Hypertensive disorders of pregnancy (HDP), including gestational hypertension, preeclampsia, eclampsia, and chronic hypertension with or without superimposed preeclampsia, are among the most common medical complications of pregnancy and contribute substantially to maternal and perinatal morbidity and mortality worldwide.^1-8 Women with HDP face elevated risks of adverse pregnancy outcomes such as preterm birth,⁹ fetal growth restriction,¹⁰ and perinatal death,^11,12 as well as increased long-term risk of cardiovascular disease (CVD) later in life.^13-17 Given the burden of HDP in many populations and the limited number of effective preventive strategies, identification of modifiable risk factors beyond known clinical and behavioral characteristics is needed in order to inform primary prevention and risk stratification in clinical practice. Emerging evidence points to an important role of broader spatial and contextual environmental factors in identifying risk of HDP, with large geographic disparities in HDP incidence suggesting that where women live may be a key determinant of disease risk.^18,19

A growing body of work has evaluated specific environmental factors in relation to HDP, including ambient air pollution,²⁰ heavy metals,²¹ temperatures,^22,23 neighborhood deprivation,²⁴ crime,²⁵ walkability,²⁶ green space,²⁷ and food access.²⁸ These studies have provided important evidence that spatial and contextual factors contribute to HDP risk, but have generally focused on a limited number of preselected exposures independently, without considering the broader “exposome”—the totality of non-genetic exposures across the life course.²⁹ The spatial and contextual exposome, which encompasses features of the natural, built, and social environments that can be assessed from preexisting spatial and contextual data linked to individuals,^30,31 offers a powerful framework for identifying novel environmental determinants of HDP at scale.

Our prior work has demonstrated the feasibility of conducting large-scale external exposome-wide association studies (ExWAS) for HDP using statewide birth records data,³² incorporating hundreds of spatially and temporally resolved exposure variables and agnostic screening strategies. However, our study and others investigating this question have treated HDP as a composite outcome, combining gestational hypertension and preeclampsia-eclampsia, and did not distinguish between specific HDP subtypes and severity. HDP with different subtypes and severity (eg, gestational hypertension Versus severe preeclampsia) may reflect distinct underlying pathophysiological processe.^33,34 For example, some studies suggest that certain environmental exposures, such as air pollution, may be more strongly associated with milder forms of preeclampsia than with severe disease.^35,36 Aggregating all HDP into a single composite outcome may therefore obscure important subtype-specific associations and limit our ability to identify high-risk women and tailor prevention strategies. Yet, due in part to limitations of commonly used data sources such as birth records and administrative claims—which often lack detailed clinical information—most environmental studies have not been able to differentiate HDP subtypes and severity.

Another key methodological challenge is how to accurately characterize residential exposures during pregnancy. Most population-based studies rely on a single residential address at delivery to assign environmental exposures, implicitly assuming that women remain at the same location throughout pregnancy. However, several studies have shown that ignoring residential mobility during pregnancy can result in exposure measurement error and subsequently biased effect estimates.^37-39 Leveraging data sources that include detailed residential histories, such as electronic health records (EHR), is therefore critical to improve exposure assessment and reduce measurement error in studies of environmental determinants of HDP.

Beyond exposure assessment, there are important analytic limitations in current exposome literature. ExWAS approaches typically rely on high-dimensional yet essentially conventional regression models, applied exposure-by-exposure, with a common set of covariates assumed to adequately control confounding for all exposures under study.³⁰ Adjusting for a single uniform confounder set across diverse environmental exposures may be problematic because different exposures arise from distinct causal pathways and may therefore have different confounding structures. For example, air pollution is strongly influenced by traffic density and meteorology,⁴⁰ whereas green space is largely shaped by neighborhood and land-use context.⁴¹ Although both exposures may relate to health outcomes, the upstream determinants and potential confounders differ. Directed acyclic graphs (DAGs) provide a formal framework to represent hypothesized causal relationships and identify minimally sufficient adjustment sets tailored to each exposure, thereby reducing the risk of residual confounding or over-adjustment.⁴² However, even with DAG-informed confounder selection, modeling the complex, high-dimensional, and correlated nature of exposome data remains challenging. Traditional causal approaches, including parametric g-computation and propensity score-based methods, rely on correct model specification and may not perform well in settings characterized by high dimensionality, multicollinearity, and nonlinear relationships.^43-45 Double machine learning (DML) offers an attractive alternative: by combining flexible machine-learning algorithms for nuisance prediction with orthogonalized estimation of exposure effects, DML can accommodate high-dimensional confounding, complex functional forms, and correlated exposures while preserving valid statistical inference under appropriate assumptions.^46,47 Yet, to our knowledge, formal DML-based causal inference has not been applied in the context of the spatial and contextual exposome or used to investigate HDP outcomes.

