnostic codes for the preclinical stages of dementia. In addition, PM2.5 can pass through the lung-78 gas-blood barrier, the gut-brain axis or directly enter brain tissue via the olfactory nerve to promote 79 oxidative stress and inflammation, processes directly related to the characteristic pathology of AD. 6, 7 80 Since PM2.5 is a heterogeneous mixture, different sources of exposure often have varying degrees of 81 toxicity and most existing studies have focused on ambient exposure, rather than those merely from 82 traffic-related emissions. Understanding the impact of PM2.5, from both ambient and traffic-related 83 sources, on preclinical stages of dementia in older at-risk adults is crucial from a public health 84 perspective, as it will improve the estimation of the burden of disease in association with air pollution 85 by identifying more affected individuals.

87
Recent systematic reviews highlight the need to expand the scope of current studies on air pollution 88 and dementia risk to include neuropathologically relevant outcomes, such as early indicators of 89 dementia, including AD cerebrospinal fluid (CSF) and plasma-based biomarkers, which can be 90 assessed at the late stage of the preclinical phase. 1, 4, 8 A recent study found positive associations 91 between PM2.5 and amyloid β-protein (Aβ)1-40 from plasma in longitudinal analyses, but no 92 associations were detected between PM2.5 and Aβ1-42 or the ratio of Aβ1-42/Aβ1-40. 9 Blood-based 93 biomarkers of protein pathology are less predictive of brain pathology and have shown insignificant 94 changes in Aβ levels as compared to CSF Aβ biomarkers in AD patients. 10 In this light, three 95 clinically validated CSF biomarkers: Aβ42, total tau (tTau), and phosphorylated tau (pTau) have been 96 noted as valid proxies for neuropathological changes of AD, 8,11 and specifically, Aβ42 has been 97 linked to the abnormal pathologic state of cerebral Aβ in both animal and human models. Since 98 pathological changes related to AD can begin decades before symptoms appear, 12 quantifying the 99 relationship between both ambient and traffic-related PM2.5 pollution and CSF biomarkers reflective 100 of AD-positive changes will elucidate how exposure influences dementia risk. 13 101 102 So far, two epidemiological studies have reported associations between PM2.5 and Aβ42 in cognitively 103 healthy individuals whereas no relationships with tTau or pTau have been noted. 14, 15 However, there 104 are several limitations with these prior studies, including 1) the exposure assessments in Alemany et 105 al. (2021) and Li et al. (2022) only focused on 1-or 2-year average PM2.5 prior to the biomarker 106 assessment; 2) the relatively small sample size (N=156) in Alemany et al. (2021), 15 and 3) the 107 outcome assessment in Li et al. (2022) 14 which relied on the Innotest-AMYLOID(1-42) ELISA assay, 108 an unstandardized manual method, showing a high correlation with the automated Elecsys® method 109 but higher intra-and inter-laboratory variations. 16, 17 110 111 To address these limitations in prior studies and grow our understanding of the impact of air pollution 112 on preclinical dementia risk, here, we characterized the cross-sectional association between long-113 term ambient and traffic-related PM2.5 exposure (1, 3, and 5 years prior to biomarker assessment) 114 and CSF biomarker composition (Aβ42, tTau, and pTau, assessed with Elecsys ® AD CSF assays) in 115 a dementia-free, aging population, as part of the Emory Healthy Brain Study (EHBS). We also tested 116 for effect modification by several well-known risk factors for AD/ADRD-related outcomes, including 117 APOE-ε4 status, the strongest genetic risk factor for AD. 118 119

121
Study Design and Population 122 123 The EHBS is a gerontology-based prospective research study focusing on the cognitive health of 124 older adults. The EHBS is nested within the Emory Healthy Aging Study (EHAS) and includes 125 participants from the metro-Atlanta region in Georgia, USA. Our cross-sectional analysis includes 126 data from the baseline visits, which were conducted between 2016 and 2020. The primary aim of the 127 EHBS is to characterize psychological and psychosocial factors associated with normal and 128 abnormal aging through assessment of the Due to evidence suggesting a relationship between both ambient PM2.5 and traffic-related PM2.5 147 exposure and cognitive decline, and because traffic-related PM2.5 is a major exposure source in 148 urban environments like Atlanta, GA, 19 we used both measures of PM2.5 in our analyses.

