Linking Ambient NO2 Pollution Measures with Electronic Health Record Data to Study Asthma Exacerbations

Abstract

Electronic health record (EHR)-derived data can be linked to geospatially distributed socioeconomic and environmental factors to conduct large-scale epidemiologic studies. Ambient NO2 is a known environmental risk factor for asthma. However, health exposure studies often rely on data from geographically sparse regulatory monitors that may not reflect true individual exposure. We contrasted use of interpolated NO2 regulatory monitor data with raw satellite measurements and satellite-derived ground estimates, building on previous work which has computed improved exposure estimates from remotely sensed data. Raw satellite and satellite-derived ground measurements captured spatial variation missed by interpolated ground monitor measurements. Multivariable analyses comparing these three NO2 measurement approaches (interpolated monitor, raw satellite, and satellite-derived) revealed a positive relationship between exposure and asthma exacerbations for both satellite measurements. Exposure-outcome relationships using the interpolated monitor NO2 were inconsistent with known relationships to asthma, suggesting that interpolated monitor data might yield misleading results in small region studies.

Introduction

Electronic health records (EHRs) have become invaluable to biomedical research since their use in clinical settings became nearly ubiquitous in the last decade¹, due in part to their ability to provide low-cost, rich data for large patient populations at a scale that is inaccessible for traditional epidemiologic studies². However, although EHRs capture clinical and demographic information on large and diverse patient populations, they lack important socioeconomic and environmental data that influences health. We and others have shown how EHR data can be useful in environmental and social epidemiology studies when its scope is expanded with external data on the physical, built, and social environment via linkage with patient addresses or other geographic information^3–8. Air pollution, the largest environmental contributor to all-cause mortality that is estimated to account for one out of every nine deaths worldwide⁹, is a prominent risk factor whose integration with EHR data can improve the conduct of research studies related to many health outcomes.

Health studies that include pollution data often rely on measures of six criteria pollutants obtained by the United States Environmental Protection Agency (EPA) at thousands of fixed, quality-controlled, regulatory monitors distributed throughout the country. Because these monitors are geographically sparse relative to the spatial variation desired for most local studies, subjects are matched to measures taken by the nearest monitor or assigned a value computed with a spatial smoothing method such as inverse-distance weighting interpolation (IDW), techniques that imprecisely reflect true exposure and introduce measurement error^10–12. To overcome barriers related to sparse geographic coverage of regulatory monitors in a region, alternatives such as the use of low-cost, portable pollution sensors¹³ and measures captured by instruments deployed with satellites^14,15 have been pursued. Although using wearable sensors makes possible the fine tracking of exposure throughout an individual’s day, this approach is difficult to implement at the scale needed for most EHR-based studies.

Nitrogen dioxide (NO₂) is a gaseous toxicant that has been associated with respiratory disease, cardiovascular disease, and lung cancer¹⁶. Because it is a component of traffic-related air pollution (TRAP) with a short atmospheric lifetime, NO₂ exhibits high intra-urban variability and is suitable for assessing fine-scale health risk. EPA regulatory monitors collect ground data on this criteria pollutant, offering valuable information on its changes over time and permitting comparison of its levels across broad geographic areas. The TROPOspheric Monitoring Instrument (TROPOMI) aboard the Sentinel-5 Precursor satellite has provided an alternative source of NO₂ measures with unprecedented spatial resolution since its launch in 2017¹⁷. Whereas EPA in situ NO₂ measurements are collected in parts per billion (ppb), TROPOMI measures NO₂ vertical column density (VCD), that is, the number of pollutant particles per unit area between Earth’s surface and the tropopause. Prior work has shown that TROPOMI NO₂ VCD strongly correlates with EPA ground measurements^18,19 and that this correlation can be improved further by using chemical transport models to estimate satellite-derived ground levels of NO₂^20,21.

