The relative importance of neighborhood environment features in explaining preeclampsia risk using machine learning

Abstract

Introduction

Disentangling the role of the neighborhood environment in preeclampsia pathogenesis is crucial for addressing social and structural determinants of pregnancy health. Building on epidemiologic studies demonstrating environmental associations with preeclampsia, we used a machine learning approach to determine the relative importance of multiple simultaneous neighborhood features to facilitate prioritization of community pregnancy health initiatives.

Methods

We linked 26 features from the neighborhood environment, encompassing social vulnerability, built environment, physical environment, and health vulnerability features, to geocoded residential addresses of participants selected for a matched, nested case‐control study from two Philadelphia hospitals. We modeled individual associations of neighborhood features with preeclampsia using conditional logistic regression models. We then built XGBoost models trained on the neighborhood features predicting preeclampsia and applied explainable artificial intelligence (XAI) to disentangle the relative importance of the features associated with preeclampsia.

Results

Among 18,754 participants (4689 preeclampsia cases and 14,065 controls), we observed significant associations of neighborhood health and social vulnerability features with preeclampsia. From the XGBoost models, three neighborhood health vulnerability features—prevalence of obesity, prevalence of high blood pressure, and prevalence of short sleep duration among adults—were the most important neighborhood features in predicting preeclampsia.

Conclusion

Our findings that neighborhood features vary with respect to their relative importance in predicting preeclampsia demonstrate the value of using XAI to contribute new insights into the vulnerability of pregnant individuals to specific neighborhood environmental features and to inform policy‐making priorities for community‐level pregnancy health interventions. Specifically, the association between neighborhood hypertension prevalence and preeclampsia, suggests that communities with worse cardiovascular health have a higher preeclampsia risk, which warrants further investigation as a potential avenue for preeclampsia prevention.

1. INTRODUCTION

Pregnancies can be complicated by a myriad of adverse outcomes, leading to serious immediate and long‐term consequences for the pregnant individual and their offspring. One such severe outcome is preeclampsia, a hypertensive disorder of pregnancy (HDP) that originates from a dysfunctional placenta [1] and affects around 1 in 30 pregnancies in the United States [2] and 4 million pregnancies worldwide each year [3]. Preeclampsia is characterized by the sudden onset of hypertension and at least one associated complication after the 20th week of gestation [4] and can lead to immediate and long‐term life‐threatening complications, with the American Heart Association recognizing a history of preeclampsia as a risk factor for cardiovascular disease [5] and stroke [6]. More than 70,000 maternal deaths per year globally are attributed to preeclampsia and there are an estimated >300 million women and children at risk of future complications [3]. Many clinical and maternal risk factors have been associated with preeclampsia; however, they do not have strong predictive power [7] and studies exploring different approaches to predicting preeclampsia, including identifying genetic risk factors [8] and environmental risk factors [9, 10, 11], have not been able to precisely predict preeclampsia.

Research on environmental risk factors often focuses on one exposure at a time, as opposed to in combination, which limits the interpretability of effect estimates, as environmental stressors often coexist and can have cumulative and multiplicative harmful health effects [12, 13, 14]. Recent work has shown that community health as a whole is a significant predictor for healthy pregnancies [15], pointing to a need to pay attention to the larger community and context in which pregnant individuals live, rather than singular risk factors and exposures. Incorporating neighborhood context in association studies looking at maternal outcomes and an emphasis on the places and spaces in which pregnant individuals live is increasing, as geospatial neighborhood information becomes more readily available. Such studies have found that a pregnant individual's neighborhood context can have protective or harmful effects, mitigating or exacerbating risks of adverse pregnancy outcomes, including HDPs [16, 17, 18]. This research is especially valuable as it can lead to targeted place‐based interventions [19] to improve the health of pregnant individuals through investments in neighborhoods, which may also improve the health of communities overall.

A comprehensive characterization of the neighborhood environment and its effects on pregnancy outcomes is rendered difficult by the countless exposures that combine to create the context in which pregnant individuals live. However, machine learning approaches are ideal for studying numerous and diverse exposures as these methods benefit from the ability to incorporate high‐dimensional data from different domains and correctly handle correlations between model features (e.g., decision tree algorithms) [20], allowing for expressive and interpretable representations of complex, real‐world data. While there are many studies that have used machine learning approaches to predict preeclampsia using clinical and maternal features [21, 22, 23], we are not aware of any that have used solely neighborhood features to predict preeclampsia. Here, we use a machine learning‐based approach to enable the generation of hypotheses associating neighborhood features to the risk of preeclampsia and interpretable feature importances, helping to disentangle the dynamics of the neighborhood environment and preeclampsia.

