Practical considerations for characterization of the external exposome in Alzheimer's disease and related dementias cohorts

Abstract

External exposome research offers opportunities to evaluate complex relationships between multiple environmental determinants of Alzheimer's disease and related dementias (ADRD). However, selecting specific exposure domains, identifying metrics, and downloading, processing, and managing big data are key challenges in translating the potential of exposome research into empirical data analyses. To overcome these issues, we have established a reproducible system that protects data privacy, setting the stage for an unparalleled examination of associations between environmental exposures and ADRD risk within an exposome framework. In this perspective, we provide a comprehensive guide, including examples, to linking natural, built, and social environment metrics to nationwide time‐varying participants’ residential address data. Further, as a proof of concept, we report descriptive statistics and spatial distribution of external exposome metrics across the contiguous United States in 2015. The external exposome can be securely and comprehensively characterized across multiple cohorts using this highly scalable approach.

Highlights

• Guidance to establish, estimate, and link external exposome metrics is provided.

• Exposome domains and metrics for ADRD research are identified.

• Data size, computational complexity, and secure linkage are key challenges.

• We describe methods to link 30+ distinct exposome metrics to multiple cohorts.

• Spatial and statistical patterns of the US external exposome were characterized.

1. INTRODUCTION

The number of cases of Alzheimer's disease and related dementias (ADRD) is projected to increase to 152 million worldwide by 2050. ¹ Mid‐life individual risk factors (e.g., smoking, alcohol consumption, hypertension) have been identified as key points for ADRD interventions to reduce disease incidence. ² However, ubiquitous and modifiable environmental exposures have emerged as important potential contributors to ADRD and ADRD risk factors. ³ , ⁴ Environmental exposures may not only determine who is exposed to individual risk factors (i.e., living in neighborhoods with more liquor stores may increase alcohol consumption), but also how severely those risk factors increase ADRD risk. ⁵ , ⁶ Prior evidence examining single environmental factors, such as particulate matter air pollution and lack of green space, suggests adverse associations with ADRD. ² , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ Yet, environmental exposures are experienced in combination, rather than in isolation. It is critical to evaluate the joint impact of multiple domains of environmental exposures, as features of the external exposome, on ADRD.

External exposome (hereafter “exposome”) research is defined by an all‐encompassing characterization of key environmental exposures across the life course. The potential for exposome research applications has been described for a range of outcomes, including cancer risk, ¹² pediatrics, ¹³ and dementia. ¹⁴ These previous perspectives have highlighted the distinctions between exposures defined by individual biomarker measures (e.g., gene expression, protein expression, metabolomics) and large‐scale geospatial data linkage of key environmental exposures (the exposome). For the purposes of this perspective, we focus on fundamental practical considerations to collect, prepare, and analyze large‐scale geospatial data to study exposomic associations with ADRD.

Prior work on individual exposures highlighted several challenges to implementing large‐scale geospatial linkage, including the size of data, computational complexity, and inconsistency in the metrics analyzed. ¹⁵ , ¹⁶ However, to our knowledge, there is no guidance on how to implement large‐scale linkages for exposome research. To conduct exposomic research on ADRD in secure environments approved for the use of individual address and health data, there is a need for new solutions to data and computational complexity. For example, assessing the impact of exposures in early life and at mid‐life is crucial because ADRD risk accumulates throughout the lifespan. However, this emphasis requires a longitudinal approach including data inputs across multiple decades. The need for life course analyses increases data storage requirements, and its analytic complexity raises study design concerns due to the paucity of follow‐up data available consistently available from childhood through older adulthood. ¹⁴ Much of the exposome literature has focused on the promise of understanding the impacts of environmental mixtures on ADRD and challenges of implementation. Key questions raised by this existing work include what input datasets can be used to represent the “totality” of exposures in an exposome, what specific metrics should be used to represent complex environmental constructs, and how to address data size, complexity, and privacy concerns. These questions highlight the fact that the practical matters of metric selection and data linkage represent major barriers to realizing the potential of exposome research.

In this perspective, we present lessons learned from exposome linkage to multiple US‐based longitudinal cohorts. Through this project, our team aimed to link multiple environmental factors to longitudinal residential addresses from cohort participants to support analyses of the impact of the exposome on aging outcomes. We describe key decisions made throughout the process, including identifying candidate exposures within the exposome, defining specific metrics for each exposure, developing a system for geospatial data linkage, and ensuring the linkage can be feasibly carried out in line with the requirements of each cohort's local Institutional Review Board (IRB) requirements for the storage of identifiable participant information. Finally, we present an example illustrating the value of incorporating an exposome approach into a nationwide cohort (using simulations) and visualize the correlation structure and spatial distribution of environmental data across the natural, built, and social environments as a proof of concept. This perspective advances exposome research by providing actionable guidance on cohort integration strategies and solutions to the practical and methodological barriers that have limited exposome epidemiology to date.

