Machine learning-driven identification of air toxic combinations associated with asthma symptoms among elementary school children in Spokane, Washington, USA

Abstract

Air toxics are atmospheric pollutants with hazardous effects on health and the environment. Although methodological constraints have limited the number of air toxics assessed for associations with health and disease, advances in machine learning (ML) enable the assessment of a much larger set of environmental exposures. We used ML methods to conduct a retrospective study to identify combinations of 109 air toxics associated with asthma symptoms among 269 elementary school students in Spokane, Washington. Data on the frequency of asthma symptoms for these children were obtained from Spokane Public Schools. Their exposure to air toxics was estimated by using the Environmental Protection Agency’s Air Toxics Screening Assessment and National Air Toxics Assessment. We defined three exposure periods: the most recent year (2019), the last three years (2017-2019), and the last five years (2014-2019). We analyzed the data using the ML-based Data-driven ExposurE Profile (DEEP) extraction method. DEEP identified 25 air toxic combinations associated with asthma symptoms in at least one exposure period. Three combinations (1,1,1-trichloroethane, 2-nitropropane, and 2,4,6-trichlorophenol) were significantly associated with asthma symptoms in all three exposure periods. Four air toxics (1,1,1-trichloroethane, 1,1,2,2-tetrachloroethane, BIS (2-ethylhexyl) phthalate (DEHP), and 2,4-dinitrophenol) were associated only in combination with other toxics, and would not have been identified by traditional statistical methods. The application of DEEP also identified a vulnerable subpopulation of children who were exposed to 13 of the 25 significant combinations in at least one exposure period. On average, these children experienced the largest number of asthma symptoms in our sample. By providing evidence on air toxic combinations associated with childhood asthma, our findings may contribute to the regulation of these toxics to improve children’s respiratory health.

Graphical abstract

1. Introduction

Asthma is the leading cause of emergency department visits, hospitalization, and absenteeism among children in the United States (U.S.).^1–3 In 2019, 5.1 million U.S. children aged 18 years and younger had asthma.⁴ In the same year, the prevalence of asthma attacks among children with asthma was 44.3%, the rate of emergency department visits was 104.7 per 10,000 children with asthma, and the rate of hospitalization was 4.5 per 10,000 children with asthma.⁴

Air toxics are pollutants in the atmosphere that have hazardous effects on health and the environment.⁵ Children’s immature and underdeveloped immune, respiratory, and metabolic systems make them particularly vulnerable to the adverse effects of air toxic exposure.^6–8 Prenatal, early life, and childhood exposure to air toxics is associated with increased risk of respiratory or asthma symptoms and reduced lung function among children.^9–18 Yet these associations have been studied for a limited number of air toxics, whereas children are typically exposed to a large number of toxics simultaneously.^{15,16,19–21}

Several studies have proposed methods to analyze multi-dimensional exposure data in relation to health outcomes, including asthma.^19,20,22,23 In particular, advances in machine learning (ML) have enabled examination of the associations of an exponentially larger set of combinations of environmental exposures than was formerly possible.^22,24,25 For instance, classification and regression trees have been used to examine the association of exposure to multiple air pollutants with children’s emergency department visits related to asthma.^26–27 More recent methods, such as Data-driven ExposurE Profile (DEEP) extraction,²⁰ have improved on earlier approaches. The strength of DEEP derives from its two-stage approach. First, the Extreme Gradient Boosting (XGBoost)²⁸ algorithm identifies combinations of air toxics associated with a given outcome. Second, established epidemiologic methods control for confounding and identify significantly associated air toxic combinations and synergistic interactions.

DEEP was originally applied to identify combinations of air toxics associated with asthma outcomes in a cohort of children residing in the greater New York area.²⁰ In that study, DEEP found a significant association between asthma and exposure to combinations of acrylic acid, ethylidene dichloride, and hydroquinone. However, exposure was assessed on the basis of participants’ ZIP codes, which define large geographic units with limited value in exposure assessment.^29–30 Additionally, participants resided in one of the ten most densely populated urban areas of the U.S., raising questions about the generalizability of these findings to more sparsely populated urban areas. Notably, 80% of the U.S. population lives in urban areas but only 12% of this population lives in the ten most densely populated urban areas with a minimum population density of 4,550 people per square miles.³¹ Furthermore, the study’s asthma outcomes were collected in healthcare settings, which may not be equitably accessible to all children.³² Finally, study outcomes were cross-sectional, and thus did not address the impact of longer-term exposure to air toxics on childhood asthma.

In the current work, we expanded the examination of associations between childhood asthma and combinations of air toxic exposures to include a broader population living in a sparsely populated urban area and presenting in non-healthcare settings. We conducted a retrospective study of air toxic combinations associated with worsening asthma symptoms among elementary school students in Spokane, Washington. Our findings offer evidence pertinent to future regulation of air toxics present in schools and surrounding neighborhoods with the goal of improving children’s respiratory health.

2. Materials & Methods

The overall workflow of our study is shown in Graphical Abstract. Individual components of the workflow are detailed below.

