Abstract
Objective:
Children may be exposed to numerous in-home environmental exposures (IHEE) that trigger asthma exacerbations. Spatially linking social and environmental exposures to electronic health records (EHR) can aid exposure assessment, epidemiology, and clinical treatment, but EHR data on exposures are missing for many children with asthma. To address the issue, we predicted presence of indoor asthma trigger allergens, and estimated effects of their key geospatial predictors.
Study Design and Setting:
Our study samples were comprised of children with asthma who provided self-reported IHEE data in EHR at a safety-net hospital in New England during 2004–2015. We used an ensemble machine learning algorithm and 86 multilevel features (e.g., individual, housing, neighborhood) to predict presence of cockroaches, rodents (mice or rats), mold, and bedroom carpeting/rugs in homes. We reduced dimensionality via elastic net regression and estimated effects by the G-computation causal inference method.
Results:
Our models reasonably predicted presence of cockroaches (Area under receiver operating curves [AUC]=0.65), rodents (AUC=0.64), and bedroom carpeting/rugs (AUC=0.64), but not mold (AUC=0.54). In models adjusted for confounders, higher average household sizes in census tracts were associated with more reports of pests (cockroaches and rodents). Tax-exempt parcels were associated with more reports of cockroaches in homes. Living in a White-segregated neighborhood was linked with lower reported rodent presence, and mixed residential/commercial housing and newer buildings were associated with more reports of bedroom carpeting/rugs in bedrooms.
Conclusion:
We innovatively applied a machine learning and causal inference mixture methodology to detail IHEE among children with asthma using EHR and geospatial data, which could have wide applicability and utility.
Keywords: Exposure assessment, asthma triggers, housing, neighborhoods, electronic health records
INTRODUCTION
Children spend approximately 75% of an average day indoors at home [1]. Among children with asthma, exposure to indoor environmental allergens has been associated with phenotype-specific disease development, sensitivity, symptom exacerbation, and subsequent morbidity [2–4]. In-home environmental exposures (IHEE) include allergenic particles from cockroaches [5], mice [6], rats [7], dust mites [5], dogs [8], cats [8], and mold [9]. Reducing or eliminating exposures to these allergens is an effective way to reduce asthma exacerbations [10,11], but clinicians often lack a way to identify and characterize in-home exposures among patients. As direct measurement of IHEE is difficult for a large clinical population [12,13], clinicians often rely on questions asked to patients, but these questions are currently not asked universally to all children with asthma.
Allergens aggregate indoors in organism-specific microenvironmental reservoirs, including carpeting, bedding, wallboard, upholstered furniture, settled dust, wall cavities, cabinets, air ducts, and crawlspaces [14,15]. Homes at risk for prevalent indoor allergens tend to be old and in poor condition, characterized by leaky roofs, leaky and broken windows, and carpeting presence [5]. These poor housing conditions have often been the result of historical discriminatory housing policies, such as redlining, that have targeted racio-ethnic minority groups and households with low income [16,17]. As such, factors predictive of residential indoor allergens could have geospatial components, including from neighborhoods and housing.
Electronic health records (EHR) are a source of rich, individual level data useful to a wide range of research and treatment-related objectives [18–20]. EHR have recently been used to develop asthma severity phenotypes [21], detail predictors of exacerbations [22–24], and even surveil asthma prevalence on a large scale [25]. However, IHEE data in EHR are limited or nonexistent, hindering researchers’ abilities to assess these exposures for patient-level asthma outcomes. While recent studies have spatially linked EHR to environmental exposures to analyze asthma outcomes [20], no such research has included reported indoor exposure information, nor have information on IHEE been characterized in a manner that would allow for inclusion in clinical decision-making.
We linked publicly available data on the built environment, neighborhood characteristics, and housing attributes to EHR data to predict self-reported indoor allergen presence for children with asthma attending primary care and specialty care visits at a large safety net hospital in Massachusetts (MA), USA. Our study had two objectives: 1) assess the feasibility of predicting presence of cockroach, rodent (mouse or rat), and mold allergens and bedroom carpeting/rug reservoirs using primarily geospatial features; and 2) estimate effect sizes of key predictive features.
MATERIALS AND METHODS
Data
The Boston Children’s Health Disparities Repository (BCHDR) includes data for all children (ages 0 to 18) seen at Boston Medical Center (BMC) from 2003 to 2020. The BMC is located in Boston, MA, and is the largest safety net hospital in New England, primarily serving a low-income and racio-ethnically diverse population [26]. The study was approved by the Boston University Medical Campus Institutional Review Board.
BCHDR is a deidentified dataset that includes individual patient demographics (e.g., age, sex) from the EHR linked to geospatial data (e.g., housing parcel and neighborhood characteristics) via geocoded home addresses [27]. Racio-ethnic options in intake forms were White, Black, Asian, American Indian, and Hispanic. Children in the BCHDR were classified as having asthma of varying severities as described in previous work [21]. Addresses recorded at patient registration during clinical visits were geocoded in a multi-step process using Environmental Systems Research Institute’s (ESRI) ArcGIS Desktop 10.7 (Redlands, CA). Addresses were first geocoded using publicly available MassGIS master address points [28], and unmatched addresses were then geocoded against the ESRI world geocoder to increase match precision. Addresses were then attached to a geographic identifiers database which included housing parcel identifiers [29] and 2010 US Census tracts [30]. Geospatial features were at the housing (e.g., housing parcel real estate type), custom buffer (e.g., restaurants within 500 ft), or census tract level (e.g., percent of homes with gas as main heating source). Housing and neighborhood features were obtained from 5-year 2015 US Census Bureau American Community Survey (ACS) estimates [31], 2020 ESRI Business Analyst [32], the Center for Research on Environmental and Social Stressors in Housing Across the Life Course (CRESSH) [33,34], and the Massachusetts Land Parcel Database (MLPD) [35]. Implausible or noticeably incorrect values in the MA parcel data were set to missing. Afterward, approximately 10% of the 61,158,748 values for 28 MLPD features included in our analyses were missing across the 2,184,241 residential parcels in MA. To reconstruct the MLPD prior to this and future analyses, missing values were replaced by multiple imputation using random forests in the missForest package in R [36–38]. Missing values were then imputed 10 times, with the median imputed value selected for continuous attributes and the mode selected for categorical attributes.