2. MATERIALS AND METHODS

2.1. Study cohort

We constructed a geocoded preeclampsia nested case‐control study from GeoBirth [24, 25, 26, 27, 28], a manually curated and annotated retrospective database of 114,748 births from January 1, 2008 to December 31, 2020, at the Hospital of the University of Pennsylvania (HUP) and Pennsylvania Hospital (PAH) in Philadelphia, PA—a large, socioeconomically diverse city in the United States’ Mid‐Atlantic region. For both cases and controls, we used the inclusion criteria of: (1) live‐born singleton gestations; (2) maternal age less than 50 years old; (3) residence in Philadelphia for the entirety of the participant's pregnancy; and (4) parity data available. For participants with multiple births in GeoBirth, we randomly selected one birth to ensure independence among individuals. We defined cases as pregnant individuals with a diagnosis of preeclampsia (see Section 2.3) and no other hypertensive disorder during their pregnancy. We defined controls as pregnant individuals with normotensive pregnancies and full‐term (39+ weeks) births to eliminate individuals with competing outcomes of spontaneous preterm birth or early‐term birth and thus not eligible to develop HDPs throughout their pregnancy prior to a full‐term birth as defined by the American College of Obstetricians and Gynecologists (ACOG) [29]. We constructed the study cohort by including all cases of preeclampsia who met inclusion criteria and selecting for each case three controls matched on age ± 5 years, nulliparity (first birth vs. not), and last menstrual period (LMP) year. LMP was calculated using the best obstetric estimate of the due date (EDD) minus 40 weeks. This study was approved with a waiver of informed consent.

2.2. Neighborhood features

We collected a range of area‐level data, hereafter referred to as neighborhood features, totaling 26 distinct features (Table 1) [30, 31, 32, 33, 34, 35, 36, 37, 38, 39]. For the American Community Survey (ACS) features, we linked participants using the year of their LMP. For all other features, we used the year of data that were available, as indicated in Table 1. Additional information on the neighborhood features is provided in Appendix S1.

TABLE 1.

Features from the neighborhood environment.

Domain	Features	Data sources	Spatial resolution
Social vulnerability	People below the federal poverty threshold (%)	US Census Bureau ACS (2011, 2016, 2021)	Census tract
	People employed (%)
	People with bachelor's degree or higher (%)
	Index of Concentration at the Extremes (ICE) Race‐Income
Built environment	Housing units built before 1960 (%)	US Census Bureau ACS	Census tract
	National Walkability Index score	US EPA National Walkability Index (2019)	Census block group
Physical environment	Tree canopy cover (%)	Philadelphia's Parks and Recreation open access database on tree canopy (2015)	Census tract
	Daily PM2.5	XIS_PM2.5, Just et al.	Census block
	Daily mean temperature	XIS_Temperature, Just et al.	Census block
Health vulnerability	Obesity (%)	US CDC PLACES (2024)	Census tract
	Coronary heart disease (%)
	High cholesterol (%)
	High blood pressure (%)*
	Diabetes (%)
	Cancer (%)
	Stroke (%)
	Asthma (%)
	Chronic obstructive pulmonary disease (COPD) (%)
	Arthritis (%)
	Fair or poor self‐rated health status (%)
	Any disability (%)
	Frequent physical distress (%)
	No leisure‐time physical activity (%)
	Short sleep duration (%)
	Frequent mental distress (%)
	Depression (%)

Note: ICE Race‐Income is a measurement of spatial social polarization; the equation used to calculate this measure is provided in Appendix S1. All health vulnerability features are among adults.

Abbreviations: ACS, American Community Survey; ICE, Index of Concentration at the Extremes.

^*Pregnant individuals with high blood pressure only during pregnancy were not included in this feature (see PLACES documentation for more information).

We linked each neighborhood feature—using residential addresses from every day of pregnancy—to the study cohort detailed above, which yielded a high‐dimensional dataset of pregnant individuals and their timestamped neighborhood features from LMP to birth (details on linkage in Appendix S2). Participant residential addresses were geocoded using ArcMap Version 10.8 and the ArcGIS Street Map Premium North America 2021.1 address locator using a minimum match score of 75 [24]. Residential addresses are confirmed by participants at each healthcare encounter in the health system and were used to reconstruct an address history. For downstream analyses, we averaged each neighborhood feature for every participant over the LMP‐to‐preeclampsia diagnosis period (ending when the first preeclampsia ICD‐9/ICD‐10 code was recorded), determined from the case participant in the case‐control matched set, to ensure exposure lengths were the same for cases and controls even when gestational length differed.

2.3. Outcome of interest

We ascertained preeclampsia diagnosis by checking that a relevant 9th revision of the International Classification of Diseases (ICD‐9) code (646.2, 642.7) or 10th revision of the International Classification of Diseases (ICD‐10) code (O11, O14) was noted between the participant's LMP and 6 weeks postpartum at least twice and occurred at least 1 day apart.