2.1. Study setting

This study was conducted in 10 elementary schools in the Spokane Public Schools system.³³ Spokane is a metropolitan area in Washington State with a population of 228,989 people and population density of 3,330 people per square miles in 2020.³⁴ It is a regional hub for healthcare, education, transportation, and financial services.³⁵ Around 80% of the urban population is non-Hispanic White and 6.8% is Hispanic. In 2021 the median household income was $56,977, and 15.6% of the population lived below U.S. federal poverty guidelines.³⁴

Spokane Public Schools provided deidentified data on the demographic characteristics (age, gender, and race/ethnicity) of children with an asthma plan in our 10 elementary schools, as well as the frequency of asthma symptoms for each child. Data on school-level percentages of students eligible for free or reduced meals were publicly available from the Office of the Superintendent of Public Instruction.³⁶ These percentages were used as a surrogate for the socioeconomic status of children attending the 10 schools.

The Washington State University Institutional Review Board determined that our study satisfied the criteria for Exempt Research. Data on children were obtained through a data sharing agreement with Spokane Public Schools.

2.2. Measures

2.2.1. Asthma outcome.

Asthma is treated as a life-threatening event in elementary schools in the Spokane Public Schools system. Children who suffer from asthma are required to have a health plan and an inhaler in school before attending. Such children are required to see the school nurse whenever asthma symptoms occur. This policy enables documentation of the number of asthma symptoms experienced over time. In schools with a full-time nurse, all asthma events are captured by in-house software accessible to the nurse. In schools with a part-time nurse, events are documented by using in-house software in conjunction with paper records. The latter are used by school staff to document health symptoms when the school nurse is not present. For the present study, paper records were digitized by school staff in collaboration with the study team, and the resulting data were combined with existing data in the software system.

To quantify disease management and severity, we calculated the total number of asthma symptoms occurring during school hours between September 2019 and February 2020 among children with an asthma plan.^37–39 Asthma symptoms included wheezing, coughing, chest tightness, dyspnea, breathlessness, and use of an inhaler. We applied the log₁₀(x+1) transformation to the original value of x (number of asthma symptoms) to account for skewness and zero values and yielded the final outcome value analyzed here. We excluded two children who were enrolled in more than one school during the study period.

2.2.2. Air toxic exposures.

We estimated children’s exposure to air toxics by using two assessments from the U.S. Environmental Protection Agency (EPA): the National Air Toxics Assessment (NATA – prior to 2017) and the Air Toxics Screening Assessment (AirToxScreen – 2017 onwards).⁴⁰ Each NATA or AirToxScreen cycle provides estimates for the annual ambient concentrations of a maximum of 181 air toxics at the Census tract level in the continental U.S. and Puerto Rico. These estimates are based on emission inventories and computer simulation models, with ambient temperature, meteorology, precipitation, and solar radiation incorporated in modeling. Data are currently available for the 1999, 2002, 2005, 2011, 2014, 2017, 2018, and 2019 cycles.^40,41

We geocoded the locations of schools attended by children in 2019 and estimated exposure to air toxics by using intersecting U.S. Census tracts in which the schools were located with the air toxics data. Using NATA/AirToxScreen data from the 2014, 2017, 2018, and 2019 cycles, we defined three exposure periods: the most recent year (2019), the last three years (2017-2019), and the last five years (2014-2019). This approach enabled us to examine the associations of both short- and long-term toxic exposures with childhood asthma outcomes. Exposure estimates for the last three and five years comprised the average annual concentrations of air toxics across the years included in each exposure period.

2.2.3. Covariates.

Because sociodemographic covariates can confound associations between exposure to air toxics and asthma^20,42–45 we included age, gender, race/ethnicity, and school-level percentage of students eligible for free or reduced meals (Section 2.1) as covariates in regression models to correct for confounding in the second stage of DEEP (Section 2.3).

2.3. Data analysis

We analyzed data on air toxic exposure, outcome, and covariates by using DEEP.²⁰ In the first stage, 100 XGBoost²⁸ models were learned to identify air toxic combinations predictive of the number of asthma symptoms experienced by children in our sample. In this process, the full dataset was randomly split 100 times into training and test sets in an 80:20 ratio. For each training split, an XGBoost model consisting of 100 decision trees was learned to predict the number of asthma symptoms. Each model was then evaluated on the test set in terms of the concordance index,⁴⁶ and scores were averaged across the 100 training and test splits to assess the overall predictive performance of DEEP. The decision trees included in the XGBoost models were constituted of nodes and edges.²⁶ Each node represented a single air toxic with its threshold concentration, as well as the subpopulation of children who were exposed or not exposed to the combination of air toxics on the path taken to reach the node. Our goal was to identify combinations where in which increased levels of air toxic concentrations (i.e., levels higher than the thresholds identified by the decision trees) were associated with higher numbers of asthma symptoms. To identify the most reliable combinations, we ranked the paths from roots to leaves in the inferred decision trees in terms of the number of XGBoost models (out of the 100 trained) that included them. We limited further consideration to combinations found in at least 10 XGBoost models.²⁰

In the second stage of DEEP, the air toxic combinations identified by the XGBoost models were regressed on the outcome by using multivariable linear regression models adjusted for our covariates. One regression model was built for each identified air toxic combination, yielding a beta coefficient representing the strength of the association between that combination and the outcome. Significant associations were identified as those with a false discovery rate lower than 0.05 after applying the Benjamini-Hochberg method to correct for multiple hypothesis testing.⁴⁷ We also used multivariable linear regression models to examine the interactions within combinations and the outcome by including the levels of constituent air toxics and their product in the models. A significant interaction between the air toxics in a combination was identified if the p-value of the corresponding interaction term was lower than 0.05.