Information on IHEE was available for only some patients and obtained from a structured Asthma Template (Figure 1) captured within the EHR from 2004–2015, which included presence of indoor allergen sources. Data from the form were included in our analytical samples if a clinician recorded data for the relevant question within the Asthma Template. Because some questions went unanswered, sample sizes consequently varied by IHEE. Radio buttons (Figure 1) and check boxes used to capture information that included options for “no” or “none” indicated interaction with specific parts of the template. The IHEE outcomes in this study were reported presence of cockroaches, rodents (mice or rats), mold, and bedroom carpeting/rugs.
Figure 1.
The in-home environmental exposures (IHEE) section of the Asthma Template used to capture data on children with asthma within the electronic health record (EHR).
Statistical Analysis
Prediction models of the reported presence of cockroaches, rodents, mold, and bedroom carpeting were built using an ensemble machine learning package in R, SuperLearner, which stacks and weights over 40 candidate machine learning algorithms to build the most accurate ensemble algorithm (SuperLearner algorithm) possible [39]. The SuperLearner algorithm performs asymptotically as good or better than any single algorithm; consequently, it yields the closest respective approximation to the real data generating mechanism, making it robust to model misspecification [39,40]. The prediction models were developed with 5-fold cross validation, and criteria for including algorithms were convergence and either improvement of classification accuracy by area under the receiver operating characteristic curve (AUC, ROC) or improvement of algorithmic diversity (e.g., regression, random forests, boosting, bagging, etc.). Inclusion of a diverse ensemble of algorithms can best approximate true exposure-outcome relationships [41]. AUC values can range from 0.5 (chance) to 1.0 (perfect classifier), but interpretation is dependent on context. For creation of a predictive diagnostic test, AUCs in the range of 0.7–0.8 are barely acceptable, while the same range in other fields, such as psychology, are considered strongly predictive [42]. For algorithms that performed poorly on the full set of features (p=86), we reassessed them after reducing feature dimensionality using screening functions available in SuperLearner. We furthermore tested prediction accuracy on subsets of features after screening, as well as subsets of only features with significant effects in subsequent analyses.
Prior to effect estimation, we first reduced dimensionality by an elastic net regression screening algorithm (screen.glmnet) in the SuperLearner package. All continuous features and confounders were then standardized to z-scores. To estimate continuous average causal effects (ACE) of key geospatial features in each model, we simulated exposure scenarios using our fitted models and a maximum likelihood-based substitution estimator from a causal inference framework, the G-formula, available in the “expose” package in R [43–45]. For each respective outcome, we kept the minimum set of features that provided AUCs consistent with the full predictive models. To focus effect estimation on geospatial features, we considered individual demographic features (e.g., age, sex) and location (e.g., latitude, longitude) as confounders, adjusting our models for them accordingly. Effects were estimated by simulating a change in z-score from zero to one, equivalent to an increase of one standard deviation (SD) in each standardized feature for continuous features, or the effect of having versus not having the respective indicator for binary features. Confidence intervals were estimated by bootstrapping (n=50). Statistical analyses are presented conceptually in Figure 2.
Figure 2.
Conceptual analytic approach.
a MLPD: Massachusetts Land Parcel Database
b BMC EHR: Boston Medical Center electronic health records
c ACS: American Community Survey
d ESRI: Environmental Systems Research Institute
e CRESSH: Center for Research on Environmental and Social Stressors in Housing Across the Life Course
f ROC AUC: area under receiver operating characteristics curve
1 missForest: R-package for random forests
2 SuperLearner: R-package for SuperLearner ensemble machine learning prediction algorithm
3 nnls: R-package for non-negative least squares
4 ranger: R-package for random forests and high dimensional data
5 randomForest: R-package for random forests
6 bartMachine: R-package for Bayesian additive regression trees
7 ipredbagg: function in R-package “ipred” for indirect classification and bagging
8 xgboost: R-package for extreme gradient boosting
9 glmnet: R-package for lasso and elastic-net regularized generalized linear models
10 step.forward: function in R-package SuperLearner defined for generalized linear models
11 gam: R-package for generalized additive models
12 expose: R-package for multiple effect estimation
* Run on subset of features and/or full feature set
RESULTS
Descriptive Statistics
Table 1 shows that children that answered questions about an IHEE had similar characteristics, as they were majority male, Black, and on public insurance, consistent with the population visiting the BMC. The most common parcel real estate type was duplex/triplex (36.7–38.5% of parcels) (Table 2). Cockroaches (279/1,744 [16.0%] of homes) and mold (190/583 [32.6%] of homes) were more frequently reported for children that were male, on public insurance, younger, and living in housing with higher parcel value metrics, area metrics, and certain ratio metrics (e.g., total value per acre) such as large, tax-exempt parcels. Locally, tax-exempt parcels are indicative of housing stock comprised of large public housing properties owned by housing authorities (e.g., Boston Housing Authority) (Supplementary Figures 1–3). Rodents (567/1,744 [32.5%] of homes) and bedroom carpeting/rugs (359/852 [42.1%] of homes) occurred more frequently in children’s housing parcels having lower value measures (e.g., total value), but rodents were reported on smaller parcels, as indicated by many metrics (e.g., total residential area). Bedroom/carpeting rugs were reported most in housing on larger parcels. Descriptive statistics for all 86 features are included in Supplementary Table 1.
Table 1.