150
We obtained ambient PM2.5 exposure data from the publicly available Socioeconomic Data and 151 Application Center (SEDAC) air quality data set for health-related applications. 20 The data set 152 consists of yearly ambient PM2.5 levels (in µg/m 3 ) estimated at a 1 km spatial resolution using a well-153 validated ensemble-based prediction model for the contiguous United States (2000-2016). As 154 described by Di et al. (2019), three machine-learning algorithms: random forest, neural network, and 155 gradient boosting, were used to predict ambient PM2.5 and included a variety of predictor variables 156 from satellite data, land use, meteorological variables, and chemical transport model simulations. 21 The ensemble model then combined these PM2.5 predictions with a generalized additive model that 158 allowed for the contribution of each machine-learning algorithm to vary by location. 21 The ensemble 159 model was trained on PM2.5 levels measured at 2,156 U.S. EPA monitors, validated with 10-fold 160 cross-validation, and produced high-resolution annual PM2.5 predictions with an average R 2 of 0.89. 21 161 162 As Covariates 205 206 Sources of potential confounding were identified with a directed acyclic graph (DAG; Figure S1). 207 Individual-level confounders were conceptualized as factors impacting both residential PM2.5 208 exposure and the outcome measure. Potential confounding factors included in the analysis were 209 gender, age, neighborhood socioeconomic status (N-SES), race/ethnicity, educational attainment, 210 and body mass index (BMI). Due to historic racism and discriminatory land-use practices such as 211 redlining, environmental exposures disproportionately affect low-income and minority populations. 212 For this reason, neighborhood deprivation characteristics were also included as potential 213 confounding variables and effect modifiers as done in our previous work. 30 Race has also been 214 noted as an important factor when interpreting CSF biomarker results. 31 In addition, BMI influences 215 biomarker concentrations and is also related to N-SES through characteristics such as neighborhood 216 walkability, greenspace, and food access. 32, 33 217 218 Due to the presence of multi-ancestral groups and small categories in self-reported race/ethnicity, 219 we used a 3-level race variable in the analysis: White/Caucasian, Black/African American, and 220 Other, as well as a dichotomous ethnicity variable indicating Hispanic origin. Similarly, educational 221 attainment was included as a 3-level variable: master or higher, college, less than college. Height 222 and weight measurements were used to calculate body mass index (BMI, weight in kilograms 223 divided by height in meters squared) which was used as a continuous variable in all models. N-SES 224 for each participant was established in this study with census-tract level American Community 225 Survey (ACS) defined principal components of neighborhood deprivation (see Li et al. 30 for details) 226 and the Area Deprivation Index (ADI). As described previously in Li et al., 30 three principal 227 components of neighborhood deprivation were calculated based on estimates for 5-year ACS 228 census-tract-level data, including 16 indicators of six socioeconomic domains (poverty/income, racial 229 composition, education, employment, occupation, and housing properties) ( Figure S2). 30 The ADI is 230 provided in national percentile rankings at the block group level from 1 to 100, where 100 represents 231 the most deprived neighborhood, and was calculated using census block group-level indicators and 232 factor analysis to cluster indicators based on their ability to explain the variance between block 233 groups. 