Asthma, a chronic disease characterized by inflammation and narrowing of the airways that affects 21 million adults in the United States²², is known to worsen with NO₂ exposure²³. Asthma management focuses on controlling symptoms and preventing exacerbations, episodes of worsening symptoms that require treatment with systemic steroids, via avoidance of patient-specific triggers and treatment with drugs²⁴. Due to a complex combination of biological and social factors²⁵, however, exacerbations remain common in some patients and are major contributors to asthma-related morbidity and mortality^26,27. Efforts to reduce the impact of exacerbations using EHR-derived data include creating predictive models of asthma exacerbations^28,29, deriving a computable asthma severity phenotype³⁰, assessing the geospatial distribution of asthma risk factors^6–8, and better understanding disease associations in real life populations^31–33. In addition to the contribution by pollution, exacerbation risk differs by race and ethnicity and has been linked to indicators of socioeconomic disadvantage, including low income³⁴, housing instability³⁵, and high crime rates³⁶. Here, we sought to determine whether EPA monitor, raw satellite, or satellite-derived NO₂ pollution measures were most suitable for the study of asthma exacerbation risk in Philadelphia.

Methods

Study population.

De-identified EHR data corresponding to Penn Medicine patient encounters from January 1, 2015 to April 8, 2021 was obtained from Penn Data Store (PDS), a clinical data warehouse, and utilized to identify subjects for a retrospective cohort analysis. Patient-level information included age at first encounter, sex, race and ethnicity, health insurance type, and geocodes corresponding to most recent patient address. Encounter-level information included dates, respiratory medication prescriptions, height, weight, smoking status, and all ICD-10 codes associated with the encounter. A years followed variable for each patient was calculated by subtracting the date of the first encounter from the last. Patient age was grouped into four categories: 18-34, 35-54, 55-74, and 75+ years. Health insurance type was grouped as private, Medicare, and Medicaid, and patients with multiple health insurance types listed were grouped as their highest frequency type. Body mass index (BMI) was computed for each encounter and categorized as: not overweight or obese (<25.0 kg/m²), overweight (25.0 to <30.0 kg/m²), grade 1 obese (30.0 to <35.0 kg/m²), grade 2 obese (35.0 to <40.0 kg/m²), and grade 3 obese and above (≥40.0 kg/m²). Height and weight entries outside of the four to seven foot and 80 to 700 pound range, respectively, were excluded to reduce bias from entry errors. Per-patient BMI was calculated as the mean of all available values. Smoking status was assigned based on frequency and recency of encounter-level responses, and was grouped into current, never, and history. Passive and past smoking were combined under history because of the small size of the passive level. Age, sex, race and ethnicity, BMI, health insurance type, and smoking status are hereinafter referred to as “EHR-derived variables.”

Inclusion criteria for the study were: (1) at least 18 years old at first encounter, (2) at least one instance of an asthma ICD-10 code (i.e., J45) in any encounter type, (3) at least one short-acting beta2-agonist (SABA) prescription, in accordance with United States national asthma management guidelines recommending short-acting beta2-agonist (SABA) treatment for managing both intermittent and persistent asthma in individuals 12 and over^37,38, (4) at least one year followed between January 2017 and December 2020, and (5) residence within the city of Philadelphia to maximize the likelihood that Penn Medicine was the main point of care for asthma. Exclusion criteria were: (1) presence of ICD-10 codes for chronic obstructive pulmonary disease (i.e., J41, J42, J43, J44) or cystic fibrosis (i.e., E84), and (2) patients with incomplete covariates (i.e., missing BMI, insurance type, smoking status, geocode).

Asthma exacerbations.

We defined our study period as January 1, 2017 to March 17, 2020, a time period that ends on the date when COVID-19 restrictions were put in place in Philadelphia, to minimize the effects of the pandemic on healthcare utilization. We identified asthma exacerbations occurring during this period as encounters that had (1) a primary asthma ICD-10 code (i.e., J45), and (2) at least one oral corticosteroid (OCS) prescription, criteria derived from national asthma management guidelines recommending OCS as part of the clinical course for mild, moderate, and severe exacerbations^37,38. For encounters without primary asthma ICD-10 codes, nonprimary codes were used to determine exacerbations. Total exacerbations per patient over this study period were categorized as: 0, 1, 2-3, and 4+.

Area Deprivation Index (ADI).