Finally, given the stratified nature of the decision trees underlying DEEP, we were able to identify the sociodemographic characteristics of children affected by specific air toxic combinations and compare them to those of children who were not exposed to the same combinations. To perform these comparisons, we used Independent Samples t-test for age (normally distributed), Mann-Whitney U-test for percentage of students eligible for free or reduced meals (non-normally distributed), and chi-squared tests for sex and race/ethnicity (categorical) variables.

3. Results

3.1. Sociodemographic characteristics of the study population

Table 1 shows the characteristics of the 269 children with asthma in our sample. Their mean age was 9.7 years (standard deviation [SD] = 3.1 years). Females and males were evenly represented. Most children (59%) were non-Hispanic White; 22% were non-Hispanic multi-racial and 12% were Hispanic. On average, 75% (SD = 17%) were eligible for free or reduced meals. The average number of asthma symptoms experienced during the study period was 9.6 (SD = 16.9). In the full sample, 103 children (38%) experienced no asthma symptoms at school during the study period.

Table 1.

Sociodemographic characteristics of our sample, which consisted of children with a registered asthma plan enrolled in the 10 elementary schools in Spokane, Washington.

Characteristics	Statistics
Age, years	9.7 (2.1)
Sex (n (%))
Female	138 (51.3)
Male	129 (48.0)
Other / unknown	2 (0.7)
Race and ethnicity
Non-Hispanic (NH) American Indian/Alaska Native	Suppressed
NH Asian or Pacific Islander	Suppressed
NH Black	10 (3.7)
NH Multi-racial	59 (21.9)
NH White	158 (58.7)
Hispanic	32 (11.9)
Unknown	Suppressed
% eligible for free and reduced meals*	75.1 (17.1)
Number of asthma symptoms	9.6 (16.9)

^*School-level percentage of students eligible for free or reduced meals.

Mean (standard deviation) and number (percentage) are shown for continuous and categorical variables respectively.

For confidentiality reasons, all counts less than 10 and rates derived from them are suppressed.

3.2. Exposure to air toxics

Our final exposure data included 109 air toxics with no missing estimates across the NATA/AirToxScreen 2014, 2017, 2018, and 2019 cycles. These air toxics included carbonyls (such as acetaldehyde and formaldehyde), heavy metals (such as arsenic and lead), chlorophenols (such as 2,4,6-trichlorophenol and phenol), and volatile organic compounds (such as 1,3-butadiene and benzene) among others. Appendix A shows the average concentration of these air toxics for our three exposure periods.

3.3. Air toxic combinations associated with asthma symptoms

Our XGBoost models were more predictive than their random counterparts inferred from permuted versions of the true outcome (concordance index = 0.59 versus 0.50, respectively). This result confirms the ability of the first-stage decision trees in DEEP to identify associations between combinations of air toxics and asthma symptoms.

In the first stage of DEEP, 373 combinations of air toxics occurring in at least 10 XGBoost models were identified across all three exposure periods. These combinations included both single and multiple air toxic combinations, although the method placed no restrictions on the number of toxics included in a combination. In the second stage of DEEP, 162 of these combinations were significantly associated with the outcome after correction for multiple hypothesis testing. Among the combinations, 14 single and 11 multiple air toxic combinations with increased levels of constituent air toxics were significantly associated with worsening asthma symptoms in at least one exposure period (Figure 1).

Figure 1. Summary of air toxic combinations identified as significantly associated with the number of asthma symptoms - individually, in combination with other air toxics, and with statistically significant interactions within the combination - across three exposure periods.

The three exposure periods were the most recent year (2019), the last three years (2017-2019), and the last five years (2014-2019) to respectively examine the associations of short- and long-term exposure to combinations of air toxics with childhood asthma outcomes.

The DEEP method identified 14 single- and 11 multi-air toxic combinations with increased levels of the constituent air toxics that were significantly associated with worsening asthma symptoms. The combinations are shown in terms of the primary air toxic in the first column, and the rest in other columns, with colored cells indicating the combinations. Significant synergistic interactions within members of six multi-air toxic combinations were also identified, and are marked in this figure by asterisks (*) in the corresponding cells.

Columns 2-4 indicate combinations that were identified in all three exposure periods, the air toxics that have been determined to pose the greatest potential health threat by the U.S. Environmental Protection Agency, and those identified in an earlier application of DEEP in the greater New York area.