2. AN APPLIED EXAMPLE OF EXTERNAL EXPOSOME CHARACTERIZATION

2.1. Identifying candidate exposures within external exposome

Linking every conceivable feature of the environment to individuals for health analyses is unattainable. For each additional environmental exposure considered, there are constraints in data storage, processing time, and complexity in the final analysis. Given these challenges, it is crucial to comprehensively consult the literature to understand existing associations demonstrated between key environmental exposures and outcomes of interest. We used existing systematic reviews to identify key environmental exposures, with the strongest evidence supporting an association with ADRD. Zhao et al. conducted a 2021 review of the modifiable natural, physical, and social environmental factors from observational studies associated with dementia and related outcomes. ¹⁷ Their meta‐analysis identified greenness, temperature, air pollution, noise, road proximity, rurality, neighborhood socioeconomic status (SES), community engagement, and social contact as key exposures associated with ADRD risk. A more recent 2024 review from Zhang et al. highlighted a similar set of environmental exposures associated with ADRD: air pollution, green space, access to goods and services, neighborhood deprivation, and neighborhood amenities (including access to recreation and public spaces). ¹⁸

Based on previous literature findings, the key exposures included in this perspective (Table 1) were air pollution (PM_2.5, NO₂, and O₃), green space, noise, ambient temperature, neighborhood deprivation (including SES), neighborhood amenities (including access to social and physical activities), street connectivity, and rurality. While noise and ambient temperature were not identified as having significant associations in the Zhao et al. meta‐analysis, noise was included as it has been highlighted as a spatially patterned environmental risk factor for ADRD. ¹⁹ We also decided to include measures of ambient temperature given studies identifying associations between short‐term heat and increased mortality, ²⁰ emergency department visits, ²¹ and reduced cognitive test performance. ²² Neighborhood amenities and street connectivity were included as representations of community engagement and social contact.

TABLE 1.

Exposure domain and metric selection for external exposome identification in Alzheimer's disease and related dementia cohorts.

	Methods + considerations	Results
Determination of exposure domains	Identify exposure domains with demonstrated association with outcome of interest. Identify exposure domains associated with relevant precursors of outcome of interest.	Green space	Air pollution	Neighborhood deprivation	Neighborhood amenities
		Noise	Ambient temperature	Street connectivity	Rurality
Selection of specific metrics	Consider tradeoffs between spatial/temporal resolution and data management storage cost. Ensure metrics effectively represent exposure as documented in existing literature for ADRD.	Google Street View Buffers NLCD tree canopy NDVI and EVI (90, 270, 1230 m)	50 m PM2.5 1 km NO2 50 m NO2 (LUR) 1 km O3	Index of Concentration of Extremes Socioeconomic status	Tract Healthy + Unhealthy Food Tract Health Care Tract Social and Physical Activities
		Aircraft Noise Contours Anthropogenic noise levels from National Park Service	Seasonal Mean Temperature and Heat Index Heat Wave Days by Month	Tract‐level walkability Google Street View Sidewalks Distance to bus and train access	Population density

We considered well‐studied mechanisms explaining observed associations with ADRD as an additional criterion for inclusion in our analysis. For greenness, mechanisms include reductions in stress and reduced exposure to other environmental hazards (such as air pollution and heat) and/or increases in physical activity and social connection. ¹⁰ Air pollution may impact ADRD through direct toxicity, neuroinflammation, oxidative stress, and accumulation of proteins. ⁷ Demonstrated pathways for noise impacts include hearing loss, stress, mental health, sleep disruption, and neuropathological damage. ¹⁹ Hypothesized mechanisms for the impacts of heat include increased core body temperature, dehydration, and changes in blood flow. ²³ Neighborhood deprivation, amenities, street connectivity, and rurality are hypothesized to act through a shared mechanism of reduced access to goods and services that support cognitive resilience, as well as through downstream effects of structural racism. ²⁴ , ²⁵ Associations between neighborhood amenities and ADRD may be explained by poorer food quality, food insecurity, and stress driven by limited food access. ²⁵

Given the infeasibility of linking the totality of environmental exposures and the practical restrictions (outlined below) in data storage and the computational cost of linkage, the criteria for inclusion should be clearly defined. Selection of environmental exposures in this study was based on strong evidence of association with ADRD and well‐defined mechanisms based on prior literature.