Socioeconomic disadvantage for each patient was summarized as the ADI, a single validated score that is based on American Community Survey data on education, income, housing, poverty, and employment to measure neighborhood disadvantage³⁹. The 2018 ADI values for Pennsylvania were extracted from the Neighborhood Atlas on the census block level and converted to a raster^40,41. ADI was assigned to each patient using bilinear interpolation from the values of the four nearest cells to their geocode using the R package terra⁴² and then grouped into four categories: 0-25, 25-50, 50-75, and 75-100.

Ambient NO₂ Measures.

Three rasters of NO₂ measures were obtained:

1. Interpolated EPA NO₂ This measure was derived from 2017-2019 averaged AirData annual concentrations by regulatory monitors in the southeastern Pennsylvania region⁴³, including four NO₂ monitors in Philadelphia⁴⁴. Inverse distance weighting (IDW) was used to create a grid of interpolated values for the entire region at ~385m² resolution. The grid was then rasterized and clipped to the Philadelphia bounding box region.

2. Raw satellite NO₂ This measure was derived from TROPOMI satellite observations on Google Earth Engine⁴⁵, a cloud-based platform designed to make remote sensing analysis widely accessible. Observations for the Philadelphia bounding box region were downloaded using the rgee package⁴⁶ with a resolution of ~1km². Only 2019-averaged data was downloaded as 2019 is the first year for which Google Earth Engine had complete observations. Pixels with quality assurance values less than 0.75 were excluded.

3. Satellite-derived ground NO₂ This measure was downloaded from a resource reported in Cooper et al. and averaged over 2017-2019⁴⁷. The NO₂ column densities were based on both TROPOMI and the Ozone Monitoring Instrument, a satellite with more years of historical data but lower resolution²¹. Column densities were then converted to ground estimates using GEOS-Chem, a chemical transport model. The downloaded rasters had a resolution of ~1km².

These measures were assigned to each patient using bilinear interpolation from the values of the four cells nearest to their geocode using the R package terra⁴².

Comparison of NO₂ Measures.

Monthly 2019 raw satellite NO₂ measures were compared to monthly 2019 interpolated EPA data at the geocoordinates of three EPA regulatory monitors (i.e., Camden Spruce Street, Car-Barn Montgomery I-76, and Torresdale Station). A linear regression model was fitted for raw satellite NO₂ versus interpolated EPA NO₂, the Pearson correlation coefficient r was estimated, and points were joined with a linear smooth. Scatter plots of raw and interpolated EPA NO₂ over 2019 at the coordinates of each regulatory monitor were smoothed using local polynomial regression.

Statistical analyses.

Chi-squared tests were used in bivariate analysis to assess relationships between patient characteristics (i.e., EHR-derived covariates and ADI) and asthma exacerbation level during the study period. Proportional odds logistic regression models were used in multivariable analysis to compute adjusted odds of exacerbation. All multivariable models included EHR-derived covariates and ADI as independent predictors. One NO₂-unadjusted model contained no additional predictors. Three NO₂-adjusted models included one of the three NO₂ measurements of interest as an additional predictor. Ethnicity was not included in multivariable analysis since only 5.7% of patients identified as Hispanic. Stratified binomial models were used to test the proportional odds assumption that the relationship between all outcome levels is similar for each predictor. Statistical analyses were conducted using the R MASS package⁴⁸.

Geospatial analyses of asthma exacerbations.

Hotspots for risk of asthma exacerbation were identified using previously described methods and the MapGAM R package^6,49. Briefly, generalized additive models (GAMs) were used to estimate local odds of exacerbation, which was releveled into a binary outcome variable, where controls were patients with no exacerbations and cases were patients with one or more exacerbations. The log odds of this outcome were estimated as a function of location while simultaneously adjusting for covariates. EHR-derived variables (excluding ethnicity), and ADI were included as predictors in four GAMs that differed as follows: one did not include additional variables, and three were adjusted using one NO₂ measurement type each. A global test of the null hypothesis that odds of exacerbation were not spatially correlated was performed by permuting the assignment of cases and controls over patient geocodes 1000 times. Log odds were converted to local odds ratios (ORs) using the odds of exacerbation for the whole cohort as a reference. A distribution of log odds at each point across all permutations was used to define “hotspots” and “coldspots”: points ranking in the upper 2.5% were termed “hotspots”, and the lower 2.5% were “coldspots.”

Results

Subject characteristics.