Individual characteristics by reported presence of in-home environmental exposures (IHEE) for children with asthma.
| Cockroaches | Rodents | Mold | Bedroom carpeting/rugs | |||||
|---|---|---|---|---|---|---|---|---|
| n (%) | Yes (% or sd) | n (%) | Yes (% or sd) | n (%) | Yes (% or sd) | n (%) | Yes (% or sd) | |
| Feature description | n=1744 | n=279 (16.0) | n=1744 | n=567 (32.5) | n=583 | n=190 (32.6) | n=852 | n=359 (42.1) |
| Sex | ||||||||
| Male | 979 (56.1) | 162 (16.5) | 979 (56.1) | 329 (33.6) | 320 (54.9) | 108 (33.8) | 471 (55.3) | 199 (42.3) |
| Female | 765 (43.9) | 117 (15.3) | 765 (43.9) | 238 (31.1) | 263 (45.1) | 82 (31.2) | 381 (44.7) | 160 (42.0) |
| Race | ||||||||
| Black | 1009 (57.9) | 158 (15.7) | 1009 (57.9) | 365 (36.2) | 381 (65.4) | 129 (33.9) | 515 (60.4) | 225 (43.7) |
| Other | 434 (24.9) | 80 (18.4) | 434 (24.9) | 124 (28.6) | 117 (20.1) | 31 (26.5) | 179 (21.0) | 60 (33.5) |
| Hispanic | 141 (8.1) | 23 (16.3) | 141 (8.1) | 40 (28.4) | 32 (5.5) | 12 (37.5) | 68 (8.0) | 34 (50.0) |
| White | 115 (6.6) | 10 (8.7) | 115 (6.6) | 20 (17.4) | 34 (5.8) | 14 (41.2) | 63 (7.4) | 33 (52.4) |
| Asian | 45 (2.6) | 8 (17.8) | 45 (2.6) | 18 (40.0) | 19 (3.3) | 4 (21.1) | 27 (3.2) | 7 (25.9) |
| Language | ||||||||
| English | 1373 (78.7) | 203 (14.8) | 1373 (78.7) | 445 (32.4) | 493 (84.6) | 163 (33.1) | 697 (81.8) | 304 (43.6) |
| Non-English | 371 (21.3) | 76 (20.5) | 371 (21.3) | 122 (32.9) | 90 (15.4) | 27 (30.0) | 155 (18.2) | 55 (35.5) |
| Insurance Type | ||||||||
| Public/Other | 1416 (81.2) | 241 (17.0) | 1416 (81.2) | 471 (33.3) | 468 (80.3) | 162 (34.6) | 673 (79.0) | 285 (42.3) |
| Commercial/Private | 328 (18.8) | 38 (11.6) | 328 (18.8) | 96 (29.3) | 115 (19.7) | 28 (24.3) | 179 (21.0) | 74 (41.3) |
| Age | 1744 | 7.17 (4.39) | 1744 | 7.34 (4.44) | 583 | 7.50 (4.65) | 852 | 7.45 (4.64) |
Continuous values showing mean (standard deviation, sd); otherwise, categorical count (percent of total, %)
Table 2.
Key housing and neighborhood characteristics by reported presence of in-home environmental exposures (IHEE) for children with asthma.
| Cockroaches | Rodents | Mold | Bedroom carpeting/rugs | |||||
|---|---|---|---|---|---|---|---|---|
| n (%) | Yes (% or sd) | n (%) | Yes (% or sd) | n (%) | Yes (% or sd) | n (%) | Yes (% or sd) | |
| Feature description | n=1744 | n=279 (16.0) | n=1744 | n=567 (32.5) | n=583 | n=190 (32.6) | n=852 | n=359 (42.1) |
| Housing parcel | ||||||||
| Affordable housing | ||||||||
| Yes | 1320 (75.7) | 193 (14.6) | 1320 (75.7) | 441 (33.4) | 433 (74.3) | 131 (30.3) | 655 (76.9) | 279 (42.6) |
| No | 424 (24.3) | 86 (20.3) | 424 (24.3) | 126 (29.7) | 150 (25.7) | 59 (39.3) | 197 (23.1) | 80 (40.6) |
| Total value (by $1M) | 1744 | 131.70 (478.01) | 1744 | 50.84 (293.65) | 583 | 103.30 (458.12) | 852 | 60.99 (414.96) |
| Total residential area (by 1000 sq ft) | 1744 | 50.12 (97.07) | 1744 | 29.18 (76.00) | 583 | 49.26 (124.92) | 852 | 39.49 (131.06) |
| Year built | 1744 | 1928 (32) | 1744 | 1926 (31) | 583 | 1932 (35) | 852 | 1934 (39) |
| % impervious (% building + % pavement) | 1744 | 74.41 (19.40) | 1744 | 72.54 (19.34) | 583 | 67.70 (21.70) | 852 | 67.07 (23.06) |
| Total value (by $1M) per ac | 1744 | 131.70 (478.01) | 1744 | 50.84 (293.65) | 583 | 103.30 (458.12) | 852 | 60.99 (414.96) |
| Real estate type (1–12) | ||||||||
| Single family home | 179 (10.3) | 17 (9.5) | 179 (10.3) | 44 (24.6 | 62 (10.6) | 20 (32.3) | 98 (11.5) | 47 (48.0) |
| Duplex or triplex | 640 (36.7) | 93 (14.5) | 640 (36.7) | 229 (35.8) | 219 (37.6) | 60 (27.4) | 328 (38.5) | 118 (36.0) |
| Small apartment (<=8 units) | 113 (6.5) | 18 (15.9) | 113 (6.5) | 41 (36.3) | 38 (6.5) | 11 (28.9) | 58 (6.8) | 24 (41.4) |
| Large apartment (>8 units) | 410 (23.5) | 53 (12.9) | 410 (23.5) | 132 (32.2) | 138 (23.7) | 45 (32.6) | 189 (22.2) | 93 (49.2) |
| Mixed use: more than 1/2 residential | 1 (0.1) | 0 (0.0) | 1 (0.1) | 1 (100.0) | 1 (0.2) | 1 (100.0) | 1 (0.1) | 1 (100.0) |
| Mixed use: more than 1/2 commercial | 111 (6.4) | 16 (14.4) | 111 (6.4) | 41 (36.9) | 33 (5.7) | 13 (39.4) | 49 (5.8) | 29 (59.2) |
| Tax exempt | 260 (14.9) | 76 (29.2) | 260 (14.9) | 71 (27.3) | 88 (15.1) | 38 (43.2) | 118 (13.8) | 41 (34.7) |
| Other | 30 (1.7) | 6 (20.0) | 30 (1.7) | 8 (26.7) | 4 (0.7) | 2 (50.0) | 11 (1.3) | 6 (54.5) |
| Longitude (decimal degrees) | 1744 | −71.07 (0.04) | 1744 | −71.08 (0.06) | 583 | −71.07 (0.04) | 852 | −71.08 (0.08) |
| Latitude (decimal degrees) | 1744 | 42.31 (0.09) | 1744 | 42.31 (0.07) | 583 | 42.30 (0.10) | 852 | 42.29 (0.12) |
| Custom buffer | ||||||||
| Grocery stores within 1000 ft | 1744 | 0.49 (0.90) | 1744 | 0.42 (0.77) | 583 | 0.26 (0.59) | 852 | 0.33 (0.76) |
| Restaurants within 1000 ft | 1744 | 2.03 (3.01) | 1744 | 1.90 (2.86) | 583 | 2.11 (5.25) | 852 | 1.88 (3.84) |
| Census tract | ||||||||