34 234 235 Statistical Analyses 236 237 We implemented multiple linear and logistic regression models to estimate the relationship between 238 residential PM2.5 exposure and AD CSF biomarker levels at enrollment. In these models, biomarker 239 concentrations (linear regression models) or dichotomized AD positive variables (logistic regression 240 models) were assigned as dependent variables and PM2.5 exposures along with the selected 241 confounding variables as independent variables. Since the biomarkers were measured on the same 242 scales, but with different ranges, we standardized all continuous biomarker measures by converting 243 them to z-scores prior to employing regression analysis to increase comparability of results across 244 different biomarkers. Z-scores were computed for each observation by subtracting the sample mean 245 from each individual value and subsequently dividing by the sample standard deviation. Further, we 246 standardized the PM2.5 estimates according to its distribution by dividing all PM2.5 exposures of 247 interest by the IQR of 1-year ambient or traffic-related PM2.5 exposure respectively. The general form 248 of the model for all analyses appears below: 249 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Where PM2.5 represents either ambient or traffic-related PM2.5 averages 1, 3, or 5 years prior to 254 specimen collection and ε represents the random error term, with an assumed mean of zero and 255 constant variance ~N(0, σ 2 ). 256 257 Effect Modification Analyses 258 259 We tested for effect modification by several well-established risk factors for AD, adding an 260 interaction term between PM2.5 and each risk factor in individual regression models. were localized exclusively to the city of Atlanta. Annual ambient and traffic-related PM2.5 exposure 294 concentrations were weakly correlated (Pearson correlation = 0.36). More details on the distribution 295 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted June 20, 2023. ; https://doi.org/10. 1101/2023 of and relation between ambient and traffic-related PM2.5 exposure concentrations are provided in 296 the supplemental material ( Figures S3 and S4).

298
We observed a wide spread of concentrations for CSF Aβ42 in the study population (median Aβ42 299 level = 1210, IQR = 692.3). Aβ42 concentrations did not show a major departure from normality, 300 except for the highest level which had a very high frequency, indicating normal Aβ42 concentrations 301 in most participants, but tTau and pTau distributions were skewed ( Figure S5). After log 302 transformation, Tau concentrations were approximately normally distributed ( Figure S6). 303 Approximately 36% of participants had Aβ42 concentrations less than or equal to 1030 pg/mL which 304 corresponds, on average, to a positive reading for AD as indicated by Elecsys ® AD CSF portfolio 305 positive (+) cut-offs. We observed AD positive readings for tTau and pTau cut-offs in 6% of the study 306 population. Based on the pTau/Aβ42 ratio AD (+) positive cut-off, we detected amyloid-positivity in 307 10.6% of participants. Details on the distributions of AD CSF biomarker concentrations and the 308 frequency of biomarker-positivity detected among participants are provided in Table 2. 309 310 PM2.5 and AD CSF biomarkers 311 312 Among the continuous AD CSF biomarkers, higher levels of 1-and 3-year ambient PM2.5 exposures 313 were associated with lower Aβ42 CSF concentrations at baseline after adjusting for potential 314 confounding variables ( Figure 2A). Specifically, an IQR (0.845 μg/m 3 ) increase in the 1-or 3-year 315 ambient PM2.5 exposure was significantly associated with a -0.09 (95% CI: -0.15, -0.02) and -0.07 316 (95% CI: -0.13, -0.005) decrease in Aβ42 CSF z-score, respectively, after confounder-adjustment. 317 The associations of 5-year ambient PM2.5 ( Figure 2A) and traffic-related PM2.5 ( Figure 2F) and Aβ42 318 CSF were similar, but not significant. No significant associations were detected between ambient 319 ( Figure 2B-E) or traffic-related ( Figure 2G-J) PM2.5 exposures and tTau, pTau, tTau/Aβ42, or 320 pTau/Aβ42 CSF concentrations at enrollment, but associations with tTau/Aβ42 and pTau/Aβ42 were 321 similar to those of Aβ42 and consistent with AD-related pathology ( Figure 2I-J).

336
The associations between ambient PM2.5 and Aβ42 CSF concentrations and AD positive (+) Aβ42 337 portfolio reading remained significant even after restricting our sample size to only those located in 338 the metro-Atlanta area ( Figure S7, Tables S1, S2, and S3), which is the subsample for which traffic-339 related PM2.5 exposures estimates were available (sample size reduced from Nambient=1,113 to 340 Ntraffic=1,080).