Our study population consisted of 16,744 people with asthma, 13,924 with no exacerbations during the study period and 2,820 with at least one exacerbation. Age, race, ethnicity, BMI, health insurance type, and ADI were associated with exacerbations according to bivariate analyses (p < 0.001) (Table 1). These relationships were consistent with known asthma disparities⁵⁰: 66% of patients with no exacerbations were Black compared to 81% of patients with 4+ exacerbations, and 40% of patients with no exacerbations had Medicaid insurance compared to 53% of patients with 4+ exacerbations.

Table 1.

Patient characteristics by exacerbation count levels.

	Number of Exacerbations
Characteristic	Overall, N = 16,7441	0, N = 13,9241	1, N = 1,9181	2-3, N = 6391	4+, N = 2631	p-value2
Patient Age						<0.001
18-34	6,727 (40%)	5,739 (41%)	700 (36%)	195 (31%)	93 (35%)
35-54	5,871 (35%)	4,752 (34%)	744 (39%)	275 (43%)	100 (38%)
55-74	3,526 (21%)	2,928 (21%)	402 (21%)	132 (21%)	64 (24%)
75+ 620	(3.7%)	505 (3.6%)	72 (3.8%)	37 (5.8%)	6 (2.3%)
Sex						0.001
Male	3,893 (23%)	3,264 (23%)	412 (21%)	134 (21%)	83 (32%)
Female	12,851 (77%)	10,660 (77%)	1,506 (79%)	505 (79%)	180 (68%)
Race						<0.001
White	4,558 (27%)	3,977 (29%)	422 (22%)	117 (18%)	42 (16%)
API	450 (2.7%)	386 (2.8%)	43 (2.2%)	18 (2.8%)	3 (1.1%)
Black	11,131 (66%)	9,042 (65%)	1,395 (73%)	480 (75%)	214 (81%)
Other/Unknown	605 (3.6%)	519 (3.7%)	58 (3.0%)	24 (3.8%)	4 (1.5%)
Ethnicity						<0.001
Non-Hispanic	15,793 (94%)	13,097 (94%)	1,818 (95%)	624 (98%)	254 (97%)
Hispanic	951 (5.7%)	827 (5.9%)	100 (5.2%)	15 (2.3%)	9 (3.4%)
BMI (kg/m2)						<0.001
Not overweight or obese	3,526 (21%)	3,034 (22%)	360 (19%)	88 (14%)	44 (17%)
Overweight	4,232 (25%)	3,538 (25%)	479 (25%)	149 (23%)	66 (25%)
Class 1 obese	3,529 (21%)	2,929 (21%)	398 (21%)	147 (23%)	55 (21%)
Class 2 obese	2,488 (15%)	2,051 (15%)	301 (16%)	100 (16%)	36 (14%)
Class 3 obese	2,969 (18%)	2,372 (17%)	380 (20%)	155 (24%)	62 (24%)
Health Insurance Type						<0.001
Private	7,941 (47%)	6,695 (48%)	858 (45%)	291 (46%)	97 (37%)
Medicaid	6,645 (40%)	5,443 (39%)	807 (42%)	255 (40%)	140 (53%)
Medicare	2,158 (13%)	1,786 (13%)	253 (13%)	93 (15%)	26 (9.9%)
Smoking History						0.004
Never	10,211 (61%)	8,539 (61%)	1,161 (61%)	372 (58%)	139 (53%)
Current	2,294 (14%)	1,862 (13%)	299 (16%)	96 (15%)	37 (14%)
History	4,239 (25%)	3,523 (25%)	458 (24%)	171 (27%)	87 (33%)
ADI						<0.001
1-24	2,437 (15%)	2,100 (15%)	244 (13%)	70 (11%)	23 (8.7%)
25-49	2,820 (17%)	2,415 (17%)	284 (15%)	84 (13%)	37 (14%)
50-74	3,207 (19%)	2,650 (19%)	361 (19%)	137 (21%)	59 (22%)
75-100	8,280 (49%)	6,759 (49%)	1,029 (54%)	348 (54%)	144 (55%)

¹n (%), ²Corresponding to Pearson’s Chi-squared test

Comparison of ambient NO₂ measures.