| Median household income (by $1k) | 1744 | 42.69 (19.46) | 1744 | 44.93 (20.21) | 583 | 43.93 (19.97) | 852 | 49.48 (24.71) |
| % habitable structures vacant | 1744 | 7.73 (3.20) | 1744 | 8.29 (3.37) | 583 | 7.95 (3.18) | 852 | 7.58 (3.53) |
| % habitable structures single family house | 1744 | 11.71 (14.66) | 1744 | 13.02 (15.36) | 583 | 14.26 (16.43) | 852 | 19.18 (21.83) |
| % habitable structures duplex | 1744 | 14.50 (8.87) | 1744 | 15.62 (8.60) | 583 | 14.23 (9.74) | 852 | 14.77 (9.58) |
| % habitable structures tri/quad-plex | 1744 | 30.41 (16.94) | 1744 | 33.65 (17.26) | 583 | 29.90 (17.34) | 852 | 26.83 (18.35) |
| % habitable structures 5–9 units | 1744 | 12.11 (10.24) | 1744 | 10.27 (7.82) | 583 | 11.25 (9.58) | 852 | 8.93 (7.02) |
| % habitable structures 10–19 units | 1744 | 8.42 (7.41) | 1744 | 6.88 (6.69) | 583 | 7.95 (7.77) | 852 | 6.89 (6.36) |
| % habitable structures >20 units | 1744 | 14.53 (15.16) | 1744 | 12.64 (14.09) | 583 | 14.28 (16.21) | 852 | 15.10 (16.43) |
| % habitable structures built before 1939 | 1744 | 49.54 (20.04) | 1744 | 52.51 (18.79) | 583 | 48.85 (18.82) | 852 | 46.74 (22.08) |
| % occupied housing renters | 1744 | 70.98 (17.46) | 1744 | 69.05 (17.15) | 583 | 69.11 (18.22) | 852 | 64.19 (20.86) |
| % occupied housing ≥1 people per room | 1744 | 4.80 (3.92) | 1744 | 4.44 (3.51) | 583 | 3.65 (2.51) | 852 | 4.03 (3.42) |
| Average people per household | 1744 | 2.70 (0.43) | 1744 | 2.74 (0.41) | 583 | 2.67 (0.39) | 852 | 2.65 (0.40) |
| Heat vulnerability index (greater values are more vulnerable) | 1744 | 2.46 (1.48) | 1744 | 2.26 (1.44) | 583 | 2.16 (1.39) | 852 | 1.78 (1.66) |
| Normalized difference vegetation index | 1744 | 0.36 (0.08) | 1744 | 0.37 (0.08) | 583 | 0.38 (0.09) | 852 | 0.40 (0.12) |
| Residential segregation: black (negative values less, positive values more) | 1744 | −0.13 (0.47) | 1744 | −0.19 (0.48) | 583 | −0.12 (0.49) | 852 | −0.02 (0.54) |
| Economic segregation (negative values less, positive values more) | 1744 | −0.20 (0.24) | 1744 | −0.18 (0.23) | 583 | −0.19 (0.24) | 852 | −0.13 (0.26) |
Continuous values showing mean (standard deviation, sd); otherwise, categorical count (percent of total, %)
Indoor allergen prediction
Figure 3 shows that models can predict cockroach, rodent, and bedroom rugs/carpeting with reasonable accuracy (AUCs: 0.63–0.65). For example, for the cockroach prediction model, results suggest that there is a 64.9% chance that the prediction model will accurately distinguish cockroach presence/absence for a randomly selected set of values for the 86 features included in the model. Accuracy of mold presence classification was poor (AUC: 0.54), and thus was excluded from further effect estimation. While AUC values generally improved by adding more geospatial features (though dependent on each model), each model’s performance tended to asymptotically approach the AUCs reported in the final prediction models using the full feature set. Algorithm-specific performance within each ensemble model are presented in supplemental material (Supplementary Figures 4–7).
Figure 3.
Area under receiver operating curves for fitted ensemble machine learning models of presence of A) cockroaches, B) rodents, C) mold, and D) bedroom carpet/rug on full multilevel feature set (p=86).
Risk factor effect estimates
For reported cockroach presence, 15 geospatial features were selected (Figure 4): eight from housing parcel data, and seven from census tract data. Significant features included percent of housing parcel area comprised of a building footprint (percent building) (ACE: 0.032; 95% CI: 0.003, 0.060), housing parcel real estate type tax exempt (dichotomous ACE: 0.041; 95% CI: 0.012, 0.070), census tract average household size (continuous ACE: 0.017; 95% CI: 0.002, 0.031), and census tract percent occupied structures with 10–19 units (continuous ACE: 0.018; 95% CI: 0.000, 0.036).
Figure 4.
Average causal effect estimates for cockroach presence for a simulated increase in continuous feature z-score from zero to one or binary absence-presence, adjusted for individual confounders sex, race, primary language, insurance type, age, latitude, and longitude.
Prefix P:* indicates housing parcel level feature.
Prefix T:* indicates census tract level feature.
For reported rodent presence, 17 geospatial features were selected (Figure 5), two at the custom buffer level, eight at housing parcel level, and seven at census tract level. Census tract White residential segregation (value >0.3 of White/Black residential segregation index ranging from −1 [Black residential segregation] to 1 [White residential segregation]) had a statistically significant negative estimate (dichotomous ACE: −0.019; 95% CI: −0.038, −0.000), and census tract average household size (continuous ACE: 0.021; 95% CI: 0.004, 0.0038) had a statistically significant positive association.
Figure 5.