Among the 14 single air toxic combinations that we identified (Figure 2), 2-nitropropane and 2,4,6-trichlorophenol were significantly associated with the outcome in all three exposure periods, indicating their sustained impact on asthma symptoms. Other air toxics individually associated with asthma symptoms included 1,2,4-trichlorobenzene, 1,3-butadiene, acetaldehyde, acetonitrile, benzene, cumene, diesel PM, and ethylbenzene.

Figure 2. Beta coefficients and 95% confidence intervals for single-air toxic combinations associated with an increased number of asthma symptoms among children from 10 elementary schools in Spokane, Washington in our sample (n = 269) across the three exposure periods.

The strength of an association is shown in terms of its beta coefficient and 95% confidence interval (CI) adjusting for age, gender, race and ethnicity, and percentage of students eligible for free or reduced meals. These values are shown numerically, as well as visually in a forest plot in the right side of the figure.

P-values were calculated from multivariable linear regression models, and adjusted for multiple hypothesis testing using the Benjamini-Hochberg method. All these computations were carried out in the second stage of DEEP. ** p < 0.05; *** p < 0.001

A unique strength of DEEP is its ability to identify multiple air toxic combinations associated with the outcome of interest. Eleven combinations, 10 consisting of two air toxics each and one consisting of three air toxics, were found to be significantly associated with childhood asthma symptoms in at least one exposure period (Figure 3). 1,1,1-Trichloroethane appeared in five, 1,3-butadiene in four, and 2,4,6-trichlorophenol in three combinations. Four air toxics in these combinations, namely 1,1,1-trichloroethane, 1,1,2,2-tetrachloroethane, BIS (2-ethylhexyl) phthalate (DEHP), and 2,4-dinitrophenol, appeared only in combinations. These air toxics were not individually associated with the outcome and would not have been identified without the use of DEEP. Two air toxics in these combinations, namely 1,1,1-trichloroethane and 2,4,6-trichlorophenol (the latter was also identified as a single air toxic combination), were significantly associated with asthma symptoms in all three exposure periods, illustrating DEEP’s ability to find pollutants with sustained effects over time.

Figure 3. Beta coefficients and 95% confidence intervals for multi-air toxic combinations associated with an increased number of asthma symptoms among children from 10 elementary schools in Spokane, Washington in our sample (n = 269) across the three exposure periods.

The strength of an association is shown in terms of its beta coefficient and 95% confidence interval (CI), adjusting for age, gender, race and ethnicity, and percentage of students eligible for free or reduced meals. These values are shown numerically, as well as visually in a forest plot in the right side of the figure.

P-values were calculated from multivariable linear regression models, and adjusted for multiple hypothesis testing using the Benjamini-Hochberg method. All these computations were carried out in the second stage of DEEP. ** p < 0.05; *** p < 0.001

We also found significant synergistic interactions within constituents of six multiple air toxic combinations in at least one exposure period (Table 2). Among these, 2,4,6-trichlorophenol was the primary air toxic (i.e., primary branch point in decision trees) in three combinations. Two of these three were identified in combination with 1,1,1-trichloroethane. The latter also appeared in five combinations with significant interactions. 2,4,6-Trichlorophenol was identified in interactions in all three exposure periods and 1,1,1-trichloroethane in two exposure periods, highlighting their sustained impact on asthma symptoms.

Table 2.

Air toxic combinations identified with statistically significant interactions within members of the combination found to be associated with asthma symptoms across the three exposure periods.

Air toxic combinations	p-value
Most recent year (2019)
ACETALDEHYDE >= 9.168e-01 & 1,1,1-TRICHLOROETHANE >= 2.040e-02	0.0036
2,4,6-TRICHLOROPHENOL >= 6.940e-07 & 1,1,1-TRICHLOROETHANE >= 2.039e-02	0.0036
Last 3 years (2017-2019)
DIESEL PM >= 4.085e-01 & 1,1,1-TRICHLOROETHANE >= 3.981e-02	0.0036
2,4,6-TRICHLOROPHENOL >= 6.265e-07 & 1,1,1-TRICHLOROETHANE >= 3.883e-02	0.0036
Last 5 years (2014-2019)
ETHYLBENZENE >= 1.656e-01 & 1,1,1-TRICHLOROETHANE >= 6.484e-02	0.0036
2,4,6-TRICHLOROPHENOL >= 4.545e-06 & ACETONITRILE >= 1.836e-02	0.0042

Multivariable linear regression models were used to examine the interactions within the XGBoost-identified multi-air toxic combinations and the outcome by using the levels of the constituent air toxics and their product in the models. A significant interaction was identified if the p-value of the corresponding interaction term was lower than 0.05.

3.4. Sociodemographic characteristics of the most vulnerable subgroups

By comparing the demographic characteristics of children affected by significant air toxic combinations with those of children who were not exposed, we identified a subgroup of 28 children who were exposed to 13 of the 25 significant combinations in at least one exposure period (Table 3). Children exposed to these combinations were slightly older (9.9 versus 9.7 years, p < 0.001) and attended the school with the lowest socioeconomic profile (i.e., the largest percentage of free or reduced meal students in our sample: 91% versus 73%, p < 0.001). Figure 4 shows sample decision trees, including a significantly associated combination (highlighted in red) to which this population of children was exposed (leaf node in each tree) in each exposure period. Note that these trees were learned from 80% of the full sample (i.e., the training test).