358
In the present study, we examined the impacts of both ambient PM2.5 exposure and traffic-related 359 PM2.5, a major source of ambient PM2.5 in urban environments, on CSF biomarkers of AD in 1,113 360 cognitively healthy individuals. Our findings show associations between long-term ambient PM2.5 361 concentrations and decreased Aβ42 AD CSF biomarker concentrations (significant for 1-and 3-year 362 average exposures), as well as increased likelihood of an Aβ42 AD (+) positive portfolio reading 363 (significant for 1-, 3-and 5-year average exposures). We further found significant associations 364 between 3-and 5-year traffic-related PM2.5 exposure estimates and a pTau/Aβ42 (+) positive portfolio 365 reading at enrollment, but not with ambient PM2.5 exposure. We found no associations between 366 ambient or traffic-related PM2.5 exposures and pTau or tTau continuous concentrations or their ratios 367 with Aβ42, however, the directions of effect for pTau/Aβ42 and tTau/Aβ42 continuous ratio outcomes 368 were consistent with AD-related amyloid pathology. Further, while not statistically significant, the 369 strength of the association between annual ambient PM2.5 exposure and Aβ42 AD CSF 370 concentrations differed by age and was particularly pronounced for individuals over the age of 60.

372
The observed associations between PM2.5 exposure and the Aβ42 AD CSF biomarker as well as 373 pTau/Aβ42 positive portfolio readings, which are equally predictive of amyloid PET status (+/-) as Aβ 374 ratio outcomes, 35 among cognitively healthy older adults is consistent with evidence from existing 375 literature. Signs of AD can be detected in the early stages of the AD continuum, 17 and decreases in 376 CSF concentrations of Aβ42 (a marker of amyloidosis) and elevation in tau species (phosphorylated 377 and total tau) are well-established as pathogenic biomarkers in AD diagnosis. 36 To date, there have 378 been few studies estimating the effects of PM2.5 exposure on certified biomarkers of AD in healthy, 379 aging populations. One study found a similar relationship between air pollution exposure and Aβ42, 380 although they used CSF Aβ42/40 ratio to reflect Aβ pathology rather than the individual biomarker 381 measurements. 15 Their estimates were similarly negative, but did not reach significance, likely owing 382 to the relatively small sample size (N=147). 15 Another study 14 found a statistically significant total 383 effect of ambient PM2.5 on Aβ42 CSF as well as pTau/Aβ42 concentrations among N=1,131 384 cognitively healthy older individuals, which was further mediated by a CSF biomarker of 385 neuroinflammation, sTREM2. 14 386 387 We found mostly null associations between 1-, 3-, and 5-year average ambient and traffic-related 388 PM2.5 exposures and tTau as well as pTau concentrations at enrollment, except for the model with 3-389 year ambient PM2.5 and tTau levels, in which we found a significant negative association. The 390 direction of estimates for tau biomarkers are expected to be positive for AD-related change. 391 Associations between PM2.5 and tau biomarkers diminished when taking AD CSF cut-offs into 392 account, suggesting that the negative association with the continuous tTau levels was a false 393 positive finding. Other studies examining the associations between air pollution and CSF biomarkers 394 of AD in cognitively healthy adults report similar findings, where both Alemany et al. (2021) and Li et 395 al. (2022) found null associations between PM2.5 and pTau as well as tTau CSF concentrations. 14, 15 396 397 Stronger associations were detected between ambient PM2.5 and AD CSF biomarkers as compared 398 with traffic-related PM2.5 exposure. Ambient PM2.5 contains emissions from traffic, industry, domestic 399 fuel burning, natural sources including soil dust and sea salt, as well as unspecific sources of human 400 origin. 37 On the other hand, traffic-related PM2.5 is a source of ambient PM2.5 that includes emissions 401 of organic and inorganic gaseous PM precursors from the combustion of fuels and lubricants. 37 402 Since both sources contain organic and often toxic particles, we expected to see relationships 403 between both sources of PM2.5 and AD CSF biomarkers, and while not significant for Aβ42, the 404 associations between traffic-related PM2.5 and AD CSF biomarkers were similar to associations with 405 ambient PM2.5. More research needs to be done to determine which PM2.5 mixtures are particularly 406 harmful to the central nervous system. 