Average causal effect estimates for rodent presence for a simulated increase in continuous feature z-score from zero to one or binary absence-presence, adjusted for individual confounders sex, race, primary language, insurance type, age, latitude, and longitude.
Prefix B:* indicates custom buffer level feature.
Prefix P:* indicates housing parcel level feature.
Prefix T:* indicates census tract level feature.
For reported bedroom rug/carpeting presence (Figure 6), 18 geospatial features were selected, eight at the housing parcel level and 10 at the census tract level. Housing parcel real estate mixed use – predominantly commercial (dichotomous ACE: −0.032; 95% CI: −0.065, −0.000) and housing parcel year built (continuous ACE: 0.043; 95% CI: 0.004, 0.081) had statistically significant positive associations.
Figure 6.
Average causal effect estimates for bedroom carpeting/rug presence for a simulated increase in continuous feature z-score from zero to one or binary absence-presence, adjusted for individual confounders sex, race, primary language, insurance type, age, latitude, and longitude.
Prefix P:* indicates housing parcel level feature.
Prefix T:* indicates census tract level feature.
DISCUSSION
We used ensemble machine learning methods, EHR data, and spatially resolved housing and neighborhood geospatial features to develop prediction models for the reported presence of IHEE among children with asthma. The prediction models leveraged partial information capture in the EHR, which can be used in the future to estimate IHEE for a much larger clinical cohort, as well as to generate hypotheses about why specific children were at higher risk of exposure to IHEE. Linking environmental exposures to EHR can potentially lead to novel epidemiologic analyses [20], but no previous research has aimed to predict IHEE among an EHR cohort with objective clinical measurements (e.g., spirometry) in the same population of children with asthma.
Social determinants of health, such as structural racism, are key drivers of environmental injustices in housing and neighborhoods that cause disparities in multiple exposures and associated asthma outcomes [17,46,47]. Racist policies like redlining resulted in historical disinvestment in housing, decreasing its value, quality, and stability, as well as neighborhood resources, among non-White groups of people [48]. Predicting allergen presence for people with asthma has been limited to date, especially on a population scale. A large-scale study linked housing characteristics, such as higher housing code violation density, to increased asthma emergency department (ED) visits [49], but presence of allergens was only posited and not measured or reported directly. Smaller scale studies predicted allergen presence by analyzing indoor dust [50] and housing characteristics [51]. However, the exposures in these studies were not able to be linked with lung function or other asthma-related outcomes within an EHR cohort.
Cockroach allergen is a key environmental trigger for people with asthma [5], a disease that is disproportionately impacts groups having lower socioeconomic status [52]. We estimated positive relationships between reported cockroach presence and the proportion of a parcel taken up by the building footprint (housing parcel percent building), tax exempt housing parcels, percent of structures with 10–19 units in a census tract, and census tract average household size. These features are indicative of children with asthma living in medium-large and public housing complexes and similar areas. Even though integrated pest management (IPM) programs have been found highly effective at reducing pests and subsequent allergen presence in public housing [53,54], and the Boston Housing Authority has IPM educational material and services available to residents [55], such housing was still associated with greater reported frequency of cockroaches present. Similar to other researchers that found that cockroach allergic sensitivity was positively associated with the number of people per room in the National Health and Nutrition Examination Survey [56], we found a positive association between area level average household occupancy and reported cockroach presence. Given that census tract average household size was positively correlated with both single-family housing and features such as census tract percent houses vacant (Supplementary Figure 1), reports of cockroaches present may also be influenced by crowding within single family homes, as well as by neighborhoods with many vacant homes. Cockroach presence was previously associated with housing in poor condition, open-style kitchens, material hardship, rented housing, area poverty, and area housing code violations [16,50,57–60], but we were not able to measure such features.
Similar to cockroach allergens, mouse and rat allergens are well documented asthma triggers [6,7] that have been linked with homes in poor condition and areas with high poverty and housing code violations [16,59–61]. As average household size was positively associated with reported cockroach presence, so too was it positively associated with reported rodent presence. Larger average household sizes in neighborhoods may act as an indicator of reports of common household pests that generate allergens indoors, which are triggers for children with asthma. We further found that census tracts that were residentially White-segregated, or neighborhoods with mostly White residents, were associated with fewer reports of rodents present in this study population, indicating the effects of current and historical legacies of discrimination in housing and related policies [48].
Rugs and carpeting are constructed of similar material, methods, and forms [62], and both are reservoirs for allergens [63]. Positive associations were found with parcel real estate type mixed use – predominantly commercial and year built. As such, mixed use parcels and parcels with newer structures tended to have more reports of carpeting or rugs in bedrooms. While research on “healthy buildings” in mixed-use communities has identified flooring as one of many components to consider for improving indoor air quality [64], no previous research to date has linked mixed-use building developments to flooring type, particularly in the bedroom.
The model of reported mold presence performed poorly. Molds and fungi require cellulose and moisture to grow, and therefore are affected indoors by ventilation and building enclosures [65]. While the ventilation and building enclosures would theoretically have geographic dimensions potentially able to be captured by the proxy geospatial features we included (e.g., parcel housing age), there are also resident behavioral dimensions to ventilation (e.g., opening windows, using exhaust fans, presence of central heating, ventilation, and air conditioning) that we were unable to capture.
There is growing recognition of the limits of EHR data in areas like assessment of IHEE since the data are often not captured systematically within the systems and subject to informative presence biases [66]. But, the approach used in this study can leverage short term surveys and partial information in EHR to predict the likelihood of IHEE for other patients of the same catchment population. This approach could guide development of clinical tools that automate predictions within the EHR using an evidence-based system to prompt clinicians to record additional patient level IHEE information. The additional patient-specific information could then theoretically guide clinical decision making for determining options for exposure reduction or prevention, alongside drug-based treatments. Mitigating IHEE has been shown to be an effective preventive strategy among people with asthma [10,11], and these results further support adding preventive components to holistic treatment of asthma by “prescribing” exposure reduction or prevention, such as an IPM plan or hard surface flooring. Linked indoor exposure information could be highly useful to epidemiologic research, such as analyses of IHEE and lung function. This approach is not limited only to studying asthma, as it could be tailored to other EHR health endpoints such as injuries, obesity, and diabetes that may be affected by complex mixtures of geographical factors comprising patients’ surrounding social and physical environments. Although the population used in this study was mostly disadvantaged and therefore unable to be contrasted with higher socioeconomic groups, it is also a vulnerable population that disproportionately experiences health disparities and for whom addressing modifiable factors in the home environment is a high priority.