2.4. Covariates

We obtained demographics, birth information, and maternal comorbidities from the electronic health record (EHR) and birth certificate data. We categorized maternal age as less than 20 years old, 20 to less than 35 years old, or 35+ years old; diabetes as pre‐existing diabetes, gestational diabetes, or no diabetes; and smoking status as current smoker, former smoker, or never smoker. We dichotomized parity as nulliparous or parous; pre‐existing hypertension as pre‐existing hypertension or no pre‐existing hypertension; and hospital as HUP or PAH. We used prepregnancy body mass index (BMI) as a continuous variable.

2.5. Statistical analysis

2.5.1. Bivariate analyses

We compared participant characteristics stratified by preeclampsia diagnosis with a bivariate analysis using two‐sample t‐tests for continuous variables and chi‐squared tests for categorical variables. We also performed a bivariate analysis using two‐sample t‐tests on the neighborhood features to describe associations with preeclampsia. We used test p‐values and standardized mean differences (SMDs) to quantify differences between cases and controls, using an SMD of >0.1 as a determination of substantial imbalance between groups [40].

2.5.2. Conditional logistic regression models

We modeled associations of neighborhood features with preeclampsia using both univariable and multivariable conditional logistic regression models to account for matching. We normalized each feature using the interquartile range (IQR) and chose covariates a priori based on literature demonstrating associations with preeclampsia [7, 41]. In the first set of adjusted models, we included maternal age, smoking, and hospital. In the second set of adjusted models, we further included variables that could be on the causal pathway connecting neighborhood features with preeclampsia (prepregnancy BMI, pre‐existing hypertension, diabetes). In all models, we included a random effect for census tract (determined from the first address during pregnancy) to account for the inherent spatial autocorrelation between tracts. For both the unadjusted and adjusted models, we ran separate models for each neighborhood feature to account for the collinearity present between some of the features. To quantify the strength and direction of the association, we calculated odds ratios (ORs) per IQR. Statistical analyses were performed in R 4.2.1 [42]. Conditional logistic regression models were built using mclogit 0.9.6 [43].

2.5.3. Machine learning models

We trained an XGBoost classification algorithm [44] (best performing algorithm out of traditional classification algorithms and a decision tree‐based algorithm that is robust to correlated features) to predict preeclampsia using the neighborhood features, with our focus on model explainability rather than model performance to detect patterns of associations between the neighborhood environment and preeclampsia. We performed hyperparameter tuning using a randomized cross‐validation search across parameters. We then applied 5‐fold cross‐validation with an 80/20 train‐test split for model evaluation, which enabled us to output performance metrics and their variability across the folds. We evaluated model performance on the test datasets using the mean and 95% confidence interval (CI) of the Area Under the Receiver Operating Characteristic curve (AUROC) across the five cross‐validation folds.

We generated feature importances through SHAP (SHapley Additive exPlanations) analyses on all five cross‐validation test datasets to account for potential variance due to collinearity [45, 46]. In a SHAP analysis, the contribution of each feature to the model prediction is computed by calculating a SHAP value for each sample's features, which represents the contribution of that feature's value to the difference observed between the model's prediction for the sample and the model's mean prediction over all samples. A feature is labeled as having a high or low overall impact on the model prediction by taking the mean absolute value of all SHAP values from every sample for the specific feature. Machine learning models were built using Python 3.9.12 with libraries Scikit‐learn (sklearn 1.5.2) [47] and shap (version 0.46.0) [46].

3. RESULTS

3.1. Participant characteristics

Our final preeclampsia case‐control study consisted of 18,754 participants (4689 cases and 14,065 controls) (Figure 1). There was one case for which only one control could be matched (instead of three controls). Table 2 details descriptive participant characteristics stratified by preeclampsia diagnosis. Cases and controls differed substantially with respect to baseline characteristics except for parity, which was an exact matching factor. Cases were more likely (SMD > 0.1) to be <20 or ≥35 years old, identify as non‐Hispanic Black, have lower educational attainment, have public/self‐pay/other insurance, have higher BMIs, and to have pre‐existing hypertension and diabetes. The study cohort covered almost all Philadelphia census tracts (379 out of 384 census tracts, using the 2010 census tract boundaries) (Appendix S3).

FIGURE 1.

Study cohort development.

TABLE 2.

Characteristics of preeclampsia cases and normotensive controls from the GeoBirth cohort (Philadelphia, 2008–2020).