2.2. Defining specific metrics for each exposure

Environmental exposure assessment approaches have seen rapid advancement in recent years driven by a combination of improved technology (satellite measures, pollution, and noise sensors) and growing use of machine learning and artificial intelligence. ²⁶ , ²⁷ These advances can lead to disconnects between the metrics commonly used in the existing literature and cutting‐edge metrics expected to most comprehensively represent the construct of interest (using the newest sensors, analytic methods). Metric selection, more so than the identification of exposure domains, requires considering a combination of trade‐offs between metric validation, accuracy (spatial resolution, model performance, temporal range), and practical considerations of data size and linkage complexity.

For each of the environmental exposures identified above, we weighed these tradeoffs, and an overview of key decisions and the final suite of metrics is described below (Tables 1 and 2). In the case of air pollution, we used commonly applied, high‐resolution metrics from the literature (i.e., PM_2.5 and NO₂ models commonly used in national cohort studies ²⁸ , ²⁹ ). In studies where PM_2.5 is the sole exposure of interest, exploring new approaches to more precisely characterize PM_2.5 may be warranted. ³⁰ In an exposome analysis with many different environmental exposures, using more commonly applied and well‐understood spatial representations of exposure such as air pollution models from national cohort studies is preferable. Use of more spatially precise exposures is most relevant where the exposure of interest varies substantially at a smaller scale (i.e., greenness), as opposed to more regional exposures (i.e., PM_2.5).

TABLE 2.

Metrics selected for external exposome characterization by buffer size, temporal, and spatial resolution.

Exposure	Metric	Metric subsets	Buffer sizes	Spatial resolution	Time span	Temporal resolution	Single raster	Iterations	Total size
Green space	Tree canopy	National Land Cover Database Tree Canopy	–	30 m	2011 to 2021	Annual	3 GB	Single raster * 11 years	30 GB
	Google Street View	1. Trees 2. Grass 3. Plants 4. Flowers 5. Field 6. Palms 7. Pavement	1. 0 m 2. 500 m 3. 1000 m	100 m	2007 to 2020	Annual	140 MB	Single raster * 7 metrics * 3 buffer sizes * 14 years	41 GB
	NDVI	Normalized Differentiated Vegetation Index	1. 90 m 2. 270 m 3. 1230 m	30 m	1985 to 2024	1. Fall 2. Spring 3. Summer 4. Winter	18 GB	Single raster * 3 buffer sizes * 40 years * 4 temporal resolutions	10 TB
	EVI	Enhanced Vegetation Index	1. 90 m 2. 270 m 3. 1230 m	30 m	1985 to 2024	1. Fall 2. Spring 3. Summer 4. Winter	18 GB	Single raster * 3 buffer sizes * 40 years * 4 temporal resolutions	10 TB
Air pollution	PM2.5	Schwartz US PM2.5	–	50 m	2000 to 2019	Annual	3 GB	Single raster * 17 years	50 GB
	NO2	1. Schwartz US NO2	–	1000 m	2000 to 2016	Annual	30 MB	Single raster * 17 years	0.5 GB
		2. Larkin Global LUR NO2	–	50 m	2005 to 2024	Annual	0.6 GB	Single raster * 20 years	12 GB
	O3	Schwartz US O3	–	1000 m	2000 to 2016	Annual	30 MB	Single raster * 17 years	0.5 GB
Neighborhood deprivation	Socioeconomic variables	1. NSES (neighborhood SES index) 2. Index of Concentration at the Extremes	‐	NA—Tract Variable	1980 to 2020	Decennial	–	Tabular data/Shapefile merge	1 GB
Neighborhood amenities	Neighborhood level Access to services	1. Activity 2. All physicians 3. Unhealthy food sources 4. Unhealthy activities 5. Food stores 6. Social activities 7. All health practitioners 8. Outpatient clinics 9. Pharmacies	‐	NA—Tract Variable	1990 to 2021	Annual	–	Tabular data/Shapefile merge	1 GB
Noise	Anthropogenic noise	Nighttime sound level	–	270 m	2000 to 2014	Annual	30 MB	Single raster * 15 years	0.5 GB
	FAA aircraft noise	1. Nighttime aircraft noise levels 2. Day‐night average sound level	‐	NA—Polygon Contour	1995 to 2015	5‐Year	–	Polygon data	3.5 GB
Ambient temperature	Temperature	1. Average Minimum 2. Average Mean 3. Average Maximum	‐	800 m	1981 to 2023	Monthly	17 MB	Single raster * 12 temporal resolutions * 42 years * 3 metrics	26 GB
	Heat index	1. Average Minimum 2. Average Mean 3. Average Maximum	–	800 m	1981 to 2023	Monthly	17 MB	Single raster * 12 temporal resolutions * 42 years * 3 metrics	26 GB
	Degree days	1. Heating Degree Days 2. Cooling Degree Days	–	800 m	1981 to 2023	Monthly	17 MB	Single raster * 12 temporal resolutions * 42 years * 2 metrics	17 GB
	Heat wave days	1. Days over 10‐year 97.5th percentile 2. Cumulative 3‐day average max temperature degrees over 10‐year 97.5th percentile	–	800 m	1981 to 2023	Monthly	17 MB	Single raster * 12 temporal resolutions * 42 years * 2 metrics	17 GB
Street Connectivity	Walkability	1. Walkability Index	–	NA—Tract Variable	2021	Annual	0.3 GB	Rasterized polygon data	0.3 GB
		2. Google Street View Sidewalks	1. 0 m 2. 500 m 3. 1000 m	100 m	2007 to 2020	Annual	140 MB	Single raster * 3 buffer sizes * 14 years	5.8 GB
	Transit Access	Distance to bus and train access	–	100 m	2020	Annual	0.1 GB	Single raster	0.1 GB
Rurality	Population Density	Population per square kilometer	–	NA—Tract Variable	2020	Decennial	–	Tabular data/Shapefile merge	1 GB