In this study, we address the aforementioned limitations by leveraging a large statewide linked EHR-vital statistics birth and fetal death records dataset from Florida, with detailed clinical information from EHR to distinguish HDP subtypes. We then reconstructed residential histories during pregnancy to assign time-resolved exposures across multiple address records. We define confounder sets tailored to different classes of spatial and contextual factors using exposure-specific causal diagrams and apply a DML framework to estimate associations between these exposures and HDP outcomes while flexibly adjusting for high-dimensional covariates. By integrating rich environmental data, granular HDP phenotyping, residential mobility, and modern causal machine-learning methods, this DML-based analysis aims to provide a more nuanced understanding of how the spatial and contextual exposome contributes to the development of specific HDP subtypes, and to identify potential targets for prevention and policy interventions.

Table 2 presents maternal characteristics for the overall study population and according to HDP subtype. Among 686 412 singleton pregnancies with conception dates between 2013 and 2018 in the linked Florida EHR-vital statistics birth and fetal death records dataset, 66 095 HDP cases were identified (9.6%). Of these, 41 865 were gestational hypertension (6.1%), 10 566 were mild preeclampsia (1.5%), 11 811 were severe preeclampsia (1.7%), and 1853 were eclampsia (0.3%). Compared with pregnancies without HDP, pregnancies affected by HDP were more likely to occur among women aged < 20 years (9.6% Versus 7.4%) or ≥40 years (2.8% Versus 2.3%), and among non-Hispanic Black women (26.0% Versus 21.1%), while Hispanic women comprised a smaller proportion of HDP cases (30.2% Versus 36.7%). Women with HDP were also more likely to have pre-pregnancy obesity (34.6% Versus 21.8%) and to be covered by Medicaid (71.7% Versus 69.8%), and they were more often nulliparous (52.5% Versus 37.7%) and had a higher prevalence of pre-pregnancy depression (3.0% Versus 2.6%). Across HDP subtypes, severe preeclampsia and eclampsia showed the highest proportions of non-Hispanic Black women (31.1% and 35.5%, respectively) and were less likely to have prenatal care initiated in the first trimester (46.2% and 49.9%, respectively); eclampsia had the highest prevalence of pre-pregnancy depression (7.3%). Table S7 shows the distribution of spatial and contextual exposome measures included in the analyses, along with the percentage of missing values for each measure.

Volcano plots shown in Figures 2 and 3 summarize the Phase 1 results based on the discovery and replication sets, respectively. Tables S8–S12 present the estimated effect sizes for HDP subtypes and overall HDP in Phase 1. As shown in Table S8, 91 and 65 exposome measures were significantly associated with gestational hypertension in the discovery and replication sets, respectively. After accounting for multiple comparisons, 26 measures that remained significant in both sets were retained in Phase 2. Similarly, 34 exposome measures were significantly associated with overall HDP in both the discovery and replication sets after multiple-comparison adjustment, as shown in Table S12 Although no exposome measures remained significant after multiple-comparison adjustment for mild preeclampsia (Table S9), severe preeclampsia (Table S10), or eclampsia (Table S11), three measures were significant for severe preeclampsia in both sets (P < 0.05) (Table S10), including two measures negatively associated with severe preeclampsia (ie, minimum relative humidity, natural A-weighted L50 sound pressure level), and arsenic compounds (inorganic including arsine), which are positively associated with severe preeclampsia.