Individual characteristic	Cases (n = 4689) n (%)	Controls (n = 14,065) n (%)	p value	SMD
Age (years)
<20	429 (9.1)	1160 (8.2)	<0.001	0.115
20 to <35	3228 (68.8)	10,393 (73.9)
35+	1032 (22.0)	2512 (17.9)
Race and ethnicity
Non‐Hispanic Black	3114 (66.4)	6656 (47.3)	<0.001	0.409
Non‐Hispanic White	1056 (22.5)	4837 (34.4)
Asian	152 (3.2)	1106 (7.9)
Hispanic	341 (7.3)	1319 (9.4)
Multiracial/Other	26 (0.6)	147 (1.0)
Educational attainment
Some college or less	3484 (74.3)	8669 (61.6)	<0.001	0.274
College degree or higher	1205 (25.7)	5396 (38.4)
Insurance type
Private	2079 (44.3)	7210 (51.3)	<0.001	0.139
Public, self‐pay, other	2610 (55.7)	6855 (48.7)
Parity
Nulliparous	2675 (57.0)	8025 (57.1)	1	0
Parous	2014 (43.0)	6040 (42.9)
Birth outcome
Preterm	1338 (28.5)	0 (0)	<0.001	0.894
Term	3351 (71.5)	14,065 (100)
Prepregnancy body mass index (kg/m2)
Mean (SD)	29.5 (8.17)	26.0 (6.40)	<0.001	0.472
Pre‐existing hypertension
Pre‐existing hypertension	1481 (31.6)	494 (3.5)	<0.001	0.794
No pre‐existing hypertension	3208 (68.4)	13,571 (96.5)
Diabetes
Pre‐existing diabetes	173 (3.7)	74 (0.5)	<0.001	0.253
Gestational diabetes	356 (7.6)	685 (4.9)
No diabetes	4157 (88.7)	13,289 (94.5)
Smoking status
Current smoker	274 (5.8)	630 (4.5)	<0.001	0.088
Former smoker	195 (4.2)	428 (3.0)
Never smoker	4205 (89.7)	12,950 (92.1)
Residential mobility
One or more moves	1133 (24.2)	3059 (21.7)	<0.001	0.057
No move	3556 (75.8)	11,006 (78.3)
Hospital
PAH	2721 (58.0)	8462 (60.2)	0.0104	0.043
HUP	1968 (42.0)	5603 (39.8)

Note: BMI had 83 (1.8%) cases and 267 (1.9%) controls with missingness. Diabetes had 3 (0.1%) cases and 17 (0.1%) controls with missingness. Smoking status had 15 (0.3%) cases and 57 (0.4%) controls with missingness.

Abbreviations: BMI, body mass index; HUP, Hospital of the University of Pennsylvania; PAH, Pennsylvania Hospital; SMD, standardized mean difference.

3.2. Features from the neighborhood environment spatially vary across Philadelphia

Features from the neighborhood environment vary throughout Philadelphia. Notably, West Philadelphia has an average prevalence of high blood pressure among adults of 42.5%, a percentage of people with a bachelor's degree or higher of 13.2%, and Index of Concentration at the Extremes (ICE) Race‐Income (i.e., a lower number represents a higher proportion of Black households with income <$20,000) of −0.38. Central Philadelphia has an average prevalence of high blood pressure of 20.7%, a percentage of people with a bachelor's degree or higher of 71.4%, and ICE Race‐Income of 0.21 (Figure 2).

FIGURE 2.

Spatial variability of three features from the neighborhood environment. Philadelphia planning districts are outlined in green with the West and Central planning districts labeled and highlighted in dark blue. (A) Prevalence of high blood pressure among adults. (B) Percentage of people with bachelor's degree or higher. (C) ICE Race‐Income. Gray census tracts represent missing data due to no or low population counts, for example, Northeast Philadelphia (PNE) airport, Philadelphia International (PHL) airport, etc. Pink map uses PLACES data and 2020 census tract boundaries; purple maps use ACS 2016 5‐year data and 2010 census tract boundaries. HBP, high blood pressure; ICE, Index of Concentration at the Extremes.

3.3. Statistical analysis of neighborhood features shows association with preeclampsia risk

The incidence of preeclampsia was higher (SMD > 0.1) among residents of census tracts with a higher percentage of people below the federal poverty threshold and prevalence of all health vulnerability features except cancer. Residing in census tracts (census block group for the National Walkability Index score) with a lower percentage of people employed, percentage of people with bachelor's degree or higher, ICE Race‐Income, National Walkability Index score, and prevalence of depression was associated with developing preeclampsia (Table 3).

TABLE 3.

Bivariate associations of neighborhood features with preeclampsia ordered by SMD.