Compared to other environmental exposures, heat presented more challenges in metric selection. Most studies of heat and ADRD assess short‐term exacerbations of cognitive symptoms or health care utilization, with exposure defined by daily time series of temperature or heat index. ²¹ We aimed to use a metric more relevant to long‐term impacts on ADRD incidence. Previous work included temperature variability, ³¹ extreme heat days, ³² or annual average temperature metrics ³³ for characterizing long‐term exposure. Since no metric has been established as a standard, we examined a suite of multiple metrics that would allow for sensitivity analyses of the impact of heat on ADRD as part of the exposome. However, the decision to include several metrics increased the size of data shared to cohorts and the time required to complete linkage. Compared to using a single annual exposure metric, using six distinct metrics at a daily resolution requires about 2000 times as much data storage (Table 2). To this end, we made several modifications, such as reducing the temporal precision from daily to monthly metrics and selecting two measures of heat wave days as opposed to exploring many feasible permutations of heat wave thresholds, durations, and reference periods. The final selected heat wave metrics included discrete number of days per month exceeding the 97.5th percentile temperature, as well as an excess heat factor that encompassed exceedance of thresholds and heat wave intensity. ³⁴ If data size and computational linkage were of no concern, then linking daily temperature data for the full time series would be preferable to create an opportunity to derive any heat‐related metric in future analyses.

2.3. Developing a system for geospatial linkage

The barriers to linkage of large geospatial datasets present a major barrier to exposome research. The complexity of performing exposome characterization across multiple cohorts was largely driven by the ethical considerations and restrictions required when working with personally identifiable residential address data. In working across institutions, our team did not have direct access to any individual residential address data. Of note, this perspective works under the assumption that individual address data have already been documented and geocoded. Reliable and precise address cleaning and geocoding are crucial precursors to environmental data linkage and an ongoing area of research and open‐source software development. ³⁵ All spatial linkages of addresses were conducted by cohort‐specific staff in local IRB‐compliant computing environments. Cohorts involved in this project were not involved in the construction of code for geospatial linkage and had limited experience in developing geospatial processing for address data. To facilitate the linkage process, we developed open‐source R scripts (https://github.com/zpopp/enrich_exposome_adrd) that could be run locally by the health data management teams to protect privacy. The practical barriers to spatial linkage can be reduced by the development of open‐source code, including robust testing and de‐identified quality control assessment as described below.