While allergen presence outcomes were self-reported and may not correspond precisely with measured allergen concentrations, this is reflective of information available in a clinical encounter and able to be gathered within EHR; future research could help determine associations with measured concentrations and inform the best approaches to ascertain IHEE information. A limitation of using publicly available geographic data as features was the inability to specifically measure many salient spatio-temporally varying influences, such as housing condition, IPM programs, trigger reduction education, and personal behaviors, such as housekeeping. However, these are difficult to measure systematically on a population scale. Our study was cross-sectional and did not capture temporal variation. However, by integrating complex multilevel data in our models, we were able to account for some of the spatial structure of exposures at varying geographic scales.
CONCLUSION
In this study, we innovatively predicted reported presence of IHEE known to be asthma triggers using partial information capture in EHR linked to multilevel geospatial features among children visiting a large safety net hospital. Prediction models can identify patients in the population who did not get asked about indoor allergens but who are at higher risk of IHEE, providing additional information to improve clinical care. Results from such approaches can further facilitate clinical follow-up, exposure assessment, and epidemiologic hypothesis generation and analyses. The approach using ensemble machine learning and causal inference mixture methods can be tailored and applied to other health outcomes in EHR.
Supplementary Material
ACKNOWLEDGEMENTS AND FUNDING
This work was supported by grant R01 ES027816 from the National Institute of Environmental Health Sciences (NIEHS), National Institutes of Health (NIH), the National Center for Advancing Translational Sciences (NCATS, NIH) and Boston University Clinical and Translational Science Institute (CTSI) grant 1UL1TR001430, and training grants from National Science Foundation NRT (DGE 1735087) and NIEHS T32 (T32 ES014562). This research was enabled by resources and personnel from the Biostatistics and Epidemiology Data Analytics Center (BEDAC) at Boston University.
LIST OF ABBREVIATIONS AND ACRONYMS
- ACE
average causal effect
- ACS
American Community Survey
- AUC
area under the curve
- BCHDR
Boston Children’s Health Disparities Repository
- BMC
Boston Medical Center
- CRESSH
Center for Research on Environmental and Social Stressors in Housing Across the Life Course
- ED
emergency department
- EHR
electronic health records
- ESRI
Environmental Systems Research Institute
- IHEE
in-home environmental exposures
- IPM
integrated pest management program
- MA
Massachusetts
- MLPD
Massachusetts Land Parcel Database
- ROC
receiver operating characteristic curve
- SD
standard deviation
- USA
United States of America
Footnotes
The authors have no competing interests to declare.
REFERENCES
- [1].U.S. EPA. Child-Specific Exposure Factors Handbook. 2008.
- [2].Matsui EC. Environmental exposures and asthma morbidity in children living in urban neighborhoods. Allergy 2014;69:553–8. 10.1111/all.12361. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [3].Schulten V, Frazier A, Calatroni A, Kattan M, Bacharier LB, O’Connor GT, et al. The association of allergic sensitization patterns in early childhood with disease manifestations and immunological reactivity at 10 years of age. Clin Exp Allergy 2019;49:cea.13406. 10.1111/cea.13406. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [4].Fitzpatrick AM, Teague WG, Meyers DA, Peters SP, Li X, Li H, et al. Heterogeneity of severe asthma in childhood: Confirmation by cluster analysis of children in the National Institutes of Health/National Heart, Lung, and Blood Institute Severe Asthma Research Program. J Allergy Clin Immunol 2011;127:382–389.e13. 10.1016/j.jaci.2010.11.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [5].Halken S Environmental causes of asthma in children. Pediatr Allergy Immunol 1994;5:57–60. 10.1111/j.1399-3038.1994.tb00350.x. [DOI] [PubMed] [Google Scholar]
- [6].Matsui EC, Eggleston PA, Buckley TJ, Krishnan JA, Breysse PN, Rand CS, et al. Household mouse allergen exposure and asthma morbidity in inner-city preschool children. Ann Allergy, Asthma Immunol 2006;97:514–20. 10.1016/S1081-1206(10)60943-X. [DOI] [PubMed] [Google Scholar]
- [7].Perry T, Matsui E, Merriman B, Duong T, Eggleston PA. The prevalence of rat allergen in inner-city homes and its relationship to sensitization and asthma morbidity. J Allergy Clin Immunol 2003;112:346–52. 10.1067/mai.2003.1640. [DOI] [PubMed] [Google Scholar]
- [8].Nwaru BI, Suzuki S, Ekerljung L, Sjölander S, Mincheva R, Rönmark EP, et al. Furry Animal Allergen Component Sensitization and Clinical Outcomes in Adult Asthma and Rhinitis. J Allergy Clin Immunol Pract 2019;7:1230–1238.e4. 10.1016/j.jaip.2018.12.018. [DOI] [PubMed] [Google Scholar]
- [9].Portnoy JM, Jara D. Mold allergy revisited. Ann Allergy, Asthma Immunol 2015;114:83–9. 10.1016/j.anai.2014.10.004. [DOI] [PubMed] [Google Scholar]
- [10].D’Amato G, Ortega OPM, Annesi-Maesano I, D’Amato M. Prevention of Allergic Asthma with Allergen Avoidance Measures and the Role of Exposome. Curr Allergy Asthma Rep 2020;20. 10.1007/s11882-020-0901-3. [DOI] [PubMed] [Google Scholar]
- [11].Cipriani F, Calamelli E, Ricci G. Allergen avoidance in allergic asthma. Front Pediatr 2017;5. 10.3389/fped.2017.00103. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].U.S. Department of Housing and Urban Development. The Healthy Homes Program Guidance Manual 2012.