407 408 While we did not find effect modifications by APOE-ε4 carriership or other common risk factors for 409 AD, the association between ambient PM2.5 exposure and Aβ42 CSF became stronger with 410 increasing age (though not statistically significant). These results could suggest that AD CSF 411 biomarkers might not be sensitive enough to detect AD-related changes in participants < 60 years 412 old, but more research in the population will clarify the most clinically relevant age for biomarker 413 measurement. Previous research suggests that biomarker patterns of Aβ42 consistent with stage 1 414 AD (amyloid pathology only) are first detectable during early middle age (45-54 years), while 415 increases in tTau and pTau are typically not apparent until later (ages ≥ 55 years). 38 However, this 416 previous study used an unstandardized assay, the INNOTEST ELISA, which often yields systematic 417 variability in comparison to the Elecsys assay. Another potential explanation for the stronger 418 associations among participants older than 60 years could be the higher accumulative PM2.5 419 exposure over the lifetime among older individuals. In line with this hypothesis, one study examining 420 the relationship between PM2.5 exposure and AD prevalence found a stronger effect of PM2.5 on AD 421 prevalence among those at or above 70 years of age. 3 422 423 There are several strengths to be noted, such as, the exposure assessment which included two 424 sources of PM2.5, ambient and traffic-related, which were estimated at a high spatial resolution of up 425 to 200m; our outcome assessment which relied on a recommended assay for AD CSF biomarker 426 measurement, 16 and for which we observed consistent associations using continuous biomarker 427 concentrations as well as AD positivity cut-offs; our inclusion of several well-known confounders and 428 methods to reduce confounding by neighborhood-level characteristics; and our relatively large 429 sample size (N=1,113) of CSF measurements from cognitively healthy older adults free of chronic 430 illness.

432
In addition to its strengths, this study has several limitations. Given that ADRD progresses over the 433 course of several years or decades, we evaluated the associations with 3-and 5-year average PM2.5 434 concentrations prior to enrollment in addition to the 1-year averages. However, given that exposure 435 was assigned based on the baseline residence and some participants could have relocated in the 436 years prior to the study, the 3-and 5-year estimates may be affected by exposure misclassification. 437 Such exposure misclassification is also a potential explanation for the weaker associations between 438 the 3-and 5-year PM2.5 exposures and Aβ42 CSF concentrations in comparison to the 1-year 439 exposure concentrations. Our study also only used cross-sectional CSF measurements; longitudinal 440 repeated measures analyses may provide a better understanding of the long-term effect of air 441 pollution on CSF biomarker trajectories of AD. Further, our sample was not representative of the 442 Atlanta metropolitan area, the target population, as it was mainly high SES and white, which limits 443 both the generalizability and transportability of our estimates. Finally, while we looked at two different 444 sources of PM2.5, we did not examine the relationship between AD pathology and specific 445 components of PM2.5. Future studies should consider the components of PM2.5 as they are dynamic 446 between ambient and traffic-related sources with different toxicity 39 and could reveal important and 447 undiscovered relationships between exposure and disease pathogenesis. 448 449 In conclusion, our results suggest that exposure to ambient and traffic-related PM2.5, even at levels 450 below current primary and secondary standards defined by the Environmental Protection Agency 451 (EPA) for PM2.5 (annual average standards with levels of 12.0 µg/m 3 and 15.0 µg/m 3 , respectively), 452 increases the risk of future AD development. Additionally, our results add to the growing body of 453 evidence which suggests that air pollution directly contributes to neurodegeneration by accelerating 454 Aβ42 accumulation in the brain. 2, 40 455 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Mean (SD) 9.85 (0.832) All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  regression. The overall effect in Figure 4A (N=855) differs slightly from Figure 4B-D (N=1113) due 598 to the decreased sample size after including only participants with APOE genotype data. 599 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.