In Phase 2, the exposome measures retained from Phase 1 were simultaneously included in a multi-treatment DML model after adjusting for confounders selected based on the DAG. Table 3 presents the adjusted odds ratios (ORs) for gestational hypertension per standard deviation increase in each exposome measure, along with 95% confidence intervals (CIs). Among the 26 measures included in Phase 2, 12 remained significantly associated with gestational hypertension, spanning air toxicants (n = 5), blue space (n = 1), crime and safety (n = 2), meteorology (n = 3), and ultraviolet radiation (n = 1). Specifically, significant associations were observed for hydrogen fluoride (hydrofluoric acid) (OR: 0.99, 95% CI: 0.98, 1.00), methyl bromide (bromomethane) (OR: 0.98, 95% CI: 0.98, 0.99), methyl isobutyl ketone (hexone) (OR: 0.97, 95% CI: 0.95, 1.00), 2,2,4-trimethylpentane (OR: 1.05, 95% CI: 1.03, 1.07), and catechol (OR: 1.03, 95% CI: 1.01, 1.05). The blue space measure proximity to coastline was also positively associated with gestational hypertension (OR: 1.04, 95% CI: 1.03, 1.05). For crime and safety, both burglary rate (OR: 1.02, 95% CI: 1.00, 1.04) and forcible sex offenses rate (OR: 1.01, 95% CI: 1.00, 1.02) were significantly associated with higher odds of gestational hypertension. Meteorological measures showing significant associations included maximum relative humidity (OR: 1.02, 95% CI: 1.01, 1.04), minimum relative humidity (OR: 0.95, 95% CI: 0.93, 0.98), and maximum temperature (OR: 0.92, 95% CI: 0.86, 1.00). Finally, a higher erythemal UV index was associated with lower odds of gestational hypertension (OR: 0.94, 95% CI: 0.90, 0.98).

Table 4 presents the Phase 2 results for overall HDP. Among the 34 measures included, 11 remained significantly associated with overall HDP in the multi-treatment DML model, including 7 air toxicants, 1 blue space measure, 1 meteorology measure, 1 noise measure, and 1 ultraviolet measure. Specifically, overall HDP was positively associated with the air toxicants selenium compounds (OR: 1.04, 95% CI: 1.03, 1.05), 2,2,4-trimethylpentane (OR: 1.04, 95% CI: 1.02, 1.06), and glycol ethers (OR: 1.03, 95% CI: 1.00, 1.05), as well as the blue space measure proximity to coastline (OR: 1.02, 95% CI: 1.02, 1.03) and the meteorology measure maximum relative humidity (OR: 1.02, 95% CI: 1.00, 1.03). Inverse associations were observed for the air toxicants hydrogen fluoride (hydrofluoric acid) (OR: 0.99, 95% CI: 0.98, 0.99), methyl bromide (bromomethane) (OR: 0.99, 95% CI: 0.98, 0.99), xylenes (mixed isomers) (OR: 0.95, 95% CI: 0.92, 0.98), and 1,1,1-trichloroethane (OR: 0.97, 95% CI: 0.95, 0.99), as well as the noise measure natural A-weighted L50 sound pressure level (OR: 0.98, 95% CI: 0.97, 0.99) and the ultraviolet measure erythemal UV index (OR: 0.94, 95% CI: 0.91, 0.98).

Results from the sensitivity analyses (Tables S13 and S14) were largely consistent with the primary analyses after additional adjustment for the number of residential moves and urbanicity. Minor differences were observed: methyl isobutyl ketone (primary P-value: 0.04; sensitivity P-value: 0.13) and maximum temperature (primary P-value: 0.04; sensitivity P-value: 0.22) were no longer statistically significant for gestational hypertension, and glycol ethers (primary P-value: 0.04; sensitivity P-value: 0.25) was no longer significant for overall HDP, although effect estimates remained similar. Conversely, several exposures that were not statistically significant in the primary analyses became significant in the sensitivity analyses, including hexane (primary P-value: 0.05; sensitivity P-value: 0.003), quinone (primary P-value: 0.33; sensitivity P-value: 0.02), 2,4-dinitrotoluene (primary P-value: 0.18; sensitivity P-value: 0.003), catechol (primary P-value: 0.06; sensitivity P-value: 0.02), forcible sex offense rate (primary P-value: 0.06; sensitivity P-value: 0.04) and murder rate (primary P-value: 0.14; sensitivity P-value: 0.01). Overall, these changes largely reflected shifts from marginal non-significance to statistical significance (or vice versa), while the magnitude and direction of the effect estimates remained stable.

In this large Florida cohort with linked birth/fetal death records and EHR data, we found that multiple early-pregnancy environmental exposures were associated with HDP. Using a two-phase DML-based approach, we identified 12 exposures associated with gestational hypertension and 11 exposures associated with overall HDP. These exposures spanned diverse domains, including ambient air toxicants, blue space, neighborhood crime and safety, meteorological factors, environmental noise, and ultraviolet radiation. While the effect sizes were modest in magnitude and specific to subtype of HDP, our findings suggest that the total environment in which women live—in terms of natural, built, and social characteristics—represents a modifiable factor which could potentially influence their risk of gestational hypertension or HDP.