- [13].Rosofsky A, Reid M, Sandel M, Zielenbach M, Murphy J, Scammell MK. Breathe Easy at Home: A Qualitative Evaluation of a Pediatric Asthma Intervention. Glob Qual Nurs Res 2016;3. 10.1177/2333393616676154. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [14].Ciaccio CE, Kennedy K, Portnoy JM. A new model for environmental assessment and exposure reduction. Curr Allergy Asthma Rep 2012;12:650–5. 10.1007/s11882-012-0297-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [15].Lewis RD, Breysse PN, Lees PSJ, Diener-West M, Hamilton RG, Eggleston P. Factors Affecting the Retention of Dust Mite Allergen on Carpet. Am Ind Hyg Assoc J 1998;59:606–13. 10.1080/15428119891010776. [DOI] [PubMed] [Google Scholar]
- [16].Call RS, Smith TF, Morris E, Chapman MD, Platts-Mills TAE. Risk factors for asthma in inner city children. J Pediatr 1992;121:862–6. 10.1016/S0022-3476(05)80329-4. [DOI] [PubMed] [Google Scholar]
- [17].Bryant-Stephens TC, Strane D, Robinson EK, Bhambhani S, Kenyon CC. Housing and asthma disparities. J Allergy Clin Immunol 2021;148:1121–9. 10.1016/j.jaci.2021.09.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [18].Cowie MR, Blomster JI, Curtis LH, Duclaux S, Ford I, Fritz F, et al. Electronic health records to facilitate clinical research. Clin Res Cardiol 2017;106:1–9. 10.1007/s00392-016-1025-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [19].Casey JA, Schwartz BS, Stewart WF, Adler NE. Using Electronic Health Records for Population Health Research: A Review of Methods and Applications. Annu Rev Public Health 2016;37:61–81. 10.1146/annurev-publhealth-032315-021353. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [20].Schinasi LH, Auchincloss AH, Forrest CB, Diez Roux AV. Using electronic health record data for environmental and place based population health research: a systematic review. Ann Epidemiol 2018;28:493–502. 10.1016/j.annepidem.2018.03.008. [DOI] [PubMed] [Google Scholar]
- [21].Peer K, Adams WG, Legler A, Sandel M, Levy JI, Boynton-Jarrett R, et al. Developing and evaluating a pediatric asthma severity computable phenotype derived from electronic health records. J Allergy Clin Immunol 2021:1–9. 10.1016/j.jaci.2020.11.045. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [22].Martin A, Bauer V, Datta A, Masi C, Mosnaim G, Solomonides A, et al. Development and validation of an asthma exacerbation prediction model using electronic health record (EHR) data. J Asthma 2020;57:1339–46. 10.1080/02770903.2019.1648505. [DOI] [PubMed] [Google Scholar]
- [23].Greenblatt RE, Zhao EJ, Henrickson SE, Apter AJ, Hubbard RA, Himes BE. Factors associated with exacerbations among adults with asthma according to electronic health record data. Asthma Res Pract 2019;5:1–11. 10.1186/s40733-019-0048-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [24].Sundaresan AS, Schneider G, Reynolds J, Kirchner HL. Identifying Asthma Exacerbation-Related Emergency Department Visit Using Electronic Medical Record and Claims Data. Appl Clin Inform 2018;9:528–40. 10.1055/s-0038-1666994. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [25].Tarabichi Y, Goyden J, Liu R, Lewis S, Sudano J, Kaelber DC. A step closer to nationwide electronic health record-based chronic disease surveillance: Characterizing asthma prevalence and emergency department utilization from 100 million patient records through a novel multisite collaboration. J Am Med Informatics Assoc 2020;27:127–35. 10.1093/jamia/ocz172. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [26].Boston Medical Center. Office of Minority Physician Recruitment 2021. https://www.bmc.org/medical-professionals/office-minority-physician-recruitment (accessed April 10, 2021).
- [27].Milando CW, Fabian MP. Using public data to customize multi-zone housing models of indoor exposure to air pollutants for a cohort of pediatric patients with asthma. 2020 Expo. Symp, New York, NY: 2020. [Google Scholar]
- [28].MassGIS. Massachusetts Master Address Points 2019. https://www.arcgis.com/home/item.html?id=4743bdb1af9b4787978f4b283a9913ba.
- [29].Metropolitan Area Planning Council (MAPC). Land Parcel Database 2019. https://www.mapc.org/learn/data/#landparceldb (accessed January 4, 2021).
- [30].US Census Bureau. 2010 US Census 2010.
- [31].US Census Bureau. 2011–2015 American Community Survey 5-Year Estimates 2015.
- [32].Environmental Systems Research Institute (ESRI). ESRI Buesiness Analyst 2020.
- [33].Center for Research on Environmental and Social Stressors in Housing Across the Life Course (CRESSH). CRESSH MAP-EHD Indexes. 2019. n.d.
- [34].Krieger N, Feldman JM, Waterman PD, Chen JT, Coull BA, Hemenway D. Local Residential Segregation Matters: Stronger Association of Census Tract Compared to Conventional City-Level Measures with Fatal and Non-Fatal Assaults (Total and Firearm Related), Using the Index of Concentration at the Extremes (ICE) for Racial, Econ. J Urban Heal 2017;94:244–58. 10.1007/s11524-016-0116-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [35].Metropolitan Area Planning Council. Land Parcel Database 2019.
- [36].Stekhoven DJ, Bühlmann P. Missforest-Non-parametric missing value imputation for mixed-type data. Bioinformatics 2012;28:112–8. 10.1093/bioinformatics/btr597. [DOI] [PubMed] [Google Scholar]
- [37].Stekhoven DJ. missForest: Nonparametric missinv value imputation using random forest 2013.
- [38].R Core Team. R: A language and environment for statistical computing 2020.