To ensure our processes were generalizable across cohorts, each script generated by the team was extensively tested using a simulated time series of random address coordinates distributed across the contiguous United States. The simulated address data consisted of 300,000 unique locations to approximate the processing burden of the largest cohort included in the study. Each linkage of environmental exposures included code to generate deidentified quality control outputs that could be referenced to identify potential failures or issues in linkage without revealing sensitive information. For example, histograms, time series plots, and summary statistics for each environmental exposure were created to identify outliers, or values out of range, and saved for comparison with test data. Prior to running linkages with real addresses, the team met with each cohort to ensure assumptions about the processing time and memory capacity of local computing environments held across cohorts. A key challenge in the spatial and temporal linkage of environmental data to residential addresses is linkage for participants that move addresses over the course of follow‐up and to residential history addresses prior to baseline, if available. For the purposes of this study, each environmental exposure metric for the full time span available was linked to each individual address from any time point (e.g., 2000 to 2016 PM_2.5 was linked to a participant's address in both 2005 and 2010). This approach enables investigators to implement exposure assessment methods optimized for their research questions and available resources.

Several alternative approaches could have been used to complete the exposome linkage process across cohorts. One option would be to acquire IRB approval across institutions for individual address data to be shared with our geospatial linkage team. Attempting to receive permission to store address data in a new place or to obtain access to cohort‐level IRB‐compliant systems can prove costly due to the independent institutional procedures in place to protect these data. It is also unlikely to be acceptable to most review boards. Another alternative would be to hire staff at each cohort to conduct the spatial linkage, but this would prove substantially more costly than providing the tools for existing staff to perform linkage. A more efficient and scalable version of the linkage applied for this exposome characterization would be the use of a cloud system where both exposure and address data could be safely stored. This approach would require institutional approval across cohorts of a cloud system for storage of sensitive data, which was not available at the outset of this study.

Identifying a cost‐effective system for cross‐institution, IRB‐compliant linkage of exposome metrics should be a priority for the field moving forward. Several platforms support Health Insurance Portability and Accountability (HIPAA)‐compliant linkage (e.g., Google Cloud Storage, Amazon Web Services), but institutions must independently approve the use of these services, and large‐scale processing can be cost‐prohibitive. The approach applied by our team offers strengths in enabling cohorts to conduct linkages with existing resources but is limited by the redundant efforts required to prepare, download, and store exposure data across cohorts, which could be avoided by the use of a centralized cloud environment.

2.4. Spatiotemporal resolution, data storage, and computational processing

Data storage and processing present a key tradeoff between exposure precision and ease of exposure data management. We took several steps to reduce the storage and computation required. For a given environmental exposure, the key levers that impact data size and processing speed include spatial resolution (i.e., grid size), temporal resolution (day, month, year), time span (number of years), data precision (integers vs floating point values), and the permutations applied for sensitivity analyses for a given exposure metric (threshold for extreme heat definition). Spatial environmental exposures are typically represented as either vector or raster data. ³⁶ Vector data are points, lines, or polygons. Raster data are spatially uniform and distributed on a spatial grid with a resolution defined by the size of each grid cell (e.g., 30 × 30 m). The majority of the metrics used in this project were linked using a raster format. Prior to distributing data to the cohorts, our team conducted several spatial processing steps at the raster grid cell level to ensure the environmental exposure data could be linked efficiently to residential address data (Figure 1). These same processes could be applied for many other exposome factors and reduce the practical barriers to exposome research implementation.

FIGURE 1.

Spatial analysis and data reduction techniques applied for efficient external exposome derivation. Efficient external exposome characterization required the implementation of strategies to derive relevant metrics and reduce the data size prior to linkage. The application of these methodologies varied based on the exposure of interest. (A) Temporal aggregation. For measures of extreme heat, daily time series were assessed against the historical climate normals for an area of interest. A historical record of daily raster data was used to calculate a threshold (i.e., 95th percentile of 10‐year period). For subsequent years, the number of days exceeding this threshold was calculated and then aggregated to the number of days in each month. (B) Buffer generation. For measures of green space, there was interest in exposure at the immediate residence vicinity (90 m), a walkable distance out to the neighborhood (270 m), and community level (1230 m). These buffers were estimated at a 30‐m resolution, representing the mean of all 30‐m pixels at each buffer. (C) Vector‐to‐raster conversion. Changing the form of the spatial data for 50‐m resolution points substantially reduced data size, alleviating the burden on cohorts for linkage. (D) Data size and complexity reduction. Conversion of decimals to integer formats, rounding of exposures to reduce integer size, and compression and tiling were used reduce the size of raster data.