Table 3. Sociodemographic characteristics of a vulnerable subpopulation of children in our sample, who were exposed to 13 of the 25 significant air toxic combinations found to be associated with asthma symptoms.

This subpopulation was exposed to three single- and 10 multi-air toxic combinations consisting of 1,1,1-Trichloroethane, 1,1,2,2-Tetrachloroethane, 1,2,4-Trichlorobenzene, 1,3-Butadiene, 2,4,6-Trichlorophenol, 2,4-Dinitrophenol, Acetaldehyde, Acetonitrile, Benzene, Bis(2-Ethylhexyl)Phthalate (DEHP), Diesel PM, and Ethylbenzene.

Characteristics	Exposed (n=28)	Not exposed (n=241)	p-value
Age, years	9.9 (2)	9.7 (2.1)	<0.001
Gender
Female	15 (53.6)	114 (47.3)	1.00
Male	13 (46.4)	125 (51.9)
Unknown	Suppressed	Suppressed
Race and ethnicity
Non-Hispanic (NH) American Indian/Alaska Native	Suppressed	Suppressed	0.89
NH Asian or Pacific Islander	Suppressed	Suppressed
NH Black	Suppressed	Suppressed
NH Multi-racial	Suppressed	53 (22)
NH White	16 (57.1)	142 (58.9)
Hispanic	Suppressed	29 (12)
Unknown	Suppressed	Suppressed
% eligible for free and reduced meals*	90.6 (0)	73.3 (17.2)	<0.001

^*School-level percentage of students eligible for free or reduced meals.

Mean (standard deviation) and number (percentage) are shown for continuous and categorical variables, respectively. Sociodemographic characteristics were compared using Independent Samples t-test for age (normally distributed), Mann-Whitney U-test for percentage of students eligible for free or reduced meals (non-normally distributed), and chi-squared tests for sex and race/ethnicity (categorical) variables.

For confidentiality reasons, all counts less than 10 and rates derived from them are suppressed.

Figure 4. Sample decisions trees learned in the first stage of DEEP for each exposure period showing air toxic combinations associated with worsening asthma symptoms.

These trees are shown for the most vulnerable subpopulation of children (n=28) in our study that was exposed to 13 of the 25 significant air toxic combinations. Note that these trees were learned from 80% of the full cohort (i.e., the training test), and consist of nodes and edges.

Each node represented a single air toxic with its threshold concentration, as well as the subpopulation of children in the overall cohort that were exposed or not exposed to the combination of air toxics on the path taken to reach the node. Potentially associated air toxic combinations, as well as their exposed subpopulations, were identified as paths from the roots to leaves of these trees.

Highlighted here in red are sample combinations that were found to be associated with significantly associated with worse (increased number of) asthma symptoms that this vulnerable population was exposed to.

4. Discussion

This study applied the DEEP method to identify the associations of short- and long-term exposure to combinations of air toxics with asthma symptoms among elementary school students in Spokane, Washington. We identified 25 air toxic combinations associated with worsening asthma symptoms in at least one of the three exposure periods we studied. Three air toxic combinations were associated in all three periods. Four air toxics appeared only in combination with other pollutants, and would not have been identified by traditional statistical methods. These methods may include bivariable or multivariable regression techniques that assess the association between an outcome and one or more independent variables but are not intended to identify combinations of independent variables associated with an outcome. Six multiple air toxic combinations exhibited significant synergistic interactions among their constituents.

Three of the air toxics identified in combinations, namely 1,1,1-trichloroethane, 2-nitropropane, and 2,4,6-trichlorophenol, were significantly associated with asthma symptoms in all three exposure periods. 1,1,1-Trichloroethane and 2,4,6-trichlorophenol appeared in 11 of the 25 combinations, and synergistic interactions were observed within the combination consisting of these two air toxics. 1,1,1-Trichloroethane, also known as methyl chloroform, is widely used in industry as a solvent, and was formerly used in household products such as cleaners, glues, and aerosol sprays that can still be found in homes.⁴⁸ Its highest intake occurs through the lungs and gastrointestinal tract, and as much as 30% of inhaled 1,1,1-trichloroethane can be retained in the lungs.⁴⁹ Exposure to this air toxic is associated with irritation of the nasal and respiratory mucous membranes, resulting in symptoms such as shortness of breath, cough, and chest tightness.^49,50 1,1,1-Trichloroethane was also associated with daily use of asthma control medication in a previous study in the New York metropolitan area,²⁰ suggesting a broad geographical association with childhood asthma.