- [39].Van Der Laan MJ, Polley EC, Hubbard AE. Super learner. Stat Appl Genet Mol Biol 2007;6. 10.2202/1544-6115.1309. [DOI] [PubMed] [Google Scholar]
- [40].Pirracchio R, Petersen ML, Van Der Laan M. Improving propensity score estimators’ robustness to model misspecification using Super Learner. Am J Epidemiol 2015;181:108–19. 10.1093/aje/kwu253. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [41].Oulhote Y, Coull B, Bind M-A, Debes F, Nielsen F, Tamayo I, et al. Joint and independent neurotoxic effects of early life exposures to a chemical mixture. Environ Epidemiol 2019;3:e063. 10.1097/ee9.0000000000000063. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [42].Rice ME, Harris GT. Comparing effect sizes in follow-up studies: ROC area, Cohen’s d, and r. Law Hum Behav 2005;29:615–20. 10.1007/s10979-005-6832-7. [DOI] [PubMed] [Google Scholar]
- [43].Oulhote Y, Bind M-A, Coull B, Patel C, Grandjean P. Combining ensemble learning techniques and G-computation to investigate chemical mixtures in environmental epidemiology studies. BioRxiv 2017. https://doi.org/ 10.1101/147413. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [44].Hernan MA. Estimating causal effects from epidemiological data. J Epidemiol Community Heal 2006;60:578–86. 10.1101/147413. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [45].Robins J A new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor effect. Math Model 1986;7:1393–512. 10.1016/0270-0255(86)90088-6. [DOI] [Google Scholar]
- [46].Khan S, Bajwa S, Brahmbhatt D, Lovinsky-Desir S, Sheffield PE, Stingone JA, et al. Multi-Level Socioenvironmental Contributors to Childhood Asthma in New York City: a Cluster Analysis. J Urban Heal 2021;98:700–10. 10.1007/s11524-021-00582-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [47].Grant T, Croce E, Matsui EC. Asthma and the social determinants of health. Ann Allergy, Asthma Immunol 2022;128:5–11. 10.1016/j.anai.2021.10.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [48].Swope CB, Hernández D. Housing as a determinant of health equity: A conceptual model. Soc Sci Med 2019;243:112571. 10.1016/j.socscimed.2019.112571. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [49].Beck AF, Huang B, Chundur R, Kahn RS. Housing code violation density associated with emergency department and hospital use by children with asthma. Health Aff 2014;33:1993–2002. 10.1377/hlthaff.2014.0496. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [50].Perry TT, Wood RA, Matsui EC, Curtin-Brosnan J, Rand C, Eggleston PA. Room-specific characteristics of suburban homes as predictors of indoor allergen concentrations. Ann Allergy, Asthma Immunol 2006;97:628–35. 10.1016/S1081-1206(10)61092-7. [DOI] [PubMed] [Google Scholar]
- [51].Wilson J, Dixon SL, Breysse P, Jacobs D, Adamkiewicz G, Chew GL, et al. Housing and allergens: A pooled analysis of nine US studies. Environ Res 2010;110:189–98. 10.1016/j.envres.2009.10.007. [DOI] [PubMed] [Google Scholar]
- [52].Kattan M, Mitchell H, Eggleston P, Gergen P, Crain E, Redline S, et al. Characteristics of inner-city children with asthma: The National Cooperative Inner-City Asthma Study. Pediatr Pulmonol 1997;24:253–62. . [DOI] [PubMed] [Google Scholar]
- [53].Zha C, Wang C, Buckley B, Yang I, Wang D, Eiden AL, et al. Pest Prevalence and Evaluation of Community-Wide Integrated Pest Management for Reducing Cockroach Infestations and Indoor Insecticide Residues. J Econ Entomol 2018;111:795–802. 10.1093/jee/tox356. [DOI] [PubMed] [Google Scholar]
- [54].Kass D, McKelvey W, Carlton E, Hernandez M, Chew G, Nagle S, et al. Effectiveness of an integrated pest management intervention in controlling cockroaches, mice, and allergens in New York City public housing. Environ Health Perspect 2009;117:1219–25. 10.1289/ehp.0800149. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [55].Boston Housing Authority. Keep It Pest Free 2021. https://www.bostonhousing.org/en/For-Public-Housing/Your-Healthy-Home/Keep-It-Pest-Free.aspx.
- [56].Stevenson LA, Gergen PJ, Hoover DR, Rosenstreich D, Mannino DM, Matte TD. Sociodemographic correlates of indoor allergen sensitivity among United States children. J Allergy Clin Immunol 2001;108:747–52. 10.1067/mai.2001.119410. [DOI] [PubMed] [Google Scholar]
- [57].Olmedo O, Goldstein IF, Acosta L, Divjan A, Rundle AG, Chew GL, et al. Neighborhood differences in exposure and sensitization to cockroach, mouse, dust mite, cat, and dog allergens in New York City. J Allergy Clin Immunol 2011;128:284–92. 10.1016/j.jaci.2011.02.044. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [58].Chew GL. Assessment of environmental cockroach allergen exposure. Curr Allergy Asthma Rep 2012;12:456–64. 10.1007/s11882-012-0287-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [59].Camacho-Rivera M, Kawachi I, Bennett GG, Subramanian SV. Associations of neighborhood concentrated poverty, neighborhood racial/ethnic composition, and indoor allergen exposures: A cross-sectional analysis of Los Angeles households, 2006–2008. J Urban Heal 2014;91:661–76. 10.1007/s11524-014-9872-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [60].Rosenfeld L, Rudd R, Chew GL, Emmons K, Acevedo-García D. Are Neighborhood-Level Characteristics Associated with Indoor Allergens in the Household? J Asthma 2010;47:66–75. 10.3109/02770900903362676. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [61].Berg J, McConnell R, Milam J, Galvan J, Kotlerman J, Thorne P, et al. Rodent allergen in Los Angeles inner city homes of children with asthma. J Urban Heal 2008;85:52–61. 10.1007/s11524-007-9232-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [62].Carpet and Rug Institute. The Carpet Primer 2003.
- [63].Becher R, Øvrevik J, Schwarze PE, Nilsen S, Hongslo JK, Bakke JV. Do carpets impair indoor air quality and cause adverse health outcomes: A review. Int J Environ Res Public Health 2018;15. 10.3390/ijerph15020184. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [64].Loftness V, Hakkinen B, Adan O, Nevalainen A. Elements that contribute to healthy building design. Environ Health Perspect 2007;115:965–70. 10.1289/ehp.8988. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [65].Lstiburek J. Moisture, Building Enclosures, and Mold. HPAC Eng 2001:1–4. [Google Scholar]
- [66].McGee G, Haneuse S, Coull BA, Weisskopf MG, Rotem RS. On the nature of informative presence bias in analyses of electronic health records. Epidemiology 2021;Publish Ah. 10.1097/ede.0000000000001432. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.