At the location of EPA regulatory monitors, EPA ground monitor measurements and raw satellite NO₂ exhibited high correlation (r = 0.77) (Figure 1A). Descriptive spatio-temporal plots comparing EPA ground monitor measurements and raw satellite NO₂ at three Philadelphia monitor locations indicated that raw TROPOMI observations were sensitive to fluctuations in ground-level NO₂ and that both exhibited similar annual variability, validating raw satellite measurements for use in further analysis to represent exposures at the ground level (Figure 1B-D). However, the correlation between interpolated EPA and raw satellite measurements decayed as the distance from a monitor site increased. Interpolated EPA NO₂ peaked in Northeast Philadelphia (Figure 2A), whereas raw satellite NO₂ concentrations were highest in parts of West, South, and Southeast Philadelphia (Figure 2B). Satellite-derived NO₂ concentrations closely matched raw satellite measurements but displayed more overall heterogeneity (Figure 2C).

Figure 1.

Comparison of raw satellite and regulatory monitor NO₂ measures. 2019 monthly averaged TROPOMI observations were compared to 2019 monthly averaged EPA measurements at three Philadelphia regulatory monitors. TROPOMI monthly averages were extracted from the pixel in which each monitor was located. A) Scatter plot for all three Philadelphia monitors combined with a linear smooth (r = 0.77). Descriptive time plots for monitors at B) Camden Spruce Street, C) Car-Barn Montgomery I-76, and D) Torresdale Station.

Figure 2.

The geospatial distribution of the three NO₂ measurement types considered. A) 2017-2019 EPA AirData computed using IDW interpolation at ~385m² resolution. B) 2019 raw TROPOMI Google Earth Engine data at~1km² resolution. C) 2017-2019 ground-level data derived from TROPOMI and OMI satellites at ~1km² resolution, as presented in Cooper et al (2021).

Factors associated with asthma exacerbations.

In each multivariable model, all age categories (p<0.001), Black race (p<0.001), and class 3 obesity (p<0.001) were significantly associated with being in a higher exacerbation category (Table 2). Sex, health insurance type, smoking history, and ADI were not significant in any model. Addition of each of the NO₂ measurements to the model resulted in a significant association with exacerbations, but raw satellite and satellite-derived NO₂ conferred an increased risk (OR = 1.08, 95% CI [1.02, 1.15] and OR = 1.14, 95% CI [1.05, 1.24], respectively) while interpolated EPA NO₂ was associated with decreased risk (OR = 0.94, 95% CI [0.91, 0.97]).

Table 2.

Factors associated with asthma exacerbations. ORs were computed from proportional odds models with exacerbation counts as the outcome. All models included EHR-derived covariates and ADI as independent variables. NO₂-adjusted models additionally included EPA interpolated, raw satellite, or satellite-derived NO₂ measures. Adjusted ORs and 95% confidence intervals (CIs) are shown. ***p<0.001, **p<0.01, *p<0.05