Neighborhood feature	Cases (n = 4689) Median (IQR)	Controls (n = 14,065) Median (IQR)	p value	SMD a
Short sleep duration (%)	45.3 (38.1–49.2)	42.0 (34.8–47.8)	<0.001	0.314
Obesity (%)	37.7 (28.3–41.8)	33.2 (25.9–40.7)	<0.001	0.306
High blood pressure (%)	35.8 (27.4–41.9)	31.0 (22.2–39.3)	<0.001	0.303
Asthma (%)	12.3 (10.5–13.3)	11.3 (9.80–13.0)	<0.001	0.302
People with bachelor's degree or higher (%)	15.2 (9.44–31.4)	20.3 (10.9–52.3)	<0.001	0.298
Diabetes (%)	16.4 (10.6–19.2)	13.8 (8.10–18.5)	<0.001	0.288
ICE Race‐Income	−0.215 (−0.365 to 0.0406)	−0.107 (−0.318 to 0.126)	<0.001	0.282
Stroke (%)	4.70 (3.10–6.10)	4.00 (2.20–5.70)	<0.001	0.282
Any disability (%)	36.2 (28.3–42.0)	33.2 (23.1–40.5)	<0.001	0.275
COPD (%)	8.07 (5.70–9.80)	7.10 (4.10–9.30)	<0.001	0.271
No leisure‐time physical activity (%)	29.2 (20.9–34.4)	26.0 (16.4–33.2)	<0.001	0.27
Fair or poor self‐rated health status (%)	26.4 (17.8–32.5)	23.2 (13.5–30.9)	<0.001	0.269
Arthritis (%)	28.4 (23.1–31.6)	26.5 (19.8–30.5)	<0.001	0.267
Frequent physical distress (%)	15.4 (12.1–18.3)	13.9 (9.50–17.4)	<0.001	0.264
People employed (%)	50.8 (42.8–59.7)	54.4 (45.2–64.6)	<0.001	0.243
Coronary heart disease (%)	6.90 (5.40–7.90)	6.40 (4.30–7.60)	<0.001	0.241
Frequent mental distress (%)	19.8 (17.3–21.8)	18.5 (16.3–21.2)	<0.001	0.24
High cholesterol (%)	31.2 (28.7–32.7)	30.6 (26.5–32.4)	<0.001	0.213
People below the federal poverty threshold (%)	27.4 (16.6–36.8)	24.4 (13.7–35.6)	<0.001	0.171
National Walkability Index score	15.2 (14.0–16.3)	15.3 (14.2–16.8)	<0.001	0.15
Depression (%)	22.7 (21.5–24.2)	23.0 (21.7–24.4)	<0.001	0.116
Housing units built before 1960 (%)	82.2 (71.6–88.2)	80.5 (68.4–87.8)	<0.001	0.1
Cancer (%)	4.90 (4.40–5.50)	5.00 (4.40–5.80)	<0.001	0.094
Tree canopy cover (%)	14.5 (9.34–18.3)	13.7 (8.45–18.4)	0.00328	0.049
Daily mean temperature (°C)	13.6 (10.9–16.4)	13.7 (10.9–16.5)	0.112	0.027
Daily PM2.5 (ug/m3)	9.95 (8.63–10.9)	9.95 (8.64–11.0)	0.614	0.008

Note: People below the federal poverty threshold had 3 (0.1%) cases and 13 (0.1%) controls with missingness. People employed had 3 (0.1%) cases and 13 (0.1%) controls with missingness. People with bachelor's degree or higher had 3 (0.1%) cases and 12 (0.1%) controls with missingness. ICE Race‐Income had 5 (0.1%) cases and 13 (0.1%) controls with missingness. Housing units built before 1960 had 5 (0.1%) cases and 12 (0.1%) controls with missingness. All health vulnerability features had 8 (0.2%) cases and 18 (0.1%) controls with missingness.

Abbreviations: COPD, chronic obstructive pulmonary disease; ICE, Index of Concentration at the Extremes; SMD, standardized mean difference.

^a We use the absolute value of the SMD.

In the first set of conditional logistic regression models adjusted for maternal age, smoking, and hospital (separate models were run for each neighborhood feature), several neighborhood features were significantly associated with preeclampsia (Table 4). Per IQR increment of each feature, individuals had higher odds of preeclampsia in neighborhoods with higher proportions of residents with incomes below the federal poverty threshold (adjusted odds ratio [aOR], 1.26; 95% CI, 1.17, 1.36), proportions of housing units built before 1960 (aOR, 1.11; 95% CI, 1.05, 1.17), and prevalence of all health vulnerability features except cancer and depression. Individuals had lower odds of preeclampsia in neighborhoods with higher proportions of residents employed (aOR, 0.65; 95% CI, 0.60, 0.69), proportions of residents with education higher than college (aOR, 0.51; 95% CI, 0.48, 0.55), ICE Race‐Income (aOR, 0.51; 95% CI, 0.47, 0.55), National Walkability Index score (aOR, 0.86; 95% CI, 0.82, 0.92), prevalence of cancer (aOR, 0.91; 95% CI, 0.87, 0.96), and depression (aOR, 0.88; 95% CI, 0.82, 0.94). In models further adjusted for BMI, pre‐existing hypertension, and diabetes, effect estimates attenuated but all retained statistical significance except for housing units built before 1960 and prevalence of cancer. We did not detect significant associations of tree canopy cover, PM_2.5, or temperature with preeclampsia.