The specific geospatial processing applied varied across exposures. We conducted temporal aggregation for temperature‐related metrics. For every grid cell in the raster data, we computed a threshold of local extreme heat, then cross‐referenced the daily time series of data against this threshold and calculated the sum of “extreme heat days” across months (Figure 1A). This process enhanced our metric of heat to account for exposure in exceedance of local acclimatization and reduced the amount of data required for linkage from 365 days of data each year to 12 monthly sums. For greenness measures, we computed multiple buffer (radii of a certain distance around a point or grid cell) distances for exposure at each grid cell, increasing the amount of data linked (Figure 1B). Following previous literature, distinct buffers help to understand greenness exposure at the immediate residence vicinity (90‐m buffer), a walkable distance out to the neighborhood (270‐m buffer), and at the community level (1230‐m buffer). The increase in data size and computation was determined to be acceptable in this case given previous work on this and the unique context of exposure at the different buffer distances.

In addition to these approaches modifying our representations of exposure, we made several technical adjustments to reduce the size of data for linkage, which can be used to make exposome characterization more feasible. For PM_2.5, the raw source data were distributed as a grid of points. The data were transformed from this point (vector) format to a raster, as the raster data could be more efficiently stored and linked to addresses (Figure 1C). Across exposures, the raster data underwent cleaning to ensure the variable precision (number of decimal points) and type (integers [INT] reduce data size over floating‐point [FLT] types) were optimized (Figure 1D). In the case of the Normalized Differentiated Vegetation Index, we converted from decimal (FLT) measures scaled from −1 to +1 to scaled integers from −10,000 to 10,000. Compression was also optimized for reducing storage and computation while maintaining accuracy. These steps were all conducted prior to linkage.

The process of spatial linkage itself largely involved extraction of raster data at the point level (Figure 2A). This process assigns each spatial point to the grid cell in which it lies. For polygon data, spatial intersection was used to link points to the geography they lay within (Figure 2B). The R terra and sf packages were used for linkage across environmental exposures. ³⁷ , ³⁸ , ³⁹ The R amadeus package offers many of the same processing capacities described here in a format dedicated to reducing the learning curve for environmental health linkages. ⁴⁰ The timeline of this project and the specific datasets required did not allow for the application of amadeus, but the ongoing open‐source development of this R tool for the download, processing, and covariate calculation of large‐scale environmental data aims to enhance the reproducibility of data acquisition and cleaning moving forward.

FIGURE 2.

Spatial analysis and intersection techniques applied for efficient external exposome linkage. Efficient external exposome characterization required the implementation of strategies to link multiple forms of exposure data to residential address data. We provide example applications with random spatial subsets of exposure data and randomly distributed points for demonstration. After reducing data size and deriving all metrics of interest, linkage to address points was conducted by overlaying raster (A) or vector (B) data with the point locations. (C) Spatial intersection and tabular merge. For exposures at an aggregate administrative boundary, an intermediate step was required to intersect address point locations with a vector shapefile representing the administrative boundaries, after which a tabular merge was used to link exposures to address data.

Spatial pre‐processing steps such as those listed here result in some reduction in the resources and time required for linkage. For some environmental exposure, the value of enhanced precision in time (e.g., monthly heat measures) or space (e.g., multiple buffers of high‐resolution greenness) outweighs the burdens of cohort‐level storage and linkage. For raster data, the environmental exposure precision is best defined by the spatial resolution of grid cells. Enhancing spatial resolution corresponds to exponential growth in data storage needs. Increasing spatial resolution from 1 km to 50 m represents an increase of 400 times in the data size and subsequently increases download and processing time substantially. Available resolutions vary substantially by exposure and metric of interest, and the relevance of high spatial resolution depends on the exposure, health endpoint, and mechanism of interest. ⁴¹ Community‐level measures like neighborhood SES or residential segregation are best measured across an aggregate spatial resolution. Levels of greenspace exposure, NO₂, or temperature have been shown to vary substantially even within neighborhoods, making prioritization of spatial resolution pertinent. The same is true of temporal resolution. Annual average air pollution metrics are standard practice for studies of long‐term outcomes. For temperature metrics, the lack of available standardized metrics and the relevance of extreme heat exposures outside the bounds of typical acclimatization highlight the need for use of data at higher temporal resolution (such as month). For outcomes like cognitive symptom exacerbation or hospitalizations, shorter‐term exposures like daily or weekly temperature may be more relevant.