Our findings are broadly consistent with prior work linking spatial and contextual environmental exposures to HDP, while extending earlier exposome-wide studies through EHR-linked phenotyping, subtype-specific analyses, and multi-treatment estimation. In our prior ExWAS based on only vital statistics birth records in Florida,³² several air toxicants, meteorological variables, and neighborhood indicators were associated with HDP; we replicate key signals such as 2,2,4-trimethylpentane and neighborhood crime measures, supporting their robustness across data sources and analytic strategies. In addition to identifying similar environmental signals in both of our approaches, the present study expands on our prior work to identify heterogeneity in effects by HDP subtype.

Among air toxicants, several compounds were positively associated with gestational hypertension and/or overall HDP in Phase 2. 2,2,4-trimethylpentane and catechol were consistently associated with higher odds of gestational hypertension, while 2,2,4-trimethylpentane, glycol ethers, and selenium compounds were associated with higher odds of overall HDP. These compounds likely reflect complex industrial or traffic-related mixtures and may influence HDP risk through oxidative stress, systemic inflammation, endothelial dysfunction, and placental vascular dysregulation—pathways strongly implicated in the pathophysiology of pregnancy hypertension.^20,75,76 The strong association observed for selenium compounds is notable, as selenium is an essential micronutrient but may be harmful at elevated environmental levels; ambient selenium may also act as a marker of broader industrial emissions rather than a direct causal agent.^77-79 Nevertheless, positive associations should be interpreted cautiously, as air toxicants often co-occur with other environmental and socioeconomic conditions, and residual confounding or correlated exposures may contribute to observed relationships despite adjustment. In contrast, several air toxicants showed inverse associations, including hydrogen fluoride (hydrofluoric acid) and methyl bromide (bromomethane) for both gestational hypertension and overall HDP, methyl isobutyl ketone (hexone) for gestational hypertension, and xylenes (mixed isomers) and 1,1,1-trichloroethane for overall HDP. These inverse associations should be interpreted cautiously. Similar paradoxical inverse associations have long been observed between maternal cigarette smoking and preeclampsia,^80,81 where proposed biological mechanisms (eg, carbon monoxide-mediated effects on placental angiogenic balance) coexist with strong concerns about residual confounding and selection bias. In the exposome context, such inverse associations may reflect complex mixture effects, correlated unmeasured exposures, or differential measurement error rather than true protective effects. Differences from prior studies may also reflect variation in spatial scale, exposure definitions, regional emission profiles, or analytic approaches. Accordingly, these findings should be viewed as hypothesis-generating.

Meteorological and ultraviolet exposures also showed consistent and biologically plausible patterns. Maximum relative humidity was positively associated with both gestational hypertension and overall HDP, whereas minimum relative humidity and maximum temperature were inversely associated with gestational hypertension. In addition, higher erythemal UV index was associated with lower odds of both gestational hypertension and overall HDP. These findings are consistent with prior evidence of seasonal variation in HDP and hypotheses linking sunlight exposure, vitamin D metabolism, and vascular function.^82-87

Several built and social environment measures were independently associated with HDP outcomes. Living closer to the coastline was associated with lower odds of both gestational hypertension and overall HDP, potentially reflecting broader coastal or blue space-related environmental and contextual characteristics rather than a single etiologic factor.⁸⁸ In the crime and safety domain, higher burglary rates and forcible sex offenses rates were associated with increased odds of gestational hypertension, aligning with prior literature linking neighborhood stressors to elevated risk of preeclampsia and gestational hypertension.^25,32,89,90 Chronic psychosocial stress related to neighborhood safety may contribute to dysregulation of the hypothalamic-pituitary-adrenal axis and vascular function,⁹¹ thereby increasing susceptibility to gestational hypertension.^92,93 Similarly, environmental noise exposure—another stress-related urban hazard—has been implicated in HDP. While impact-weighted noise measures did not remain significant after mutual adjustment, higher natural A-weighted L50 sound pressure levels were inversely associated with overall HDP, suggesting that quieter, less anthropogenically disturbed environments may be beneficial for maternal cardiovascular health. This finding is consistent with growing evidence linking chronic noise exposure to hypertension and cardiometabolic risk.^94,95