TABLE 4.

Unadjusted and adjusted models for preeclampsia using one IQR increment increase of the exposure variables ordered by OR.

Neighborhood feature	Unadjusted models 0 OR (95% CI)	Adjusted models 1 aOR (95% CI)	Adjusted models 2 aOR (95% CI)
Short sleep duration (%)	2.22 (2.04, 2.41)	2.37 (2.17, 2.58)	1.76 (1.60, 1.93)
Obesity (%)	2.15 (1.98, 2.35)	2.25 (2.06, 2.46)	1.69 (1.53, 1.86)
Asthma (%)	2.05 (1.89, 2.23)	2.17 (1.99, 2.36)	1.66 (1.51, 1.82)
High blood pressure (%)	1.95 (1.81, 2.11)	2.04 (1.88, 2.21)	1.61 (1.48, 1.75)
Diabetes (%)	1.89 (1.75, 2.05)	1.94 (1.79, 2.11)	1.55 (1.42, 1.69)
Stroke (%)	1.85 (1.71, 2.00)	1.91 (1.76, 2.07)	1.54 (1.41, 1.68)
Any disability (%)	1.82 (1.68, 1.98)	1.87 (1.72, 2.03)	1.49 (1.37, 1.63)
Fair or poor self‐rated health status (%)	1.80 (1.66, 1.95)	1.83 (1.68, 1.99)	1.48 (1.35, 1.61)
No leisure‐time physical activity (%)	1.78 (1.64, 1.93)	1.82 (1.67, 1.97)	1.47 (1.35, 1.58)
COPD (%)	1.77 (1.64, 1.92)	1.81 (1.67, 1.96)	1.48 (1.36, 1.61)
Frequent physical distress (%)	1.73 (1.60, 1.88)	1.76 (1.62, 1.91)	1.44 (1.33, 1.57)
Arthritis (%)	1.70 (1.57, 1.84)	1.73 (1.59, 1.87)	1.45 (1.34, 1.57)
Coronary heart disease (%)	1.55 (1.44, 1.68)	1.58 (1.46, 1.70)	1.36 (1.26, 1.47)
Frequent mental distress (%)	1.50 (1.40, 1.60)	1.52 (1.41, 1.63)	1.30 (1.21, 1.40)
High cholesterol (%)	1.40 (1.30, 1.50)	1.41 (1.31, 1.52)	1.27 (1.18, 1.36)
People below the federal poverty threshold (%)	1.25 (1.16, 1.35)	1.26 (1.17, 1.36)	1.16 (1.07, 1.25)
Housing units built before 1960 (%)	1.11 (1.05, 1.17)	1.11 (1.05, 1.17)	1.04 (0.99, 1.10)
Tree canopy cover (%)	1.03 (0.97, 1.09)	1.04 (0.97, 1.10)	1.03 (0.98, 1.09)
Daily PM2.5 (ug/m3)	0.98 (0.88, 1.10)	0.97 (0.87, 1.09)	1.00 (0.88, 1.13)
Daily mean temperature (°C)	0.97 (0.91, 1.02)	0.96 (0.90, 1.02)	0.95 (0.88, 1.01)
Cancer (%)	0.92 (0.87, 0.96)	0.91 (0.87, 0.96)	0.96 (0.92, 1.00)
Depression (%)	0.88 (0.83, 0.94)	0.88 (0.82, 0.94)	0.90 (0.85, 0.95)
National Walkability Index score	0.87 (0.82, 0.92)	0.86 (0.82, 0.92)	0.89 (0.84, 0.94)
People employed (%)	0.66 (0.62, 0.71)	0.65 (0.60, 0.69)	0.76 (0.70, 0.81)
People with bachelor's degree or higher (%)	0.52 (0.48, 0.56)	0.51 (0.48, 0.55)	0.65 (0.60, 0.70)
ICE Race‐Income	0.53 (0.50, 0.57)	0.51 (0.47, 0.55)	0.64 (0.59, 0.70)

Note: All models are run separately for each neighborhood feature. Adjusted models 1 are adjusted for maternal age, smoking, and hospital. Adjusted models 2 are adjusted for maternal age, smoking, hospital, BMI, pre‐existing hypertension, and diabetes. Since there was less than 0.3% missingness in the exposure variables (Table 3), models included only complete cases and controls (participants without missingness). As a sensitivity analysis, we ran the unadjusted models with all participants (including those with missingness), and ORs did not differ.