Characteristic	NO2-Unadjusted Model	Model with EPA-Interpolated NO2	Model with Raw Satellite NO2	Model with Satellite-Derived NO2
Patient Age
18-34	Reference	Reference Reference	Reference
35-54	1.36 (1.24, 1.5)***	1.36 (1.24, 1.5)***	1.37 (1.24, 1.51)***	1.37 (1.24, 1.51)***
55-74	1.24 (1.1, 1.4)***	1.23 (1.09, 1.39)***	1.25 (1.11, 1.42)***	1.25 (1.11, 1.42)***
75+	1.58 (1.22, 2.03)***	1.54 (1.2, 1.99)***	1.59 (1.23, 2.05)***	1.59 (1.24, 2.05)***
Sex
Male	Reference	Reference Reference	Reference
Female	0.99 (0.9, 1.1)	1 (0.9, 1.1)	1 (0.9, 1.1)	1 (0.9, 1.1)
Race
White	Reference	Reference Reference	Reference
API	1.18 (0.89, 1.56)	1.15 (0.87, 1.52)	1.17 (0.89, 1.55)	1.17 (0.88, 1.55)
Black	1.47 (1.3, 1.65)***	1.42 (1.25, 1.6)***	1.5 (1.33, 1.7)***	1.52 (1.34, 1.71)***
Other/Unknown	1.12 (0.87, 1.43)	1.11 (0.87, 1.42)	1.15 (0.9, 1.47)	1.15 (0.9, 1.48)
BMI (kg/m2)
Not Overweight or Obese	Reference	Reference	Reference	Reference
Overweight	1.14 (1, 1.29)*	1.14 (1.01, 1.3)*	1.14 (1.01, 1.3)*	1.14 (1.01, 1.3)*
Class 1 Obese	1.12 (0.98, 1.28)	1.13 (0.99, 1.3)	1.13 (0.99, 1.29)	1.13 (0.99, 1.29)
Class 2 obese	1.13 (0.98, 1.31)	1.14 (0.98, 1.32)	1.13 (0.98, 1.31)	1.14 (0.98, 1.31)
Class 3 Obese	1.32 (1.15, 1.52)***	1.33 (1.16, 1.53)***	1.32 (1.15, 1.52)***	1.32 (1.15, 1.52)***
Health Insurance Type
Private	Reference	Reference	Reference	Reference
Medicaid	1.05 (0.95, 1.15)	1.05 (0.95, 1.16)	1.03 (0.94, 1.14)	1.03 (0.93, 1.14)
Medicare	0.92 (0.79, 1.07)	0.92 (0.79, 1.08)	0.91 (0.78, 1.06)	0.91 (0.78, 1.06)
Smoking History
Never	Reference	Reference	Reference	Reference
Current	1.07 (0.95, 1.21)	1.06 (0.93, 1.2)	1.06 (0.94, 1.2)	1.06 (0.94, 1.2)
History	1 (0.9, 1.1)	0.99 (0.9, 1.1)	1 (0.9, 1.1)	0.99 (0.9, 1.1)
ADI
1-24	Reference	Reference	Reference	Reference
25-49	0.95 (0.81, 1.12)	0.97 (0.83, 1.14)	1 (0.85, 1.18)	1.02 (0.86, 1.2)
50-74	1.05 (0.9, 1.24)	1.11 (0.94, 1.3)	1.12 (0.95, 1.32)	1.15 (0.97, 1.36)
75-100	1.04 (0.89, 1.21)	1.08 (0.93, 1.26)	1.09 (0.93, 1.28)	1.13 (0.96, 1.33)
Interpolated EPA NO2		0.94 (0.91, 0.97)***
Raw Satellite NO2			1.08 (1.02, 1.15)*
Satellite-Derived NO2				1.14 (1.05, 1.24)**

Geospatial distribution of asthma exacerbation risk.

GAM analysis adjusting for all EHR-derived covariates and ADI resulted in a significant global test statistic (p < 0.001) rejecting the null hypothesis of no spatial trend in asthma exacerbations (Figure 3A). The ORs for this GAM showed a single hotspot capturing parts of West Philadelphia and Southeast Philadelphia (Figure 3A). Three additional NO₂-adjusted GAMs including all EHR-derived covariates, ADI, and one of the three NO₂ measurements of interest also resulted in significant global test statistics (p < 0.001), indicating that spatial correlation persists after adjusting for NO₂. However, the hotspots and coldspots for each model differed from each other and from those of the model without NO₂. The GAM that included EPA interpolated NO₂ had less spatial variation and the size of hotspots diminished substantially (Figure 3B). The coefficient for NO₂ was negative, indicating an inverse association between exacerbations and exposure. The GAM that included raw satellite NO₂ had a single hotspot nearly identical to that of the GAM that did not include an NO₂ term (Figure 3C), suggesting that the spatial variation was unexplained by the inclusion of raw satellite NO₂. The GAM that included satellite-derived NO₂ had similar hotspots as those of the GAM that did not include NO₂ but with markedly reduced size (Figure 3D), indicating that this measure of NO₂ partially explained the spatial variation in exacerbations.

Figure 3.

Spatial odds ratios (ORs) of exacerbation before and after integrating NO₂ data. Significant hotspots and coldspots (p < 0.001) are indicated by black contour lines. A) NO₂-unadjusted model included EHR-derived covariates and ADI. NO₂-adjusted models included all covariates and B) interpolated EPA NO₂, C) raw TROPOMI NO₂, and D) satellite-derived ground NO₂.