2,4,6-Trichlorophenol, another air toxic identified in all three exposure periods, is a chlorophenol, a group of chemicals in which hydrogens are replaced by chlorines on phenol, with the latter being an aromatic compound derived from benzene.⁵¹ 2,4,6-Trichlorophenol was widely used as an antiseptic and anti-mildew agent in the U.S., and can still be found in pesticides and preservatives produced before it was discontinued in the 1980s.⁵² Increased risk of asthma and respiratory issues, such as cough, chronic bronchitis, chest wheezing, altered pulmonary function, and pulmonary lesions, have been reported as a result of chronic exposure.^51,53–55

Four air toxics in the combinations we identified, namely 1,3-butadiene, benzene, acetaldehyde and 1,1,2,2-tetrachloroethane, have been characterized by the EPA as pollutants that pose a serious threat to human health in urban areas.⁵⁶ More than 60% of environmental releases of 1,3-butadiene and benzene are caused by mobile sources such as vehicles.⁵⁷ 1,3-Butadiene is also released by manufacturing, forest fires, and tobacco smoke, and it is used in the production of rubber and plastics.⁵⁸ Common sources of benzene include burning coal and oil, vehicle service stations, and tobacco smoke.⁵⁹ As much as 50% of the benzene in the environment is absorbed through the lungs and gastrointestinal tract and then distributed throughout the body, where it accumulates in fatty tissues.⁶⁰ Exposure to 1,3-butadiene is associated with respiratory tract irritation,⁶¹ and exposure to benzene with mucous membrane irritation, acute granular tracheitis, laryngitis, bronchitis, and pulmonary edema.^60,62–65

Acetaldehyde, another air toxic, is commonly used as a chemical intermediate in perfumes, polyester resins, and dyes, and as a solvent in rubber, tanning, and paper manufacture.⁶⁶ Acetaldehyde primarily targets the upper respiratory mucosa,⁶⁷ triggering cough, irritation of the respiratory tract, and pulmonary edema.^67,68 The latter condition may result in deeper penetration of acetaldehyde in the respiratory system, causing bronchoconstriction among people with asthma.⁶⁷ Notably, acetaldehyde was also associated with children’s asthma-related emergency department visits in a previous study in the New York metropolitan area,²⁰ indicating broad geographical relevance.

In addition to finding novel combinations of air toxics associated with worse childhood asthma symptoms, we also identified a vulnerable subgroup of 28 children who were exposed to 13 of the 25 combinations, including five with significant mutual interactions. These children reported the largest number of asthma symptoms, and all of them attended the same school, whose students had the lowest socioeconomic status in our sample. This finding supports previous research associating lower socioeconomic status with greater severity and less effective management of asthma symptoms in children.^42–45 Moreover, U.S. public schools with a higher percentage of free or reduced meal students are exposed to higher concentrations of several air toxics, such as 1,1,1-trichloroethane, 2,4-dinitrotoluene, acrylamide, and lead, than schools with a lower percentage of free or reduced meals.⁶⁹ Future research should examine the extent to which students’ environmental and socioeconomic characteristics are associated with asthma symptoms.

As most children spend one-third of their waking hours in school,⁷⁰ it is vitally important to monitor and reduce air toxic concentrations in and around schools. The Clean Air Act identified 187 air toxics that are required to be controlled by the EPA.⁵ In partnership with state and local organizations, the EPA monitored air toxic concentrations outside 62 schools in 22 states to identify areas where extended monitoring and other actions were needed.⁷¹ Because more than 98,000 public schools exist in the U.S., and because schools with minority students and students with low socioeconomic status are exposed to higher concentrations of air toxics,⁶⁹ an expansion of this monitoring effort is needed to identify additional vulnerable school sites. The highly vulnerable subpopulation of children that we found can serve as a case study for such an expansion.

We acknowledge certain limitations in our study. First, we used data from the EPA’s NATA/AirToxScreen assessments to estimate exposure to air toxics. These estimates are based on validated deterministic dispersion models, but exposures may be under- or over-estimated.^40,41 Nevertheless, these data are commonly used for exposure assessment, given the challenges of collecting data on 180 air toxics in more than 70,000 U.S. Census tracts. Second, we estimated exposure by assessing elementary schools’ locations, whereas asthma symptoms may have been triggered at other sites to which we had no access for reasons of confidentiality. Third, several of the air toxics we identified have already been associated with asthma and other respiratory issues in previous research; however, such studies typically examined individuals rather than populations and often excluded children. Fourth, our results do not provide evidence for a causal effect of air toxics on asthma symptoms. Finally, our study was conducted in a sparsely populated urban area in Washington State, so our findings may not generalize to other regions in the U.S. Nonetheless, our findings confirmed a broad geographical association of air toxics with childhood asthma.

In conclusion, we identified several air toxics and combinations of air toxics associated with childhood asthma symptoms that would likely not have been found by using traditional statistical methods. Our findings suggest that adverse environmental health outcomes, such as worsening asthma symptoms, should be studied by obtaining multi-dimensional exposure data, especially in settings with vulnerable populations. Our results provide a basis for regulating the air toxics we identified in elementary schools. Future research should examine the extent to which exposure to air toxics affects children’s overall and respiratory health, along with the biological mechanisms that drive those effects.

Highlights.