Abbreviations: CI, confidence interval; COPD, chronic obstructive pulmonary disease; ICE, Index of Concentration at the Extremes; IQR, interquartile range; OR, odds ratio.

3.4. Machine learning models find variation in relative importance of neighborhood features associated with preeclampsia

XGBoost models trained on the 26 neighborhood features detect an association between the features of the neighborhood environment and preeclampsia, with a mean AUROC of 0.58 (95% CI, 0.54–0.61) across the five folds. It should be noted that high predictive accuracy was not a goal of this study—since only a portion of risk for preeclampsia is governed by environmental exposures, we expect a relatively low AUROC, and use the AUROC to show that the model is learning patterns by performing better than random chance (i.e., an AUROC near 0.50). Given the low AUROC, the model has limited predictive utility and the SHAP results should be interpreted as model‐based predictive importance within the specific study context and not as causal evidence. From the SHAP analysis we find that the top three features’ mean absolute SHAP values contribute a mean of 44.8% across the five cross‐validation folds to the model prediction (prevalence of obesity among adults with 16.1%, prevalence of high blood pressure among adults with 15.7%, and prevalence of short sleep duration among adults with 13%) (Figure 3). We find that pregnant individuals who live in census tracts with a higher prevalence of obesity, prevalence of high blood pressure, and prevalence of short sleep duration among adults are more likely to be predicted as having preeclampsia (Figure 4). The top three features are the same across the five cross‐validation folds with a 0.6% standard deviation for the top three features’ mean absolute SHAP values contribution to the model prediction.

FIGURE 3.

Feature contribution to the best‐performing XGBoost model prediction. Plot shows features sorted by the feature's mean absolute SHAP value divided by the sum of mean absolute SHAP values for all features (normalized mean absolute SHAP value), using the second cross‐validation fold. ICE, Index of Concentration at the Extremes; SHAP, SHapley Additive exPlanations.

FIGURE 4.

SHAP beeswarm plot analyzing the top three features with the highest normalized mean absolute SHAP value from the second cross‐validation fold. Each dot on the plot represents a participant with the color of the dot indicating whether that participant's feature value was high (yellow) or low (blue). The X‐axis indicates the SHAP value for each participant feature, with negative values indicating a negative impact (no preeclampsia) and positive values indicating a positive impact (preeclampsia). SHAP, SHapley Additive exPlanations.

4. DISCUSSION

This study generated a rich set of social vulnerability, built environment, physical environment, and neighborhood health vulnerability features from the neighborhood environment. We found spatial variation in features across Philadelphia neighborhoods associated with preeclampsia, including neighborhood poverty, employment, education, ICE Race‐Income, walkability, and all health vulnerability features. Using a machine learning approach we also observed that a group of features from the neighborhood environment has a quantifiable association with preeclampsia, with specific features driving the association. These features—neighborhood prevalences of obesity, high blood pressure, and short sleep duration among adults—are all facets of the health vulnerability of a community that may help to predict preeclampsia in communities.

Our finding of the association between neighborhood features and preeclampsia is supported by the existing literature. Neighborhood deprivation indices containing facets of socioeconomic vulnerability (e.g., neighborhood‐level educational attainment) and individual‐level educational attainment have independently been associated with adverse pregnancy outcomes such as preeclampsia [17, 48, 49]. The Environmental Protection Agency's (EPA) National Walkability Index has previously been associated with HDPs [17] and the ICE Race‐Income has been used as a proxy for structural racism and shown to be not only associated with adverse birth outcomes but also that exposure to structural racism can increase the probability of chronic health conditions before pregnancy, which in turn can lead to adverse pregnancy outcomes [50, 51, 52, 53, 54]. Individual‐level chronic hypertension and obesity are well‐documented risk factors for preeclampsia [7] and associations between individual‐level short sleep duration and preeclampsia have previously been reported [55], along with an increased risk of diagnosis of adult hypertension [56]. We note that we did not find associations for previously documented associations of tree canopy cover [16] and PM_2.5 [9, 57, 58] with preeclampsia. We suspect differences in findings are due to study design and population differences, respectively.