One approach to substantially reducing the complexity of exposome linkage would be to pre‐aggregate all data to some geographic unit, like a census tract, and then conduct a series of tabular merges. Aggregation of raster data to a geographic unit can be time‐consuming but could be done in a cloud computing environment to generate nationwide aggregate data for sharing with any cohort. This approach offers the advantage of all metrics being assessed at a consistent geographic unit, as opposed to a combination of tract aggregates, point‐based measures, and raster data at several different spatial resolutions. Avoiding the mix of raster and vector data eases processing and can improve the interpretability of findings given the uniform spatial scale across exposures. We used a tract‐aggregation approach for a handful of exposures (amenities, neighborhood access measures, SES) as these data are only estimated using tract‐level inputs and represent community exposures. With tract data, the linkage could be completed with a single spatial intersection between the address coordinate points and tract polygon data (Figure 2C). Any number of metrics can then be linked in a single step using a unique geographic identifier consistent between the polygon dataset and the exposure data. The tradeoff with this approach is the bias introduced in aggregation from gridded data to a geographic unit. The modifiable area unit problem can bias estimates, as the varying size and spatial composition of census tracts impacts exposure metrics. ⁴² Further, a key advantage of exposome research is to assess the mixture of exposures at the residential level alongside the community level. Several key exposures that contribute to ADRD vary substantially even across small distances, and tract aggregation removes this exposure precision.

2.5. Exploring national correlations between external exposome components

The goal of this paper is to highlight approaches to overcoming practical barriers in identifying and linking environmental factors for health analyses. After the linkage process, many challenges remain in terms of how to best prepare the data for statistical analysis. ¹⁵ , ¹⁶ We do not fully explore the environmental epidemiology of ADRD risk factors as a part of this paper; however, we use a 300,000‐point simulation dataset to evaluate the spatial and statistical distribution of our exposome metrics across the contiguous United States (CONUS) in 2015 as a proof of concept (Figure 3A). The simulated dataset was created by randomly placing points within ZIP Code tabulation areas (ZCTA) across CONUS based on the ZCTA population.

FIGURE 3.

Spatial distribution of select metrics for 2015 external exposome characterization. External exposome metrics for 2015 were extracted at 300,000 points distributed randomly across United States ZIP Code tabulation areas. The decile sum across all external exposome components was estimated and visualized across the contiguous United States (A) and within four large metropolitan areas – Los Angeles, California; New York, New York; Dallas–Fort Worth, Texas; and Chicago, Illinois (B).

To evaluate the distribution of our exposome spatially, we derived a decile sum across metrics. Unweighted decile sums are insufficient for studying the exposome in health research contexts. Our exposome includes multiple metrics of air pollution, greenness, amenities, and temperature. Deriving a sum across metrics without addressing the different number of factors included across these domains leads to bias in aggregation. We selected an unweighted decile sum as a simple data reduction technique, as we did not aim to assess a specific research question. Zhu et al. provided overviews of several methods that can be applied for data reduction prior to health analyses, such as principal component analysis, factor analysis, and k‐means cluster analysis. ⁴³ These methods can be used to enhance the interpretability of analyses and to address challenges in dimensionality and multicollinearity that are present in environmental mixtures.

For this demonstration, the unweighted decile sum was calculated by adding together all metrics, after re‐estimating each metric to a decile with respect to the full 300,000‐point simulated cohort. For the suite of 34 metrics, a decile sum of 340 would represent a point in the highest decile of exposure (most hazardous) across all metrics. A score of 34 would represent a point in the lowest decile (least hazardous) of exposure across metrics. No point in our dataset was in the lowest or highest decile for all exposures, with the actual distribution of the decile sum ranging from 100 to 280. For metrics that represent hypothesized protective relationships (i.e., green space, access to resources), the decile was calculated as the inverse of the exposure where the 10th decile represents the lowest amount of greenspace exposure (most hazardous) or lowest access to resources. The spatial distribution of the decile sum indicates that the areas with the highest exposure (hazard) appear to surround urban areas, which could be expected given that several of our exposome metrics are likely to vary across urban and rural contexts (i.e., noise, access to green space). Some variance can be seen across domains of exposure (temperature metrics are higher in the southern United States as opposed to northern states). While heterogeneity appears to be driven by urbanicity at the national scale, focusing on specific urban hubs highlights that heterogeneity in the exposome decile sum is also present at local scales (Figure 3B).