A central motivation for our study was to examine whether different HDP subtypes have distinct environmental risk factor profiles. Indeed, our findings suggest notable heterogeneity between gestational hypertension and the preeclampsia spectrum. Gestational hypertension—often considered a more benign or milder condition—showed the richest exposome signal, with 12 exposures replicating and remaining independently associated in Phase 2. In contrast, preeclampsia and eclampsia had no exposure passing our discovery and replication thresholds. These differences could have several explanations. First and mostly likely, due to different sample sizes for HDP subtypes, the statistical power to detect exposome-wide associations was much greater for gestational hypertension than for severe preeclampsia or eclampsia. Therefore, it is possible that meaningful but modest associations for severe preeclampsia were missed. Second, there may be true etiological differences: gestational hypertension may be more influenced by environmental stressors, whereas severe preeclampsia could be more strongly driven by placental genetic/immune factors that are less tied to the external environment. This concept has been supported by the literature: eg, one study found air pollution exposure was associated with higher risk of mild preeclampsia but not with severe disease.⁹⁶ It is also possible that our stringent analysis (controlling false discovery and using replication) was too conservative for the smaller subtype strata. For example, if we relaxed the multiple-testing correction, we might find that severe preeclampsia was nominally associated with three exposome measures (ie, minimum relative humidity, natural A-weighted L50 sound pressure level and arsenic compounds). Our inability to detect such links in an agnostic screen likely reflects a combination of limited power and the challenge of multiple comparisons. Our results highlight that future studies should stratify HDP by subtype to unmask associations that might be diluted in composite outcomes. Larger exposome studies or targeted environmental analyses for severe preeclampsia may eventually identify exposures that we were underpowered to detect.

This study has several strengths. First, we combined statewide vital statistics with EHR data to improve both scale and phenotype specificity, enabling ascertainment of HDP subtypes and reconstruction of pregnancy residential histories. We evaluated a broad set of spatial and contextual exposome measures across multiple domains and used a rigorous two-phase design (temporal discovery/replication followed by multi-treatment estimation) to reduce false positives and identify exposome measures independently associated with the outcomes after accounting for co-exposures. Finally, our analytic framework was grounded in causal inference principles and modern machine learning. We departed from the standard exposome study practice of using one-size-fits-all covariate adjustment for every exposure.³⁰ Instead, we developed exposure-specific confounder sets guided by DAG, recognizing that different exposures have different confounding structures. To handle the high dimensionality of covariates and account for potential non-linearities, we applied DML for effect estimation, which uses machine learning to flexibly model nuisance variables, and then estimates the effect size orthogonal to those nuisances. This approach provides efficient adjustment for many confounders without manual model specification, and importantly, yields asymptotically unbiased effect estimates with valid confidence intervals under moderate assumptions. To our knowledge, this is the first spatial and contextual exposome study to apply a formal DML approach. The successful implementation of DML in our study demonstrates its feasibility for exposome analyses, addressing a key analytic challenge in this field: how to adjust for the dense web of correlated covariates and exposures without overfitting or introducing bias. We see this as a methodological contribution that can be extended to other exposome-health investigations to improve causal inference rigor.

Several limitations should be considered. First, exposure assignment relied on secondary, model-based or contextual-level measures, which may not fully capture individual time-activity patterns and can introduce measurement error (likely nondifferential, and therefore potentially biasing effects toward the null). Second, residential histories were derived from encounter-based EHR address records; while address start and end dates were available, they may not reliably reflect true move-in and move-out dates, and address update frequency may vary across individuals. We did not have information on workplace locations or commuting patterns; thus, exposures experienced outside the home environment were not captured. In addition, address histories were derived from in-state EHR records and may not capture periods of residence outside Florida. Third, housing instability is not captured in our data, which may introduce exposure misclassification for a small subset of women. Fourth, although EHR linkage improved outcome definitions, some misclassification likely remains (eg, undiagnosed chronic hypertension or incomplete documentation). While sensitivity analyses further adjusting for residential mobility and urbanicity yielded similar effect estimates, residual confounding cannot be excluded—particularly for individual behaviors and health status factors that may be imperfectly captured. Fifth, generalizability may be limited to Florida’s climate, geography, and healthcare context, and selection into the linked EHR network could introduce bias. Finally, stringent multiple-comparison control and replication improve credibility but may increase false negatives, especially for rarer subtypes (eg, eclampsia).

In conclusion, our exposome-wide analysis of a statewide pregnancy cohort identified a set of diverse environmental factors—ranging from air toxicants and meteorology to neighborhood built and social characteristics—that are independently associated with HDP. These findings support the premise that the spatial and contextual exposome contributes to HDP risk and highlight modifiable domains that could inform prevention strategies.