The neighborhood environment can influence risk of preeclampsia through multiple pathways; we focus here on the influence it has on two preeclampsia risk factors—high blood pressure and obesity—as these were the neighborhood features that had the highest relative importance with regards to preeclampsia from the SHAP analysis. Moving from a neighborhood with a high level of poverty to one with a lower level has been associated with a reduction in obesity [59]. Obesity is associated with preeclampsia through metabolic changes that lead to insulin resistance, inflammation, and oxidative stress [60]. Neighborhood deprivation and lower levels of neighborhood educational attainment have both been associated with higher systolic blood pressure (SBP) [61, 62]. Neighborhood‐level racial residential segregation has been associated with an increased risk of hypertension for Black and Hispanic residents [63]. Living in less walkable neighborhoods has been associated with a high predicted 10‐year cardiovascular disease risk and higher mean SBP [64], while living in more walkable neighborhoods has been associated with a lower likelihood of hypertension [65] and pregnancy‐induced hypertension [66]. This decreased risk of pregnancy‐induced hypertension may be due to increased physical activity, as aerobic exercise during pregnancy has been linked to a reduced risk of HDPs [67].

4.1. Strengths and limitations

Strengths of this study include exploring the relative importance of a diverse set of features from the neighborhood environment using a machine learning‐based approach in a large sample of preeclampsia cases. Our use of daily residential addresses instead of birth or first‐trimester addresses alone may help reduce exposure misclassification. We show the utility of using XAI methods to screen neighborhood features for hypothesis generation to inform policy‐making priorities for pregnancy health interventions that can be further validated using epidemiological or experimental approaches. However, our study should be interpreted with the following limitations in mind: the modest predictive performance of the XGBoost model was expected as preeclampsia is a complex, heterogeneous, and multifactorial disorder, and the model is trained only on neighborhood environmental factors unlike previous preeclampsia prediction models that are trained on clinical and maternal factors that have larger risk contributions. As mentioned above, we used the AUROC metric to calibrate whether the model was learning patterns that are present in the data rather than to demonstrate the ability of the model to yield high‐confidence predictions. The goal of this study was not a clinical prediction tool but rather to investigate the neighborhood environment contribution to preeclampsia using interpretable machine learning models to prioritize community‐level interventions, and we note that our findings are hypothesis‐generating results and work remains to be done to achieve a fuller view of the neighborhood environment and improve model performance. Although we used a rigorous approach to identify preeclampsia cases, some misclassification may still be present in the cases and controls as we relied on EHR data and billing codes. The temporal heterogeneity of neighborhood feature availability could lead to misclassification; however, we would not expect this misclassification to be differential between cases and controls as we matched on LMP year. The US CDC PLACES uses small area estimates and while these estimates are not direct survey estimates, they are an appropriate proxy when local health data is not available, as is the case in our study [68, 69, 70]. The case‐control design restricts generalizability to preeclampsia and full‐term births. Additionally, the study participants come from two Philadelphia hospitals (within the same health system) and from one geographic location with a specific distribution of neighborhood features and population, which may limit generalizability to cities with majority Non‐Hispanic Black or White residents. However, restricting to Philadelphia residents, who use these academic hospitals as their community hospitals, overcomes some of the selection bias that can occur in tertiary care centers. Furthermore, understanding local neighborhood environmental contributions to preeclampsia risk in a single city may facilitate health system and community collaboration to improve pregnancy health for a specific population.

5. CONCLUSION

Our findings highlight significant associations of neighborhood health and social vulnerability with preeclampsia risk and the importance of additional studies on the impact of community health investments for improving pregnancy outcomes. The variation in the relative importance of neighborhood features could be used to inform place‐based understandings of pregnancy health and interventions to address structural obstacles to positive pregnancy outcomes. Continued research characterizing the complex relationship between place and preeclampsia is necessary to inform structural interventions reducing adverse maternal outcomes.

AUTHOR CONTRIBUTIONS

Chloé F. Paris: Conceptualization; project administration; data curation; methodology; investigation; formal analysis; software; visualization; writing—original draft; writing—review & editing. Rachel Ledyard: Data curation; writing—review & editing. Allan C. Just: Data curation; writing—review & editing. Eugenia C. South: Conceptualization; writing—review & editing. Max Jordan Nguemeni Tiako: Conceptualization; writing—review & editing. Silvia P. Canelón: Data curation; writing—review & editing. Heather H. Burris: Conceptualization; funding acquisition; methodology; project administration; resources; supervision; writing—review & editing. Joseph D. Romano: Conceptualization; funding acquisition; methodology; project administration; resources; supervision; writing—review & editing.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

ETHICS STATEMENT

The University of Pennsylvania Institutional Review Board approved this study with a waiver of informed consent (Penn IRB protocol 829674).

Supporting information

Supplementary material is available at Pregnancy online.

DATA AVAILABILITY STATEMENT

Participant data from GeoBirth cannot be shared publicly due to patient privacy. All the necessary code (without patient data) to reproduce our analyses is available on GitHub (https://github.com/RomanoLab/geospatial‐preeclampsia‐geobirth) with an archived version on Zenodo (https://doi.org/10.5281/zenodo.17309276) [71].