• Children are vulnerable to the adverse effects of exposure to air toxics

• We used machine learning to examine air toxics associated with asthma symptoms

• Both known and novel associations with symptoms were identified

• We found a subgroup of children particularly vulnerable to air toxic exposure

• Results may inform regulation of air toxics to improve children’s health

Abbreviations:

ML

Machine learning

XGBoost

EXtreme Gradient Boosting

DEEP

Data-driven ExposurE Profile extraction

NATA

National Air Toxics Assessment

AirToxScreen

Air Toxics Screening Assessment

Appendix A. Average concentrations of 109 air toxics across the 10 elementary schools in Spokane, Washington.

The US Environmental Protection Agency’s National Air Toxic Assessment (NATA) and Air Toxics Screening Assessment (AirToxScreen) databases from the 2014, 2017, 2018, and 2019 cycles were utilized. Three exposure periods were defined, namely the most recent year (2019), the last three years (2017-2019), and the last five years (2014-2019). The exposure estimates for the last three and five years were the average annual concentrations of air toxics across the years included in the exposure period being considered. Exposure was estimated by identifying the census tracts where the schools were located, and mapping them to the NATA/AirToxScreen databases.

Units of measure for the standards are micrograms per cubic meter of air (μg/m³).

Air toxic	2019	2017-2019	2014-2019
1,1,1-TRICHLOROETHANE	0.020320336	0.035946768	0.056375674
1,1,2,2-TETRACHLOROETHANE	0.000052720	0.000052810	0.000049230
1,1,2-TRICHLOROETHANE	0.000000033	0.000000031	0.000000032
1,2,4-TRICHLOROBENZENE	0.000002561	0.000002320	0.000004434
1,3-BUTADIENE	0.048788730	0.052689210	0.059707774
1,3-DICHLOROPROPENE	0.024403077	0.010983650	0.008742348
1,4-DICHLOROBENZENE	0.001294972	0.000802937	0.000933313
1,4-DIOXANE	0.000028070	0.000018140	0.000024710
2,2,4-TRIMETHYLPENTANE	0.240433994	0.264921642	0.390715899
2,4,6-TRICHLOROPHENOL	0.000000615	0.000000562	0.000004287
2,4-DINITROPHENOL	0.000000016	0.000000018	0.000000019
2,4-DINITROTOLUENE	0.000001241	0.000001124	0.000001190
2,4-TOLUENE DIISOCYANATE	0.000164581	0.000136533	0.000106412
2-CHLOROACETOPHENONE	0.000000046	0.000000041	0.000000052
2-NITROPROPANE	0.000000025	0.000000020	0.000000026
4,4′-METHYLENEDIPHENYL DIISOCYANATE (MDI)	0.000188765	0.000120278	0.000164679
4-NITROPHENOL	0.000105917	0.000100507	0.000117421
ACETALDEHYDE	0.856636910	1.046843400	1.149019795
ACETAMIDE	0.000000379	0.000000458	0.000000427
ACETONITRILE	0.007048869	0.016941878	0.018207840
ACETOPHENONE	0.000076700	0.000069850	0.000166074
ACROLEIN	0.033024721	0.042088090	0.046778019
ACRYLIC ACID	0.000979804	0.001722174	0.001815737
ACRYLONITRILE	0.000097968	0.000101319	0.000093750
ALLYL CHLORIDE	0.000000501	0.000000453	0.000000479
ANTIMONY COMPOUNDS	0.000001827	0.000002137	0.000002778
ARSENIC COMPOUNDS(INORGANIC INCLUDING ARSINE)	0.000099097	0.000092469	0.000081079
BENZENE	0.423565193	0.479230772	0.527670988
BENZYL CHLORIDE	0.000004770	0.000004303	0.000005431
BERYLLIUM COMPOUNDS	0.000002225	0.000001988	0.000004943
BIPHENYL	0.000146686	0.000139590	0.000377319
BIS(2-ETHYLHEXYL)PHTHALATE (DEHP)	0.000132564	0.000112777	0.000593784
BROMOFORM	0.000000254	0.000000229	0.000000291
CADMIUM COMPOUNDS	0.000012554	0.000011833	0.000013633
CARBON DISULFIDE	0.000119669	0.000111080	0.000117512
CARBON TETRACHLORIDE	0.490040994	0.446706329	0.463407792
CARBONYL SULFIDE	0.000001077	0.000001271	0.000001909
CHLORINE	0.000576014	0.000625168	0.000504239
CHLOROBENZENE	0.000607307	0.000459604	0.000654167
CHLOROFORM	0.086058916	0.069005668	0.070355707
CHLOROPRENE	0.000000336	0.000000310	0.000000296
CHROMIUM VI (HEXAVALENT)	0.000012074	0.000012810	0.000014770
COBALT COMPOUNDS	0.000000572	0.000000669	0.000000700