Discussion

The characteristics of people with more asthma exacerbations in the adjusted proportional odds models—Black race, class 3 obesity, Medicaid insurance—are known asthma exacerbation risk factors⁶. In contrast, we did not find that sex, health insurance type, smoking history, or ADI were significant predictors of asthma exacerbations in our adjusted models, despite reports suggesting that they play a role⁶ but consistent with our previous observations using EHR data³². The geospatial distribution of asthma exacerbations based on EHR data and ADI was found to be heterogeneous, with hotspots in West and South Philadelphia, regions of the city with a high proportion of residents who are Black and/or of low socioeconomic status.

Our results comparing ambient NO₂ measurement types in Philadelphia found high correlation across time in measures taken at the location of EPA regulatory monitors, suggesting that EPA and satellite measures are consistent. However, review of interpolated EPA data versus that provided by satellite measures showed that regional variation was inadequately captured by the EPA measures. For example, examination of Figure 2 shows that EPA monitors failed to detect the largest NO₂ hotspot in the city near Southeast Philadelphia. Further, the statistically significant inverse relationship observed between interpolated EPA NO₂ exposure and asthma exacerbations in the proportional odds and spatial GAM models raises questions about the adequacy of interpolated EPA data to predict health risk because the inverse relationship is inconsistent with known associations between NO₂ and asthma risk. Associations between the raw satellite NO₂ and satellite-derived ground NO₂ measures with asthma exacerbations, on the other hand, were consistent with the known relationship between NO₂ and exacerbation risk.

Comparison of raw satellite data versus satellite-derived ground NO₂ levels showed that although the two measures were similar, the ground level estimates were more heterogenous. However, we are limited in our ability to compare these two satellite measurements because they were based on different averaging periods (2019 for the raw satellite and 2017-2019 for the satellite-derived estimates) due to the limited availability of the relatively new TROPOMI instrument. The satellite-derived ground NO₂ estimates introduced by work in Cooper et al. were originally validated using annual averages at 3,977 monitor sites worldwide and found to be consistent with in situ observations (r = 0.71)²¹. These measures provide improved exposure estimates because they address the limitations of both regulatory monitor and raw satellite measurements: poor spatial resolution and limited ability to relate vertical column density (VCD) to real in situ exposures, respectively. In our association models, satellite-derived NO₂ had a larger odds ratio than the raw satellite NO₂ (1.16 versus 1.09, respectively) for contributing to asthma exacerbations. In the spatial models, inclusion of satellite-derived NO₂ reduced the size and effect of exacerbation hotspots, whereas inclusion of the raw satellite NO₂ increased the size of the exacerbation hotspot while increasing its odds ratio slightly. As more TROPOMI measures become available, researchers studying time periods 2019 or later will be able to easily download and average multiple years of data. This will allow for detailed comparisons between raw satellite observations and satellite-derived estimates for use in exposure assessment studies. In addition, as other satellite-based measures of pollution beside NO₂ become available, these will permit the comparison of contributions of other pollutants.

Our study is subject to limitations, including some related to use of EHR data: missingness, entry error, and phenotype misclassification resulting from the criteria used to classify asthma and asthma exacerbations. Penn Medicine is not the sole provider of care in Philadelphia, and hence, our patient density is greatest near Penn Medicine facilities. Although the GAM accounts for spatial variation in observation density, residual confounding may still be present given the uneven distribution of patients in our study area. Finally, the geocodes we used were the most current at the time of the data pull, but they may not have represented place of residence during the entire study period, nor does the location of residence enable a full assessment of environmental exposures. In future studies, our approach can be extended to include pediatric patients and to consider changes during the COVID-19 pandemic.

Conclusion

By comparing three publicly available ambient NO₂ measurement types, we found that satellite data captured more spatial variability than EPA regulatory monitor data for local health studies, and that inclusion of ground-level satellite estimates of NO₂ into multivariable models of asthma exacerbations revealed a significant association between NO₂ exposure and asthma exacerbations in Philadelphia. In contrast, use of spatially smoothed EPA regulatory monitor data might yield misleading results when used in studies of small regions. Our findings demonstrate the potential of satellite-based pollution measurements to gain insights on local disease patterns, especially as these data become more widely available.