One key analytic consideration for exposome research is the correlation between components. High correlation between features can require data reduction techniques or the use of mixture methods to effectively handle these inputs. To provide an overview of the relationship between exposome components, we present a correlation plot with select metrics across the domains of heat, greenness, air pollution, noise, access to amenities, and neighborhood SES (Figure 4A). Clusters of high correlations appear within exposure domains and across metrics – for instance, Street View‐based metrics of green space (grass, pavement, plants, and trees) show mild to strong positive correlations with one another. NO₂ and PM_2.5 metrics show strong correlations with one another and a negative correlation with O₃. Anthropogenic noise demonstrates positive correlations with the Street View‐based metrics of the built environment and with PM_2.5 and NO₂. The income and race indices of concentration at the extremes (ICE), which represent neighborhood residential segregation, show strong correlations with one another and with neighborhood socioeconomic status but weak correlations with most other exposome components. Measures of heat (summer heat index, temperature, heat waves) are strongly positively correlated and strongly negatively correlated with measures of cold exposure (heating degree days). When aggregating across these metrics to our unweighted decile sum, we find a symmetric, unimodal distribution (Figure 4B). These results serve as a demonstration of the variability in a sample of pooled exposure domains but will vary substantially based on the specific metrics considered and based on the weighting scheme applied to exposome summarization.

FIGURE 4.

Correlation and histogram of select metrics for 2015 external exposome characterization. External exposome metrics for 2015 were extracted at 300,000 points distributed randomly across United States ZIP Code tabulation areas. The correlation across all external exposome components was visualized to identify high (red) and low correlation (blue) patterns between metrics (A). The overall distribution of an unweighted decile sum was also visualized (B).

The high correlations detected here underscore the need for applying analytic frameworks suitable to the unique dimensionality and collinearity of environmental mixtures in health analyses. One approach to analyzing these data would be to apply dimension reduction methods like factor analyses, which reduces the many exposure factors down to uncorrelated aggregate variables. ⁴³ , ⁴⁴ In addition, several statistical models can incorporate highly correlated inputs. Methods like weighted quantile sum regression, Bayesian kernel machine regression (BKMR), and Bayesian hierarchical models can be used to analyze correlated data and assess both summed exposome effects and individual exposure effect contributions. ⁴⁵ Joubert et al. provide an overview of the statistical workflow for analyzing mixtures. ⁴⁴ They document how model selection varies based on the number of available time points for exposures and outcomes, the sample size, and the presence of interactions and mediation identified in analyses. Careful consideration of these study characteristics are crucial for the effective epidemiologic analysis of the exposome.

2.6. Limitations

In this perspective, we have outlined key challenges and the decisions we made to characterize the exposome for use in ADRD research. Our approach does not account for key limitations in defining the exposome for ADRD. Our exposome measures consist of a subset of the true totality of exposures experienced by an individual across the life course that might influence ADRD risk. We base our exposure selection on the existing literature linking environmental factors with ADRD. Our approach excluded several of the factors included in previous reviews (i.e., pesticides, solvents) due to the lack of reliable spatial representations of exposure, and our reliance on existing reviews precludes the inclusion of novel exposure metrics that have emerged in the growing environmental health and ADRD literature. We expect that the current set of exposures and metrics described in this manuscript will change over time, in line with advances in the understanding of environmental influences on ADRD. Disease incidence is impacted by factors experienced across the life course and across generations. The ability to characterize exposure across the entirety of the life course is restricted both by the availability of address data and the consistency of environmental exposure data into the historical past. Historical exposure data are especially limited for novel metrics (i.e., high resolution air pollution, noise data), many of which are only available for the 21st century. We prioritize datasets available nationwide for our exposure metrics, but many of these measures cannot be directly reproduced on a global scale. Global standardization of metrics will be relevant to ensuring the generalizability of exposome research beyond the United States. Our approach aims to minimize the burden of linkage for cohorts; however, this process included the download and storage of multiple 100 GB data subsets for linkage. This time and effort may not be available in all contexts.

3. CONCLUSION

Environmental exposure metrics have demonstrated both hazardous (air pollution, extreme heat) and protective (green space, neighborhood amenities) effects on the incidence of ADRD. The environment is experienced as a mixture of these interlinked factors. Epidemiologic analyses of how environmental factors impact ADRD should account for the reality that no exposure is experienced in isolation. We have described the process of developing a representation of the external exposome for linkage to US‐based longitudinal cohorts. Challenges are present across the pipeline of exposome characterization from exposure definition to metric selection and ethical storage and linkage of environmental data to protected residential address data. Standardization of these steps across cohorts would be very valuable for the reproducibility and generalizability of study results.

The process of describing environmental exposure identification, metric selection, and geospatial data linkage can support ADRD cohorts in further integration of the exposome into epidemiologic analyses. Given the ubiquity of environmental exposures and the complex mixture in which they are experienced, scalable exposome characterization will be crucial to advancing ADRD prevention at the population level.

CONFLICTS OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

Supporting information

Supporting Information