CRESOL_CRESYLIC ACID (MIXED ISOMERS)	0.060332915	0.056589622	0.071443680
CUMENE	0.000561059	0.000740559	0.000942402
CYANIDE COMPOUNDS	0.025795137	0.024398199	0.035267165
DIBENZOFURAN	0.000012143	0.000010984	0.000084860
DIBUTYLPHTHALATE	0.082002309	0.052600261	0.039544068
DIESEL PM	0.335469369	0.319553470	0.369345359
DIETHANOLAMINE	0.005184761	0.002978236	0.002233901
DIMETHYL FORMAMIDE	0.000000285	0.000000213	0.000000303
DIMETHYL PHTHALATE	0.000044090	0.000033260	0.000030350
DIMETHYL SULFATE	0.000000347	0.000000312	0.000000391
EPICHLOROHYDRIN	0.000000117	0.000000105	0.000000112
ETHYL ACRYLATE	0.000000045	0.000000041	0.000000964
ETHYL CHLORIDE	0.000019340	0.000023600	0.000024460
ETHYLBENZENE	0.109273245	0.118406724	0.135486016
ETHYLENE DIBROMIDE (DIBROMOETHANE)	0.000006254	0.000005653	0.000005484
ETHYLENE DICHLORIDE (1,2-DICHLOROETHANE)	0.000289014	0.000269249	0.000249323
ETHYLENE GLYCOL	0.281036046	0.213958803	0.249126127
ETHYLENE OXIDE	0.000005106	0.000004816	0.000005055
ETHYLIDENE DICHLORIDE (1,1-DICHLOROETHANE)	0.000041530	0.000052810	0.000055840
FORMALDEHYDE	1.098845803	1.262435382	1.323219222
GLYCOL ETHERS	0.066718672	0.070518556	0.094497633
HEXACHLOROBENZENE	0.000000127	0.000000116	0.000000501
HEXACHLOROBUTADIENE	0.000000019	0.000000017	0.000000018
HEXACHLOROCYCLOPENTADIENE	0.000000015	0.000000014	0.000000014
HEXAMETHYLENE DIISOCYANATE	0.000000000	0.000000000	0.000000001
HEXANE	0.181869482	0.185447557	0.202136418
HYDROCHLORIC ACID (HYDROGEN CHLORIDE [GAS ONLY])	0.080082703	0.064631613	0.074278468
HYDROGEN FLUORIDE (HYDROFLUORIC ACID)	0.002848985	0.003151003	0.002954682
HYDROQUINONE	0.000059750	0.000046910	0.000039600
ISOPHORONE	0.000037800	0.000033430	0.000222218
LEAD COMPOUNDS	0.000614412	0.000542994	0.000510852
MANGANESE COMPOUNDS	0.000661549	0.000644111	0.000606508
MERCURY COMPOUNDS	0.000291186	0.000681400	0.000970762
METHANOL	0.698811050	1.037408605	1.124426202
METHYL BROMIDE (BROMOMETHANE)	0.028705795	0.029289680	0.029880650
METHYL CHLORIDE (CHLOROMETHANE)	1.188146314	1.178934670	1.197300374
METHYL IODIDE (IODOMETHANE)	0.000000708	0.000000764	0.000000713
METHYL ISOBUTYL KETONE (HEXONE)	0.025149750	0.021949675	0.024638610
METHYL METHACRYLATE	0.004309175	0.003570701	0.004005586
METHYL TERT-BUTYL ETHER	0.000002066	0.000001834	0.000002039
METHYLENE CHLORIDE	0.369941293	0.204999140	0.208696257
METHYLHYDRAZINE	0.000001105	0.000000997	0.000001269
N,N-DIMETHYLANILINE	0.000018187	0.000019600	0.000017016
NAPHTHALENE	0.049509341	0.058157122	0.064905818
NICKEL COMPOUNDS	0.000292542	0.000287834	0.000238659
NITROBENZENE	0.000000169	0.000000153	0.000000162
O-TOLUIDINE	0.000000045	0.000000041	0.000000043
PAHPOM	0.024072992	0.025633461	0.020452996
PENTACHLOROPHENOL	0.000000022	0.000000020	0.000001014
PHENOL	0.068964982	0.064642502	0.084478005
PHOSPHORUS	0.000001965	0.000002309	0.000002456
POLYCHLORINATED BIPHENYLS (AROCLORS)	0.000000308	0.000000417	0.000054150
PROPIONALDEHYDE	0.017087585	0.018297922	0.024870274
PROPYLENE DICHLORIDE (1,2-DICHLOROPROPANE)	0.000018050	0.000018370	0.000017570
PROPYLENE OXIDE	0.000108134	0.000085240	0.000069200
SELENIUM COMPOUNDS	0.000010724	0.000011109	0.000039580
STYRENE	0.000052031	0.000025072	0.000094874
TETRACHLOROETHYLENE	0.019994338	0.019326281	0.020269695
TOLUENE	0.785699692	0.874705873	0.977555835
TRICHLOROETHYLENE	0.002185116	0.007899219	0.011057465
TRIETHYLAMINE	0.000198701	0.000143280	0.000121947
TRIFLURALIN	0.000148152	0.000226091	0.000196399
VINYL ACETATE	0.001553415	0.001070443	0.000867556
VINYL CHLORIDE	0.000121559	0.000121826	0.000113733
VINYLIDENE CHLORIDE	0.000014374	0.000014222	0.000015030
XYLENES (MIXED ISOMERS)	0.467974749	0.499372207	0.547353462

Data availability:

Air toxics data are available from the US Environmental Protection Agency. Data on the frequency of asthma symptoms among children are confidential and cannot be shared.